PCB或BGA,使用CT/CL成像,检测,无监督或半监督检测

Element detection is an important step in the process of the non-destructive testing of printed circuit boards (PCB) based on computed tomography (CT). Compared with the traditional manual detection method, the image semantic segmentation method based on deep learning greatly improves efficiency and accuracy. However, semantic segmentation models often require a large amount of data for supervised training to generalize better model performance. Unlike natural images, the PCB CT image annotation task is more time-consuming and laborious than the semantic segmentation task. In order to reduce the cost of labeling and improve the ability of the model to utilize unlabeled data, unsupervised pre-training is a very reasonable and necessary choice. The masked image reconstruction model represented by a masked autoencoder is pre-trained on the unlabeled data, learning a strong feature representation ability by recovering the masked image, and shows a good generalization ability in various downstream tasks. In the PCB CT image element segmentation task, considering the characteristics of the image, it is necessary to use a model with strong feature robustness in the pre-training stage to realize the representation learning on a large number of unlabeled PCB CT images. Based on the above purposes, we proposed a contrastive dual-masked autoencoder (CD-MAE) pre-training model, which can learn more robust feature representation on unlabeled PCB CT images. Our experiments show that the CD-MAE outperforms the baseline model and fully supervised models in the PCB CT element segmentation task.

Computed tomography (CT) is a prominent technology for nondestructive quality control and is already used in industry for defect detection. However, as quality control is shifting towards a full in-line inspection, automatic CT analysis is required to meet the tight production time. Nonetheless, in settings where a high amount of data is produced, a robust fully automatic defect detection is essential In the past years, deep learning (DL) has been extensively used to perform vision tasks in an automatic way, and given its promising results, has been successfully applied in CT settings. Most of the recent work is based on supervised DL often adapted from results in the medical field. Supervised DL, although extremely powerful, has the drawbacks of requiring a high amount of labeled data done by experts and is biased to the specific dataset used. Therefore, an unsupervised DL model is presented. A two stages network formed by an auto-encoder and an autoregressive model, originally implemented for image generation, is adapted for volume segmentation. The network is trained on the specific task of defect segmentation of cast aluminum parts. CAD models of such parts are gathered, and corresponding simulated CT scans are acquired. Results show that the architecture, although originally implemented for data generation, can be adapted for CT volume segmentation.

For successful printed circuit board (PCB) reverse engineering (RE), the resulting device must retain the physical characteristics and functionality of the original. Although the applications of RE are within the discretion of the executing party, establishing a viable, non-destructive framework for analysis is vital for any stakeholder in the PCB industry. A widely regarded approach in PCB RE uses non-destructive x-ray computed tomography (CT) to produce three-dimensional volumes with several slices of data corresponding to multi-layered PCBs. However, the noise sources specific to x-ray CT and variability from designers hampers the thorough acquisition of features necessary for successful RE. This article investigates a deep learning approach as a successor to the current state-of-the-art for detecting vias on PCB x-ray CT images; vias are a key building block of PCB designs. During RE, vias offer an understanding of the PCB’s electrical connections across multiple layers. Our method is an improvement on an earlier iteration which demonstrates significantly faster runtime with quality of results comparable to or better than the current state-of-the-art, unsupervised iterative Hough-based method. Compared with the Hough-based method, the current framework is 4.5 times faster for the discrete image scenario and 24.1 times faster for the volumetric image scenario. The upgrades to the prior deep learning version include faster feature-based detection for real-world usability and adaptive post-processing methods to improve the quality of detections.

Automatic optical inspection for manufacturing traditionally was based on computer vision. However, there are emerging attempts to do it using deep learning approach. Deep convolutional neural network allows to learn semantic image features which could be used for defect detection in products. In contrast to the existing approaches where supervised or semi-supervised training is done on thousands of images of defects, we investigate whether unsupervised deep learning model for defect detection could be trained with orders of magnitude smaller amount of representative defect-free samples (tenths rather than thousands). This research is motivated by the fact that collection of large amounts of defective samples is difficult and expensive. Our model undergoes only one-class training and aims to extract distinctive semantic features from the normal samples in an unsupervised manner. We propose a variant of transfer learning, that consists of combination of unsupervised learning used upon VGG16 with pre-trained on ImageNet weight coefficients. To demonstrate a defect detection, we used a set of Printed Circuit Boards (PCBs) with different types of defects - scratch, missing washer/extra hole, abrasion, broken PCB edge. The trained model allows us to make clusters of normal internal representations of features of PCB in high-dimensional feature space, and to localize defective patches in PCB image based on distance from normal clusters. Initial results show that more than 90% of defects were detected.

Training data is crucial for any artificial intelligence model. Previous research has shown that various methods can be used to enhance and improve AI training data. Taking a step beyond previous research, this paper presents a method that uses AI techniques to generate CT training data, especially realistic, artificial, industrial 3D voxel data. This includes that material as well as realistic internal defects, like pores, are artificially generated. To automate the processes, the creation of the data is implemented in a 3D Data Generation, called SPARC (Synthetized Process Artificial Realistic CT data). The SPARC is built as a pipeline consisting of several steps where different types of AI fulfill different tasks in the process of generating synthetic data. One AI generates geometrically realistic internal defects. Another AI is used to generate a realistic 3D voxel representation. This involves a conversion from STL to voxel data and generating the gray values accordingly. By combining the different AI methods, the SPARC pipeline can generate realistic 3D voxel data with internal defects, addressing the lack of data for various applications. The data generated by SPARC achieved a structural similarity of 98% compared to the real data. Realistic 3D voxel training data can thus be generated. For future AI applications, annotations of various features can be created to be used in both supervised and unsupervised training.

As technology evolves, more components are integrated into printed circuit boards (PCBs) and the PCB layout increases. Because small defects on signal trace can cause significant damage to the system, PCB surface inspection is one of the most important quality control processes. Owing to the limitations of manual inspection, significant efforts have been made to automate the inspection by utilizing high resolution CCD or CMOS sensors. Despite the advanced sensor technology, setting the pass/fail criteria based on small failure samples has always been challenging in traditional machine vision approaches. To overcome these problems, we propose an advanced PCB inspection system based on a skip-connected convolutional autoencoder. The deep autoencoder model was trained to decode the original non-defect images from the defect images. The decoded images were then compared with the input image to identify the defect location. To overcome the small and imbalanced dataset in the early manufacturing stage, we applied appropriate image augmentation to improve the model training performance. The experimental results reveal that a simple unsupervised autoencoder model delivers promising performance, with a detection rate of up to 98% and a false pass rate below 1.7% for the test data, containing 3900 defect and non-defect images.

X-ray inspection on connectors becomes more and more important in these years as more connectors are mounted on PCBA due to their apparent benefits1-2. In this paper, we would like to share our studying experience for PCB Cavity Press-Fit connector on cavity PCB area. Our assembly with two press-fit connectors are inspected with the AXI, in order to make sure there is no broken pins enclosed inside the cavity, especially at the connector pins area. We have been working with AXI3, AXI5 and AXI6, and looking for good solutions for the last year. AXI3 is normal AXI (Automated X-ray Inspection); AXI5 and AXI6 are special AXI machines which have CT (Computed tomography) capabilities. Both AXI5 and AXI6 have capabilities to detect broken pin in cavity connector pins area with very clear X-ray images, and pass/fail judgments are independent of their operator skills. However AXI5 and AXI6 testing times are longer than AXI3. Finally we has developed good method to solve its issue at AXI after more than one year studying: the Algorithm and Threshold setting of AXI3 will be discussed in the paper. Recently AXI3 has capabilities to detect defect with PCB Cavity Press-Fit connector with very good balance between defects escaped percent and false call PPM. With joined studies and collaboration between Flextronics AXI team and AXI 3 R&D for eight months, outcome of the collaboration is very promising. Defects escaped rate has reduced from 20% to 0%, while false call PPM has reduced from >10,000 PPM to <;500 PPM.

Solder interconnects in several packaging architectures are located at the bottom of the package and are attached to PCB using surface mount technique and are often not available for direction visual inspection without crosssectioning the package. In this paper, the potential of x-ray micro-computed tomography (μCT) to fulfill the need for non-destructive three dimensional imaging of electronic assemblies has been evaluated. A number of defect seeded assemblies have been used for detection of a number of common solder interconnect defects and failure modes. In addition, the ability of the x-ray μCT for examination of complete products has been examined. Three-dimensional rendered versions of the board assemblies have been constructed for visualization of the defects and failure modes. Void sizes have been measured using Volume Graphics reconstruction and Matlab modules. In each case, the assemblies have been cross-sectioned after imaging by x-ray μCT to ascertain the morphology of the defect or failure mode using optical imaging. Results indicate that xray μCT is capable of providing high resolution imaging of the common defect types and failure modes in electronic assemblies and has potential for risk mitigation in sustainment of long-life high-rel systems.

Quad Flat No Lead (QFN) is extremely popular because of their low cost, low stand-off height and excellent thermal and electrical properties. The challenge for the industry is to achieve the best possible QFN solder joint quality. Here are some very important topics. What is the best way to use test and inspection techniques? How to minimize voiding in thermal pads by changing design rules in order to meet stringent customer requirements? How to reduce the use of Mechanical Cross Sectioning as it destroys the valuable PCBs and is time consuming? In this project, we focus on developing the best test procedure for identifying solder joint defects in QFNs and improve our SMT process. The test methods include SPI, AXI, and 3D CT Board Level X-Ray Inspection, also called Limited Angle or Partial CT (PCT). All these methods are completely non destructive. Our test assembly has a large number and variety of QFN devices. We analyzed data from Solder Paste Inspection (SPI), AXI (5DX), and Board Level CT X-Ray (Virtual Cross Sectioning) using nine QFN package types (24 components). The goal is to look for a correlation between SPI and AXI using more than 6,400 attribute and variable data points and also to compare the data with 2DX and Board Level CT. This is a very challenging project requiring advanced SPC techniques. Finally, based on the test results, prepare an action plan for SMT process improvements intended to take place within several months. Our QFN SPI-AXI Pearson Correlation does not show a strong relationship, however, attribute data has shown good results: 100 % of the packages show correlation for more than 90% of the pins, while 80.6% of the packages showed correlation for 100% of the pins. SPI is the first test machine in the SMT process -- implementing SPI in the SMT line results in significant reduction of components to the repair line.

With more and more BGA and PoP packages used on contemporary PCB boards, achieving the best possible BGA, and PoP solder joint quality becomes extremely important for the PCBA manufacturer. Most of the solder joints of these types of devices are invisible by optical means, thus AXI (Automated X-ray Inspection) is becoming more frequently used on the SMT lines. However, based on our experience, it is very challenging for current AXI machines to achieve defect escape rate of better than 2%, and false call PPM less than 2000 for BGAs, PoP s and some critical packages. Naturally we ask the following questions: What are the best ways to utilize the X-ray inspections techniques currently available? How to optimize the AXI algorithms settings? What is the correlation between AXI and 2D/3D CT X-ray inspection results? Within this project we focus on two main subjects: 1. Use 2D/3D X-ray inspection to identify solder joint Head-In-Pillow (HIP) defects in BGA devices, and utilize the collected data to improve our AXI programming process. 2. Use 2D/3D X-ray inspection for identifying solder’s joint quality for PoP on cavity board, and improve our SMT process. The 3D X-ray test method used is 3D X-ray Board Level Computer Tomography (Limited Angle Computer Tomography or PCT). All of the above X-Ray inspection methods are completely non-destructive, thus the expensive board does not need to be destroyed during the inspection process. We analyzed data from four different AXI machines at four Flextronics sites using boards known to have HIP (Head-In-Pillow) defects 1 . Initially all HIP pins were identified using 2D X-Ray inspection. Recently we used 3D X-ray Large Board CT (or PCT) to produce 3D X-Ray models and images of the areas of interest on the PCB, in order to check and confirm the 2D X-ray results, and would like to see current AXI can detect what kind of HIP type. We used the 2D/3D CT data to monitor the level of success for current AXI machines in finding various types of HIP defects. As part of our new product development process, we use 2D and 3D large board CT X-ray inspection to identify solder joint defects of PoP packages mounted within a cavity on the PCB. The two X-ray techniques working in tandem let us determine precisely the type and the location of the defects, thus providing extremely useful data during the SMT development process. We will also report the improvement data for the novel cavity PoP process and discuss our process development methodology.

High performance X-ray inspection systems are widely used in the PCBA manufacturing process. These include 2DX systems with sub-micron feature recognition that are equipped with 3D micro-CT (Computer Tomography) and Large Board CT 1-2 (also called Partial Angle CT and PCT). The Large Board CT functionality is extremely important in order to be able to perform 3D CT inspection of regions of interest within a large PCBs in a completely non-destructive way and still provide very high resolution and detail. The micro-CT is mostly effective for small samples and not for large PCBAs. In this paper we share our experience for verifying the quality of a potting process using completely nondestructive X-ray methods (2DX and Large Board CT/PCT). The defects we verify for are voiding and incomplete potting curing. Traditionally, the standard way of controlling the potting process has been a destructive sampling (mechanical cross-sectioning) that results in board damage and scrapping. We compared our X-ray inspection results (2DX and PCT) with mechanical cross section results and proved that the non-destructive X-ray techniques provide completely satisfactory testing solution that gives good confidence to our customers. The X-ray inspection provides clear 2D and 3D images of the defects and it is easy to optimize. In addition, the X-ray inspection (2DX and PCT) of potting process resulted in a significant cost savings for our customers. This is due to the fact that the X-ray techniques are completely non-destructive and also much faster, compared to the mechanical cross-sectioning. In addition, using X-ray inspection we can offer 100% testing solution that is not possible using mechanical cross section. Naturally, as a future project, we will be looking to fully automate the X-ray inspection and provide a fully automatic, extremely fast and completely non-destructive solution for PCBA potting process verification.

With PCB complexity and density increasing and also wider use of 3D devices, tougher requirements are now imposed on device inspection both during original manufacture and at their subsequent processing onto printed circuit boards. More complicated and dense packages have more opportunities to exhibit defects both internal to the package as well as to the PCB. As components increase in complexity their cost increases, making counterfeiting them a potentially lucrative business for unscrupulous individuals and organizations. Recent years have brought significant improvement in the capabilities in the 2D/3D X-Ray Inspection systems. New X-ray sources, detectors, and ergonomic features improve the efficiency and productivity of the inspection process. This paper reviews the methods of finding defects in BGAs, QFNs, and 3D packages using X-Ray inspection with real-life examples provided. Voiding, cracks, shorts, open joints, and head in pillow (HIP) will be discussed. Comparison of the relative merits of the 2D and 3D (CT) X-Ray inspection for investigating 3D packages is presented with examples. Using X-Ray inspection for detecting counterfeits is discussed at the end.

Head in Pillow (HiP) defects are the most undesirable occurrences in PCB assembly process. HiP occurs when a solder ball of an area array component (BGA or LGA) rests on the mound of solder without forming a good metallurgical bond. These defects are very difficult to detect by standard inspection techniques and may escape most standard electrical tests (e.g. bed-of-nails test). Prevalent hypothesis says that HiP defects form when flux in the solder paste used for joining, is allowed to dry out before the joint is formed, and does not adequately remove all the oxidation from the solder ball, for instance, due to extended dwell in flux activation temperature range[1]. One way to reduce HiP occurrence during reflow cycle is to heat the PCBA at a high ramp rate. For very thick multi-layered PCBAs however, achieving high ramp rates may be very difficult, especially in localized rework situations. This paper describes a methodical approach adopted to study the underlying causes that can lead to HiP formation. Wetting characteristics of the solder paste were studied using a wetting balance. Dynamic warpage profiles of the BGA component and the circuit board were obtained using Moiré Interferometry, for different reflow cycles. A proper selection of paste chemistry and the most optimum profile were used to make assemblies, which were then analyzed with CT Scan (3D X-ray) techniques to detect HiP type of defects. Inspection of several assemblies with CT scan has shown no presence of HiP formation.

Computer Tomography (CT) is a powerful inspection technique used widely in the electronics industry, especially for the analysis of multi-layered devices and joint interconnections. As the resolution required to be able to inspect today's devices is in the micron range, this aspect of CT is often referred to as μCT so as to differentiate it from medical and industrial CT applications where the same level of resolution is not possible or required. The μCT technique permits different layers / slices of the device to be isolated and examined individually, so practically providing an electronic, or virtual, cross-sectioning within the sample. The benefits of an ‘electronic cross-section’ compared to traditional mechanical cross-sectioning are many. These include that the electronic cross-sectioning is reversible - you cannot over polish and go too far into the sample - the cutting plane can be positioned in any orientation within the 3D space of the CT model and no additional defects are introduced or existing defects concealed compared with the process of mechanically cutting, polishing and preparing the sample for a cross-section. One of the limitations of traditional μCT is that there is a restriction to the maximum sample size that can be used to produce a μCT model with reasonable speed, quality and analytical value. Usually, the maximum practical size for a regular μCT is ~ 2" x 2" (50 x 50 mm). Thus, it is not possible to use the μCT technique on a large PCB unless you are willing to cut around the device / region of interest to be examined to make it small enough for analysis, but in so doing destroying the board. In order to overcome the sample size limitation of ‘full μCT’, a ‘limited angle’ or ‘partial’ μCT technique has been developed and used in some X-Ray systems. This permits a 3D model to be created from devices / regions of interest anywhere within a board without the need to destroy it. This paper will explain the mechanism and differences between full μCT and the various types of limited angle μCT and compare different applications where one or the other technique is applicable, backed by real life cases and examples.

As the use of 3D devices, such as package-in-package (PiP), package-on-package (PoP) and system-in-package (SiP), continues to increase, a greater burden is now placed upon device inspection both during original manufacture and at their subsequent processing onto printed circuit boards. This is because these more complicated packages now have more opportunities to exhibit defects both internal to the package as well as to the PCB. Such defects can include interfacial joint cracks, missing connections, excessive joint voiding, wire and wire bonding defects and die / interface issues. The use of x-ray inspection provides a vital non-destructive method for 3D package investigation and quality control for these potential problems. This paper will provide a comparison of the relative merits of using 2D and 3D (or CT) x-ray inspection techniques for investigating 3D packages as well as showing typical examples.

We propose a defect detection method of solders on a printed circuit board using X-ray CT inspection system and Adversarial Autoencoder (AAE)[1] . We obtain sliced images of the solder using X-ray CT and extract their features that follow the standard normal distribution by using AAE. Then, the solder defects are detected by Hotelling's T square[2]. As a result of experiments, we show that we can classify normal and anomalous data samples completely on the condition of training with large normal samples and small anomalous samples.

Mission-critical electronic systems demand early and accurate detection of solder joint degradation to ensure reliability. Quad Flat No-Lead (QFN) packages, widely used in automotive and industrial applications, are especially prone to vibration-induced solder fatigue. However, traditional failure analysis methods (e.g., dye-and-pry, cross-sectioning, manual Xray inspection) are labor-intensive and often insufficient to detect early-stage cracks. This paper presents an automated inspection framework that combines high-resolution 3D X-ray tomography with a YOLOv11-based deep learning model to detect and segment vibration-induced cracks in QFN solder joints. The pipeline achieves precise localization of cracks in volumetric data, discriminates them from voids, and extracts morphological descriptors through parametric fitting. By statistically correlating these image-derived crack features with electrical resistance measurements recorded in situ during vibration tests, we establish a direct link between physical crack evolution and functional degradation of the joint. The results demonstrate that our AI-driven method can automatically identify tiny solder cracks and reliably offer predict impending interconnect failures in comparable granularity of traditional inspection techniques, surpassing them in speed. This approach offers a powerful prognostic health monitoring tool for electronic packaging, and it is extensible to other package types and stress conditions.

With the improvement of electronic circuit production methods, such as reduction of component size and the increase of component density, the risk of defects is increasing in the production line. Many techniques have been incorporated to check for failed solder joints, such as X-ray imaging, optical imaging and thermal imaging, among which X-ray imaging can inspect external and internal defects. However, some advanced algorithms are not accurate enough to meet the requirements of quality control. A lot of manual inspection is required that increases the specialist workload. In addition, automatic X-ray inspection could produce incorrect region of interests that deteriorates the defect detection. The high-dimensionality of X-ray images and changes in image size also pose challenges to detection algorithms. Recently, the latest advances in deep learning provide inspiration for image-based tasks and are competitive with human level. In this work, deep learning is introduced in the inspection for quality control. Four joint defect detection models based on artificial intelligence are proposed and compared. The noisy ROI and the change of image dimension problems are addressed. The effectiveness of the proposed models is verified by experiments on real-world 3D X-ray dataset, which saves the specialist inspection workload greatly.

Internal defect detection constitutes a critical process in ensuring component quality, for which anomaly detection serves as an effective solution. However, existing anomaly detection datasets predominantly focus on surface defects in visible-light images, lacking publicly available X-ray datasets targeting internal defects in components. To address this gap, we construct the first publicly accessible component X-ray anomaly detection (CXR-AD) dataset, comprising real-world X-ray images. The dataset covers five industrial component categories, including 653 normal samples and 561 defect samples with precise pixel-level mask annotations. We systematically analyze the dataset characteristics and identify three major technical challenges: (1) strong coupling between complex internal structures and defect regions, (2) inherent low contrast and high noise interference in X-ray imaging, and (3) significant variations in defect scales and morphologies. To evaluate dataset complexity, we benchmark three state-of-the-art anomaly detection frameworks (feature-based, reconstruction-based, and zero-shot learning methods). Experimental results demonstrate a 29.78% average performance degradation on CXR-AD compared to MVTec AD, highlighting the limitations of current algorithms in handling internal defect detection tasks. To the best of our knowledge, CXR-AD represents the first publicly available X-ray dataset for component anomaly detection, providing a real-world industrial benchmark to advance algorithm development and enhance precision in internal defect inspection technologies.

X-ray metrology for solder joint voids faces challenges in precise measurement due to limited annotated data, resulting in uncertainty and reduced accuracy. To address this, we propose ultra self-training (UST), an improved semi-supervised semantic segmentation framework for accurate geometric quantification of solder joint voids. UST features a progressive semi-supervised training (PSST) module, which gradually incorporates unlabeled data based on confidence scores, reducing noise while maintaining convergence speed. It also employs an adaptive pseudo-label quality assessment (APQA) mechanism to enhance pseudo-label filtering and quality control. For segmentation, we replace the backbone with ICSA-UNet to improve feature representation of small, dense defects. Experiments show UST achieves strong performance on the solder joint X-ray dataset, with only 18.7M parameters and a mean intersection over union (mIoU) of 77.61%. Generalization results on Cityscapes and PIDRAY dataset confirm UST’s efficiency and reliability.

No abstract available

ABSTRACT This study combines machine learning (ML) and X-ray imaging to evaluate the health of solder joints in printed circuit boards (PCBs). A convolutional neural network (CNN) served as the base framework, with CNN-LSTM and CNN-CapsNet models added to enhance performance. Pre-training with the CNN facilitated feature extraction, boosting the subsequent performance of LSTM and CapsNet models. The research focused on three objectives: identifying the best ML model for limited datasets, addressing class imbalance in defective solder samples with data augmentation, and using image manipulation to assess model strengths and limitations. Data augmentation significantly improved model accuracies, with CNN, LSTM, and CapsNet achieving 87.05%, 91.29%, and 94.65%, respectively, compared to 76.23%, 83.32%, and 88.05% without augmentation. CapsNet outperformed other models, leveraging its dynamic routing mechanism to preserve feature hierarchies and maintain stable performance. LSTM demonstrated rapid learning through memory cells, while CNNs were prone to overfitting. CapsNet also excelled in balancing classification across solder types, highlighting its ability to handle complex feature relationships. Robustness tests showed CapsNet’s resilience to image transformations like rotation, scaling, and flipping, though extreme deformations remained challenging. These results underscore CapsNet’s potential for accurate and reliable solder joint classification in diverse scenarios.

Ball Grid Array (BGA) packaging technology has become the mainstream solution for integrated circuit packaging due to its high-density integration characteristics‎[1]‎[2]. However, process fluctuations during the soldering process can easily lead to void defects inside the solder joints, affecting the reliability of electronic devices. The traditional Otsu method based on threshold segmentation has problems such as high threshold sensitivity and insufficient noise resistance when used for quantitative detection of BGA solder joints by CT. Therefore, an intelligent detection method integrating the two-dimensional Otsu algorithm and Particle Swarm Optimization (PSO) algorithm is proposed. By obtaining three-dimensional solder joint structure data through X-ray computed tomography (CT), a two-dimensional Otsu adaptive threshold optimization model based on PSO is established to achieve precise segmentation and quantitative characterization of voids inside the solder joints, effectively improving the segmentation accuracy and suppressing the noise interference of CT images. This provides a high-precision non-destructive testing solution for BGA packaging quality assessment and has important engineering application value for enhancing the reliability of microelectronic devices.

No abstract available

This paper demonstrates a non-destructive technique to evaluate the internal microstructure in the Sn-Ag-Cu (SAC) solder joint through synchrotron X-ray radiation tomography. Synchrotron X-ray tomography is increasingly utilized for characterizing the internal microstructure of materials in 3D images. A 3D model is reconstructed from a set of 2D projection images taken from different angles and angular position during the sample rotation, thus it could provide a more comprehensive description of the microstructure of an alloy compared to 2D images. In this paper, it is successfully observed and evaluated the internal microstructure of a 900 μm solder joint sample. The key principles and methods of synchrotron X-ray tomography are briefly described. Examples of quantitative and qualitative assessments on the grain refinement effect of Mg addition to SAC35 solder joint are also presented in this paper.

Anomaly detection in X-ray images of electronic components plays a vital role in quality control for advanced manufacturing. During the industrial production of electronic components, internal microdefects often arise due to material fatigue or process variations. Automatically identifying these defective regions remains a significant and challenging task in the field. However, existing methods typically rely on prior knowledge of normal categories and often suffer from limited generalization across diverse component types. To address these limitations, we propose UAD-X, a universal anomaly detection framework tailored for X-ray inspection of electronic components. Our method leverages a pretrained vision Transformer encoder (DINOv2) to extract robust image representations and introduces a normal reference representation (NRR) extractor that dynamically learns intrinsic normal patterns from a single image. A bottleneck module is employed to compress semantic information, and a reference-guided decoder reconstructs the normal regions, thereby suppressing anomalous features effectively. To enhance the detection of subtle defects, we further design an entropy-aware mining loss (EAML), which adaptively amplifies training signals based on the regional uncertainty. Extensive experiments conducted on our curated X-ray dataset of electronic components and the MVTec AD benchmark demonstrate that our approach significantly outperforms state-of-the-art (SOTA) methods in both pixel-level localization and image-level classification tasks. Specifically, on our electronic component X-ray dataset, the proposed method achieves an image-level area under the receiver operating characteristic (AUROC) of 92.5%, a pixel-level AUROC of 98.0%, and a pixel-level area under the per-region overlap (AUPRO) of 96.3%. Moreover, it exhibits strong cross-category generalization, achieving high performance without requiring retraining on new categories.

The voids of solder joints have a very important impact on the reliability of the integrated circuit. X-ray detection technology is commonly used in the inspection of integrated circuit packaging solder joints for its non-destructive characteristics. Accurate and efficient calculation of void ratio(VR) of X-ray images is of great significance for the quantitative assessment of solder joint quality and the study of its correlation with electrical and thermal parameters of devices. It is different to calculate void rate accurately and automatically because of the noise, inconsistent X-ray and overlap by adjacent components. In this paper, an algorithm for calculating void ratio of various solder joint from X-ray image is proposed. The steps of the algorithm including four major steps: preprocessing, morphological processing, void extraction and VR calculation. The core algorithm is K-means clustering which use gray value, eccentricity and area as the features of void. The typical results of VR algorithm on different test sample of X-ray image like LED, connector, power device etc. shows that the algorithm proposed in this paper can accurately mark the position of the void and obtain the accurate void rate of solder joint from the X-ray image.

A deep learning-based nondestructive approach for void segmentation in BGA solder balls using 3D x-ray microscopy is presented.

The inexorable trend towards heterogeneous integration in high-bandwidth optical transceivers presents a formidable challenge for in-line metrology and yield enhancement. This work introduces a novel, end-to-end automated inspection framework leveraging high-resolution X-ray computed tomography to overcome the limitations of conventional surface-based techniques. Our dataset consists of 3D X-ray scans of diced packages from 300 mm reconstituted wafers, each integrating silicon photonics integrated circuits, electronic integrated circuits, through-mold interconnects, and multi-layer redistribution layers, encapsulated within epoxy mold compounds. Traditional inspection methods, such as optical microscopy and electrical testing, are fundamentally blind to critical buried defects like solder voids at micro-bump interfaces, pad misalignment, and other related defects. Our automated workflow addresses this critical visibility gap. The system demonstrates exceptional performance, achieving a dice score of 0.799 with training on just 3 bumps. This automated analysis reduces the inspection cycle time from several hours per unit for manual review to under 10 minutes, enabling statistically significant process control.

This research proposes a predictive anomaly detection (AD) framework for solder joints. In surface mount technology (SMT), anomalous solder joints reduce the reliability of printed circuit boards (PCBs), which raises reworking expenses for PCB assembly lines. Therefore, predictive AD is essential to prevent solder joints with anomalies. The solder joint formation consists of three primary phases: solder paste printing, pick and place, and solder reflow. This research aims to predict the solder joint’s quality before the solder reflow phase by a novel framework, predictive AD for solder joints, which is called PADS. PADS first extracts 65 solder-joint-related features from datasets, then learns normal solder joints’ patterns by reconstructing these features, and finally identifies a sample as an anomaly if its reconstruction error exceeds a designated threshold. The uniqueness of PADS is the utilization of novel features generated from interpreting the existing physics-based models and substantial real-world data acquired from SMT inspection machines, i.e., solder paste inspection (SPI) and automatic optical inspection (AOI). PADS has been extensively evaluated with commercial chip resistors R0402M (<inline-formula> <tex-math notation="LaTeX">$0.4 \times 0.2$ </tex-math></inline-formula> mm), R0603M (<inline-formula> <tex-math notation="LaTeX">$0.6 \times 0.3$ </tex-math></inline-formula> mm), and R1005M (<inline-formula> <tex-math notation="LaTeX">$1.0 \times 0.5$ </tex-math></inline-formula> mm), as well as SAC305 solder paste. The experimental results indicate that these novel features enable PADS to perform better in anomaly prediction for solder joints, and PADS outperforms many competitive baselines in prediction accuracy.

Defects within chip solder joints are usually inspected visually for defects using X-ray imaging to obtain images. The phenomenon of voids inside solder joints is one of the most likely types of defects in the soldering process, and accurate detection of voids becomes difficult due to their irregular shapes, varying sizes, and defocused edges. To address this problem, an X-ray void image segmentation algorithm based on improved PCB-DeepLabV3 is proposed. Firstly, to meet the demand for lightweight and easy deployment in industrial scenarios, mobilenetv2 is used as the feature extraction backbone network of the PCB-DeepLabV3 model; then, Attentional multi-scale two-space pyramid pooling network (AMTPNet) is designed to optimize the shallow feature edges and to improve the ability to capture detailed information; finally, image cropping and cleaning methods are designed to enhance the training dataset, and the improved PCB-DeepLabV3 is applied to the training dataset. The improved PCB-DeepLabV3 model is used to segment the void regions within the solder joints and compared with the classical semantic segmentation models such as Unet, SegNet, PSPNet, and DeeplabV3. The proposed new method enables the solder joint void inspection to get rid of the traditional way of visual inspection, realize intelligent upgrading, and effectively improve the problem of difficult segmentation of the target virtual edges, to obtain the inspection results with higher accuracy.

Ring artifacts are prevalent in 3D cone-beam computed tomography (CBCT) due to non-ideal responses of X-ray detectors, substantially affecting image quality and diagnostic reliability. Existing state-of-the-art (SOTA) ring artifact reduction (RAR) methods rely on supervised learning with large-scale paired CT datasets. While effective in-domain, supervised methods tend to struggle to fully capture the physical characteristics of ring artifacts, leading to pronounced performance drops in complex real-world acquisitions. Moreover, their scalability to 3D CBCT is limited by high memory demands. In this work, we propose Riner, a new unsupervised RAR method. Based on a theoretical analysis of ring artifact formation, we reformulate RAR as a multi-parameter inverse problem, where the non-ideal responses of X-ray detectors are parameterized as solvable physical variables. Using a new differentiable forward model, Riner can jointly learn the implicit neural representation of artifact-free images and estimate the physical parameters directly from CT measurements, without external training data. Additionally, Riner is memory-friendly due to its ray-based optimization, enhancing its usability in large-scale 3D CBCT. Experiments on both simulated and real-world datasets show Riner outperforms existing SOTA supervised methods.

Defect detection in semiconductor packaging, particularly for 3D High Bandwidth Memory (HBM) packages, is critical for ensuring manufacturing quality and long-term reliability. However, accurate detection of fine structural defects in complex 3D-stacked HBM packages demands high spatial resolution scanning, which in turn requires long acquisition times. Such requirements make conventional high-resolution 3D X-ray imaging impractical for high-throughput industrial settings. To overcome this limitation, we introduce a deep learning-based framework that enables resolution-enhanced 3D X-ray imaging from low-fidelity acquisitions. Our approach leverages super-resolution networks to enhance low-fidelity volumes from fewer scans, thereby reducing dependency on expensive high-resolution imaging systems. This computational enhancement facilitates scalable, non-destructive inspection of advanced semiconductor packages without compromising defect detectability. Extensive experiments on industrial HBM datasets show that our method reduces acquisition time by up to 80 % while maintaining comparable quality of spatial resolution and improving sample quality for downstream defect analysis. By significantly lowering both imaging and computational costs, our work offers a practical pathway towards the widespread deployment of AI-driven, high-throughput semiconductor inspection systems.

Defect detection in 2.5D-3D packages is key in the manufacturing process to ensure product quality and reliability. With the increasing miniaturization of microbumps in integrated circuits (IC), this task has been more challenging and time-consuming. With recent advances in deep learning AI and generative AI, the inspection process can be further enhanced. Deep learning models provide fast and accurate defect detection and metrology, and generative AI provides insights and meaningful information on defects to the operator. In this paper, we propose a framework that uses generative AI to leverage on multiple sources of data such as fabrication parameters, technical documents or historical analysis to provide a highly accurate and comprehensive report of the analysed samples. We integrate the hierarchical retrieval-augmented generation framework to build better prompts and queries to large language models. With this framework, the large language model can search in specific documents with better context and accurate information and prioritize relevant information related to a specific defect or sample. We demonstrate that our approach provides faster and more accurate inspection report with 11% improvement compared to a basic report.

3D X-ray imaging plays an important role in package level failure analysis. Like most other microscopies, Xray microscopes (XRM) generally have field of view (FOV) limits for high-resolution imaging. As precise fault isolation becomes more challenging in large and complex IC packages, acquiring numerous high-resolution images to search for defects in a large FOV is required. For example, if a suspect fault region is 785 mm3 in a cylinder with 10mm in diameter and 10mm in height, more than 125 high-resolution scans with 6.28 mm3 each are required to cover the volume. This significantly diminishes XRM imaging throughput. In this report, we propose a new deep learning reconstruction method to address the issue of achieving high-resolution at large FOVs. This AI powered technique and workflow can be used to restore the resolution from a large FOV scan.

As Moore's Law slows, the electronics industry increasingly relies on advanced IC packages to deliver the necessary interconnect density required for continued performance improvements. This trend is particularly pronounced in AI data centers, where massive Large Language Models (LLMs) are driving memory and bandwidth demands to unprecedented levels. Over the next decade, peak package interconnect density is expected to grow significantly - rising from approximately $400 \mathrm{I} / \mathrm{O}$ per square millimeter (IO/sqmm) to nearly $\mathbf{1 0, 0 0 0 ~ I O} /$ sqmm, representing a 25-fold increase. This paper introduces a novel 3D X-ray-based methodology for extracting metrology data from buried micro-connections. The proposed approach combines high resolution with fast throughput, making it suitable for real-time, inline applications. The theory of operation, along with metrology data from advanced IC packages, will be discussed in detail.

No abstract available

As a non-destructive detection method, X-rays are widely used in the field of electronic component inspection. However, the subsequent defect detection needs to be completed manually, which leads to poor efficiency and low reliability due to a large number of components. To solve the above problems, we propose a component X-ray image defect detection method based on deep learning. On the one hand, we have designed an algorithm for the segmentation and correction of X-ray images. On the other hand, in the case of fewer defect samples and variable defect forms, we only use defect-free samples for training. We propose an unsupervised learning model based on variational autoencoder and add a convolutional attention module to realize automatic reconstruction of defects. In addition, we combine gradient magnitude similarity and absolute error of the input image and the reconstructed image to detect and locate the defect region. The effectiveness of the proposed method is verified through experiments on a typical X-ray dataset, and the accuracy of defects detection reaches about 99%.

In the testing of chips, defect diagnostics in X-ray images of packaging chips is mainly performed by humans, which is time-consuming and inefficient. To overcome the abovementioned problems, a novel intelligent defect diagnostics system based on hybrid deep learning for chip X-ray images was proposed. The system consists of four successive stages: image segmentation and normalization, image reconstruction and defect detection, contour matching, and qualification diagnosis. The first stage is used to localize the external contours of the target chip and remove extraneous backgrounds through the improved UNet. Then, considering the variety of defects and the complexity of labeling, an unsupervised learning model is designed to reconstruct defect-free images to detect defects, which requires only normal samples for training. Third, the multicomponent template matching based on structural prior is used to localize the internal contours of the chip. In the final stage, the qualification is diagnosed based on the previous results through the Floyd–Warshall algorithm. The effectiveness and robustness of the proposed methods are verified by experiments on real-world inspection lines. The experimental results demonstrate that the developed system can successfully perform fault diagnostics tasks, achieving a judgment accuracy of 92.5%.

To satisfy the ongoing need for improved product performance, the semiconductor industry has begun to utilize complex 3D IC packages, where chips are stacked on top of each other and are linked electrically by high-density interconnects. To enable ever-higher communication rates between the chips, the electrical interconnects are shrinking aggressively in size and at the same time increasing in density; use of these ultra-small, high-density electrical interconnects comes with the need for enhanced package assembly process controls. Since by their very nature the interconnects are embedded within the IC packages, the only practical non-destructive method to inspect them is via X-rays. In FA labs, high-resolution X-rays microscopes are commonly used to detect defects in faulty packages. However, typically only coupon-sized samples are scanned, and scan times can range in the order of hours with larger scan times for bigger samples. For inline process control of 300mm wafers such long scanning time cannot be accepted. By resorting to a laminographic scanning geometry and by making use of sophisticated proprietary reconstruction algorithms combined with extensive use of AI, we will demonstrate that scan times can be reduced from hours to minutes, independent of the lateral sample size. To this end we have developed a prototype tool dedicated to in-line X-ray inspection of 300mm wafers, which we refer to as ILX (In-Line X-ray) prototype.

This article shows how 3D XRM can be applied to nondestructively detect non-optimized assembly processes that can influence local stresses and overall device reliability. This makes it useful for process development and failure analysis. When used along with AI training models, 3D XRM can achieve analysis of highly integrated packaging structures with reasonable throughput for process validation and error correction guidance.

We introduce a fully unsupervised framework designed to reconstruct X-ray CT images from truncated projections without requiring prior truncation correction. By incorporating a Radon projection layer as the final layer of a deep learning model and using a projection-based loss function, our method effectively removes truncation-related artifacts, particularly ring artifacts. The framework is demonstrated on small-scale images and further extended to large-scale or arbitrary-scale images. For large-scale reconstruction, fully connected layers are applied in a distributed manner, enabling memory-efficient processing even with limited GPU resources. The effectiveness of the framework is evaluated using PSNR, SSIM, and MAE ± SD metrics. In cases of high-degree truncation, the method achieves consistently higher PSNR and SSIM values and lower MAE ± SD, showing its ability to reduce ring artifacts while preserving reconstruction quality.

The emerging era of 3D Heterogeneous Integration (3DHI) in Advanced Packages and Wafer level packaging of ICs introduce significant challenges for inline defect inspection and offline failure analysis techniques. Primarily, 3D stacking and wafer bonding result in optically opaque systems that require approaches such as X-rays to see through multiple layers of buried structures for defect detection. However, with the continual shrinkage in device features in 3DHI (e.g., microbumps are scheduled to shrink to <10 μm diameter and TSVs interconnects are scaling to single digit micrometers), non-destructive techniques are facing a technological brick wall.This includes 3D X-ray approaches, which need to have higher resolution than currently available in order to meet the evolving requirements. Furthermore, the acquisition time for sub-one micron imaging using conventional X-ray tomography even at a single location within a large 300 mm wafer may take hours or is outright impossible. To address these metrology gaps, we have developed two complementary groundbreaking 3D X-ray inspection tools:1.High throughput (3D data in minutes): The first tool is designed for rapid inspection of 300 mm wafers during wafer level packaging and bonding which can resolve various 3D defects to 0.5 μm resolution automatically and in minutes. This tool will also address board level FA, such as PCB at high resolution.2.High resolution (300 nm): The other complementary tool is designed to address the limits of resolution of the existing leading high resolution 3D X-ray and X-ray Microscopes (XRM) in Failure Analysis of Advanced Heterogeneous Packages. The system delivers true 300 nm spatial resolution (<50 nm voxel) for characterizing submicron defects in microbumps, delamination, voids interfacial cracks and RDL that cannot be seen and measured by existing XRMs.

This paper presents an innovative methodology for automated defect recognition in X-ray images by leveraging advanced Instance Segmentation techniques. The proposed approach accurately identifies and delineates defect regions, which are subsequently processed by a deep learning-based classification model to categorize defects based on their distinct morphological and textural characteristics. To enhance the model's robustness and mitigate false positives, a series of morphological operations, including Dilation, Erosion, and filtering, are applied to pre-process the data. Dilation is used to expand the boundaries of detected defect regions, ensuring that small gaps and holes within the defects are filled. Erosion, on the other hand, helps in removing small noise and separating closely located defect regions. Filtering techniques, such as Gaussian or median filtering, are employed to smooth the image and further reduce noise, thereby improving the quality of the input data for the segmentation and classification models. Furthermore, Image Augmentation techniques, such as rotation, scaling, translation, and flipping, are utilized to artificially expand the dataset. This augmentation not only increases the volume of training data but also introduces variability in the defect appearances, which helps in improving the model's generalization capabilities and feature diversity. By exposing the model to a wider range of defect scenarios, the accuracy and robustness of the defect recognition process are significant enhanced. The integrated approach combines the strengths of Instance Segmentation and classification models, along with robust pre-processing and augmentation strategies, to offer a substantial improvement in the efficiency and reliability of X-ray image-based defect analysis. This methodology has significant implications for quality control in industrial applications, where accurate and efficient defect detection is critical for maintaining product standards and reducing manufacturing cost.

Failure analysis is crucial in improving semiconductor manufacturing yields. Yield improvement is done by collecting, analyzing, identifying the causes of defects, and applying corrective actions to resolve the root causes. With the ongoing miniaturization of TSVs, micro-bumps, RDLs, and other package interconnects [1], detecting defects in these buried interconnects is becoming more difficult as well as more important. Traditionally semiconductor packages are cross-sectioned to identify internal process defects such as unsolders, solder shorts, and pad misalignment. Cross-sectioning is a destructive approach, is difficult to do, and provides information in a single 2D plane only. Due to the large effort and the destructive nature of this approach, the amount of data that can be generated is typically quite limited. The development of 3D x-ray microscopy provides industry with the capability to image and analyze buried features such as micro-bumps, TSVs, and other metallic structures using a non-destructive, 3-dimensional technology [2]. At the same time, deep learning has revolutionized other technologies such as visual surveillance, predictive maintenance [3], object detection [4], and is now revolutionizing defect detection in semiconductors. When used together, the combination of 3D x-ray microscopy and deep learning is establishing a new paradigm in package inspection and metrology. In this paper, we will present a novel method for automatically detecting internal anomalies in semiconductor packages and using deep learning to assess the attributes of these interconnects. Chips representative of stacked 2.5D packages were fabricated and assembled using thermo-compression bonding. Bonding parameters were varied in order to create packages with different bond line thickness, different solder fillet shapes, and various pad alignment scenarios. A commercial 3D x-ray imaging tool was used to create high-quality tomographies of these packages. Deep-learning and computer vision-based methods were employed to automatically detect internal features and measure attributes. A three-step procedure was used for data analysis. In the first step, a bounding box was detected for each region of interest (Copper Pillar, Pad, etc.) using a modified single shot detector object model. In the second step, we isolated features within the region of interest and performed 3D segmentation on them. The third and final step utilized automated 3D metrology using the segmented regions. Robust 3D computer vision techniques were deployed to measure the extent of voids which are key attributes for the chip fabrication and process control step. This is the first part of a multi-part paper.

For over 40 years lithographic silicon scaling has driven circuit integration and performance improvement in the semiconductor industry. As silicon scaling slows down, the industry is increasingly dependent on IC package technologies to contribute to further circuit integration and performance improvements. This is a paradigm-shift and requires the IC package industry to reduce the size and increase the density of internal interconnects on a scale which has never been done before. Traditional package characterization and process optimization relies on destructive techniques such as physical cross-sections and delayering to extract data from internal package features. These destructive techniques are not practical with today's advanced packages. In this paper we will demonstrate how data acquired nondestructively with a 3D X-ray microscope can be enhanced and optimized using machine learning, and can then be used to measure, characterize and optimize the design and production of buried interconnects in advanced IC packages. Test vehicles replicating 2.5D and HBM construction were designed and fabricated, and digital data was extracted from these test vehicles using 3D X-ray and machine learning techniques. The extracted digital data was used to characterize and optimize the design and production of the interconnects and demonstrates a superior alternative to destructive physical analysis. We report a mAP of 0.96 for 3D object detection, a dice score of 0.92 for 3D segmentation and an average of 2.1 um error for 3D metrology on the test dataset. This paper is the first part of a multi-part report.

This article presents advanced workflows that combine 3D X-ray microscopy (XRM), nanoscale tomography, and electron microscopy to generate detailed visualization of the interior of electronic devices and assemblies to enable the study of internal components for failure analysis (FA). Recently developed techniques such as integrating deep learning (DL)-based algorithms for 3D image reconstruction are also discussed in this article. Additionally, a DL-based tool (called DeepScout) is introduced that uses high-resolution 3D XRM datasets as training data for lower resolution, higher field of view datasets and scales larger volume data using a neural network. Ultimately, these workflows can be performed independently or complementary to other multiscale correlative microscopy assessments and will provide valuable insights into the internal workings of electronic packages and integrated circuits across multiple length scales, from macroscopic features on electronic devices (e.g., hundreds of mm) to microscopic details in electronic components (in the order of tens of nm). Understanding advanced electronic systems through X-ray imaging and electron microscopy, possibly integrated with additional correlative microscopy investigations, can accelerate development time, increase cost efficiency, and simplify FA and quality inspection of electronic packaging, printed circuit boards (PCBs) and electronic devices assembled with new emerging technologies.

As packaging drives to extend the capabilities of Moore’s Law via heterogenous integration, numerous bonding technologies have emerged as potential pathways, allowing for integration of both a variety of devices in the lateral space as well as vertical integration of chips via stacking. As vertical integration on interposer becomes more complex due to shrinking CDs and increasing die layers, alignment and shape control in the bonding process becomes critical. Thermocompression bonding (TCB) is particularly well-suited for larger die, where localized reflow allows a better control of chip gap height and tilt throughout the bonding process. TCB systems require both high accuracy and repeatability in all dimensions to provide a reliable system in package. While a 2D system can provide a plan view assessment of a single bond layer between two chips, it falls short in characterizing the critical z-dimension in a highly integrated chip stacking scenario. Therefore a robust solution is needed for both system alignment validation in the x,y plane as well as a capability to explore gap height and bump quality/buried issues in high aspect ratio TSV. This paper will explore an x-ray application of a stacked 15-layer die-to-interposer structure to qualitatively validate overall health of the TSV (bond, pad, fill) via a nondestructive volumetric method as well as quantitatively evaluate stacked die for via/ball/pad alignment to determine overall alignment shift during the bonding process. In the test vehicle for a flip chip FO-WLP process flow first presented by IMEC, a silicon (Si) bridge is utilized to connect the logic and through-package via dies with a bump pitch of 20µm for the logic die interconnect. This is done by carefully aligning the logic dies on the carrier and then placing the through-package dies aligned to the carrier and logic dies. In the final step, a high accuracy placement and stacking TCB tool attaches the Si bridge. To perform an accurate assessment for bump alignment and bond quality, an initial high-resolution 3D x-ray scan is performed on a 1.5mm3 region of interest with over 2,000 individual projections at a 1.4µm voxel resolution with a 4x objective, taking approximately two hours. The dataset is put through a deep-learning algorithm which generates an AI model. This model is then applied to a faster scan of 15 minutes with a 1.4µm voxel resolution for the same field of view (FOV) on a nearby array and the data is collected and processed. To verify alignment of the entire sample, data is collected on the tightest pitch (in this case 20µm) oriented in both the vertical (x) and horizontal (y) plane in the through-silicon via (TSV) region. The z-plane is analyzed and processed by computer vision algorithms, and individual interconnects are then quantified and sorted for solder height (z), width (x,y) and TSV misalignment (deviation from expected center, x1y1 → x2y2). To validate the model, a high-resolution uncorrected 20h scan and a 15min scan based on deep learning are collected for the same region of interest (ROI). The data is comparable. Furthermore, the global misalignment data between both datasets demonstrates similar misalignment propagating through the 15-layer stack.

BACKGROUND: The rapid development of industrialization in printed circuit board (PCB) warrants more complexity and integrity, which entails an essential procedure of PCB inspection. X-ray computed laminography (CL) enables inspection of arbitrary regions for large-sized flat objects with high resolution. PCB inspection based on CL imaging is worthy of exploration. OBJECTIVE: This work aims to extract PCB circuit layer information based on CL imaging through image segmentation technique. METHODS: In this work, an effective and applicable segmentation model for PCB CL images is established for the first time. The model comprises two components, with one integrating edge diffusion and l 0 smoothing to filter CL images with aliasing artifacts, and the other being the fuzzy energy-based active contour model driven by local pre-fitting energy to segment the filtered images. RESULT: The proposed model is able to suppress aliasing artifacts in the PCB CL images and has good performance on images of different circuit layers. CONCLUSIONS: Results of the simulation experiment reveal that the method is capable of accurate segmentation under ideal scanning condition. Testing of different PCBs and comparison of different segmentation methods authenticate the applicability and superiority of the model.

Back drilling technique is useful for high-speed printed circuit board (PCB) design. Currently, the stub length is measured by the destructive method that involves cutting the PCB and obtaining the results by optical imaging of the cut surface. X-ray computed laminography (CL) is an applicable technique for high-resolution imaging of plate-like object, which greatly contributes to stub nondestructive measurement. However, the interslice aliasing artifacts in the reconstructed images resulting from incomplete projection data make it difficult to position the trace layer and drilling tip. For better application to engineering, we propose to segment reconstructed slices (cross sections) of the PCB by self-training network and automatically locate key positions to measure stub length. Experiments on both simulated data and real data are carried out to verify the accuracy of the segmentation and measurement results for the proposed method. The proposed method can not only achieve nondestructive automatic rapid measurement but obtain similar results to the state-of-the-art destructive method, which brings great value to via stub measurement in practical engineering applications.

Cone-beam computed laminography (CL) is still a very challenging problem for the inspection of thin-plate objects. Since CL projections are incomplete, the reconstructed images always suffer from severe aliasing and blurring in the z direction. To mitigate this problem, we propose an anisotropic adaptive weighted total variation (AAwTV) reconstruction model, which takes the edge properties between adjacent voxels into account and introduces different weights in different directions. In addition, we solved the proposed AAwTV using the Chambolle-Pock (CP) framework, since it has good computational efficiency and stable convergence, and is often easy to get a satisfactory reconstruction result. Experiments on simulated PCB phantom and simulated workpiece phantom show that the proposed algorithm can preserve the detailed features of the object well, and can effectively suppress inter-slice aliasing and blurring.

Cone beam computed laminography (CL) has become one of the best methods for nondestructive testing of plate objects because of its particular scanning geometry. However, a challenging task in CL imaging is to deal with inter-slice aliasing caused by incomplete projection data from CL scanning. Using the classical total variation (TV) reconstruction method can reduce aliasing to a certain extent, but it often makes the image too smooth. In order to improve the quality of reconstructed CL image, a CL reconstruction algorithm based on generalized Charbonnier penalty is proposed in this article. In detail, the TV regular term is replaced by a nonconvex generalized Charbonnier penalty, and different weights are assigned to different directions. The original problem is divided into two subproblems to solve separately. The first subproblem uses Chambolle and Pock (CP) algorithm to solve the fidelity term, and the second subproblem uses additive half-quadratic minimization algorithm to denoise the image. Experiments were conducted using simulated printed circuit board (PCB) phantom, simulated workpiece phantom, and real data to verify the effectiveness of the proposed algorithm. The experimental results through visual assessment and quantitative assessment [including root-mean-square error (RMSE), multiscale structural similarity (MSSIM), Pearson correlation coefficient (PCC), and universal image quality index (UQI) indicators] demonstrate that the proposed method is effective in preserving image features and suppressing inter-slice aliasing in reconstructed CL images. The MATLAB code for the proposed algorithm is available at https://github.com/LIUyi827728/CL-Generalize-Charbonnier

The recognition of components in a Printed Circuit Board (PCB) is the foundation for conducting PCB inspection and analysis. Computed Laminography (CL) is a new X-ray non-destructive testing technology for PCB-like plate-shaped objects, therefore, the accuracy of component detection in PCB CL images is crucial. As a general image segmentation model, SAM performs excellently in many tasks. This paper improves SAM according to the characteristics of PCB images and proposes the PCB-SAM model: it designs a multi-scale semantic segmentation head, adopts the LoRA fine-tuning strategy to fine-tune the SAM encoder, and improves the Focal loss. We constructed a PCB CL dataset, and conducted training and testing of the PCB-SAM model based on this dataset. The experimental results show that the performance of this model has been significantly improved.

 Geometrical calibration of Cone Beam Computed Tomography (CBCT) systems is essential for achieving accurate image reconstruction and enhancing spatial resolution. A modern application of these systems is planar CT imaging (Computed Laminography), which is particularly effective for imaging flat objects, such as printed circuit boards (PCBs). Among the existing calibration approaches, self-calibration, which primarily utilizing mutual information as a similarity measure, has garnered significant attention[1]. In this study, we propose the use of regional mutual information[2] as an alternative similarity metric. Our findings demonstrate that regional mutual information enhances calibration accuracy and computational cost, offering a promising advancement in this field. 

Computed tomography (CT) is a widely used nondestructive testing (NDT) technique for material research, paleontology research and other fields. However, it is difficult for CT to reconstruct flat objects at high magnification ratios. Computed laminography (CL) enables high-resolution imaging for flat objects due to its unique scanning geometry. A challenging task for CL image reconstruction is to deal with the cross-section artifacts resulting from the incomplete projection data acquired from the CL scan. An effective multi-scale fusion reconstruction algorithm of CT and CL was proposed in this paper. The algorithm combining the advantages of the two scanning geometries, low-resolution CT data was used to compensate for the data missing in CL projection domain, and the cross-section artifacts were reduced. Experiments on paleontological fossils and multilayer printed circuit boards (PCB) were performed, where CT and CL data from different systems and scanning conditions. The results showed that the method can effectively suppress the cross-section artifacts of CL and obtain high-resolution reconstructed images.

The present-day electronic supply chain is infested with adversaries who threaten the integrity of electronic circuit boards and components. A major intent of such adversaries is to manufacture counterfeit printed circuit boards (PCBs). They infiltrate the supply chain with these forged PCBs, which closely mimic the original designs, albeit jeopardizing the business cycle of electronic design. In this work, we aim to restrict the circulation of tampered and counterfeit boards in the supply chain by leveraging the inherent physical deformities of authentic PCBs, which are difficult to replicate. Such anomalies include solder defects, material deposition, and microscopic irregularities unique to each PCB. These anomalies, though imperceivable to the human eye, can be captured at a specific angle under an X-ray microscope when imaged at appropriate orientations. Our proposed framework, TREX-F, comprises an image-based authentication protocol that defines unique fingerprints of each PCB by extracting such anomalies and converting them into quick-response and data-matrix codes. We construct device-specific templates from the X-ray computed tomography slices of the PCB samples by utilizing a combination of computer vision techniques, including Canny edge detection, contour detection, principal component analysis, and scale-invariant feature transform. As a case study, we validate our framework on Arduino UNO, Raspberry Pi 4 model B, and STM32F407G boards. TREX-F enables a sustainable electronics supply chain ecosystem, wherein end users can directly verify the authenticity of the procured PCBs with an average accuracy of ≥95% across all boards, and an average false acceptance rate (FAR) of 0. 1

X-ray imaging has been a necessary tool in many industrial inspection processes for over two decades. Most recently it has gained favor in the analysis of electronic Surface Mount Technology (SMT) components. SMT circuit processes are becoming more complex day by day. In 1986, the nascent SMT techniques were making six solder joints per square centimeter (6/cm 2 ). Today, it is possible to build an automated manufacturing line that can routinely incorporate over forty joints per square centimeter (40/cm 2 ). While a tremendous increase in the ability to mount more and more complex components to boards, this development has brought with it more opportunity for drastic changes in the Probability of Detection (POD) using previous inspection/testing methods. The Process Engineer needs tools to analyze the products that failed during operational testing. With most solder joints hidden between the board and the component, microfocus x-ray imaging has been successfully exploited to reveal and analyze these hidden solder joints. For example, when gull wing leads are used, a phenomenon such as a Pressure Open Lead may occur. Indiscernible from visual inspection and undetectable from a testing probe, this anomaly is easily detected using x-ray images. One particularly vexing problem is the analysis of voids in the solder ball grid array (BGA). Knowledge of the location within the solder ball of anomalies can indicate to the Process Engineer whether sufficient contact can be made with the board and the component. Until recently, the use of micro-focus x-ray technology to image the BGA has been limited two predominant approaches. One uses Real Time transmission x-ray cabinets using 5 axis manipulators, attempting to tilt the package so the defective solder ball(s) can be viewed from a near lateral perspective. The other technique is the utilization of laminar imaging techniques to actually slice through the solder ball. This concept employs split beam sources and/or conventional tomosynthesis reconstruction algorithms to generate the data. The tilting approach may not be suitable due to package location on the board. The latter concept involves extreme precision manipulators and/or more costly x-ray sources. A novel approach to the problem, Tuned Aperture Computed Tomography® (TACT®)has been used to generate such laminar data employing simple Real Time x-ray systems and only manipulation of the x axis and y axis . This 3-D Laminography algorithm was developed and patented at the Wake Forest University School of Medicine, licensed to CERBERUS Technologies, Inc., and commercialized as N-tact®. A series of (8) images are acquired using a simple image processor with a frame grabber. The resulting digital data set is analyzed utilizing the TACT laminography algorithm operating in a Windows based Personal Computer. The laminography algorithm uses reference markers to compute a translation for each image. The data are then reconstructed and shifted by the software algorithm to generate depth data on imaged anomalies. As the various anomalies are shifted and brought into “focus”, the depth at which this focusing occurs is registered. Data on anomalies at adjacent and other levels is blurred, with the blurring becoming more graduated as the depth separation between anomalies increases. Images at various depth levels of the BGA can be generated. Standard Image Processing tools available with the software can determine the area of any voids imaged at each level. Loss of mass may be determined using area and slice thickness. The ease of implementation and the validity of the information gathered shows a high degree of promise. Using this technology can facilitate SMT professionals in their Statistical Quality Control efforts to analyze BGA voiding on the high speed assembly line.

BGA defects can be challenging to detect when images are not taken from the appropriate perspective or with the correct equipment. While IPC's recommended tests are effective in certain situations, recent improvements in CT technology offer a more comprehensive analysis of each solder ball within seconds. Utilizing a genuine 3D CT method, defects such as Voiding, Non-Wetting, and Head-in-Pillow are distinctly visible, enabling engineers to conduct accurate root-cause analyses. 3D CT X-Ray inspection offers exceptional clarity when comparing defects against IPC standards and removes many of the visible obstructions introduced by other advanced X-Ray Technologies.

Based on the reliability application requirements of space borne products for thermal insulation adhesives, the process adaptability and environmental adaptability of newly two-component thermal insulation adhesive were tested and evaluated. Through the design and manufacture of the typical BGA device process test piece of the satellite terminal machine. the micro-pore filling process technology is used to fill the thermal insulation adhesive. After the industrial CT scanning detection, the filling void rate is less than 10%. At the same time, the module simulation parts of space borne products were designed and made, and filled with thermal insulation adhesive. The acceptance level environment and reliability tests of the simulated parts were carried out, such as temperature cycle, sinusoidal vibration, random vibration and thermal vacuum. The electrical performance indicators and monitoring process of the simulator before and after the environmental test have no obvious differences. and meet the index design requirements. There are no defects introduced in the process of filling the thermal insulation adhesive in the simulator which verifies that the simulator has the ability to withstand the working stress and environmental stress, and the thermal insulation adhesive has no abnormal impact on the electrical performance index of the simulator, which meets the requirements of heat dissipation and high load resistance of space borne products. and realizes reliable application.

Anomaly detection is an active research field in industrial defect detection and medical disease detection. However, previous anomaly detection works suffer from unstable training, or non-universal criteria of evaluating feature distribution. In this paper, we introduce UTRAD, a U-TRansformer based Anomaly Detection framework. Deep pre-trained features are regarded as dispersed word tokens, and represented with transformer-based autoencoders. With reconstruction on more informative feature distribution instead of raw images, we achieve a more stable training process and a more precise anomaly detection and localization result. In addition, our proposed UTRAD has a multi-scale pyramidal hierarchy with skip connections that help detect both multi-scale structural and non-structural anomalies. As attention layers are decomposed to multi-level patches, UTRAD significantly reduces the computational cost and memory usage compared with the vanilla transformer. Experiments on industrial dataset MVtec AD and medical datasets Retinal-OCT, Brain-MRI, Head-CT have been conducted. Our proposed UTRAD out-performs all other state-of-the-art methods in the above datasets. Code released at https://github.com/gordon-chenmo/UTRAD.

Failure analysis of electronic components is a complex discipline, involving a variety of analytical tools, including optical microscopy, electron microscopy and other techniques. X-ray radiography, both 2D X-ray and 3D computed tomography (CT), have been shown to be powerful techniques in diagnosing failure modes of surface mount electronic components as well as visualizing defects in the printed wiring boards themselves. In this paper, a few case studies involving 3D CT as a non-destructive failure analysis tool of electronics will be discussed. The case studies will include analysis and discussion of ball grid array (BGA) features, such as mixed metallurgy and void content as well as diagnosing printed wiring board assembly issues, such as handling issues and overall failure analysis.

X-ray inspection systems are key tools for quality control, yield enhancement, and failure analysis of advanced printed circuits board assemblies (PCBA). The electronics industry has witnessed significant improvements in the X-ray inspection capabilities (AXI, 2DX, 3D CT, Large Board CT) during the last twenty years. These advancements have aided in the development of new high performance packages and PCBA processes. In many cases, X-Ray inspection is the only non-destructive technique to inspect optically hidden components and solder joints such as BGA, POP, QFN, flip chips, through holes, etc. Automated X-Ray inspection (AXI) has been available since 1999 at our manufacturing sites, and plays very important role to assure high quality SMT process. Along the way, manual X-ray systems (2DX and 3D CT) have been providing invaluable testing capabilities for verifying the AXI results, fine tuning the AXI parameters, process development, and X-ray inspection that is out of the scope of the AXI equipment. Through the years these capabilities have been significantly improved and best practices have been developed on how to jointly use the AXI, 2DX(MXI), and 3D CT X-ray equipment to achieve best possible testing results and speed. In this paper, we will summarise the extensive experience we have accumulated while working on real problems together with our customers, X-ray inspection vendors and R&D teams. These projects include BGA HIP (head in pillow), solder charge connector, Package on Package, package 03015. In addition, we will discuss how to optimize test coverage and eliminate defect escapes during PCBA testing. These studies have been performed using variety of X-ray inspection systems at several production sites and lasted several years. We will also discuss the X-ray inspection capabilities that are considered most important to the PCBA manufacturing process including: 1. Highest resolution X-ray images and clear distinction between good and defective pins; 2. Measurement data from inspection output; 3. Operator independent automated results for pass/fail condition; 4. Non-destructive e-cross section; 5. Easy and intuitive operation and programing; 6. Real time data feedback; 7. Flexible automated algorithms that are easily adjustable by the machine operator; 8. Zero defect escape with reasonable false call PPM.

For years industry has used traditional, two dimensional, transmission X-Ray as a means of exploring hidden structures such as BGA Solder Joints. Three-dimensional techniques such as Laminography and Tomosynthesis allow viewing of slices, or cross sections, within a 3D structure so as to eliminate confusion caused by double sided boards. Computed Tomography provides a more useful, true three-dimensional view of the entire sample including vertical location and extent of defects within a solid structure. This paper will look at the latest CT technology and present some practical applications.

PCBs are indispensable components that enable the functionality of modern electronics across numerous industries. This study presents a framework for their semi-supervised netlist extraction by addressing the challenges during X-ray tomography. Leveraging AI models during annotation saves time while allowing SMEs to fix AI-generated errors.

Deep Learning is being widely used to identify and segment various 2D and 3D structures in voxelized data in fields such as robotics and medical imaging. Automated object detection and segmentation has had a rich history in semicon inspection and defect detection technologies for past few decades. Deep learning-based object detection and image segmentation has the potential to further improve defect detection accuracy and reduce manpower required for the quality inspection process. We develop a novel framework that utilizes the advancements in deep learning-based object detection and image segmentation techniques to leverage on partial labeled data and remaining unlabeled data to significantly improve the performance of locating microscopic bumps and defects such as voids for the defect detection process. We apply our Semi-Supervised Learning approach on various buried structures such as memory bumps and logic bumps. We briefly describe our fabrication and scanning process and thereafter, explain our approach in locating these different structures in 3D scans in detail. We extract the virtual 2D slices from 3D scans, perform Semi-Supervised object detection and image segmentation to classify each pixel of these individual slices into solders, voids, Cu-Pillars, and Cu-Pads. We compare our approach with state-of-the-art fully supervised techniques and perform a thorough analysis to discuss the advantages and disadvantages of our approach in both object detection and image segmentation steps.

No abstract available

Printed circuit board (PCB) defect detection plays a crucial role in PCB production, and the popular methods are based on deep learning and require large-scale datasets with high-level ground-truth labels, in which it is time-consuming and costly to label these datasets. Semi-supervised learning (SSL) methods, which reduce the need for labeled samples by leveraging unlabeled samples, can address this problem well. However, for PCB defects, the detection accuracy on small numbers of labeled samples still needs to be improved because the number of labeled samples is small, and the training process will be disturbed by the unlabeled samples. To overcome this problem, this paper proposed a semi-supervised defect detection method with a data-expanding strategy (DE-SSD). The proposed DE-SSD uses both the labeled and unlabeled samples, which can reduce the cost of data labeling, and a batch-adding strategy (BA-SSL) is introduced to leverage the unlabeled data with less disturbance. Moreover, a data-expanding (DE) strategy is proposed to use the labeled samples from other datasets to expand the target dataset, which can also prevent the disturbance by the unlabeled samples. Based on the improvements, the proposed DE-SSD can achieve competitive results for PCB defects with fewer labeled samples. The experimental results on DeepPCB indicate that the proposed DE-SSD achieves state-of-the-art performance, which is improved by 4.7 mAP at least compared with the previous methods.

Defect inspection is essential in the semiconductor industry to fabricate printed circuit boards (PCBs) with minimum defect rates. However, conventional inspection systems are labor-intensive and time-consuming. In this study, a semi-supervised learning (SSL)-based model called PCB_SS was developed. It was trained using labeled and unlabeled images under two different augmentations. Training and test PCB images were acquired using automatic final vision inspection systems. The PCB_SS model outperformed a completely supervised model trained using only labeled images (PCB_FS). The performance of the PCB_SS model was more robust than that of the PCB_FS model when the number of labeled data is limited or comprises incorrectly labeled data. In an error-resilience test, the proposed PCB_SS model maintained stable accuracy (error increment of less than 0.5%, compared with 4% for PCB_FS) for noisy training data (with as much as 9.0% of the data labeled incorrectly). The proposed model also showed superior performance when comparing machine-learning and deep-learning classifiers. The unlabeled data utilized in the PCB_SS model helped with the generalization of the deep-learning model and improved its performance for PCB defect detection. Thus, the proposed method alleviates the burden of the manual labeling process and provides a rapid and accurate automatic classifier for PCB inspections.

No abstract available

Automated optical inspection (AOI) is widely used by manufacturers for the detection of defects in printed circuit boards (PCBs). Recent works have proposed to apply deep learning for defect detection, which is much faster and cheaper than manual inspection. However, AOI can only capture defects on the outmost layers of PCBs using cameras, while modern high-speed circuit PCBs usually have multiple internal layers that need to be inspected. Compared to optical sensors, X-ray tomography provides noninvasive imaging results of all PCB layers. Though one can directly apply an off-the-shelf deep detection model trained on optical domains for X-ray imagery, we show that it usually leads to much lower accuracies in practice. The degraded performance is mainly due to the relatively low quality of X-ray imaging results and the gaps between optical and X-ray modalities. Furthermore, no X-ray PCB image dataset is publicly available for training deep defect detectors. To this end, we propose a novel dataset for X-ray PCB defect detection, dubbed XD-PCB. In XD-PCB, we provide a benchmark for training X-ray automated defect detection models containing synthesized X-ray images and real X-ray images with real defects. However, in a practical environment, retraining the deep model for every unseen X-ray domain is inefficient due to the domain gaps created by different X-ray machine settings and the scarcity of defects. Thus, we propose a domain adaptation framework, dubbed feature-based domain adaptation X-ray (FDX), to improve the efficiency of X-ray PCB defect detection methods. By minimizing the differences between the deep features extracted from abundant training images and the scarce unseen images, we improve the model’s performance in a practical situation, thus enhancing the generalization ability and efficiency of deep detection algorithms when exposed to unseen domains. Our results demonstrate that XD-PCB provides a valuable training baseline for X-ray PCB defect detection, and our proposed FDX framework can effectively increase the popular deep learning model by achieving an increment of 10% in terms of average precisions (APs) compared to other adaptation methods.

Recent advancements in 3D deep learning have led to significant progress in improving accuracy and reducing processing time, with applications spanning various domains such as medical imaging, robotics, and autonomous vehicle navigation for identifying and segmenting different structures. In this study, we employ the latest developments in 3D semi-supervised learning to create cutting-edge models for the 3D object detection and segmentation of buried structures in high-resolution X-ray semiconductors scans. We illustrate our approach to locating the region of interest of the structures, their individual components, and their void defects. We showcase how semi-supervised learning is utilized to capitalize on the vast amounts of available unlabeled data to enhance both detection and segmentation performance. Additionally, we explore the benefit of contrastive learning in the data pre-selection step for our detection model and multi-scale Mean Teacher training paradigm in 3D semantic segmentation to achieve better performance compared with the state of the art. Our extensive experiments have shown that our method achieves competitive performance and is able to outperform by up to 16% on object detection and 7.8% on semantic segmentation. Additionally, our automated metrology package shows a mean error of less than 2 μm for key features such as Bond Line Thickness and pad misalignment.

Being able to spot defective parts is a critical component in large-scale industrial manufacturing. A particular challenge that we address in this work is the cold-start problem: fit a model using nominal (non-defective) example images only. While handcrafted solutions per class are possible, the goal is to build systems that work well simultaneously on many different tasks automatically. The best peforming approaches combine embeddings from ImageNet models with an outlier detection model. In this paper, we extend on this line of work and propose PatchCore, which uses a maximally representative memory bank of nominal patch-features. PatchCore offers competitive inference times while achieving state-of-the-art performance for both detection and localization. On the challenging, widely used MVTec AD benchmark PatchCore achieves an image-level anomaly detection AUROC score of up to 99.6%, more than halving the error compared to the next best competitor. We further report competitive results on two additional datasets and also find competitive results in the few samples regime. Code: github.com/amazon-research/patchcore-inspection.

Anomaly detection is an important application in large-scale industrial manufacturing. Recent methods for this task have demonstrated excellent accuracy but come with a latency trade-off. Memory based approaches with dominant performances like PatchCore or Coupled-hypersphere-based Feature Adaptation (CFA) require an external memory bank, which significantly lengthens the execution time. Another approach that employs Reversed Distillation (RD) can perform well while maintaining low latency. In this paper, we revisit this idea to improve its performance, establishing a new state-of-the-art benchmark on the challenging MVTec dataset for both anomaly detection and localization. The proposed method, called RD++, runs six times faster than PatchCore, and two times faster than CFA but introduces a negligible latency compared to RD. We also experiment on the BTAD and Retinal OCT datasets to demonstrate our method's generalizability and conduct important ablation experiments to provide insights into its configurations. Source code will be available at https://github.com/tientrandinh/Revisiting-Reverse-Distillation.

Self-supervised feature reconstruction methods have shown promising advances in industrial image anomaly de-tection and localization. Despite this progress, these meth-ods still face challenges in synthesizing realistic and di-verse anomaly samples, as well as addressing the feature redundancy and pre-training bias of pre-trained feature. In this work, we introduce RealNet, a feature reconstruction network with realistic synthetic anomaly and adaptive feature selection. It is incorporated with three key inno-vations: First, we propose Strength-controllable Diffusion Anomaly Synthesis (SDAS), a diffusion process-based syn-thesis strategy capable of generating samples with varying anomaly strengths that mimic the distribution of real anomalous samples. Second, we develop Anomaly-aware Features Selection (A FS), a method for selecting repre-sentative and discriminative pre-trained feature subsets to improve anomaly detection performance while controlling computational costs. Third, we introduce Reconstruction Residuals Selection (RRS), a strategy that adaptively selects discriminative residuals for comprehensive identification of anomalous regions across multiple levels of granularity. We assess RealNet onfour benchmark datasets, and our results demonstrate significant improvements in both Image AU-Rae and Pixel AUROC compared to the current state-of-the-art methods. The code, data, and models are available at https://github.com/cnulab/RealNet.

We aim at constructing a high performance model for defect detection that detects unknown anomalous patterns of an image without anomalous data. To this end, we propose a two-stage framework for building anomaly detectors using normal training data only. We first learn self-supervised deep representations and then build a generative one-class classifier on learned representations. We learn representations by classifying normal data from the CutPaste, a simple data augmentation strategy that cuts an image patch and pastes at a random location of a large image. Our empirical study on MVTec anomaly detection dataset demonstrates the proposed algorithm is general to be able to detect various types of real-world defects. We bring the improvement upon previous arts by 3.1 AUCs when learning representations from scratch. By transfer learning on pretrained representations on ImageNet, we achieve a new state-of-the-art 96.6 AUC. Lastly, we extend the framework to learn and extract representations from patches to allow localizing defective areas without annotations during training.

Visual anomaly detection is commonly used in industrial quality inspection. In this paper, we present a new dataset as well as a new self-supervised learning method for ImageNet pre-training to improve anomaly detection and segmentation in 1-class and 2-class 5/10/high-shot training setups. We release the Visual Anomaly (VisA) Dataset consisting of 10,821 high-resolution color images (9,621 normal and 1,200 anomalous samples) covering 12 objects in 3 domains, making it the largest industrial anomaly detection dataset to date. Both image and pixel-level labels are provided. We also propose a new self-supervised framework - SPot-the-difference (SPD) - which can regularize contrastive self-supervised pre-training, such as SimSiam, MoCo and SimCLR, to be more suitable for anomaly detection tasks. Our experiments on VisA and MVTec-AD dataset show that SPD consistently improves these contrastive pre-training baselines and even the supervised pre-training. For example, SPD improves Area Under the Precision-Recall curve (AU-PR) for anomaly segmentation by 5.9% and 6.8% over SimSiam and supervised pre-training respectively in the 2-class high-shot regime. We open-source the project at http://github.com/amazon-research/spot-diff .

Industrial anomaly detection (I AD) has garnered signif-icant attention and experienced rapid development. However, the recent development of I AD approach has encountered certain difficulties due to dataset limitations. On the one hand, most of the state-of-the-art methods have achieved saturation (over 99% in AUROC) on mainstream datasets such as MVTec, and the differences of methods cannot be well distinguished, leading to a significant gap between public datasets and actual application scenarios. On the other hand, the research on various new practical anomaly detection settings is limited by the scale of the dataset, posing a risk of overfitting in evaluation results. Therefore, we propose a large-scale, Real-world, and multi-view Industrial Anomaly Detection dataset, named Real- I AD, which contains 150K high-resolution images of 30 different objects, an order of magnitude larger than existing datasets. It has a larger range of defect area and ratio proportions, making it more challenging than previous datasets. To make the dataset closer to real application scenarios, we adopted a multi-view shooting method and proposed sample-level evaluation metrics. In addition, beyond the general unsupervised anomaly detection setting, we propose a new setting for Fully Unsupervised Indus-trial Anomaly Detection (FUIAD) based on the observation that the yield rate in industrial production is usually greater than 60%, which has more practical application value. Finally, we report the results of popular I AD methods on the Real- I AD dataset, providing a highly challenging benchmark to promote the development of the I AD field.

Anomaly synthesis strategies can effectively enhance unsupervised anomaly detection. However, existing strategies have limitations in the coverage and controllability of anomaly synthesis, particularly for weak defects that are very similar to normal regions. In this paper, we propose Global and Local Anomaly co-Synthesis Strategy (GLASS), a novel unified framework designed to synthesize a broader coverage of anomalies under the manifold and hypersphere distribution constraints of Global Anomaly Synthesis (GAS) at the feature level and Local Anomaly Synthesis (LAS) at the image level. Our method synthesizes near-in-distribution anomalies in a controllable way using Gaussian noise guided by gradient ascent and truncated projection. GLASS achieves state-of-the-art results on the MVTec AD (detection AUROC of 99.9\%), VisA, and MPDD datasets and excels in weak defect detection. The effectiveness and efficiency have been further validated in industrial applications for woven fabric defect detection. The code and dataset are available at: \url{https://github.com/cqylunlun/GLASS}.

Unsupervised Anomaly Detection (UAD) with incremental training is crucial in industrial manufacturing, as unpredictable defects make obtaining sufficient labeled data infeasible. However, continual learning methods primarily rely on supervised annotations, while the application in UAD is limited due to the absence of supervision. Current UAD methods train separate models for different classes sequentially, leading to catastrophic forgetting and a heavy computational burden. To address this issue, we introduce a novel Unsupervised Continual Anomaly Detection framework called UCAD, which equips the UAD with continual learning capability through contrastively-learned prompts. In the proposed UCAD, we design a Continual Prompting Module (CPM) by utilizing a concise key-prompt-knowledge memory bank to guide task-invariant 'anomaly' model predictions using task-specific 'normal' knowledge. Moreover, Structure-based Contrastive Learning (SCL) is designed with the Segment Anything Model (SAM) to improve prompt learning and anomaly segmentation results. Specifically, by treating SAM's masks as structure, we draw features within the same mask closer and push others apart for general feature representations. We conduct comprehensive experiments and set the benchmark on unsupervised continual anomaly detection and segmentation, demonstrating that our method is significantly better than anomaly detection methods, even with rehearsal training. The code will be available at https://github.com/shirowalker/UCAD.

2D-based Industrial Anomaly Detection has been widely discussed, however, multimodal industrial anomaly detection based on 3D point clouds and RGB images still has many untouched fields. Existing multimodal industrial anomaly detection methods directly concatenate the multimodal features, which leads to a strong disturbance between features and harms the detection performance. In this paper, we propose Multi-3D-Memory (M3DM), a novel multimodal anomaly detection method with hybrid fusion scheme: firstly, we design an unsupervised feature fusion with patch-wise contrastive learning to encourage the interaction of different modal features; secondly, we use a decision layer fusion with multiple memory banks to avoid loss of information and additional novelty classifiers to make the final decision. We further propose a point feature alignment operation to better align the point cloud and RGB features. Extensive experiments show that our multi-modal industrial anomaly detection model outperforms the state-of-the-art (SOTA) methods on both detection and segmentation precision on MVTec-3D AD dataset. Code at github.com/nomewang/M3DM.

This paper presents LogiCode, a novel framework that leverages Large Language Models (LLMs) for identifying logical anomalies in industrial settings, moving beyond the traditional focus on structural inconsistencies. By harnessing LLMs for logical reasoning, LogiCode autonomously generates Python codes to pinpoint anomalies such as incorrect component quantities or missing elements, marking a significant leap forward in anomaly detection technologies. A custom dataset “LOCO-Annotations” and a benchmark “LogiBench” are introduced to evaluate the LogiCode’s performance across various metrics including binary classification accuracy, code generation success rate, and precision in reasoning. Findings demonstrate LogiCode’s enhanced interpretability, significantly improving the accuracy of logical anomaly detection and offering detailed explanations for identified anomalies. This represents a notable shift towards more intelligent, LLM-driven approaches in industrial anomaly detection, promising substantial impacts on industry-specific applications. Our code are available at https://github.com/22strongestme/LOCO-Annotations. Note to Practitioners—This work introduces LogiCode, an innovative system leveraging Large Language Models (LLMs) for logical anomaly detection in industrial settings, shifting the paradigm from traditional visual inspection methods. LogiCode autonomously generates Python codes for logical anomaly detection, enhancing interpretability and accuracy. Our novel approach, validated through the “LOCO-Annotations” dataset and LogiBench benchmark, demonstrates superior performance in identifying logical anomalies, a challenge often encountered in complex industrial components like assembly and packaging. LogiCode provides a significant advancement in addressing the nuanced requirements of detecting logical anomalies, offering a robust and interpretable solution to practitioners seeking to enhance quality control and reduce manual inspection efforts.

The defect detection of printed circuit board (PCB) images faces challenges such as limited sample number, imbalanced sample types, and varying detection reliability. To address these issues, this article proposes an uncertainty-aware unsupervised detection model on PCB images, short for <inline-formula> <tex-math notation="LaTeX">$\rm { U^{2}D^{2}}$ </tex-math></inline-formula>PCB. The proposed method uses two U-Net networks to serve as the reconstructive subnetwork and the discriminative subnetwork, respectively. The former one reconstructs defect-free PCB images from defective PCB images, while the latter segments the defects and evaluates the defects uncertainty with the concatenated inputs of the defective and reconstructed images. The <inline-formula> <tex-math notation="LaTeX">$\mathbf { U^{2}D^{2}}$ </tex-math></inline-formula>PCB model is trained in an unsupervised manner with only defect-free images embedding with multiscale artificial defects. Experimental results on the public PCB defect dataset and DeepPCB dataset demonstrate the effectiveness of the proposed method. The mean average precision (mAP) is 99.29% on the PCB defect dataset, while it reaches 95.78% on the DeepPCB dataset. These results are competitive to those of state-of-the-art (SOTA) fully supervised methods. The findings of <inline-formula> <tex-math notation="LaTeX">$\mathrm { U^{2}D^{2}}$ </tex-math></inline-formula>PCB highlight the potential significance of using unsupervised learning techniques for PCB defect detection.

Printed Circuit Boards (PCBs) are fundamental components in modern electronic devices, and their quality directly affects product reliability. Conventional inspection processes often rely on supervised learning models that require a large amount of annotated defect data. However, collecting and labeling defect samples is costly and impractical due to the diverse and unpredictable nature of PCB defects. In this study, we propose an unsupervised learning-based defect detection framework that leverages PCB circuit images for automated quality inspection. The approach employs convolutional autoencoders to learn standard pattern representations without requiring labeled defect data. Anomalous regions are detected by reconstruction error analysis, and statistical thresholds are applied to classify pass/fail conditions. To evaluate the effectiveness, experiments were conducted on a dataset of PCB circuit images, including both standard and defective samples. The proposed method demonstrated high detection accuracy with minimal false positives, outperforming baseline thresholding and clustering-based approaches. Moreover, the framework showed robustness to varying defect types such as missing tracks, misalignments, and surface contamination. These results suggest that unsupervised learning can serve as a practical alternative to traditional supervised inspection methods in PCB manufacturing. The proposed method reduces dependency on labeled datasets, enhances adaptability to new defect patterns, and contributes to the realization of intelligent and automated quality inspection systems in smart manufacturing environments.

The usage of electronic devices increases, and becomes predominant in most aspects of life. Surface Mount Technology (SMT) is the most common industrial method for manufacturing electric devices in which electrical components are mounted directly onto the surface of a Printed Circuit Board (PCB). Although the expansion of electronic devices affects our lives in a productive way, failures or defects in the manufacturing procedure of those devices might also be counterproductive and even harmful in some cases. It is therefore desired and sometimes crucial to ensure zero-defect quality in electronic devices and their production. While traditional Image Processing (IP) techniques are not sufficient to produce a complete solution, other promising methods like Deep Learning (DL) might also be challenging for PCB inspection, mainly because such methods require big adequate datasets which are missing, not available or not updated in the rapidly growing field of PCBs. Thus, PCB inspection is conventionally performed manually by human experts. Unsupervised Learning (UL) methods may potentially be suitable for PCB inspection, having learning capabilities on the one hand, while not relying on large datasets on the other. In this paper, we introduce ChangeChip, an automated and integrated change detection system for defect detection in PCBs, from soldering defects to missing or misaligned electronic elements, based on Computer Vision (CV) and UL. We achieve good quality defect detection by applying an unsupervised change detection between images of a golden PCB (reference) and the inspected PCB under various setting. In this work, we also present CD-PCB, a synthesized labeled dataset of 20 pairs of PCB images for evaluation of defect detection algorithms. The sources of ChangeChip, as well as CD-PCB, are available at: https://github.com/Scientific-Computing-Lab-NRCN/ChangeChip.

No abstract available

This paper presents a robust inspection framework for detecting non-wet defects in semiconductor solder joints using 3D CT slice imaging and supervised learning. The proposed method leverages a slice-level ResNet18 classifier combined with a tunable classification confidence parameter to predict defective slices. These slice-level predictions are then aggregated to determine the volume-level defect status through a slice-counting strategy. To accommodate varying defect characteristics across semiconductor packages, we introduce an adjustable defect count threshold and validate its impact on detection performance. Experiments show that the method achieves perfect recall with zero false positives under optimal settings and maintains a stable range across thresholds, outperforming traditional unsupervised and feature-based baselines. The proposed approach is lightweight, adaptable, and requires no retraining to adjust sensitivity, making it well-suited for deployment in real-world inspection pipelines. This work demonstrates the practical synergy of 3D imaging and machine learning in enhancing reliability and efficiency in semiconductor manufacturing. Our codes and data are released at here.

Automatic anomaly detection on engineering structures is often carried out using supervised models, raising the issue of anomalous images acquisition and annotation. Unsupervised methods like normalizing flows achieve excellent results while trained with defect-free images only. However, normalizing flows methods, such as MSFlow, are generally applied on features extracted by an encoder pre-trained on datasets that may not be related to engineering structures images. Therefore, we investigate the possibility to derive more discriminative features with an additional fine-tuning of the feature extractor on images with synthetic anomalies. We consider two types of such anomalies and demonstrate their efficiency with MSFlow on the MVTec (Wood/Tile) and Crack500 datasets, with significantly improved predictions. Interestingly, both tasks produce similar results suggesting that pre-training is mainly improved by the healthy part of images and not very sensitive to anomaly realism. Additionally, when comparing our fine-tuned MSFlow with a reference supervised model, CT-CrackSeg, on the Crack500 dataset, we observe similar qualitative behaviours. This open a promising direction towards annotation-free, more scalable alternatives, in particular for anomaly detection in engineering structure applications.

As Deep Neural Network (DNN)-based algorithms are improving, pivotal changes are happening towards efficient and effective automation in the field of industrial inspection. In the scope of our project, we analyze x-ray images of metal pipelines to detect the presence of corrosion in a novel way. In our industrial scenario, a drone lands a crawler that is equipped with an x-ray system on top of insulated pipelines to perform X-ray scans which are able to penetrate only the insulation, due to power consumption limitations. In this paper, we use modern unsupervised anomaly detection algorithms to detect the presence of corrosion, yielding quite promising results. Moreover, to compare several state-of-the-art approaches in terms of robustness to image degradation, we simulate two types of degradation that can occur: (i) Poisson Noise, (ii) Motion Blur Deformation. We conclude that the problem we are dealing with can be handled sufficiently well with state-of-the-art approaches, and that in the scenario of image degradation, the most robust algorithms are based on memory banks and teacher-student architectures.

The problem of detecting dangerous or prohibited objects in luggage is a very important step during the implementation of Security setup at Airports, Banks, Government buildings, etc. At present, the most common techniques for detecting such dangerous objects are by using intelligent data analysis algorithms such as deep learning techniques on X-ray imaging or employing a human workforce for inferring the presence of these threat objects in the obtained X-ray images. One of the major challenges while using deep-learning methods to detect such objects is the lack of high-quality threat image data containing the “dangerous” objects (objects of interest) versus the non-threat image data in practical scenarios. So, to tackle this data scarcity problem, anomaly detection techniques using normal data samples have shown great promise. Also, among the available Deep Learning Strategies for anomaly detection for computer vision applications, generative adversarial networks have achieved state-of-the-art results. Considering these insights, we adopted a newly proposed architecture known as Skip-GANomaly and devised a modified version of it by using a UNet++ style generator which performed better than Skip-GANomaly, getting an AUC of 94.94% on Compass-XP, a public X-ray dataset. Finally, for targeting better latent space exploration, we combine these two architectures into an Ensemble, which gives another boost to the performance, getting an AUC of 96.8% on the same Compass-XP, a public X-ray dataset. To further validate the effectiveness of ensemble-based architecture, its performance was tested on patch-based training data on a subset of randomly chosen images of another huge public X-ray dataset named as SIXray, and obtained an AUC of 75.3% on this reduced dataset. To demonstrate the prowess of the discriminator and to bring some explainability to the working of our ensemble, we have used Uniform Manifold Approximation and Projection to plot the latent-space vectors for the dangerous and non-dangerous objects of the test-set; this analysis indicates that the Ensemble learns better features for separating the anomalous class from non-anomalous with respect to the individual architectures. Thus, our proposed architecture provides state-of-the-art results for threat object detection. Most importantly, our models are able to detect threat objects without ever being trained on images containing threat objects.

3D imaging via X‐ray microscopy (XRM), a form of tomography, is revolutionising materials characterisation. Nondestructive imaging to classify grains, particles, interfaces and pores at various scales is imperative for our understanding of the composition, structure, and failure of building materials. Various workflows now exist to maximise data collection and to push the boundaries of what has been achieved before, either from singular instruments, software or combinations through multimodal correlative microscopy. An evolving area on interest is the XRM data acquisition and data processing workflow; of particular importance is the improvement of the data acquisition process of samples that are challenging to image, usually because of their size, density (atomic number) and/or the resolution they need to be imaged at. Modern advances include deep/machine learning and AI resolutions for this problem, which address artefact detection during data reconstruction, provide advanced denoising, improved quantification of features, upscaling of data/images, and increased throughput, with the goal to enhance segmentation and visualisation during postprocessing leading to better characterisation of samples. Here, we apply three AI and machine‐learning‐based reconstruction approaches to cements and concretes to assist with image improvement, faster throughput of samples, upscaling of data, and quantitative phase identification in 3D. We show that by applying advanced machine learning reconstruction approaches, it is possible to (i) vastly improve the scan quality and increase throughput of ‘thick’ cores of cements/concretes through enhanced contrast and denoising using DeepRecon Pro, (ii) upscale data to larger fields of view using DeepScout and (iii) use quantitative automated mineralogy to spatially characterise and quantify the mineralogical/phase components in 3D using Mineralogic 3D. These approaches significantly improve the quality of collected XRM data, resolve features not previously accessible, and streamline scanning and reconstruction processes for greater throughput.