Recently, many companies have introduced automated defect detection methods for defect-free PCB manufacturing. In particular, deep learning-based image understanding methods are very widely used. In this study, we present an analysis of training deep learning models to perform PCB defect detection stably. To this end, we first summarize the characteristics of industrial images, such as PCB images. Then, the factors that can cause changes (contamination and quality degradation) to the image data in the industrial field are analyzed. Subsequently, we organize defect detection methods that can be applied according to the situation and purpose of PCB defect detection. In addition, we review the characteristics of each method in detail. Our experimental results demonstrated the impact of various degradation factors, such as defect detection methods, data quality, and image contamination. Based on our overview of PCB defect detection and experiment results, we present knowledge and guidelines for correct PCB defect detection.
In this study, we present a framework for product quality inspection based on deep learning techniques. First, we categorize several deep learning models that can be applied to product inspection systems. In addition, we explain the steps for building a deep-learning-based inspection system in detail. Second, we address connection schemes that efficiently link deep learning models to product inspection systems. Finally, we propose an effective method that can maintain and enhance a product inspection system according to improvement goals of the existing product inspection systems. The proposed system is observed to possess good system maintenance and stability owing to the proposed methods. All the proposed methods are integrated into a unified framework and we provide detailed explanations of each proposed method. In order to verify the effectiveness of the proposed system, we compare and analyze the performance of the methods in various test scenarios. We expect that our study will provide useful guidelines to readers who desire to implement deep-learning-based systems for product inspection.
Conventional multi-object tracking (MOT) methods often fail to track objects that repeatedly disappear and reappear in the scene due to occlusion or moving off the screen. Previous studies have used motion and appearance information to address these issues but have limitations in dealing with complex scenarios. To overcome these limitations, we propose a post-hoc framework for MOT, a so-called rethinking MOT. The proposed framework efficiently combines MOT and re-identification methods to handle tracking failures in conventional MOT methods. To this end, we proposed a track reliability checking method to determine unreliable tracks and performed re-identification to correct incorrect tracking results. For re-identification, we proposed appearance aspect-ratio matching that is practical and effective. In addition, we designed an online appearance model management for accurate and efficient re-identification. The proposed framework does not require any offline training stages, such as feature or metric learning, but it is very practical for comparing appearances in real time. In addition, it is a flexible and pluggable framework that can adopt any baseline methods for MOT and Re-id. To validate the proposed methods, we collected 12 new challenging indoor datasets. Unlike the existing benchmark MOT datasets, we consider complex indoor sequences where targets reappear many times in the scene. The experimental results reveal that the proposed methods are promising for MOT under various challenging sequences and outperform state-of-the-art methods. Compared to baseline of the proposed rethinking MOT, the IDF 1 score was increased by 36.9% in own dataset and 50.4% in the benchmark dataset.
In this study, we address the challenge of 3D human motion prediction from motion capture data, which has become critical in various applications such as autonomous vehicles and human-robot interaction. Previous deep learning-based methods have improved prediction accuracy, but require significant network parameters and do not effectively consider independent joint movements. To overcome the limitations, we propose two lightweight network structures for human motion prediction: LG-Net and LGT-Net, which focus on the individual movements of distinct human limbs and their inter-dependencies. The LG-Net comprises local and global networks, while the LGT-Net combines the proposed LG-Net structure with Long and Short Term Memory (LSTM) cells to exploit temporal information. Our networks, designed to be extremely lightweight with only 0.08M and 0.5M parameters, achieve higher prediction performance compared to state-of-the-art methods. In addition, this study is the first to consider the root joint to improve motion prediction performance. The proposed approach demonstrates the potential for efficient and accurate human motion prediction in various applications.
In recent years, many semantic segmentation methods have been proposed to predict label of pixels in the scene. In general, we measure area prediction errors or boundary prediction errors for comparing methods. However, there is no intuitive evaluation metric that evaluates both aspects. In this work, we propose a new evaluation measure called weighted Intersection over Union (wIoU) for semantic segmentation. First, it build a weight map generated from a boundary distance map, allowing weighted evaluation for each pixel based on a boundary importance factor. The proposed wIoU can evaluate both contour and region by setting a boundary importance factor. We validated the effectiveness of wIoU on a dataset of 33 scenes and demonstrated its flexibility. Using the proposed metric, we expect more flexible and intuitive evaluation in semantic segmentation filed are possible.