Title: Defect detection of IC substrate based on Bi-direction Feature Pyramid Network

Biography:

Hao-Wei Lin received the B.S. degree in electrical engineering in 2019 from Tamkang University, Tamsui, Taiwan. He is a master student in the Graduate Institute of Automation Technology at National Taipei University of Technology, Taipei, Taiwan. He mainly researchs machine vision and makes applications in the automation industry. In most of his research time, He spent time researching the application of deep learning in automated optical inspection.

Abstract

Automatic optical inspection is a typical technology in industrial processes. In recent years, deep learning has been widely used in automated optical inspection (AOI). Deep learning used in AOI brings benefit of defect detection performance. In this paper, the IC substrate is the inspection object. It depends on low-level visual task, which mainly involves the detection of structural features. In contrast, most of the current object detection networks deal with high-level visual tasks, and can’t achieve good results in defect detection of IC substrate. In the case of Feature Pyramid Network (FPN) has multi-scale feature map and its architecture transmits features from high-level semantics features to low-level structure features, which lead to the tiny defects in IC substrate could be ignored and missed detection. In this paper, we improved the FPN network architecture, the semantic message transmission from topdown in the original architecture is maintained in the feature pyramid, and a branch line of structural feature message transmission from bottom-up is added. Finally, multi-scale feature superposition is performed, so that the network can obtain structural feature messages from the branch line architecture to improve the problem of missing detection of tiny defects. In this paper, we use 697 training images and 129 testing images for multi-class defect detection, such as broken copper, laser marking, scratch and dust. Finally, the test results show that the proposed Bi-direction Feature Pyramid Network architecture can achieve recall of 92.7%, precision of 77.2% and F2-score of 93.3%, which are superior to the original structure's recall of 89.2%, precision of 75% and F2-score of 85.9%, respectively.