A lightweight investigation on automated visual SMD-PCB inspection based on multimodal one-class novelty detection

In recent years, with the research and development of deep learning, it has been more widely used in various fields, becoming an important productivity tool. In electronic manufacturing, an adaptive automatic optical inspection (AOI) system is proposed for defect detection of printed circuit board components (SMD-PCB), a key part of the industry chain. It is a combination of AOI based on traditional computer vision and multimodal imaging and PatchCore, a one-class novelty detection method based on deep learning. It aims to utilize one-class novelty detection method to detect and avoid defect missing due to different and variable objects and defects that require a lot of manpower to eliminate in normal AOI. Due to the unique characteristics of industrial applications, it is important to ensure inspection quality while controlling hardware and time costs. Therefore, in this adaptive AOI system, the reduction of time and hardware consumption of deep learning-based PatchCore during the training and inference process becomes an important part of putting it into practical application. In this work, we investigate the lightweighting of the PatchCore method, whose core idea is to use deep convolutional neural networks (CNN) for feature extraction to construct a feature memory bank, and then retrieve the feature memory bank to determine the novelty of the samples to be tested. According to this principle, we mainly lighten the PatchCore method in two directions: lightning the structure of the feature extractor and selecting the main features of feature memory bank, and then compare its performance with the original version. The experimental results show an order of magnitude reduction in hardware and time consumption reduction, while the performance remains almost the same.