Measuring defects in high-speed production lines—a three-phase convolutional neural network model

Abstract

Abstract Conventional automatic optical inspection (AOI) systems using rule-based image processing suffer from precision and velocity issues, particularly when simultaneously measuring multiple defects of a product in a high-speed production line. Such AOI stations usually become a bottleneck in the line. This paper presents a three-phase model for defect detection based on convolutional neural network to release the cycle time of the line. The phase I model using a deep residual network (ResNet50) performs defect classification of products with high accuracy. The phase II model is another independent ResNet50 that classifies defect products into defect categories, rescuing good products that have been mistakenly killed in the previous stage and replacing the conventional re-inspection labors. The phase III model is a you only look once—based network that detects multiple defects and their positions simultaneously in a single product, providing informative quality data for continuous improvement. The proposed model successfully resolves the issue of multiple-defect and multiple-len quality inspection in a high-speed production line. The proposed model resolved defect inspection by integrating object detection and defect classification simultaneously. By deploying the three-phase model in a tiny electronic connector component production line, the present model has demonstrated that it reaches high precision and facilitates prompt quality correction for high-speed production lines.