A Defect Inspection Method for Machine Vision Using Defect Probability Image with Deep Convolutional Neural Network

摘要

Deep learning is replacing many traditional machine vision techniques. However, defect inspection systems still rely on traditional methods due to difficulties in obtaining training data and the absence of color images. Thus, overall performance heavily depends on individual human skill in tuning hundreds of parameters. This paper presents a defect inspection technique using a defect probability image (DPI) and a deep convolutional neural network (CNN). DPIs are the estimated probability of a defect in given image and can be obtained from traditional inspection techniques. The DPI and gray image are stacked as input to the CNN. Performance was compared with a conventional CNN model using RGB or grayscale images, and ViDi, an artificial intelligence software for industry. The proposed method outperforms the other methods, works well on small dataset, and removes the requirement for human skill.