Automatic Defect Detection System Based on Deep Convolutional Neural Networks

Abstract

Deep learning has been widely used in various fields. This paper uses the supervised deep learning architecture for automatic optical inspection. The advantage of using deep learning is that it can detect the defects that cannot be detected by traditional machine vision algorithms. The dataset in this study is surface scratches on plastic housings. The shape of scratches is mostly slender. They are too narrow that the features will disappear after CNN reduces spatial resolution. For this issue, we further evaluate the input size of the defect. In addition, we also propose a mask labeling map based on pixel annotation to do the block cropping with sliding windows strategy. The results show that the influence of the feature's pixel size into the convolution calculation will enable us to more accurately locate the defect, and the pixel accuracy can reach 73.47% only with two images. The research results also provide a new consideration for defect detection and its pixel size.