Two-step Convolutional Neural Network for Image Defect Detection

Abstract

Industrial production has a high demand for the application of machine learning on defect detection based on image. The superiority of deep learning leads to its widespread use in the field of image defect detection. In this paper, we propose a two-step convolutional neural network (TSCNN) model to improve the accuracy of the defect detection on images of electronic component. The TSCNN consists of a coarse detection network (CDN) and a precise detection network (PDN). The CDN is expected to detect images with obvious class features, while the PDN with center loss is applied to reclassify images that present low class probability in the CDN. Considering that the CDN doesn’t have enough confidence in the detection of images with probability between $\alpha$ and (1-$\alpha$), ($\alpha$ is 0.25 in this paper) we integrate the output probability of CDN and PDN as the final probability for them. Experiments prove the great performance of the two-step convolutional neural network for image defect detection. Meanwhile, contrast experiences of TSCNN and other methods demonstrate the effectiveness and necessity of using PDN in TSCNN to improve the accuracy of detection in industrial production.