Electronic Product Surface Defect Detection Based on a MSSD Network

Abstract

The appearance defect detection of electronic products is a difficult problem for industrial production. The main problems are the types of defects detected and the complex background texture interference. This paper proposed a deep learning-based target defect detection method, combined with a lightweight mobile convolutional MobileNet feature extraction network. A SSD-based model can automatically perform multi-level feature extraction from defect samples without the need for manual feature extraction. For the problem of small target defects, it is difficult to detect. By adjusting the model structure, more feature information of the low-level convolutional layer is retained. The experimental results show that in the defect image with a complex background, the types of scratches, pits, bumps, and scratches can be quickly and accurately identified. The proposed method can significantly improve accuracy, efficiency, and robustness. The mAP of surface defect target detection can reach 88.6%.