Research on PCB defect detection based on SSD
Litian Kang, Yawei Ge, Hong Huang, Ming Zhao
Abstract
To solve the problem of wrong detection in PCB defect detection, a deep learning detection network based on SSD, named multi-layer SSD (mSSD), is proposed. A small target prediction feature layer module is added to this network, which can improve the perception ability of small target features. In addition, we used ResNet50 feature extraction network instead of the original VGG network to amplify the original six feature prediction layers to seven. Mosaic enhancement was also used for PCB data sets to measure the parameters of multiple images during the Batch Normalization training phase. Verified on the constructed PCB validation data set, the mAP of PCB detection network based on mSSD reached 95.91%, which improved 13.0% compared with the test result of SSD network. The experimental results show that the improved mSSD detection network greatly improves the detection accuracy of SSD in PCB defect detection.