Defect detection of printed circuit board based on small target recognition network

The quality of the printed circuit board (PCB) determines the performance of electronic products, and the existing PCB surface defect detection methods cannot meet the requirements of non-destructive testing and efficient detection. The size of the PCB defect only accounts for less than 0.1% of the entire PCB area, which poses a challenge to traditional vision based inspection methods requiring high-definition cameras to take images multiple times. This paper proposed a method for defect detection of PCB board based on deep learning system to increase the efficiency and accuracy of detection. Based on Faster-RCNN, in which resnet101 is the backbone network, combined with the feature pyramid network (FPN) it can take into account the high-level semantic features and the underlying position information, while avoiding small target defect information in the convolution process. The ROI-Pooling algorithm is replaced by the ROI-Align algorithm, and the two-line interpolation in ROI-Align is applied to avoid the error generated due to quantization when the feature map is intercepted. Eventually, the accuracy of the inspection network was improved. The proposed network is evaluated on a public dataset 98.5% mAP was achieved.