关键词: 印刷电路板瑕疵检测, Faster RCNN, 神经网络, 损失函数, 多尺度训练

Abstract: There are many electronic components on PCB，the distance between them is small，and the circuit routing color is relatively similar. The traditional method based on machine vision detection has some problems，such as slow detection speed，false detection rate，fewer types of defects that can be detected. Based on this，we propose a defect detection method of printed circuit board based on improved Faster RCNN，which can detect six kinds of defects on printed circuit board：leakage，notch，open circuit，short circuit，burr，and residual copper at the same time. Firstly，the basic detection framework is Faster RCNN，and the basic improvement measures are pyramid feature network（FPN），multiscale training，anchor box，and clustering. Secondly，we replace the smoothL1 loss function in the original algorithm with the improved loss function（DIoU）as boundary box location regression loss function. Finally，we calculate the mean precision（mAP）of the model under various experimental conditions and compare the results of different algorithms. The experimental results show that the mAP of the original Faster RCNN is 73. 7%，and the mAP of the improved Faster RCNN is 95. 1%. Compared with the original algorithm，the mAP of the printed circuit board defect detection model increases by 21. 4%，which has obvious advantages over other algorithms.

Key words: printed circuit board defect detection, Faster RCNN, neural network, loss function, multiscale training