Research On PCB Defect Detection Technology Based On Deep Learning

Printed circuit board(PCB)is the key component of electronic equipment,but its manufacturing process is affected by a variety of factors,which easily lead to defects,making PCB unusable.Automatic Optical Inspection technology is a common method to assist production personnel in defect detection,which has been widely used in PCB production.However,the method still needs to be verified by human eyes.This process has high labor cost and low efficiency.In this paper,PCB defect detection technology is studied,aiming to achieve an automatic PCB defect detection method based on deep learning,reduce manual operation and improve production efficiency.The main work of this paper is as follows:1.Defect image samples are collected from PCB production line,and a real PCB defect dataset with practical application value is established.After sorting and analyzing the dataset,it is found that PCB defects mainly have the following two characteristics: 1)small size;2)similar to the background.Feature Pyramid Networks(FPN)is constructed for PCB defect detection.The experimental results show that compared with the mainstream object detection algorithm,FPN has better effect,and m AP reaches 94.86%,but there is still room for further optimization.2.Aiming at the problem that PCB defect size is small and easy to lead to missed detection,a small target PCB defect detection network(IFP-STDDNet)based on improved feature pyramid is proposed.The main innovations include: 1)constructing a bidirectional feature pyramid structure to supplement the shallow information of the whole feature level;2)using K-means clustering algorithm to adjust the size of the anchor to generate a more accurate ROI;3)ROI Align is used instead of ROI pooling to reduce the impact of quantization error.The experimental results show that IFP-STDDNet is effective for small target PCB defect detection.Compared with the FPN,the m AP value is increased by 1.7%.3.Aiming at the problem that defects in PCB images are similar to the background,which is easy to cause false detection,a small target PCB defect detection network based on foreground information enhancement(FIE-STDDNet)is proposed,which is optimized and adjusted on the basis of IFP-STDDNet.The main innovations include: 1)introducing the SE channel attention mechanism to enhance the expression of useful features;2)adding the spatial attention mechanism to generate the foreground information;3)Online Hard Example Mining is adopted to solve the problem of imbalance between positive and negative samples.The experimental results show that FIE-STDDNet can effectively detect small target defects which are difficult to distinguish from the background.Compared with IFP-STDDNet,the m AP value increases by 1.32%.