Improved Faster-RCNN Based Defect Detection Of High Density Flexible Printed Circuit Board

Flexible Printed Circuit Board(FPC)is an important part of integrated circuits and it has been widely used in various electronic products.FPC defect detection can improve the FPC yield in the manufacturing process and is a key in the manufacturer's production process.At present,most manufacturers are still using manual visual inspection,which takes a lost of time and labor.If the traditional digital image feature technology is used,FPC pictures need to be traversed at the pixel level,making a problem of slow speed.In recent years,represented by Faster-RCNN,the algorithm based on deep convolutional neural network has begun to be applied to FPC defect detection,but some defects of FPC are relatively small,and there are few defect samples,that is,there are problems about small targets and small samples,causing the detection results not good.Therefore,on the basis of the Faster-RCNN algorithm,this paper studies these two problems,and the main work is as follows:(1)Aiming at the problem that some defects of FPC are small and the detection results are not good,an improved Faster-RCNN algorithm based on feature enhancement and multireceptive field selection is proposed.Based on the existing ideas and structure of the FasterRCNN,a feature fusion module is proposed to enhance the overall feature,and then a target feature enhancement module is proposed to separately improve the target feature intensity,and finally a multi-receptive field region selection module is introduced,respectively extract target feature of different scales.These three modules make improvements to the Faster-RCNN algorithm and are applied to FPC defect detection to solve small target problems.Compared with the original Faster-RCNN algorithm,the m AP of algorithm with three modules increased by 3.52%,proving the effectiveness of the improved method.(2)Aiming at the problem of less FPC defect samples and bad training and detection results,an improved Faster-RCNN algorithm based on multi-feature transformation and changing intraclass similarity is proposed.Based on the structure of the Faster-RCNN algorithm,the structure of the parallel spatial transform network(STN)is introduced,and adding the strategy of limiting the STN parameter transformation form and introducing increment to the network parameters,which enriches FPC defect features and enhances the ability of the network to extract FPC defects.Then,a binary loss function is introduced into the regional classification network to enhance inter-class differences and intra-class similarities.Finally,the entire improved FasterRCNN algorithm is used in FPC defect detection task to solve the problem of small samples.Compared with the original Faster-RCNN algorithm,the m AP of the algorithm with three improved strategies increased by 5.35%,proving the effectiveness of the improved method.This paper provides a reference for the defect detection technology based on improved Faster-RCNN in manufacturing process of high-density flexible boards.