PCB Defect Identification Based On Multi-scale Faster R-CNN Algorithm

With the advent of modern information technology,the microelectronics industry has developed rapidly.Printed Circuit Board(PCB)constitutes an important component in electronic components which is used in many fields such as communications,military aviation,and industrial control.However,With the advent of PCBs in the direction of high precision,miniaturization and high performance,the difficulty of PCB quality inspection has been greatly improved.Because the traditional inspection methods by people have many disadvantages which have low efficiency and high cost sometime may have missed detection and false detection,the traditional manual inspection method can not adapt to basic requirements.In order to solve those problems,this paper proposes a multi-scale Faster R-CNN algorithm.By improving this algorithm,it can effectively identify the defects of the PCB bare board.the Faster R-CNN has more advantages,Such as higher measurement accuracy and faster detection speed.But PCB board defects are small targets which have many types,and only single-scale prediction cannot meet the requirements of accurate detection.Therefore,this paper starts from the perspectives of multi-scale feature fusion and multi-scale pooling.Faster R-CNN target The detection algorithm has been improved to improve the accuracy and accuracy of the algorithm.By introducing the FPN,using ResNet101 instead of VGG16 as the feature extraction network,and replacing ROI Pooling with ROI Align for multi-scale pooling,the detection accuracy and detection efficiency of the model are improved.The specific improvements are as follows:(1)The idea of FPN network is applied to Faster R-CNN algorithm.Furthermore,ResNet101 is used as the backbone network of FPN.On the basis of this,a new multi-scale feature map model structure is constructed.On the one hand,it enhances the number and complexity of network layers,on the other hand,it fuses the high-level semantic information feature map at the lower level with the high-level feature map at the rough target location,effectively solving the problem of small and dense target proportion in the image.(2)Add structure.In order to solve the problem caused by the inconsistent size of the input image required by the deep convolution neural network.Avoid cutting or scaling the target to distort its shape and maximize the original semantic information of the image.(3)Replace ROI Pooling with ROI Align.Because ROI Pooling integrates the length of the boundary,ROI Align can use bilinear interpolation to accurately interpolate the void of a pixel point when its coordinates are floating-point numbers.Experiments show that the average accuracy mean(mAP)of the Faster R-CNN visual detection model based on multiscale fusion is 86%,which is 8%higher than before.It can detect,locate and identify surface defects such as copper,mousebite,short,circuit,spur,open in PCB bare plate with high accuracy.Solves the problem of identifying multiple classes and features of defects.