PCB Image Defect Detection Based On Image Processing And Deep Learning

The continuous improvement of industrial production and manufacturing levels has accelerated the upgrading of electronic products,and therefore put forward higher requirements for the underlying infrastructure hardware facilities.The printed circuit board is an important basic component of integrated circuits.Its quality determines the overall performance of electronic products.However,it is difficult to avoid PCB defects due to various factors during the PCB production process.Traditional PCB defect detection uses manual labeling of its defect information,which has low efficiency and low accuracy.In recent years,deep learning-based defect detection algorithms have gradually been studied,but traditional convolutional neural networks are aimed at PCB image defect detection.It is difficult to take into account the global and detailed feature information,and the accuracy of detection and the accuracy of marking are not high.Therefore,aiming at the problem of PCB image defect detection,this thesis proposes a PCB image defect detection algorithm based on image processing and deep learning.The main contents of this article are as follows:(1)First of all,in view of the fact that the characteristic information in the PCB defect image is not obvious and contains a lot of noise information that is not related to the characteristic,it is proposed to perform image preprocessing such as image grayization,grayscale histogram equalization,and Gaussian filtering on the image.The method,on the one hand,removes noise interference to highlight feature information,on the other hand,it also provides training and verification data sets for subsequent neural network training.(2)Secondly,on the preprocessed PCB defect image data set,global defect detection and positioning are realized through machine vision technology.Based on the reference comparison algorithm,SURF is used to extract the feature points of the PCB standard image and the defect image,and the feature points are matched to achieve image registration,and then the image is processed into a binary image by threshold segmentation,and then image difference is performed to obtain the defect image The defect information contained is processed by morphology to make the defects clearer,so as to realize the preliminary detection and positioning of the PCB image.(3)Then,feature extraction of PCB images based on deep learning methods.The traditional residual neural network Resnet101 effectively solves the problem of gradient dispersion caused by too deep network layers,but the low network layer high resolution is insufficient for image semantic feature extraction,and the high network layer low resolution is insufficient for image detail feature extraction.The method of multi-scale feature fusion is proposed to improve the network to improve the quality of feature extraction,forming the Resnet101-finetune network.The network is trained and verified on the PCB image data set through transfer learning.The experimental results show that the average accuracy of the improved Resnet101-finetune neural network reaches 94.2%,which is significantly improved compared to other neural networks.(4)Finally,the defect target detection of PCB images is realized based on Faster-RCNN algorithm.Improve the efficiency and accuracy of target detection by improving the FasterRCNN algorithm.YOLO,R-CNN,SSD and the algorithm proposed in this thesis are compared on the same PCB image data set.The experimental results show that the algorithm proposed in this thesis achieves an average accuracy of 95.5%,and the m AP value is also higher than other algorithms.Verification The effectiveness of the algorithm in this thesis is verified,and the versatility of the algorithm in this thesis is verified through the example detection results.It can be applied to PCB image defect detection and has high practical value.