Application Research Of Printing Defect Detection And Classification Based On Deep Learning

Defect detection is a very important application in industry,especially in the field of printing.Due to the diversity of defects,it is difficult for traditional image processing methods and traditional machine vision algorithms to complete the modeling and migration of defect features.The reusability is not great,requiring the differentiation of working conditions,which will waste a lot of labor costs.As deep learning has achieved very good results in feature extraction and location,more and more scholars and engineers begin to introduce deep learning algorithms into the field of defect detection.In this paper,the single-stage object detection algorithm YOLOv4-tiny based on convolutional neural network is introduced into the surface defect detection of printed circuit board and improved.Although there are some deficiencies,excellent results are achieved on the whole.In terms of data set making,all PCB defect images were obtained by a linear scanning CCD,which was preprocessed into 640×640 images.The experimental environment configured in this paper is as follows: The operating system is Ubuntu18.04.CPU intel Xeone5-2678v3,the main frequency is 2.5GHZ.GPU is NVIDIAGe Force GTX1080 Ti.In order to make full use of GPU-accelerated network training,CUDA10.1 and its companion CUDNN are installed in the system.The deep learning framework is Darknet.The main innovations are:(1)Considering that the printing defects on PCB surface are very small,the spatial attention module SPP is added into the backbone network structure of YOLOv4-tiny to obtain the algorithm model of YOLOv4-tiny-spp,which improves the ability to extract the characteristics of tiny target defects.The accuracy is improved by 0.8%.The FPS of video detection(how many images a target network can detect per second)is125;(2)In order to improve the detection accuracy of YOLOv4-tiny,based on the above YOLOv4-tiny-spp algorithm model,an improved algorithm based on three detectors is used to obtain the model YOLOv4-tiny-spp-3l,which reduces the missed detection of small targets and thus improves the detection ability of small targets.However,the improved detection accuracy and detection speed are decreased,which may be caused by the loss of feature information.(3)In order to maximize the feature information of feature maps,we designed auxiliary residual network blocks to extract global features by using 3x3 receptive fields and layer-hopping operations.Finally,a new backbone network YOLOv4-tiny-spp-3l-cbl is constructed by merging the auxiliary network designed by us into the backbone network.In this way,the convergence of deep and shallow networks is realized.The improved backbone network can extract global and local features of the detected object,and the detection accuracy is improved by 2.1%.Through the method of PCB surface defect detection based on deep learning,the PCB surface defect detection and classification can be intelligently,which can not only meet the demand of real-time detection in speed,improve work efficiency,but also save human resources and enterprise cost.