Research On PCB Surface Defect Detection Method Based On Improved YOLOv3

PCB defect detection based on deep learning has become research hotspot and difficult point in the field of machine vision.At present,there are still many problems of deep learning,many model parameters,large amount of calculation,too slow detection speed,high missed detection rate,resulted in high missed detection rate and low detection accuracy.In response to solve the above problems,an improved YOLOv3 PCB surface defect detection method was proposed.First of all,to solve the difficult PCB defects locations and high missed detection rate,combined the genetic algorithm and K-means algorithm,used genetic algorithm for global optimization and Kmeans algorithm for local optimization,improved the K-means calculation distance,recalculated the distance assigned to each cluster with the 1-IOU value instead of euclidean distance,generated anchor boxes better fit the dimensions of PCB defect detection,improved the accuracy;Secondly,to solve the problem that the YOLOv3 model has a large number of parameters and a large amount of calculations,combined depthwise separable convolution and traditional convolution,formed a double branch separable network structure,instead of Darknet-53 in YOLOv3 as a new backbone network for feature extraction,reduced the amount of parameters through separable convolution,improved the detection speed;then,used the double branch separable network structure,the feature information obtained by the two branches were spliced and fused,obtained more complete feature informations and improved the detection accuracy;finally,through the joint prediction of multi-scale features,outputed defects classifications and marked defects locations for detection of PCB defect targets.The PCB defect dataset published by the intelligent robot laboratory of Peking University was used to conduct experimental tests,the results showed that the precision rate,recall rate,F1 value,mean average precision,parameter quantity and FPS of the proposed method reached 97.48%,98.81%,98.46%,97.87%,5.7M,23.09f/s,at the same time cost,the performance indicators were better than the comparison methods of SSD,Faster R-CNN,YOLOv3,YOLOv4,the effectiveness of the proposed methods were verified;and further verified by ablation experiments and random tests feasibility of the proposed method.Therefore,the proposed method was able to guarantee the high detection accuracy,at the same time,reduced the amount of parameters and had better PCB defect target detection performance.The thesis has 43 pictures,17 tables,and 60 references.