Surface Defect Detection Of Hot Rolled Steel Based On Deep Learning

Due to the influence of manufacturing process and equipment,production environment,production materials and many other factors,hot-rolled steel surfaces formed different types of defects,affecting the quality of steel.The detection method was mainly manual detection,which had problems such as slow detection speed and low accuracy.Aiming at the problems existing in the detection of surface defects in hot-rolled steel,a surface defect detection algorithm for hot-rolled steel based on deep learning was proposed.The details were as follows:1)The data set was composed of images of various defects of hot-rolled steel collected in the actual production of steel mills.In view of the variety of surface defects and complex characteristics of hot-rolled steel in actual production,the collected images were made into a hot-rolled steel surface defect data set through angle rotation,brightness transformation,cutting,etc.,and artificial labeling.2)A surface defect detection algorithm for hot-rolled steel based on improved Yolov4 was proposed.The Yolov4 algorithm was selected as the detection master algorithm,and SENet network was embedded into each layer of residual network of CSPDarknet53 to distinguish the feature information of different defect types.SE-Yolov4 network realized the information interaction between channels by weighting the characteristic information of each channel,to strengthen effective information and suppress invalid information.Due to the certain similarity between crazing and grooves-and-gouges in appearance,these two types of defects were not easy to distinguish.The number of convolutional layers was increased before and after the SPP and after the backbone feature network output different feature information.The detection accuracy of crazing,grooves-and-gouges and surface-unwarping increased by 1.72%,10.07% and 2.67%respectively,and mAP increased from 86.62% to 93.13%.3)An improved lightweight SE-Yolov4 surface defect detection algorithm for hot-rolled steel was proposed.The SE-Yolov4 algorithm after increasing the number of convolutional layers slowed down the detection speed.Referring to the design idea of MobileNet series algorithms,embedded the basic block of MobileNet v3 into the backbone feature extraction network of SE-Yolov4.The time used to detect a picture was reduced from the original 0.062 s to 0.026 s.The mAP value of the improved algorithm in the test set reached 93.82%,which was 7.2% higher than that of the Yolov4 algorithm.Figure 37;Table 6;Reference 55...