Research On Surface Defect Detection Algorithm Of Hot Rolled Strip Based On Deep Learning

Hot rolled strip is an important product of current iron and steel industry,which is widely used in various industries.In the process of the strip hot rolling,strip surface defects are important factors affecting product quality.The process of traditional defect detection methods are complex,and they are difficult to meet the requirements of fast and accurate defect detection.The strip surface defect detection system based on deep learning has the advantages in accuracy and efficiency,but there are still some problems in complex defect detection and defect object location.In this thesis,the multi-label classification algorithm is used to solve the problem of composite detection on the surface of hot rolled strip.The YOLO algorithm is used to locate and classify the defects.The specific research contents are as follows:1)The multi label classification model based on neural network is used to solve the problem that it is difficult to detect all kinds of defects when they are mixed.The VGGNet,Res Net and Mobile Net are used to conduct experiments.Attention mode is used to improve the ability of multi label classification model to extract important features.The improved loss function is used to balance the quantity of positive and negative samples.Experiments show that these two improved measures can effectively enhance the detection capacity of multi label classification model.2)The modified YOLOv3 model and the modified YOLOv4 model are proposed to detect the defects,which can get the location and category information of surface defects of strip steel.The weighted K-means clustering algorithm and multi-scale are used in YOLOv3 model to detect small object.Adaptively spatial feature fusion and EIo U loss function are used in YOLOv4 model to improve the detection capacity.Experimental results show Mean Average Precision of the modified YOLOv3 model can reach 80.1%,which is 11% higher than original YOLOv3 model,and Mean Average Precision of the modified YOLOv4 model can reach 83.4%,which is 3.7% higher than original YOLOv4 model.3)Network pruning is used to reduce the scale of the modified YOLOv4 network model.The detection speed of the modified YOLOv4 model is improved on the basis of ensuring that the model still has a high detection capacity.Finally,the scale of the modified YOLOv4 model is reduced by 74%,and the detection speed of model is improved from 68 to 82.