Research On Classification And Detection Of Workpiece Surface Defects Based On Deep Learning

In the process of production and processing,surface defects will inevitably occur on workpiece due to some factors such as production equipment and human misoperation.These defects not only affect the appearance,but also reduce its performance and service life.Therefore,the workpiece with defects on the surface is detected in time to reduce cost consumption.In recent years,more and more new technologies have been combined with surface defect detection,and deep learning stands out with its powerful feature extraction capabilities.Aiming at the problem of defects on the surface of the workpiece,this paper uses convolutional neural network to extract features,classify and detect defects.The main work of this paper is as follows:(1)The classification experiment of workpiece surface defects based on improved Res Net50 was carried out,and the types and characteristics of defects were analyzed by taking hot rolled steel as the research object.First,the data is preprocessed,and then the data set is expanded to balance the number of samples between categories.By comparing the classical target classification model,Res Net50 is selected as the basic classification model.After the residual structure,the mixed attention module CBAM is introduced to make the model pay more attention to the feature information of the defect area in channel and space.Based on the idea of transfer learning,the model parameters are fine-tuned to accelerate the model convergence and alleviate the over-fitting problem.The classification accuracy of the improved model reaches 99.92 %,which is better than the original model.(2)An improved YOLOv5 s workpiece surface defect detection experiment was carried out to identify the defect type of hot rolled steel and locate the defect area.By analyzing and comparing the target detection algorithm,YOLOv5 s network is selected as the basic model of defect detection.Aiming at the low detection accuracy of the original network for small target defects such as cracks and oxide scales,based on the YOLOv5 s network,a variety of improved methods are proposed for the lightweight network Shuffle Net V2,which integrates the mixed attention module,adds the Bi FPN weighted bidirectional feature pyramid structure and replaces the CSPDarknet53 of the backbone network.After multiple sets of experimental comparisons,the effectiveness of the fusion attention module,the improved feature fusion structure and the replacement of the backbone network is verified.The detection accuracy of the final improved model reaches 78.9 %,which is 2.9 % higher than the original model.The detection accuracy of small target defects is improved without reducing the detection accuracy of other categories.(3)The workpiece surface defect detection system is built,and the defect detection system is designed based on the improved classification and detection model.The system interface is composed of login interface,defect recognition interface and defect detection interface.The feasibility of the improved model is verified by actual test.