Research On Detection Method For Steel Plate Surface Defects Based On Deep Learning

Steel plates is one of the important manufacturing materials and is widely used in aerospace,aerospace,transportation,petrochemical and industrial fields.China is the largest steel producer in the world,and the steel industry is an important industrial support for it.However,the surface of steel plates is prone to defects during production and transportation,which affects product quality.Therefore,it is of great significance to study the detection methods of steel plate surface defects.In recent years,computer vision algorithms based on deep learning have been rapidly and widely used in the field of surface defect detection.In this paper,the steel plate surface defects detection is regarded as the main research object.There are mainly three aspects researched: steel plate surface defect classification,prediction of defect location,size and class,and pixel-level accurate identification of defect areas.The main research contents and work of this paper are as follows:(1)In order to classify the surface defect categories contained in the defect image,and solve the problems of large network model,low running speed and low accuracy in defect classification,a lightweight and high-accuracy rapid steel plate surface defect classification network Res VT is proposed.Firstly,a convolutional residual network is used to extract the local features of the defect image in the early stage and generate feature maps with high-level semantic information;then a convolution self-attention block(CSB)is applied to 2D images;based on CSB,a convolution encoder block(CEB)is proposed to establish the connection between the global feature information and focus more on the defect area,which effectively solves the problem of insufficient in CNN global feature extraction and difficulty in self-attention model training.In addition,feature channels of the network are scaled to reduce feature redundancy and computational complexity.Finally,Res VT reaches 100%classification accuracy on the NEU-CLS steel surface defect dataset,and the network parameters are only 1.26 M with a running speed of 56.23 FPS.In addition,it reaches the classification accuracy of 98.40% on the CAT-CLS aluminum casting surface defect dataset.It indicates that the proposed method has generalization ability and can effectively classify surface defects.(2)In order to realize the prediction of the location,classification and size of the defect area,in this paper,a plate surface defect detector based on anchor-free object detector DCC-Center Net,is proposed to resolve the conflict between speed and accuracy.Keypoint estimation is used to locate center points and regresses other defect properties.Firstly,a dilated feature enhancement model is proposed to enlarge the receptive field of the detector and improve the ability of multi-scale defect detection.Secondly,a new centerness function center-weight is proposed to generate the heatmaps,which makes the keypoint estimation more accurate and alleviates the imbalance of positive and negative samples.Then,the CIo U loss that considers the overlap area and aspect ratio of the defect is adopted in the size regression.Finally,the accuracy of DCC-Center Net can reach 79.41% m AP,and the running speed FPS is 71.37 with input picture pixels 224×224 on the NEU-DET steel defect dataset.And it reaches 61.93% m AP on the GC10-DET steel sheet surface defect dataset at a running speed of 31.47 FPS with input picture pixels 512×512.It demonstrates that the developed detector can detect steel plate surface defects efficiently and effectively.(3)In order to achieve accurate prediction of the shape and contour of steel plate surface defects,a semantic segmentation network CT-UNet is proposed to solve the problems of irregular shape of steel plate surface defects and difficulty in dividing defect contours.Firstly,the network combines the encoder-decoder structure with the convolution self-attention encoder module to improve the capability of feature extraction and establish the pixel-level spatial information connection of the feature map.Then,the windowing strategy is used to reduce the parameters and improve the efficiency of segmentation,and the feature channel confusion module is adopted to mix the features between sub-windows to provide feature maps containing global information for the multi-head attention mechanism.Finally,the segmentation accuracy of CT-UNet reaches 84.05% m IOU on the NEU-SEG steel surface defect dataset,which has achieved the highest accuracy among the comparative semantic segmentation networks,and the running speed reaches 38.65 FPS,which can meet the needs of real-time online segmentation.