Strip Surface Defect Detection Based On Deep Learning

In today’s steel industry,strip steel is one of the products that are widely used in various industries.During production,the strip rolling process is prone to surface defects,and it is particularly important to effectively detect them.Existing conventional inspection methods are difficult to meet the requirements of accurate and efficient detection.The advantage of convolutional neural network is that it can automatically learn the features of samples,so this paper uses it as a basic network for strip steel surface defect classification and detection and conducts subsequent improvement and experimental analysis.The main research work in this paper includes:(1)To address the problems of complex structure,poor feature generalization ability,many parameters and low recognition accuracy of existing strip steel surface defect classification methods,this paper proposes a strip steel surface defect classification method I-Res Net34 based on the residual network.The method simplifies the network structure: reduces the size of convolution kernel,replaces the pooling layer with Inception structure and reduces the number of network layers;replaces the activation function: replaces Re LU with Leaky Re LU to speed up the convergence speed of the network;the loss function is optimized: a center-loss function is added to the cross-entropy-loss function to solve the problem of class spacing of the dataset.With the combination of the three,the final residual network classification model I-Res Net34 has a recognition accuracy of 95.40% and a time of 48s/e.The number of parameters and computation are better than the original residual network,which improves the classification of strip surface defects.(2)To solve the problem of insufficient classification samples,the NEU dataset was expanded using the data enhancement library imgaug.The NEU dataset for strip surface defect classification has six categories,each with only 300 samples.The images in different categories are similar,and the images in the same category do not differ significantly.The images in NEU are panned,mirrored,horizontally flipped,and brightness changed to obtain a new dataset,which enhances the generalization ability of the network and improves the classification accuracy significantly.(3)In this paper,we designed the overall experimental system for strip steel surface defect detection,and then improved each component of the Faster RCNN network model based on the characteristics of more small and medium-sized defects on the strip steel surface,easy loss of small and medium-sized targets in the feature extraction process,and feature shifts in the pooling process.The backbone network is improved: the first four convolutional modules of Res Net50 plus the Feature Pyramid Network are selected as the feature extraction network;change the anchor box size of the RPN layer: 15 anchors are reset;the pooling method is changed: ROI Align is used instead of ROI Pooling.The improved model has improved recognition rates for small and medium-sized defects,no feature shifts,and improved overall performance.(4)To address the situation that the defect detection dataset has unbalanced samples and many errors in labeling information,the sample images are screened and deleted and then expanded,and then manually labeled with Label Img one by one to obtain a new defect detection dataset.The improved backbone network,RPN layer and pooling layer are fused into a new model for strip surface defect detection,and experiments are conducted on the new defect detection dataset.The final m AP value of the defect detection model is increased from 69.2% to73.2%,and the FPS is changed from 21 to 23,which is a great improvement in speed and accuracy,proving that the improved method in this paper is practical and feasible.