Detection Of Strip Steel Surface Defects Based On Deep Learning

The steel industry is one of the pillar industries in China,however,due to issues such as raw materials and manufacturing processes,the surface of the strip often has some defects.These defects not only affect the aesthetics of the product,but also the product’s corrosion resistance,abrasion resistance,and fatigue strength.Existing defect detection methods such as manual visual inspection are highly subjective and have a high miss-detection rate;physical detection relies on equipment and the types of materials that can be detected are limited;the design of machine vision detection algorithms requires a large amount of professional domain knowledge and has poor generalization performance.In this context,this paper adopts a deep learning-based strip steel surface defect detection method to conduct in-depth research on two types of tasks:defect classification and defect detection.The main research contents and results are as follows:firstly,in the defect classification task,Combining the advantages of existing attention modules,MFAM(Matrix Fusion Attention Mechanism)和FFAM(Fourier Fusion Attention Mechanism)attention modules are proposed;aiming at the problem that the manually designed network architecture is difficult to achieve the optimal classification accuracy on a specific data set,an improved neural architecture search framework NATv2 is used,and the Actor-Critic algorithm is used to replace the REINFORCE algorithm to update the controller parameters,and search for the optimal architecture on the NEU-CLS-64 dataset,introduce Top-1 accuracy and model calculation complexity GFLOPs indicators for evaluation and analysis.secondly,in the defect detection task,aiming at the problem that the overall sample size of the NEU-DET hot-rolled strip surface defect object detection dataset is small,an online image stitching data augment method is proposed.Multiple images are scaled and stitched into one image without the need for other augment methods such as cropping,random brightness,etc.,which may cause unrealistic defective images;aiming at the problem that the hyperparameters involved in the online image stitching data augment method and the original hyperparameters of YOLOv5 are coupled with each other and it is difficult to manually search for the optimal combination,TPE(Tree-structured Parzen Estimator)is used to optimize the entire hyperparameter space of YOLOv5,introduce mAP(mean average precision)indicators for evaluation and analysis.The experimental results show that,firstly,in the defect classification task,the Top-1 accuracy of MFAM and FFAM on the NEU-CLS-64 dataset is better than other classic attention mechanisms;the top-1 accuracy of the architecture obtained by the improved NATv2 is slightly improved compared to the architecture obtained by the original NATv2,and it is better than the manually designed network ResNet.Secondly,in the task of defect detection,the mAP of the online image stitching data augment method on the NEU-DET dataset is better than the original YOLOv5 mosaic enhancement method;combining the hyperparameter combination searched by TPE(Tree Parzen Estimation Algorithm),the mAP has been improved after training;on this basis,the method of weighted feature fusion is adopted to enhance the network’s ability to capture defects of different scales.Introduce deformable convolution,so that the network can adapt to strip defect object and improve the accuracy of network location.the mAP of the improved detection algorithm reaches 87.2%,the detection speed reaches 40FPS,which meets the online detection requirements of strip steel in terms of accuracy and speed indicators.