Improved YOLOv5-based Defect Detection For Hot Rolled Strip

In order to achieve efficient and accurate steel surface defect detection in industrial scenarios,this paper selects a target detection method based on deep learning,which has higher detection accuracy and adaptability compared to traditional methods.However,due to the industrial environment,small defect targets,and poor algorithm performance,the detection of steel surface defects is difficult,with high false detection and leakage rates and low efficiency.Therefore,how to accurately and quickly identify steel surface defects is of great importance for industrial production.In this paper,to address this problem,the most widely used hot-rolled strip steel is selected as the research object,and YOLOv5 deep learning target detection algorithm is used to carry out relevant research work,which is mainly as follows:(1)To address the current problem of difficulty in extracting small target defects on the surface in hot rolled strip steel defect detection,this paper conducts improvement research in three aspects.In terms of anchor frame,this paper uses a method combining K-means++ algorithm and hyperparameters to solve the problem that the globally optimal clustering center cannot be selected when clustering,improves the selection of the initial optimal anchor frame,and reduces the positioning error of the prediction frame;in terms of network structure,this paper adds a 1/4-scale extraction feature map to the feature extraction network structure,adds an asymmetric convolution module to improve the In the optimization of the loss function,a smoother KL scatter is chosen to replace the original cross entropy as the loss function of confidence to improve the optimization of YOLOv5,which improves the detection accuracy of the model to a certain extent.The accuracy of the model is improved to some extent.The final results show that the m AP of the improved YOLOv5 network model is improved by 8.2% compared with the original YOLOv5.(2)In order to meet the current demand for real-time detection of hot-rolled strip defects on the production floor,the improved YOLOv5 model is lightened by using a deep separable convolutional module and int8 lightening,which results in a simplified YOLOv5 model that uses less computational and memory resources and is faster and easier to deploy in embedded or mobile devices.It is also faster and easier to deploy in embedded or mobile devices.Meanwhile,to address the problem of data security in collaborative training,a differential privacy stochastic gradient descent algorithm based on local random perturbation(LRDDP-SGD)is used to differentially privacy the data to prevent the model training data from being stolen by attackers.The experimental results show that the algorithm has higher detection accuracy compared to the differential privacy stochastic gradient descent algorithm(DP-SGD),which better balances the problem of model detection accuracy and privacy.