Research And Implementation Of Hot-rolled Strip Surface Defect Detection Algorithm

In the industrial production process,the occurrence of surface defects in hot-rolled steel strips is inevitable.To ensure the quality of hot-rolled steel strip products,timely detection of surface defects and problem identification have become important tasks.In traditional hot-rolled steel strip production processes,techniques such as infrared,ultrasonic,and magnetic leakage are commonly used for surface defect detection.However,these techniques have drawbacks such as reliance on human experience,high cost,and limited universality.With the development of intelligent technology,deep learning-based machine vision methods have been widely applied to surface defect detection in smart industries.This approach offers advantages such as low cost,good safety,high efficiency,and strong flexibility.In view of the above situation,Thesis aims to improve the YOLOv5 s and Effnet+Soft Max algorithms for the task of detecting surface defects in hot-rolled strip steel.The main innovations and work of Thesis are as follows:1)The YOLOv5 s algorithm was improved by optimizing the IOU loss function,enhancing the label assignment process,adding attention mechanisms,and optimizing the convolutional layer model for lightweight implementation.The improved YOLOv5 s algorithm achieved a training mean Average Precision(m AP＿0.5)of 97.2% in the task of surface defect detection in hot-rolled strip steel,which is a 10 percentage point improvement compared to the unimproved version.It also demonstrated better performance in occlusion,fusion,connection,and detection of subtle defects compared to the original algorithm.The Effnet+Soft Max algorithm was used as the classification algorithm,and the impact of different loss functions and optimizers on the algorithm’s performance was investigated.The combination of cross-entropy loss function and Adam W optimizer,which exhibited the best performance,achieved a classification accuracy of 98.2%.Finally,transfer learning was employed to address the issue of small-sample size in surface defect detection of hot-rolled strip steel.2)Based on the improved YOLOv5 s algorithm and Effnet+Soft Max,a comprehensive deep learning defect detection software for surface defects in hot-rolled strip steel has been developed.This software encompasses functions such as data annotation,model training,and defect detection.In practical applications,users can directly select pre-captured images for annotation,adjust various hyperparameters of the training algorithm to optimize training performance,and view statistical information of the training results.In terms of software performance,the Soft Max classification accuracy achieved 96.1%,while the improved YOLOv5 s algorithm achieved a m AP＿0.5 of 97.1%.These results effectively reproduce the accuracy of the algorithms studied in thesis and greatly enhance the efficiency of usage.