Research And Application Of Small Defects Detection On Steel Rolling Surface Based On Improved YOLOv6

Rolling quality is of great significance in industrial production,but surface defects will affect the rolling quality level and further seriously affect the reliability of steel products.In order to identify surface defects of rolling steel,it is necessary to do a good job of defect detection in all aspects.Previous studies mainly rely on human labor for defect detection,which is time-consuming and laborious,and the detection effect is not good.However,the existing detection algorithm has poor detection accuracy for small defects such as holes and cracks.In this paper,an improved convolutional neural network method is proposed to detect fine defects,including holes,edge cracks,etc.,and the system is designed based on python language,which not only improves the accuracy of fine defects detection,but also ensures that the detection algorithm can be applied to actual factory practice.The main contents of this paper are as follows:(1)Firstly,the research significance of early recognition of steel rolling surface defects is described.Through literature analysis,the feasibility of target detection research based on depth learning algorithm for steel rolling defect images is presented,and the research framework of this article is given.The characteristics of several typical surface defects in steel rolling are analyzed,including single type defects,planar type defects,and periodic type defects.The introduced dataset was introduced,and the filtering function of the SIFT operator was used to preprocess the data.(2)In order to further enhance the ability of the algorithm in detection of small targets and complex targets,an optimization method combining multi-head attention mechanism and YOLOv6 algorithm was proposed,and the published data set was used for comparison calculation.The comparison results showed that the introduction of multi-head attention mechanism could improve the accuracy of YOLOv6 in detection of defect information.(3)In order to improve the detection effect of rolled steel under the case that the target and background colors are confused,U-Net algorithm is introduced,and YOLOv6 and U-Net algorithm are integrated.The research shows that the attention mechanism combined with YOLO model and further combined with U-Net model has the best effect in the detection of rolled steel defects.The accuracy of complex periodic recognition is also about 95 percent.(4)In order to realize the integration of system and algorithm,a Web-based detection system was designed based on python-flask framework.The realized functions include a series of functions such as image uploading,image detection,diagnostic results,etc.The system can assist the final inspection of the factory,save labor,realize detection automation,and complete the deployment and application of optimization algorithms.A solid algorithm is fused with the final system.This paper finally realizes the theoretical research of rolling steel defects,deep learning algorithm optimization research and algorithm application research,the designed algorithm meets the needs of steel mill rolling steel defects detection.