Research On Detection Algorithm Of Steel Surface Defects Based On Deep Learning

In the process of steel production,steel surface quality not only reflects the integrity of steel surface,but also affects the quality and safety of downstream production equipment.If the steel surface damage is serious,it is easy to cause damage to the production equipment.Surface defect detection has become a crucial step in steel production.Manual detection methods are often time-consuming and subject to subjective factors.It is increasingly difficult for manual detection methods to meet the needs of rapid development of enterprises.In order to realize the classification and positioning of steel surface defects,a steel surface defect detection algorithm based on YOLOv5 is proposed in this paper.Due to the large scale variation of steel surface defects and relatively dispersed defects,the detection accuracy of YOLOv5 model is low.Therefore,an improved method based on YOLOv5 model is proposed in this paper.The improvement methods are as follows: First,a bidirectional weighted feature fusion network is constructed by combining CA attention mechanism and Bi FPN structure to improve the sensitivity of the model to the direction perception and position perception of the target,aiming at the large variation of steel defect scale.In order to address the issue of scattered steel surface flaws,the revised Non-local multi-scale feature enhancement attention mechanism is added to the backbone network and its model is contrasted with the current model.Third,for small-size steel surface defects,a detection head was added to the original output layer of YOLOv5 to deal with small-target defects.Determine the overall process of steel surface defect detection,and carry out experimental verification.Firstly,rotation,histogram normalization and vertical translation were used to enhance the data set to enrich the number of data set samples.Then,according to the steel surface image collected under the condition of non-uniform illumination,the image is preprocessed by using gray correction,image filtering and image sharpening to enhance the image quality.Finally,the four improved network models proposed in this paper were used for ablation experiments to verify the optimization effects of each improved method,and the improved network models were tested.The results show that the m AP value of the steel surface defect detection algorithm proposed in this paper increases by 11.5%.Aiming at the problems of large parameters and slow detection speed of the improved network model,this paper proposes YOLOv5 lightweight model.The lightweight model integrates the Ghost Bottleneck module and deep separable convolution into the detection network to reduce the parameters of the model.It is deployed on the Jetson Nano embedded device.Through the deployment and testing of the model in the embedded system,the detection speed reaches 152.9FPS while ensuring the m AP value of 73.4%,It meets the real-time requirements of steel surface defect detection.