Defects Inspection Of Strip Steel Surface Based On Model Compression

Hot-rolled strip plays an important role in the iron and steel industry.Its surface defects may lead to accidents,which will cause many unnecessary losses.Therefore,the strip surface defect detection system has become an indispensable tool for monitoring strip quality.Manual inspection combined with stroboscopic inspection,infrared inspection,and other inspection methods have poor sensitivity,slow detection speed,and high false negative.Therefore,they are not suitable for high-speed steel rolling production line,which makes the surface defect inspection method based on machine vision become the mainstream.However,traditional image processing methods rely on hand-craft features of professionals,which is easily affected by the surrounding environment and leads to poor detection performance or robustness.The method based on deep learning either build large-scale or highly complex models,although the detection performance is high,the detection speed is slow,thus it is difficult to deploy on the equipment with limited computing resources,which affects the industrial applications.Or build a lightweight model for detection speed,but cause a non-negligible loss in detection performance.To achieve higher detection performance and take into account higher detection speed,so that the model is easier to deploy on resource-constrained hardware platforms,model compression was born.Therefore,based on model compression technology,an algorithm of strip surface defects with high accuracy,high speed,and easy deployment is proposed.The main research results are as follows.(1)Firstly,a large-scale model with a recognition accuracy of 100% and a speed of 22 frames per second(FPS)is built to recognize the surface defects of the strip steel.Then,the compression method based on knowledge distillation is applied to extract the simplest or lightweight small model with large model recognition performance from the large model.Secondly,the impact of prior knowledge for the large model on the small model in the distillation of original knowledge has been explored,and a soft optimization scheme is proposed to get rid of the constraints of the knowledge of the large model,so that the small model can once again breakthrough itself in the recognition performance and get the performance improvement.Experimental results have shown that the small model achieves 100% recognition performance on the independent validation set,and the detection speed is about 200 FPS while the large model has only 0.12% parameters and 1.5% floating-point computation.(2)As the classification algorithm cannot recognize and locate multiple defect categories in an image,an accurate defect detection network is proposed to complete the detection of each defect area.Then,a novel,simple,and effective end-to-end pruning scheme was designed,which solves the cumbersome and redundant problems of the traditional three-level pruning process.The experimental results show that the detection speed is greatly improved without sacrificing too much detection accuracy.Specifically,79.2% detection accuracy has been achieved at 103 FPS on the mid-end GPU,and 40.1 FPS on a single low-end GPU with 0.6 M weight parameters.