Research On Surface Defect Detection Of Cold Rolled Steel Sheet Based On Deep Learning

The surface quality of steel plate is one of the most important quality factors,which directly affects the performance and quality of the final product.With the concept of “Made in China 2025” and “Industry 4.0”,more and more people pay attention to the intelligent and automatic detection of defects.However,the traditional steel plate surface defect detection method is artificially designed to extract features and easy to be disturbed by the external environment,resulting in slow detection speed and low accuracy.Therefore,according to the characteristics of steel plate surface defects,this thesis adopts a variety of deep learning algorithms,and optimizes the algorithms through comparative experiments,in order to construct a model that can detect surface defects of steel plate in a non-destructive,real-time and accurate manner,and can identify the type and location of defects.Specific research contents and results are as follows:(1)Firstly,aiming at the problem of insufficient image data of steel plate surface defects,the rolling process of cold rolled steel plate production line and the causes of defects are analyzed.The defect image data set is collected and made.The defect image is grayed and denoised,and the data set is expanded by data enhancement technology.The expanded data set is transformed into the format required by different network models to adapt to the training of neural networks.(2)Secondly,aiming at the problem that the traditional computer vision detection method has poor feature extraction effect and slow detection speed,Faster R-CNN series model and Detr series model are established for experiments.The advantages and disadvantages of using Resnet50 and Resnet101 as backbone networks are studied and analyzed to verify their imbalance in detection accuracy and detection speed.To solve this problem,this thesis establishes a Yolov3 network model to study the surface defect detection problem.The model has improved the detection speed and accuracy,and finally determines that the Yolov3 model is more suitable for this defect detection problem.(3)Finally,in order to further improve the accuracy of the detection model and improve the detection effect of the model on elongated defects,this thesis proposes a Yolov3 steel plate surface defect detection model based on Transformer,which integrates the surface defect characteristics of the steel plate through the attention mechanism to enhance the learning effect on long strip defects.The loss function is improved to enhance the ability of the model to capture the spatial context information of the defect image.The experimental results show that the detection accuracy of the improved Yolov3 steel plate surface defect detection model can reach 96.1 %,and the FPS reaches 35.7,which is suitable for the industrial production requirements of the cold rolled steel plate production line.