Research On Semi-supervised Inspection Method Of Steel Surface Defect

Steel surface quality control has always been the focus of intelligent inspection in steel production process.With the proposal of the concept of "Industry 4.0" and the development of deep learning,it has become a trend for machines to replace manual defect detection.However,at the same time,the variety of steel surface defects and the lack of labeling data are also difficult problems in the steel industry.Semi-supervised learning method is very effective for the problem of lack of labeled data,so applying semi-supervised learning method to object detection can improve the shortcomings of supervised learning object detection,and is of great significance for steel surface defect detection.Based on the deep learning method,the semi-supervised target detection method is used as the basis,Pytorch is used as the experimental basic environment,and the method of combining theoretical analysis and experimental verification is used.Aiming at the characteristics of steel surface defect detection,such as the diversity of defect categories and the lack of label data,the following work is carried out:(1)A method of combining semi-supervised object detection model with steel surface defect detection is proposed to solve the problem of lacking relevant semi-supervised model for steel surface defect detection.The semi-supervised target detection theories such as STAC model,deterministic pseudo-label and dynamic reweighting,EMA parameter update and dense pseudo-label are clarified,which lays a theoretical foundation for the construction of steel surface defect detection models.The FFRCNN network,the loss function and the frame jitter algorithm in Softteacher are systematically described,which makes a theoretical foundation for improvement.(2)Aiming at the problem that the features of NEU are not obvious when the defects are carried out in the steel surface inspection data set,an improved FFRCNN network model is proposed.The serial structure of the network is changed into a multi-branch parallel structure,which aims to extract and fuse the features multiple times to avoid the problem of feature loss caused by the inobvious data features.Aiming at the unreliable problem of semi-supervised pseudo-labels,the pseudo-label generation mechanism is improved,and the loss function is adjusted by generating pseudo-labels multiple times to reduce the error caused by model prediction and achieve the purpose of improving the reliability of pseudo-labels.A new steel surface defect detection model is constructed,and the effect is verified by experiments and analysis.(3)Focusing on the intelligent management requirements of steel production,a steel production monitoring and management system based on semi-supervised target detection is designed and implemented.The basic functions include real-time monitoring of production environment,statistics of steel plate quantity,detection of steel plate surface defects,and visual display of detection results in graphical form.