A Research Of Metal Surface Defect Detection Method Based On Semi-supervised Learning

In the production process of metal industrial products,the existence of surface defects may greatly reduce the quality and performance of products,and even lead to product damage and failure.Therefore,it is very important to find and control these defects in time by means of surface defect detection,which can not only provide product quality and safety assurance,but also reduce the losses caused by defects.At present,the mainstream surface defect detection method is to realize the detection and location of defects by combining machine vision and deep learning.However,deep learning requires the support of a large amount of labeled data to obtain high detection accuracy.In industrial scenarios,it is very difficult to obtain the labeled data required for each defect.In addition,some current defect detection methods simply design a larger network structure,fail to fully consider the actual needs of defect detection tasks,and fail to meet the real-time requirements of industrial scenarios.This thesis conducts in-depth research on the above problems,and puts forward its own research program.The research content of this thesis is as follows:(1)This thesis analyzes the problems existing in the current defect detection tasks in detail,including: the difficulty of collecting defect data sets,the lack of real-time performance of the current defect detection methods,and the weak ability of the detection network to identify some defects.In response to the above problems,this thesis proposes a semi-supervised defect detection method based on the teacher-student network and the single-stage target detection framework YOLOv5 to achieve efficient detection of defect locations and categories.This method alleviates the dependence on labeled data through a semi-supervised learning method,and uses a single-stage detection network to improve the efficiency of detection tasks.In addition,combined with the characteristics of actual tasks,this thesis introduces a coordinate attention mechanism to enhance the network’s ability to extract defect features,and an additional micro-defect detection head is added to the network to improve the detection performance of micro-defects.(2)Aiming at the problem that pseudo-labels in the semi-supervised defect detection network will introduce noise and damage the detection performance,this thesis proposes a new uncertainty-based semi-supervised defect detection method.This method starts from data uncertainty and model uncertainty,so that the network no longer learns only for deterministic pseudo-labels during the training process,and at the same time estimates the uncertainty of pseudo-labels,correcting the impact of noise that the pseudo-labels introduce on the network.The method proposed in this thesis is tested on the public data sets NEU-DET,GC10-DET,and Severstal.Experiments show that the semi-supervised defect detection method based on teacher-student network and YOLOv5 improves m AP by 9.8%,7.6%,9.0%compared to supervised learning when there is only 10% labeled data.The m AP of the semi-supervised defect detection method based on uncertainty awareness has been further improved,reaching 68.9%,65.9%,and 55.8%,respectively,which has obvious advantages over other semi-supervised methods.