Research On Strip Surface Defect Identification Method Based On Residual Shrinkage Network

Deep learning plays an increasingly important role in industrial product detection and can be applied in more and more scenarios.However,there are also problems such as "large intra-class differences,small inter-class differences" and unbalanced small data samples.Researchers attempt to apply deep learning to product defect identification to solve these problems.In this paper,the attention mechanism,soft threshold and residual shrinkage network are combined to identify the defects of strip steel by residual shrinkage network and feature fusion residual shrinkage network.The specific research contents are as follows:(1)The overall method of product surface quality defect identification and its research value are expounded.The mechanism of attention,soft threshold and residual shrinkage network are discussed.(2)Aiming at the problem of "large intra-class difference and small inter-class difference" of industrial data collected,this paper constructed an improved residual shrinkage network method for product surface defect recognition to optimize network performance.This paper introduces the selection of attention mode and the modeling process of soft attention embedding residual network,and expounds their advantages.Through experimental verification,the improved residual shrinkage network proposed in this paper has certain advantages over the traditional machine learning method and the convolutional neural network method,and its accuracy can reach 98.8%.(3)In view of the problem of small data,small samples and unbalanced samples,this paper constructed a residual shrinkage network surface defect recognition method based on feature fusion.Pyramid convolution is used as the main feature extraction technology,and it is integrated into the residual contraction network to expand the receptive field area of the residual contraction network,and feature fusion of multi-scale features is carried out in the form of convolution kernel attention.The method can still maintain good performance under small data training samples,and its accuracy is up to 97%.The method proposed in this paper has great advantages in unbalanced samples,and its accuracy can reach 98%.Compared with other methods,it has great advantages.