Research On Classification Method Of Strip Steel Surface Defects Based On Semi-supervised Learning

The steel industry is an important basic industry in our country,and strip steel is an important intermediate material.Surface defect detection technology during strip processing has always been a research hotspot.Traditional computer vision combined with deep learning strip surface defect classification methods,under supervised learning,often pay attention to the improvement of edge detection algorithms and classification networks.The strip surface defect classification task has the characteristics of a small data set sample size and difficulty in labeling samples,which limits the performance improvement of the classification model under traditional supervised learning.Semi-supervised Learning(SSL)can use both labeled and unlabeled samples to train the classification model under deep learning,thereby reducing the labeling work of training data set samples.In order to introduce the idea of semi-supervised learning into the strip surface defect classification system,and solve the problems of sample labeling that restrict the performance of traditional supervised learning,the semi-supervised learning strip surface defect classification system based on Resnet18 combines the pseudo-label idea of semisupervised learning.The classification model that originally needs to be trained under supervised learning can be trained through a data set that is a mixture of labeled samples and unlabeled samples to achieve semi-supervised learning and reduce the labeling of the initial data set in the strip surface defect classification task.The semi-supervised learning strip surface defect classification system based on DCGAN and Resnet18 introduces Generative Adversarial Networks(GAN)to generate unlabeled samples,which provides more sample support for the training of classification models.In addition,under semi-supervised learning,in order to effectively combine GAN and classification network,the system adopts a new voting mechanism,so that the samples generated by GAN can be labeled with a suitable pseudo label before being learned by the classification network.Under the idea of data enhancement,semi-supervised learning can improve the classification accuracy of the system when the initial sample size of strip steel surface defects is limited,and make the system more robust and generalizable.Experiments show that the semi-supervised learning system based on Resnet18 can reduce the workload of labeling by 75% with a loss of classification accuracy of about 2%;while the semi-supervised learning system based on DCGAN and Resnet18 can be expanded at the appropriate sample expansion multiples.The system's classification model improves the highest classification accuracy by about 2%,achieving good results.