Research On Strip Steel Surface Defects Classification Method Based On Semi-Supervised Deep Learning

As one of the important basic industries in China,strip steel is a very important intermediate material in the steel industry.In the process of processing,various kinds of defects are inevitably generated on the surface,which greatly affects the performance and lifetime of the material,so strip surface defect classification technology has been a hot research topic.Existing strip surface defect classification techniques usually combine computer vision with deep learning,and the whole system operates in a fully supervised mode.The performance of this fully supervised strip surface defect classification model is inextricably linked to the number of strip surface defect samples,all of which are required to carry labels.However,the performance of the traditional fully-supervised strip surface defect classification model is severely compromised by the small number of strip surface defect samples in industrial scenarios and the fact that most of the samples are unlabeled.Semi-supervised learning(SSL)can train classification models under deep learning using a small number of labeled samples and a large number of unlabeled samples.Therefore,this thesis applies the idea of semi-supervised learning to the strip surface defect classification system,which can train the semi-supervised strip surface defect classification model with a small number of labeled samples and a large number of unlabeled samples,and realize the model to complete the strip surface defect classification work autonomously.This can effectively alleviate the problem of poor performance of traditional supervised classification models due to the limited number of labeled samples,and reduce the workload of sample labeling.However,the image samples of the strip surface defect image dataset in industrial scenes are small and mostly unlabeled,the image quality is not high and the sample size is too large,which may cause overfitting of the semi-supervised image classification model and eventually affect the performance of the classification.Therefore,this thesis proposes a semi-supervised strip steel surface defect classification model as well as a semi-supervised strip steel surface defect classification model based on DCGAN by modifying the original semi-supervised deep learning framework.The specific innovative parts of this paper are as follows:1.data preprocessing stage: DCGAN is used to generate some virtual samples to supplement the number of unlabeled samples in the training set to improve the classification accuracy of the model;the Cow Mask method is also used to perform data enhancement on the samples,aiming to reduce the sensitivity of the model to images.2.Feature extraction stage: Combining the CBAM attention mechanism with the Dense Net network,the model aims to understand the importance of different local information of different images from both channel and spatial perspectives and learn the key information of images that are more important for the current classification task.3.Classifier part: the supervised loss of labeled samples and the unsupervised loss of unlabeled samples are reconstructed using uncertainty weighting methods,specifically to make effective use of the features of each sample,to learn the mapping relationship between unlabeled and labeled samples more effectively,and to apply semi-supervised learning methods to the classification of surface defects in strip steel.In this thesis,we compare with existing fully supervised and semi-supervised strip surface defect classification models and show that applying semi-supervised learning ideas to strip surface defect classification systems can effectively alleviate the problem of poor performance of traditional fully supervised strip surface defect classification models due to the small number of labeled samples of strip surface defects in industrial scenarios.Meanwhile,compared with the existing semi-supervised model,the semi-supervised classification model in this thesis improves the image classification accuracy by about 2% by expanding the number of unlabeled samples through DCGAN and weighting the original supervised loss as well as unsupervised loss using the uncertainty weighting method,which achieves more desirable results.