Research On Surface Defect Image Clustering Method Of Metal Strip

With the increasingly fierce competition in the metal materials industry such as steel and nonferrous metals,comsumers have raised more and more attention to the surface quality of metal strips.Therefore,the classification of surface defects on strips is crucial.At present,the common classification methods for surface defects of metal strips are mostly based on a large number of artificially labeled samples,which are expensive.It is even more necessary for the factory to identify and classify the surface defects of the metal strip without a large number of labeled samples.This thesis will focus on this need,the main work is as follows:(1)The research status of surface defect identification of metal strip products is reviewed.The shortcomings of current detection methods are analyzed.The characteristics and causes of surface defects of common metal strips are summarized.The evaluation indexes of cluster analysis and the theoretical basis of artificial neural networks are introduced.(2)The semi-supervised clustering using Bag-of-Words method is designed to realize identification of surface defects under the condition that only one mark sample is known.Based on the characteristics of scale-invariant feature transform(SIFT)and gray histogram,the fusion feature is proposed as the Bag-of-Words.Experiments were carried out using three kinds of features respectively.The experimental results showed that the identification effect of the fusion features was better than the single feature on the strip surface defect dataset.(3)In the condition of unmarked samples,the clustering analysis of the defect images is carried out to research the clustering effect based on different feature representations.Firstly,the gray histogram features combined with K-Means algorithm are used to realize the clustering of defects;then the net learns autonomously and obtains better image feature representation using the deep autoencoder.Finally,the convolutional neural network is combined with the deep autoencoder to realize the autonomic learning of the best feature representation.Experiments showed that the improved autoencoder extracted feature representation had the best clustering effect.(4)The idea of deep embedding for clustering(DEC)is used to realize the joint optimization of feature space and clustering algorithm.According to this,a clustering network based on DEC is proposed.The result of experiments showed that this method could further improve the clustering accuracy,but the robustness of this method was limited.Aiming at this problem,a new adaptive sample training method was proposed,which was more aligned with the human learning method.The result of experiments showed that the training method effectively improved the robustness of the DEC-based clustering network and could obtain more accurate clustering.In summary,the main contributions of this thesis include:(1)proposed a semi-supervised clustering method based on Bag-of-Words,which solved the defect recognition under the condition that only one labeled sample;(2)proposed an improved autoencoder,the net solved the defect clustering under unmarked samples.(3)The DEC-based clustering net was proposed,the training method of the net was improved,and the clustering precision and robustness were further improved.