Research On Three-dimensional Gray Matrix Image Recognition And Zero-shot Learning

The quality of steel plate production has always been an issue of particular concern to enterprises.It is not only related to the economic benefits and product influence of enterprises,but also plays a decisive role in the downstream manufacturing industry.But in the actual production process,the finished steel plate will inevitably have defective products,if these defective products are put on the market,it will bring huge losses.According to statistics,more than 60%of quality objections in domestic steel products are caused by steel surface defects,so it is very important to inspect the surface defects in the production process of steel plate.At present,most of the small and medium-sized steel plate manufacturers still use manual visual inspection or flicker light detection for the detection of steel plate surface defects.However,these methods are either easy to miss detection and wrong detection,and the detection efficiency is not high;or the maintenance cost is expensive,which can not achieve universal.Therefore,in view of the increasingly mature research of computer vision technology and deep learning,taking the detection and recognition technology of steel plate surface defects as the research object,this paper analyzes the advantages and disadvantages of existing computer vision technology and deep learning in the detection of steel plate surface defects,and improves the efficiency of defect detection and recognition through the design and improvement of algorithm model.The main research contents are as follows:An image segmentation algorithm based on 3D gray matrix is proposed to solve the problems of low recognition efficiency and obvious over segmentation in the steel plate defect image with complex gray structure and fuzzy edge.Firstly,the three-dimensional gray matrix is constructed by combining the spatial characteristics of the image gray matrix;secondly,the semi class variance is introduced to improve Kriging interpolation algorithm to draw the contour map of the three-dimensional gray matrix;secondly,the topological relation tree of the contour is constructed;finally,according to the combination of the self-defined global search strategy and the local search strategy,the local concave convex region is found,so as to locate the defect region Purpose of segmentation of steel plate surface defects.Based on the research and analysis of zero-shot learning theory in the field of image recognition,this paper summarized the main journal literature in recent years,and explored the possibility,advantages and disadvantages of deep learning in the task of steel plate surface defect classification and recognition.Using deep learning algorithm to train massive labeled data can achieve unprecedented recognition accuracy.However,the labeling of massive data is expensive,and it is difficult to obtain massive data for rare categories.Therefore,how to identify unknown classes that are rare or never seen during training is still a serious challenge.In response to this problem,the recent research on zero-sample image recognition technology is reviewed,and the application of zero-sample learning in image recognition is fully explained from the aspects of research background,model analysis,data set introduction,and experimental analysis.A transductive zero-shot classification model based on strengthening features is proposed to solve the problems of various surface defects of steel plate and the difficulty of obtaining annotation data set.Combined with the theory of transductive zero-shot learning and generalized settings,the model completes the task of steel plate surface defect image classification and recognition by constructing basic feature network,strengthening feature network,semantic feature network,feature embedding network and softmax classifier,and reduces the strong dependence of traditional deep learning model on massive labeled training data set.