Research On Cross-media Semantic Mining Based On Deep Canonical Correlation Analysis

From the era of portal websites in the 1990 s,through various developments,to the current era of big data,multimedia research has been given more and more scientific,social and commercial value.How to maximize the hidden information contained in different types of media information is particularly important.To overcome the semantic gap between different media and complete the semantic mining,it is necessary to map different dimensions of media into a public media space.Existing cross-media semantic mining methods,such as canonical correlation analysis,kernel canonical correlation analysis and deep canonical correlation analysis,only utilize the relevant features,but ignore the semantic feature.And without using the label information,only through unsupervised learning to complete semantic mining,which is not conducive to understanding the meaning expressed by the media itself,but also restricts the semantic understanding among various media data.In order to solve the above problems,the cross-media semantic mining algorithm proposed in this paper is based on deep canonical correlation analysis,and proposes deep semantic canonical correlation analysis algorithm.By introducing the semantic information of media data about categories into the loss function of deep canonical correlation analysis,the features containing both relevant information and semantic information are obtained.Then,we use logistic regression method to classify all kinds of media data,obtain the category features,and construct the combined semantic space based on the deep semantic correlation analysis algorithm to further mine the semantic information contained in the media data.This paper uses the effect of cross-media retrieval to verify the effectiveness of the algorithm.In two authoritative cross-media datasets,the proposed deep semantic canonical correlation analysis algorithm and public subspace algorithm are compared with four mainstream cross-media semantic mining methods.The experimental results show that the average performance of the proposed algorithm is better than that of the other four algorithms.