Research On Cross-modal Hash Retrieval Algorithm For Information Asymmetry

In recent years,the rapid development of Internet media has brought the society into the era of big data information,and massive multi-modal data has emerged accordingly.People's demand for the retrieval of these massive multi-modal data is also increasingly urgent,and cross-modal hash retrieval can not only realize the retrieval of data across modals,but also has the characteristics of retrieval efficiency and low storage.Many researchers are sought after.However,the focus of most current algorithms is on the relevance and semantic gap between modal data,without much consideration of the difference in the amount of information between different modal data.Taking image and text modal data as an example,for a computer,the amount of information contained in image data is more abundant than text data.This paper will also conduct research on cross-modal hash retrieval for this kind of information asymmetry.In view of the problem of information asymmetry between image and text data.Based on the deep semantic features and deep hashing cross-modal retrieval network framework,this paper allocates different proportions of associative losses to image and text used as the reference learning objects,according to the richness of the information contained in the image and text data.So that the cross-modal hashing network performs asymmetric reference learning.Specifically,in order to maximize the use of the rich information of image data,this paper increases the proportion of associative loss using image as the reference learning object.Therefore,in this paper,the image is defined as the primary reference learning object of the cross-modal hash network,and the text data is the secondary reference learning object due to the small amount of information.On the basis of the above algorithm,this paper further proposes a reference hash code initialization method that is semantically similar in a single modal,and replaces the reference hash code random initialization method.The specific method is to perform reference learning in each single modal before asymmetric reference learning,and finally generate the semantically similar image and text reference hash code in the single modal,which is used to replace the randomly initialized image and text reference hash code.This will better guide the cross-modal hash network for asymmetric reference learning.Finally,this paper verifies the effectiveness of the proposed algorithm and the reference hash code initialization method by conducting experimental comparative analysis on the MIRFLICKR-25K,NUS-WIDE universal cross-modal data set.