The Research On Cross-Media Retrieval Based On Deep Correlation Mining

With the arrival of the Internet plus era,the amount of multimedia information on the network is growing geometrically relative to several decades ago.People's demand for searching information they need in such complex condition is increasing at the same time.The traditional single modal data retrieval has been unable to meet people's demand for multimodal data retrieval,the concept of cross-media retrieval has emerged as the times require,and has rapidly become a hot spot of scholars all over the world in recent years.Cross-media retrieval aims at breaking the shackles of single-modal retrieval confined to the same multimedia form.It hopes to achieve mutual retrieval across media forms.The most urgent problem for cross-media retrieval is to mine the semantic relations between the underlying features and improve the similarity between them.In view of the above problems,some scholars put forward the method of canonical correlation analysis,but its performance on nonlinear problems is not so good,and the improved kernel canonical correlation analysis method is not satisfactory enough.In this paper,a deep correlation mining method is proposed,which applies deep learning to train different media features and then the correlation between the characteristics after training is fusion analyzed to solve the heterogeneity between them,making the characteristics of different multimedia data comparable.On this basis,we employ the Levenberg-Marquardt(LM)algorithm to overcome the shortcoming of getting into local minimum solution easily of the deep learning method while in gradient training process.The precision and recall rate are taken as the evaluation indicators in the experiment.The experimental results on different databases show the effectiveness of the proposed method in cross-media retrieval,and also show the superiority compared with the traditional multi-modal retrieval method.