Specific class center guided deep hashing for cross-modal retrieval

Abstract

Hashing approaches show excellent retrieval efficiency and low storage usage in search tasks. In general, most existing deep hashing approaches mainly focus on constructing the pairwise similarity matrix by exploiting the supervised information. However, they rarely construct a label network using the labels to guide hash code learning and thus cannot generate accurate hash codes in some cases. To alleviate this issue, a novel supervised hashing model, named Specific Class Center Guided Deep Hashing (SCCGDH), is proposed in this paper. The purpose of SCCGDH is to learn the specific class centers from the neural network and guide the hashing learning of multi-media data. We design three different neural networks: label network, image network and text network. Specifically, the label network outputs the hash codes of the center of each category. The hash codes from the image network and the text network are encouraged to approximate the corresponding specific centers, reducing the intraclass variation of multi-media data. Furthermore, we seek hash codes of different modalities to be consistent by minimizing the inter-modal invariance loss. We integrate three neural networks into a unified end-to-end hashing learning framework. Experimental results on three cross-modal datasets show that our proposed SCCGDH approach can obtain the better performance than other state-of-the-art hashing approaches.