Correlation Hashing Network for Efficient Cross-Modal Retrieval

Due to the storage and retrieval efficiency, hashing has been widely deployed to approximate nearest neighbor search for large-scale multimedia retrieval. Cross-modal hashing, which improves the quality of hash coding by exploiting the semantic correlation across different modalities, has received increasing attention recently. For most existing cross-modal hashing methods, an object is first represented as of vector of hand-crafted or machine-learned features, followed by another separate quantization step that generates binary codes. However, suboptimal hash coding may be produced, because the quantization error is not statistically minimized and the feature representation is not optimally compatible with the binary coding. In this paper, we propose a novel Correlation Hashing Network (CHN) architecture for cross-modal hashing, in which we jointly learn good data representation tailored to hash coding and formally control the quantization error. The CHN model is a hybrid deep architecture constituting four key components: (1) an image network with multiple convolution-pooling layers to extract good image representations, and a text network with several fully-connected layers to extract good text representations; (2) a fully-connected hashing layer to generate modality-specific compact hash codes; (3) a squared cosine loss layer for capturing both cross-modal correlation and within-modal correlation; and (4) a new cosine quantization loss for controlling the quality of the binarized hash codes. Extensive experiments on standard cross-modal retrieval datasets show the proposed CHN model yields substantial boosts over latest state-of-the-art hashing methods.