DAH: Discrete Asymmetric Hashing for Efficient Cross-Media Retrieval

Given the merits in high computational efficiency and low storage cost, hashing techniques have been widely studied in cross-media retrieval. Existing cross-media hashing algorithms usually adopt an equal-length encoding scheme to represent the multimedia data. However, the strictly equal length encoding scheme may not effectively characterize the multimedia data because the underlying dimension of different modalities is often various, challenging its flexible generalization to real-world applications. In addition, there exist other challenges in designing a cross-media retrieval system, e.g., how to address the discrete constraints, how to avoid the complexity in calculating the large
$n\times n$
similarity matrix, and how to effectively exploit the discriminative label information. To conquer the above challenges, we propose a novel model, i.e., discrete asymmetric hashing (DAH). In particular, DAH exploits a flexible model to formulate the cross-modal retrieval, which can seamlessly deal with equal or unequal hash length encoding scenarios. Moreover, DAH constructs a supervised semantic embedding framework by jointly minimizing the distance-distance difference and label reconstruction error, significantly reducing the computational complexity. An asymmetric strategy is employed to establish the connection between hash codes and the latent subspace. Furthermore, the hash codes can be learned discretely by the designed optimization algorithm, with which the large quantization error can be avoided. Besides, the developed DAH is a two-stage method, a semantic intersection scheme is proposed in the second stage, resulting in more powerful hash functions. Extensive experiments conducted on several databases show that our DAH is superior to several recent competitive methods in terms of efficiency and accuracy in equal-length encoding scenarios. Our method also achieves effective performance in unequal length encoding scenarios, improving the flexibility in the real-world retrieval process.