Visual-to-Semantic Hashing for Zero Shot Learning

Hashing-based multimedia retrieval are facing the problem of the dramatic increase of data, especially new unseen categories. It is time-consuming, expensive, and sometimes impractical to label new samples and retrain the hashing model. Recently, several zero-shot hashing methods are proposed to generate the hash function with good generalization for unseen classes, via exploring semantic information and similarity relationship. However, the performance of existing solutions is still not satisfying. Therefore, we propose a modified two-stage framework, called Visual-to-Semantic Hashing (VSH). To fully exploit the semantic information, visual feature is firstly mapped to the semantic space, and then generate its hash codes. To transfer supervised knowledge from seen classes to unseen classes, a margin-based ranking loss is employed to learn the semantic structure. To boost the discriminability of semantic mapping, a classification module is adopted to distinguish between different semantic mapping vectors. Plenty of experiments demonstrate that the proposed VSH is superior to state-of-the-art methods.