Task-adaptive Asymmetric Deep Cross-modal Hashing

Supervised cross-modal hashing aims to embed the semantic correlations of heterogeneous modality data into binary hash codes with discriminative labels. It can support efficient large-scale retrieval due fast speed and low storage cost. However, existing methods equally handle tasks, simply learn same couple functions in a symmetric way. Under such circumstances, characteristics different tasks are ignored sub-optimal performance may be brought. Motivated by this, we present Task-adaptive Asymmetric Deep Cross-modal Hashing (TA-ADCMH) method this paper. task-adaptive for two sub-retrieval via simultaneous representation asymmetric learning. Different from previous methods, our learning framework jointly optimizes preserving multi-modal features codes, regression query explicit With model, learned effectively preserve correlations, meanwhile, adaptively capture semantics. Besides, design an discrete optimization strategy directly which alleviates relaxing quantization errors. Extensive experiments demonstrate state-of-the-art proposed TA-ADCMH various aspects.