Individual mapping and asymmetric dual supervision for discrete cross-modal hashing

Cross -modal hashing has gained popularity in similarity search due to its excellent query efficiency and economical storage costs. However, current models frequently overlook the distinctive property of each modality, resulting in reduced accuracy due to inadequate utilization of these attributes. Moreover, there is a weak semantic relevance between modality attributes and multiple supervision knowledge (the labels and similarity constraints constructed by labels), accompanied by a cumulative quantization of the models. To address these issues, we propose an Individual Mapping and Asymmetric Dual Supervision method (IMADS). It merges specific and shared information to effectively learn a cross -modal representation space. Furthermore, we present an asymmetric dual supervision learning framework to produce discriminative hash codes. This framework achieves two primary goals: (1) Combing cross -modal representation and multiple supervision information to enhance the consistent relation of distinct modalities, and (2) developing a discrete optimization algorithm to mitigate the information loss caused by the hash code. Comprehensive experimental results illustrate that the introduced IMADS outperforms other stat-of-the-art hashing methods.