Multi-Granularity Based Collaborative Learning for Semi-Supervised Hashing

Shuai Cheng, Lin Wang, Xiaoshuai Hao, Wanqian Zhang, Xiaohua Chen, Wei Wang

Abstract: Recently, deep semi-supervised hashing methods have achieved remarkable success by simultaneously leveraging sufficient unlabeled data and limited labeled data. However, these methods focus on the pseudo-label semantic similarity relation of unlabeled data, while ignoring multi-granularity similarity relations. The fine-grained instance-level and neighborhood similarity relations can help to learn a more detailed data distribution in the Hamming space, which is beneficial to improve the discrimination of hash codes. We thus propose a novel Multi-Granularity Based Collaborative Learning Hashing (MGCLH), which utilizes the complementary multi-granularity similarity relations of unlabeled data to collaboratively improve the discrimination of hash codes. Specifically, we introduce an Instance-Wise Contrastive Module (ICM), which embeds instance-wise similarity relation into hash codes to achieve coarse clustering of hash codes. Moreover, we design a Neighborhood Consistent Module (NCM) to capture neighborhood similarity relations for preserving the inherent neighborhood structure of hash codes. Furthermore, the Class-Wise Contrastive Module (CCM) embeds the class-wise semantic similarity relation between unlabeled data into the hash codes to improve its inter-class separability. Extensive experimental results on four image datasets demonstrate that the proposed method outperforms several state-of-the-art semi-supervised hashing methods.

External IDs:dblp:conf/icmcs/ChengWHZCW25