Deep Collaborative Graph Hashing

The remarkable success of deep hashing in large-scale image retrieval is attributed to its potent discriminative representation through deep learning and the computational efficiency of compact hash code learning. However, existing deep semantic-preserving hashing approaches often treat semantic labels as ground truth for classification or transform them into prevalent pairwise similarities, overlooking interactive correlations between visual semantics in images and category-level labels. Moreover, these strategies utilize fixed piecewise or pairwise semantics as optimization objectives, limiting flexibility in semantic representation and adaptive knowledge communication during hash code learning. This chapter introduces a pioneering approach, Deep Collaborative Graph Hashing (DCGHDeep collaborative graph hashing (DCGH)), considering multi-level semantic embeddings, latent common space construction, and intrinsic structure mining for discriminative hash code learning in large-scale image retrieval. DCGHDeep collaborative graph hashing (DCGH) introduces the first collaborative graph hashing for image retrieval. Instead of a conventional single-flow visual network, a dual-stream feature encoding network jointly explores multi-level semantic information across visual and semantic features. A shared latent space is established based on space reconstruction to concurrently explore information and bridge the semantic gap between visual and semantic space. Additionally, a graph convolutional network preserves latent structural relations in optimal pairwise similarity-preserving hash codes. The entire learning framework is optimized end to end. Extensive experiments across diverse datasets demonstrate DCGHDeep collaborative graph hashing (DCGH)’s superb image retrieval performance against state-of-the-art supervised hashing methods. The incorporation of collaborative graph hashing facilitates capturing intricate semantic correlations for enhanced large-scale image retrieval.