Inductive Structure Consistent Hashing

Semantic-preserving hashing enhances multimedia retrieval by transferring knowledge from original data to hash codes, preserving both visual and semantic similarities. However, a significant bottleneck arises in effectively bridging trilateral domain gaps—visual, semantic, and hashing spaces—to further elevate retrieval accuracy. This chapter introduces the Inductive Structure Consistent Hashing (ISCH)Inductive Structure Consistent Hashing (ISCH) method which coordinates semantic correlations among the visual feature space, binary class space, and discrete hashing space. An inductive semantic space is formulated using a multi-layer stacking class-encoder, transforming naive class information into flexible semantic embeddings. A semantic dictionary learning model aids visual–semantic bridging and guides the class-encoder to mitigate visual–semantic bias. The visual descriptors and semantic class representations are regularized through a coinciding alignment module. To generate privileged hash codes, semantic and prototype binary code learning jointly quantifies semantic and latent visual representations into unified discrete hash codes. An efficient optimization algorithm addresses the resulting discrete programming problem. Comprehensive experiments across four large-scale datasets showcase the superiority of ISCH over state-of-the-art alternatives under different evaluation protocols, emphasizing its effectiveness in addressing the trilateral domain gaps and improving multimedia retrieval accuracy.