Scalable Supervised Asymmetric Hashing

Learning compact hash codes is widely employed for rapid similarity search, capitalizing on reduced storage requirements and efficient query speeds. However, mastering discriminative binary codes that perfectly preserve full-pairwise similarities in high-dimensional real-valued features remains a challenging task for ensuring optimal performance. To tackle this challenge, this chapter introduces a novel method, Scalable Supervised Asymmetric Hashing (SSAH)Scalable Supervised Asymmetric Hashing (SSAH) [1]. SSAH adeptly approximates the full-pairwise similarity matrix by leveraging the maximum asymmetric inner product of two distinct non-binary embeddings. To comprehensively exploit semantic information, the method simultaneously considers supervised label information and refined latent feature embeddings to construct high-quality hashing functions, enhancing the discriminative capability of learned binary codes. Specifically, SSAH learns two distinctive hashing functions by minimizing regression loss for semantic label alignment and encoding loss for refined latent features. Notably, instead of utilizing only partial similarity correlations, SSAH directly employs the full-pairwise similarity matrix to prevent information loss and performance degradation. Its optimization phase adeptly manages the cumbersome computation complexity of the \(n \times n\) matrix. Additionally, an efficient alternating optimization scheme with guaranteed convergence is designed to address the resulting discrete optimization problem. Experimental results on diverse benchmark datasets highlight the superiority of SSAH over recently proposed hashing algorithms.