Supervised topology preserving hashing

Abstract

Learning based hashing is gaining traction in large-scale retrieval systems. It aims to learn compact binary codes that can preserve semantic similarity in the hamming space. This paper presents a supervised topology hashing (SPTH) algorithm to learn compact binary codes that can exploit both the supervisory information as well as the local topology structure of datasets. To build a connection between the original space and the resultant hamming space, we minimize the quantization errors together with a classification error term and a topology preserving term. A nonlinear kernel feature space is further used to improve the generalization power. An alternating iterative algorithm is developed to minimize the complex objective function that contains both continuous and discrete variables. Experimental results on three benchmark datasets demonstrate the effectiveness of the proposed method on image retrieval tasks.