Truncated Cauchy for Robust fast supervised discrete hashing

Abstract

We propose a novel data-dependent hashing algorithm named Truncated Cauchy for Robust fast supervised discrete hashing (RFSDH) for robust subspace learning. In this paper, a Truncated Cauchy loss is proposed to measure the factorization error, which can handle outliers by truncating large errors. The proposed method can inhibit the unreliable binary codes, which generate the optimal binary codes. Our method can be expressed as a mixed-integer optimization problem, which can be to solve by iterative discrete cyclic coordinate descent. RFSDH method accomplishes outperform other hashing learning methods in image retrieval effect on two large scale datasets.