Transformed anti-sparse learning for unsupervised hashing

Abstract

Anti-sparse representation was recently considered for unsupervised hashing, due to its remarkable robustness to the binary quantization error. We relax the existing spread property [4, 22] for anti-sparse solutions, to a new Relaxed Spread Property (RSP) that demands milder conditions. We then propose a novel Transformed Anti-Sparse Hashing (TASH) model to overcome several major bottlenecks, that have significantly limited the effectiveness of anti-sparse hashing models. TASH jointly learns a dimension-reduction transform, a dictionary and the anti-sparse representations in a unified formulation. We have conducted extensive experiments on real datasets and practical settings, and demonstrate the highly promising performance of TASH.