HCCN

Abstract

Due to its computation efficiency and retrieval quality, hashing has been widely applied to approximate nearest neighbor search for large scale image retrieval, while deep hashing further improves the retrieval quality by end-to-end representation learning and hash coding. However, subject to the ill-posed gradient difficulty in the optimization with sign activations, existing pairwise similarity learning based hashing methods need to first learn continuous representations and then generate binary hash codes in a separated binarization step, which cannot learn similarity relationships adequately to generate sufficient good hash codes and suffer from substantial loss of retrieval quality. To overcome this limitation, this work presents a continuous learning method based on hash (HCCN). The main idea is to improve the recognizability of the hash code through central similarity learning. At the same time, continuous learning is used to optimize the deep network with symbolic function activation to solve the ill-posed gradient problem and reduce the feature loss of the data. Comprehensive experiments demonstrate that HCCN can generate cohesive hash codes and achieve noticeable boost in retrieval performance on three datasets, NUS-WIDE, CIFAR-10, and MS-COCO.