A Novel Unsupervised Hashing Method for Image Retrieval Based on K-Reciprocal Nearest Neighbors

Abstract

Hashing methods play an important role in large-scale image retrieval. Unsupervised hashing is popular in practical applications because it does not require labels for supervised training. However, as one of the important steps in unsupervised hashing, construction of semantic relationships between data by k-nearest neighbors has limitations due to the existence of false neighbors among the first k neighbors. In this paper, we propose a novel unsupervised deep hashing method for image retrieval. We firstly construct a semantic similarity matrix which utilizes deep features and the expanded k-reciprocal nearest neighbors to guide the learning of hash codes. After that, we design a deep neural network to preserve the structural information of original images. In addition, a weighted pairwise loss function generated by the positive pairs and negative pairs is employed to solve data imbalance problem. Extensive experiments on CIFAR-10, MIRFLICKR and NUS-WIDE datasets show that our method significantly outperforms the state-of-the-art unsupervised hashing methods.