Graph Regularized Unsupervised Deep Hashing for Large Scale Image Retrieval

Abstract

With the development of technology, the number of data is growing rapidly day by day. How to perform efficient big data retrieval becomes a critical problem. Similarity-preserving hashing has been widely used in large-scale information retrieval because of its low storage cost and high computation efficiency. It maps the data from high-dimensional feature space into binary hamming space while preserving the similarity. Particularly, deep learning based hashing methods have shown their significantly advantages in both effectiveness and accuracy. However, the performance of unsupervised deep hashing algorithms is still unsatisfactory because semantic labels are not available in unsupervised learning. In this paper, we propose an end-to-end unsupervised deep hashing model to simultaneously learn image representation and generate compact hash codes. Intuitively, we consider that the convolutional neural networks can capture high level semantic information. Therefore, we explore the semantic relations between images by utilizing data mining techniques on deep features. Specifically, we first extract the features from pretrained CNN model and conducting K-means and k-NN graph construction on these features. Then we train the deep network using the pseudo labels. In addition, due to the binary constrain of hash codes, we iteratively update the binary hash codes using cyclic coordinate descent method. Extensive experiments validate the performance of our method, which outperforms previous state-of-art methods in image retrieval task.