Hierarchical hashing for image retrieval

Abstract

Hashing has been widely used in large-scale vision problems thanks to its efficiency in both storage and speed. The quality of hashing can be boosted when supervised information is used to learn hash functions. On large-scale hierarchical datasets, hierarchical semantic information reflects the relationship between classes and their children, which however has been ignored by most supervised hashing methods. In this paper, we propose a hierarchical hashing method for image retrieval. This method models and fuses both hierarchical semantic level relationship through taxonomy structure of dataset and feature level relationship of images into an integrated learning objective, then an optimization scheme is developed to solve the learning problem. Experiments are performed on two large-scale datasets: ImageNet ILSVRC 2010 and Animals with Attributes (AWA) dataset. Besides standard evaluation criteria, we also developed hierarchical evaluation criteria for image retrieval and classification tasks. The results show that the proposed method improves the accuracy of supervised hashing in both types of criteria.