Fusing Semantic Prior Based Deep Hashing Method for Fuzzy Image Retrieval

Abstract

Fuzzy image retrieval is a novel visual application about designing a multi-modal retrieval system that supports querying across image modalities, e.g., a fuzzy type image searches for some similar images. However, most existing deep hashing methods are not suitable for obtaining a robust image hash code on multi-modal retrieval task. In this paper, we propose Fusing Semantic Prior based Deep Hashing (FSPDH) method, which is the first attempt to integrate unsupervised semantic prior into end-to-end deep architecture for fuzzy image retrieval task. The major contribution in this work is extracting the prior information from images and incorporating it effectively into hash learning process. In addition, our strategy can be usefully used in single-modal retrieval task. Extensive experiments show that our FSPDH approach yields state-of-the-art results in both multi-modal and single-modal image retrieval tasks on our image datasets.