Deep Multi-Similarity Hashing for Multi-label Image Retrieval

Abstract

mage retrieval based on deep hashing methods has attracted more and more attentions from both academic and industry, due to the out-standing performance of deep neural network in various tasks of computer vision. However, most of the hashing methods are designed to learn simple similarity only for single-label image retrieval, thus cannot work well for the multi-label cases. In this paper, we proposed a framework named Deep Multi-Similarity Hashing (DMSH) method to learn semantic binary representations for multi-label image retrieval task. In the proposed model, a convolutional architecture is incorporated with hash function to learn compact binary representations from every pair of images with multiple labels. On the purposed of learning semantic structure of multi-label images, we define the pairwise loss for multi-label image pairs, which is influenced by zero-loss interval under the control of the number of common labels. The objective loss function consists of hashing quantification loss and pairwise loss for multi-label images, which pays more attention to high-level similarity than low-level similarity during the training process. Furthermore, our proposed model is flexible to be implemented with various deep networks. Experiments on large scale dataset NUS-WIDE have proved the state-of-the-art performance of our proposed DMSH model in the task of multi-label image retrieval.