Explore instance similarity: An instance correlation based hashing method for multi-label cross-model retrieval

•The main contributions can be summarized as follows:•We define instance similarity and then utilize it to guide the cross-modal hashing learning. By using the instance similarity, the network is able to learn the minor difference between specific instances.•We propose an end-to-end cross-modal hashing framework to bridge intra-modal and cross-modal relationship, which integrates semantic feature learning and hash code learning into the same deep learning architecture.•Extensive experiments in different datasets demonstrate the outstanding performance of our methods compared to other state-of-the-art cross-modal hashing methods. With the rapid growth of multimedia data, cross-media hashing has gained more and more attention. However, most existing cross-modal hashing methods ignore the multi-label correlation and only apply binary similarity to measure the correlation between two instances. Most existing methods perform poorly in capturing the relevance between retrieval results and queries since binary similarity measurement has limited abilities to discriminate minor differences among different instances. In order to overcome the mentioned shortcoming, we introduce a novel notion of instance similarity method, which is used to evaluate the semantic correlation between two specific instances in training data. Base on the instance similarity, we also propose a novel deep instance hashing network, which utilizes instance similarity and binary similarity simultaneously for multi-label cross-model retrieval. The experiment results on two real datasets show the superiority of our proposed method, compared with a series of state-of-the-art cross-modal hashing methods in terms of several metric evaluations.