Orthogonal Embedding Hashing with Semantic Reinforcement for Modal Matching

作     者：Fang, Yixian Wang, Canwei Yuan, Feng Sun, Qian Ren, Yuwei 

作者机构：School of Information Engineering Shandong Management University Shandong Province Jinan250357 China School of Information Science and Engineering Shandong Normal University Shandong Province Jinan250014 China Faculty of Mathematics and Artificial Intelligence Qilu University of Technology Shandong Academy of Sciences Shandong Province 250353 China 

出 版 物：《SSRN》 

年 卷 期：2022年

核心收录：

主　　题：Semantics 

摘      要：Cross-modal hashing (CMH) is a fast modal association technique in multimodal learning community. Current CMH techniques, however, are susceptible to scaling in binarization, resulting in semantic deficiency of hash codes. It is mainly caused by the redundancy of raw features and the loss of binary quantization. For these occasions, a novel orthogonal embedding hash with semantic reinforcement (SR-OEH) is proposed to alleviate these limitations. To be specific, SR-OEH firstly leverages the associated tags to guide feature semantic condensation by optimized adjustment weights, which are trained by maximizing the sample edge, so as to highlight the relevant features with detrimental redundancy degradation. What s more, we construct an efficient iterative optimization algorithm to learn a strong semantic hash dictionary by supervised rotation factor step by step orthogonal, so as to effectively offset the quantization loss induced by the mixed integer optimization. Finally, two common hash functions are employed to concatenate the hash book and the deepened semantic features, thus accomplishing cross modal matching. The promising results obtained on five public data sets bear out the superiority of proposed strategy against the representative state-of-the-art cross-modal hashing baselines. © 2022, The Authors. All rights reserved.