语义嵌入重构的跨模态哈希检索  

Semantics embedding and reconstructing for cross-modal hashing retrieval

作　　者：许炫淦 房小兆 孙为军[2] 韩娜 吴惠粦 黄永慧[2] Xu Xuangan;Fang Xiaozhao;Sun Weijun;Han Na;Wu Huilin;Huang Yonghui(School of Computer,Guangdong University of Technology,Guangzhou 510006,China;School of Automation,Guangdong University of Technology,Guangzhou 510006,China;School of Computer Science,Guangdong Normal University of Technology,Guangzhou 510006,China;Guangzhou National Science&Technology Innovation Center of Modern Agricultural Industry,Guangzhou 510500,China)

机构地区：[1]广东工业大学计算机学院,广州510006 [2]广东工业大学自动化学院,广州510006 [3]广东技术师范大学计算机科学学院,广州510006 [4]广州国家现代农业产业科技创新中心,广州510500

出　　处：《计算机应用研究》2022年第6期1645-1650,1672,共7页Application Research of Computers

基　　金：广东省重点领域研发计划项目(2019B010121001);国家自然科学基金面上项目(61772141,62006048,61972102);科技部国家重点研发计划项目(2018YFB1802400);广东省科技计划项目(2021A1515012017,2019B110210002,2019B020208001)。

摘　　要：针对大多数跨模态哈希检索方法仅通过分解相似矩阵或标签矩阵,从而导致标签语义信息利用不充分、标签矩阵分解过程语义信息丢失以及哈希码鉴别能力差的问题,提出了一种语义嵌入重构的跨模态哈希检索方法。该方法首先通过最小化标签成对距离和哈希码成对距离之间的距离差,从而将标签矩阵的成对相似性嵌入哈希码;接着对标签矩阵分解并重构学得共同子空间,共同子空间再回归生成哈希码,从而将标签矩阵的类别信息嵌入哈希码,并有效地控制标签矩阵分解过程的语义信息丢失情况,进一步提高哈希码的鉴别能力。在公开的三个基准数据集上进行了多个实验,实验结果验证了该方法的有效性。In order to solve the problem that most cross modal hash retrieval method only decompose the similarity matrix or label matrix,which led to insufficient utilization of label semantic information,loss of semantic information in the decomposition process of label matrix and weak discrimination of hash code,this paper proposed a hash method named SERCH(semantics embedding and reconstructing hashing).Firstly,it minimized the distance difference between the pairwise distances of label and hash code to embed the pairwise similarity of label matrix into hash code.Then,it decomposed and reconstructed the label matrix to obtain the common subspace,and regressed the common subspace to generate the hash code,so as to embed the category information into hash code,effectively controlled the loss of semantic information in matrix decomposition process,and further improved the discrimination of hash code.Extensive experiments were carried out on three benchmark data sets,and the experimental results demonstrate the effectiveness of SERCH.

关 键 词：跨模态检索 有监督哈希 相似矩阵 标签信息 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]