相似度保持跨模态哈希检索  

Similarity Preserving Hashing for Cross-modal Retrieval

作　　者：房小兆 唐宝动 韩娜[1] 孙为军[2] 滕少华[1] FANG Xiao-zhao;TANG Bao-dong;HAN Na;SUN Wei-jun;TENG Shao-hua(Department of Computer Science,Guangdong University of Technology,Guangzhou 510006,China;School of Automation,Guangdong University of Technology,Guangzhou 510006,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

出　　处：《小型微型计算机系统》2021年第2期256-263,共8页Journal of Chinese Computer Systems

基　　金：国家自然科学基金面上项目(61772141,61972102)资助;科技部国家重点研发计划项目(2018YFB1802400)资助;广东省科技计划项目(2019B110210002,2019B010121001,2019B020208001,2019B010118001)资助;广州市科技计划项目(201903010107,201804010347)资助.

摘　　要：由于不同模态数据之间的异构性以及语义鸿沟等特点,给跨模态数据分析带来巨大的挑战.本文提出了一个新颖的相似度保持跨模态哈希检索算法.利用模态内数据相似性结构使得模态内相似的数据具有相似的残差,从而保证学习到的哈希码能够保持模态内数据的局部结构.同时利用模态间数据的标签,使得来自于不同模态同时具有相同标签的数据对应的哈希码能够紧密聚集在一起.为了进一步提高哈希码的鉴别能力,算法引入线性回归使得投影后的哈希码能够逼近样本的二值标签.在三个公开的不同跨模态检索数据集上的实验结果显示本文算法有较高的平均查准率.It is challenging toretrieve information on massive multimodal data due to the heterogeneity and semantic gap of multimodal data.This paper proposes a novel similarity preserving cross-modal hash(SPCMH)retrieval method,which uses the similarity structure of the intra-modal to make similar samples have similar residuals.This can ensure that the learned hash code can preserve the local structure of data.The labels of data from different modalities are used to guarantee that the hash codes of data from different modalities but with the same label can be closely clustered together.In order to further improve the discrimination abilityof the hash code,SPCMH introducers linear regression so that the learned hash code can approximatethe binary labels of the corresponding samples.The experimental results on three cross-modal retrieval datasets show that the performance of SPCMH has highaverage precision.

关 键 词：哈希函数 跨模态检索 语义鸿沟 线性分类 近邻图 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]