基于成对标签的深度哈希图像检索方法  

Deep hashing image retrieval based on pairwise label

作　　者：李雪 于炯[1,2] 李梓杨[2] 陈嘉颖 蒲勇霖 LI Xue;YU Jiong;LI Zi-yang;CHEN Jia-ying;PU Yong-lin(School of Software,Xinjiang University,Urumqi 830008,China;School of Information Science and Engineering,Xinjiang University,Urumqi 830046,China;Key Laboratory of Software Engineering Technology,Xinjiang University,Urumqi 830008,China)

机构地区：[1]新疆大学软件学院,新疆乌鲁木齐830008 [2]新疆大学信息科学与工程学院,新疆乌鲁木齐830046 [3]新疆大学软件工程技术重点实验室,新疆乌鲁木齐830008

出　　处：《计算机工程与设计》2021年第7期1981-1988,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(61862060、61562078、61462079、61562086);国家科技部科技支撑基金项目(2015BAH02F01)。

摘　　要：针对采用松弛-量化策略的深度哈希方法面临的二值码离散优化的难题,提出一种端到端的基于成对标签的哈希方法来学习更具有判别力的哈希码,通过优化损失函数来解决离散优化丢失信息的问题。引入锚点哈希码概念,以汉明空间中的锚点作为监督信息训练AlexNet网络,将表示图片的二值码拟合至各锚点附近,使用优化后的损失函数计算分类误差和锚点误差,使哈希函数生成具有强判别力的哈希码。在CIFOR-10数据集和ImageNet-100数据集上实验,检索精度优于当前主流方法。Aiming at the problem of discrete optimization of binary codes faced by the existing deep hashing method using relaxation-quantization strategy,an end-to-end hashing method based on paired labels was proposed to learn more discriminative hash codes.The problem of discrete information loss was solved by optimizing the loss function.The concept of anchor hash code was introduced,and the anchor point in Hamming space was used as the supervising information to train the AlexNet network to fit the binary code representing the picture to the vicinity of each anchor point,and the optimized error loss was used to calculate classification error and anchor point error,and the hash function was optimized to generate a strong discriminative hash code.Results of experiments on the CIFOR-10 dataset and ImageNet-100 dataset show that the retrieval accuracy of the proposed method is better than that of the current mainstream methods.

关 键 词：成对标签 深度哈希 图像检索 哈希学习 卷积神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]