基于大批量训练和正交正则化的跨模态哈希方法  

Cross-modal hashing method based on large batch training and orthogonal regularization

作　　者：张学旺[1,2] 周印 Zhang Xuewang;Zhou Yin(School of Software Engineering,Chongqing University of Posts&Telecommunications,Chongqing 400065,China;School of Microelectronics&Communication Engineering,Chongqing University,Chongqing 400044,China)

机构地区：[1]重庆邮电大学软件工程学院,重庆400065 [2]重庆大学微电子与通信工程学院,重庆400044

出　　处：《计算机应用研究》2021年第4期1092-1096,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(61571072);重庆市基础研究与前沿探索专项重点项目(cstc2019jcyj-zdxmX0008);重庆市重点产业共性关键技术创新专项重大主题专项(cstc2017zdcy-zdzxX0013)。

摘　　要：基于深度学习的跨模态哈希方法都使用小批量训练方式来训练模型,然而小批量方式在每次更新参数时获取样本数量有限,不能得到很好的梯度,影响最终训练的模型检索性能。针对此问题,提出了一个新的跨模态哈希方法。该方法使用大批量方式进行训练,并引入正交正则化来增加大批量训练的稳定性;同时考虑了哈希码的离散性,将哈希码与特征之间的距离加入到目标函数中,使得哈希码能够更加真实地表示数据。在两个广泛使用的跨模态检索数据集上的实验表明,该方法比现有的几种哈希方法具有更好的性能。The cross-modal hashing methods based on deep learning use the small batch training method to train their model.However,it cannot get a good gradient using this training method due to the limited number of samples in each parameter update,which affects the retrieval performance of the final trained model.To solve the problem,this paper proposed a new cross-modal hashing,which used large batch training and introduced orthogonal regularization to increase the stability of this kind of training.Considering the discreteness of hash codes,it added the distance between hash codes and features to the objective function,which made hash codes to represent data more realistically.Extensive experiments on two widely used public datasets in cross-modal hashing show that this method achieves better performance than several existing hashing methods.

关 键 词：跨模态哈希 大批量训练 正交正则化 哈希码和特征之间的距离 

分 类 号：TP391.3[自动化与计算机技术—计算机应用技术]