基于对比学习的无监督三元哈希方法  

Unsupervised ternary hash method based on contrastive learning

作　　者：李玉强[1] 陆子微 刘春[1] Li Yuqiang;Lu Ziwei;Liu Chun(School of Computer Science&Artificial Intelligence,Wuhan University of Technology,Wuhan 430070,China)

机构地区：[1]武汉理工大学计算机与人工智能学院,武汉430070

出　　处：《计算机应用研究》2023年第5期1434-1440,共7页Application Research of Computers

摘　　要：为了解决现有无监督二元哈希方法由于存在较大量化损失而导致检索精度较低的问题,在CIBHash方法的基础上,提出了一种新的基于对比学习的无监督三元哈希方法——CUTHash,将三元哈希编码用于图像检索。具体来说,首先,使用融合了解耦对比损失的对比学习框架,在目标数据集上进行无监督的图像特征学习;接着,为了得到三元哈希编码,对学习到的图像特征使用平滑函数进行量化操作,解决离散函数量化后导致的零梯度问题;最后,应用改进后的对比损失,约束同属一张图像的增强视图的特征在哈希空间中尽可能地接近,从而使得三元哈希编码具有一定的辨识力,使其更好地应用于无监督图像检索任务。在CIFAR-10、NUS-WIDE、MSCOCO以及ImageNet100数据集上进行了大量对比实验,取得了较当前主流的无监督哈希方法更好的检索性能,从而验证了CUTHash方法的有效性。To solve the problem of low retrieval accuracy of the existing unsupervised binary hashing method due to quantization loss,this paper proposed a new unsupervised ternary hash method based on contrastive learning refers to the CIBHash method——CUTHash,using ternary hash code for image retrieval.Specifically,the method used the contrastive learning framework of decoupled loss to acquire a compact and accurate feature representation for each sample.Then,to obtain the ternary hash codes,it used the smooth function after the feature representation which could solve the zero gradient problem caused by the quantification of discrete functions.Finally,the representation of the enhanced view of the same image after the application of improved contrastive loss could preserve the semantic information and improve the discriminative ability in the Hamming space.So that it can be better applied to unsupervised image retrieval tasks.It performed a large number of compa-rative experiments on the CIFAR-10,NUS-WIDE,MSCOCO,and ImageNet100 datasets,and achieved better retrieval performances than the current mainstream unsupervised hash method,thus verifying the effectiveness of the CUTHash method.

关 键 词：图像检索 无监督哈希 对比学习 三元哈希编码 哈希量化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]