摘要
针对目前深度监督哈希检索方法提高检索精度往往需要大量注释数据来提高模型获取的信息量,且大多数深度哈希方法学习的数据关系仅从局部角度捕获,导致学习到的特征判别性不强的问题,提出一种基于自监督蒸馏学习辅助的图像检索方法,能够在监督哈希网络的基础上利用自监督蒸馏网络学习数据间的全局关联信息,在数据量不变的同时增大模型学习的互信息量,实现更精准的哈希编码。在两个基准数据集NUS-WIDE和ImageNet-100上实验,检索精度优于HashNet、PSLDH等主流方法。
To improve the retrieval accuracy of the current deep supervised hash retrieval methods,a large number of annotated data are often needed to improve the information obtained by the model,and most of the data relations learned by the deep hash method are only captured from a local perspective,which leads to the problem of poor discrimination of the learned features.An image retrieval method assisted by self-supervised distillation learning was proposed.Based on supervised hash network,the global association information of self-supervised distillation network was used to learn data,the mutual information of model learning was increased while the amount of data remained unchanged,and more accurate hash coding was achieved.Experimental results on two benchmark datasets,NUS-WIDE and ImageNet-100,show that the retrieval accuracy is better than that of the current mainstream methods such as HashNet and PSLDH.
作者
李为杰
杨志景
LI Wei-jie;YANG Zhi-jing(School of Information Engineering,Guangdong University of Technology,Guangzhou 510006,China)
出处
《计算机工程与设计》
北大核心
2023年第11期3420-3426,共7页
Computer Engineering and Design
基金
国家自然科学基金项目(61972163)
广东省自然科学基金项目(2021A1515011341)
广州市科技计划基金项目(202002030386)。
关键词
图像检索
自监督学习
卷积神经网络
图像处理
哈希算法
知识蒸馏
监督学习
image retrieval
self-supervised learning
convolutional neural network
image processing
hash algorithm
know-ledge distillation
supervised learning
