基于自适应三元组网络的细粒度图像检索算法  

A Fine-grained Image Retrieval Algorithm Based on Adaptive Triplet Network

作　　者：潘丽丽[1] 陈蓉玉 雷前慧 邵伟志 黄诗祺 PAN Lili;CHEN Rongyu;LEI Qianhui;SHAO Weizhi;HUANG Shiqi(School of Computer and Information Engineering, Central South University of Forestry and Technology,Changsha 410000,China)

机构地区：[1]中南林业科技大学计算机与信息工程学院,湖南长沙410000

出　　处：《郑州大学学报（理学版）》2021年第4期36-43,共8页Journal of Zhengzhou University:Natural Science Edition

基　　金：国家自然科学基金项目(61772561);湖南省自然科学基金项目(2021JJ31164);湖南省重点研发计划项目(2018NK2012)。

摘　　要：立足于深度学习,提出面向细粒度图像的自适应三元组网络的鲁棒图像检索算法。首先,提出的视觉显著性检测方法被用来去除图像噪音,以便提取图像中目标主体辨识度更高的深度特征;然后,添加特征增强模块来提高深度特征的表征能力和鲁棒性;最后设计三元组网络,弥补传统分类模型特征判别能力不足的缺陷,获取更适用于细粒度图像检索的网络模型。经实验验证,采用视觉显著性检测、特征增强模块和自适应三元组损失函数方法构建的网络模型提取的深度特征不仅加快检索效率,同时也提高了检索精度。Based on deep learning,a robust image retrieval study based on adaptive triplet network for fine-grained image data was proposed.Firstly,a visual significance detection method was proposed to remove the image noise and extract the deep features of the target subject with higher identification before image training.Then,feature enhancement modules were added to improve the representability and robustness of deep features.Finally,a triad network was designed to make up for the lack of distinguishing ability of traditional classification models and to obtain a network model more suitable for fine-grained image retrieval.Experimental results showed that the deep features not only speeded up the retrieval efficiency,but also improved the retrieval accuracy extracted by the network model constructed by visual significance detection,feature enhancement module and adaptive triplet loss function.

关 键 词：细粒度图像检索 视觉显著性检测 卷积神经网络 特征增强模块 自适应三元组损失 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]