基于响应值中心加权卷积特征的图像检索算法  

Image retrieval algorithm based on response value center weighted convolution feature

作　　者：李晓宁 赵宏伟[1,3] 张丹阳[1] 张媛 Xiao-ning LI;Hong-wei ZHAO;Dan-yang ZHANG;Yuan ZHANG(College of Computer Science and Technology,Jilin University,Changchun 130012,China;College of Computer Science and Technology,Changchun Normal University,Changchun 130032,China;Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education,Jilin University,Changchun 130012,China)

机构地区：[1]吉林大学计算机科学与技术学院,长春130012 [2]长春师范大学计算机科学与技术学院,长春130032 [3]吉林大学符号计算与知识工程教育部重点实验室,长春130012

出　　处：《吉林大学学报（工学版）》2022年第11期2669-2675,共7页Journal of Jilin University:Engineering and Technology Edition

基　　金：吉林省省级科技创新专项项目(20190302026GX);吉林省自然科学基金项目(20200201037JC);吉林省教育厅“十三五”社会科学项目(JJKH20181219SK)。

摘　　要：提出一种新颖的基于交叉熵的损失,将该损失应用于经典的卷积神经网络上获得了更优的嵌入空间。设计了一种基于响应值中心加权的卷积特征聚合算法处理神经网络得到的三维卷积特征,该算法通过当前位置响应值以及高斯中心计算获得特征图的空间加权系数,减少了三维卷积特征图降成一维图像特征描述子时的信息丢失,并实现了目标区域的增强。最后,将得到的图像特征描述子用于检索任务。在CUB-200-2011数据集上通过消融实验分别验证了损失函数和响应值中心加权算法的有效性。本文算法在Paris6k、Oxford5k、CUB-200-2011、CARS196四个数据集上较当前已有的检索方法获得了更高的准确率和召回率。A novel cross-entropy-based loss is proposed,which is applied to a classical convolutional neural network to obtain a better embedding space. A convolutional feature aggregation algorithm based on the weighting of the center of the response value is designed to process the three-dimensional convolutional features obtained by the neural network. The algorithm obtains the spatial weighting coefficient of the feature map by calculating the current position response value and the Gaussian center. The algorithm reduces the loss of information when the 3D convolution feature map is reduced to a 1D image feature descriptor and realizes the enhancement of the target area. Finally,the obtained image feature descriptors are used for retrieval tasks. On the CUB-200-2011 dataset,the effectiveness of the loss function and the response value center weighting algorithm are respectively verified by ablation experiments. Compared with the current retrieval methods,higher precision and recall rates have been obtained on the four datasets of Paris6k,Oxford5k,CUB-200-2011 and CARS196.

关 键 词：计算机应用 图像检索 深度学习 卷积特征聚合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]