«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

33

期数:

2023年07期

页码:

34-40

栏目:

媒体计算

出版日期:

2023-07-10

文章信息/Info

Title:

Image Retrieval Technology Combining Self-attention and Convolution

文章编号:

1673-629X(2023)07-0034-07

作者:

曾凡锋¹ ; 王 祺²

1. 北方工业大学,北京 100144;

2. 北方工业大学 信息学院,北京 100144

Author(s):

ZENG Fan-feng1 ; WANG Qi2

1. North China University of Technology,Beijing 100144,China;

2. School of Information Science and Technology,North China University of Technology,Beijing 100144,China

关键词:

图像检索; 自注意力; 卷积神经网络; 特征提取; 深度学习

Keywords:

image retrieval; self-attention; convolutional neural networks; feature extraction; deep learning

分类号:

TP391. 41

DOI:

10. 3969 / j. issn. 1673-629X. 2023. 07. 005

摘要:

针对细粒度图像类别之间差异较小的问题,需要对不同区域进行特征提取,自注意力模型能够有效地获取全局特征,卷积神经网络有利于得到图像的局部细节特征。 为了实现高效的图像特征提取,提出一种融合自注意力和卷积的图像检索方法。　 自注意力和卷积是特征提取的两种强有力的方法,将两者进行有效融合,以提取更鲁棒的特征。 该方法先通过卷积层提取图像局部特征,然后再输入到自注意力模型捕获全局信息,生成质量更高的图像特征用于图像检索。 为了使自注意力模型与卷积能够有效融合,对自注意力模型进行了改进,更好地将局部特征与全局表示进行融合,实现改善图像检索效果的目的。 在 CUB-200-2011 及 CARS196 图像检索数据集上的实验结果表明,所提方法可以有效地提高检索精度。

Abstract:

In view of the small difference between fine-grained image categories,feature extraction in different regions is needed. The self-attention model is effective in obtaining global features. Convolutional neural networks are beneficial in obtaining local detailed featuresof images. In order to achieve efficient image feature extraction,an efficient image retrieval method that combines self-attention and convolution is proposed. Self-attention and convolution are two powerful methods for feature extraction,and the two are effectively fused toextract efficient and robust features. The local features of the image are extracted through the convolutional layer,and then the global information is captured by inputting into the self-attention model to generate the image features with higher quality for image retrieval. Inorder to integrate the self-attention model and convolution effectively,the self-attention model is improved to integrate the local featureswith the global representation better,so as to improve the image retrieval effect. Experiments on the CUB - 200 - 2011 and CARS196image retrieval datasets show that the proposed method can effectively improve the retrieval accuracy.

常用功能

更新日期/Last Update: 2023-07-10