基于注意力机制和软匹配的多标签遥感图像检索方法
A MULTI-LABEL REMOTE SENSING IMAGE RETRIEVAL METHOD BASED ON ATTENTION MECHANISM AND SOFT MATCHING
张永梅 1徐敏 1李小冬1
作者信息
- 1. 北方工业大学信息学院 北京 100144
- 折叠
摘要
针对卷积神经网络对于多标签遥感图像特征提取能力弱、不能准确反映遥感图像多标签复杂性的问题,提出基于注意力机制和软匹配的多标签遥感图像检索方法.在特征提取阶段,以密集卷积神经网络模型为基础,在每个密集块(Dense Block)后添加CBAM(Convolutional Block Attention Module)层,实现对多标签图像区域特征提取.在模型训练时,利用区分硬匹配与软匹配的联合损失函数,学习图像的哈希编码表示.通过评估遥感图像哈希编码间的汉明距离,实现相似图像的检索.实验结果表明,所提方法在数据集NUS-WIDE和多标签遥感图像数据集DLRSD上与其他基于全局特征的深度哈希方法相比,明显提升了检索准确率.
Abstract
To address the problems that convolutional neural networks are weak in extracting features of multi-label remote sensing images and the reflection of complexity multiple labels in remote sensing images,a multi-label remote sensing image retrieval method based on attention mechanism and soft matching is proposed.In the feature extraction stage,the method was based on the densely connected convolutional neural networks,and a CBAM(Convolutional Block Attention Module)layer was added after each dense block to achieve feature extraction of multi-label image regions.During model training,the joint loss function that distinguished hard matching and soft matching was used to learn the Hash code representation of the images.The retrieval results were obtained by evaluating the Hamming distance between the image Hash code and the retrieved image Hash code.The experimental results show the proposed method has a significant improvement in retrieval accuracy compared with other deep hashing methods based on global features on the universal dataset NUS-WIDE and multi-label remote sensing image dataset DLRSD.
关键词
遥感图像检索/密集卷积神经网络/深度哈希/多标签/软匹配Key words
Remote sensing image retrieval/Densely connected convolutional neural networks/Deep hashing/Multi-label/Soft matching引用本文复制引用
基金项目
国家自然科学基金项目(61371143)
教育部科技发展中心"天诚汇智"创新促教基金项目(2018A03029)
出版年
2024
国家科技期刊平台
NSTL
万方数据