基于改进SuperPoint与线性转换器的可见光红外匹配算法
A matching method based on improved SuperPoint and linear Transformer for optical and infrared images
伍薇 1鲜勇 1苏娟 2张大巧 1李少朋 3李冰1
作者信息
- 1. 火箭军工程大学作战保障学院,西安 710025
- 2. 火箭军工程大学核工程学院,西安 710025
- 3. 火箭军工程大学作战保障学院,西安 710025;清华大学自动化系,北京 100084
- 折叠
摘要
针对可见光和红外图像的异源图像匹配难度大、误匹配率高的问题,提出一种基于改进SuperPoint与线性转换器的深度学习匹配算法.首先在SuperPoint网络结构的基础上,引入特征金字塔的思想构建特征描述分支,基于铰链损失函数进行训练,从而较好地学习可见光与红外图像多尺度深层次特征,增大图像同名点对描述子的相似度;在特征匹配模块,利用线性转换器对SuperGlue匹配算法进行改进,聚合特征以提高匹配性能.在多个数据集上对所提算法进行实验验证,结果表明,与现有的算法相比,所提算法获得了更好的匹配效果,提高了匹配准确率.
Abstract
A deep learning matching algorithm based on improved SuperPoint and linear transformer was proposed to solve the problem of difficult matching and high mismatching rates between visible and infrared heterologous images.Firstly,based on the SuperPoint network structure,the algorithm introduced the idea of a feature pyramid to build a feature description branch and trained it based on the hinge loss function,so as to better learn the multi-scale deep features of visible and infrared images and increase the similarity of the image with correspondence points to the descriptor.In the feature matching module,SuperGlue was improved by adopting a linear transformer for aggregating features to obtain better matching results.Experiments conducted on multiple datasets demonstrate that the proposed method improves the matching precision and provides better matching performance in comparison with existing matching methods.
关键词
红外图像/可见光图像/异源图像匹配/特征金字塔/线性转换器Key words
infrared image/visible light image/heterologous image matching/feature pyramid network/linear transformer引用本文复制引用
出版年
2025
国家科技期刊平台
NSTL
维普
万方数据