基于改进Superglue的图像匹配算法研究
DOI:
CSTR:
作者:
作者单位:

华北理工大学电气工程学院

作者简介:

通讯作者:

中图分类号:

TP751.1;TN911.73

基金项目:

唐山市沉浸式虚拟环境三维仿真基础创新团队项目(No.18130221A)


Research on image matching algorithm based on improved superglue
Author:
Affiliation:

Fund Project:

Tangshan City Immersive Virtual Environment 3D Simulation Foundation Innovation Team Project(No.18130221A)

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    传统的Superglue算法在图像匹配准确度上表现优异,但是在面对经过旋转变换后且纹理较复杂的图像时,仍具备优化空间。针对上述问题,提出一种基于置信度分数更新机制的Superglue特征匹配算法。首先通过引入Top-k注意力机制,帮助模型从大量的无关噪声中筛选出最重要的图像特征;其次,提出一个可变形注意力图像增强模块,对待匹配图像进行进一步的特征增强,将可变形卷积后的特征与更新后的特征拼接;最后,提出一种置信度分数更新机制,在分数高于阈值时考虑输出,在分数低于阈值时终止匹配,通过不断迭代优化匹配分数,最终提高匹配精度。经过实验验证,改进后的算法在5°、10°、20°的阈值条件下,在经过旋转变换后且纹理较复杂的图像中,位姿估计精度达到了34.57%、52.00%和72.59%,并取得了82.25%的平均精度。本算法的实现,为图像匹配问题提供了一种新的改进方法。

    Abstract:

    The traditional superglue algorithm performs well in image matching accuracy, but it still has optimization space in the face of images with complex texture after rotation transformation.To address these issues, a Superglue feature matching algorithm based on a confidence score update mechanism is proposed.Firstly, the Top-k attention mechanism is introduced to help the model select the most important image features from a large number of unrelated noises; Secondly, a deformable attention image enhancement module is proposed, which further enhances the features of the matched image and stitches the deformable convolution features with the updated features; Finally, a confidence score updating mechanism is proposed, which considers the output when the score is higher than the threshold, terminates the matching when the score is lower than the threshold, and finally improves the matching accuracy by continuously iterating and optimizing the matching score. The experimental results show that the improved algorithm achieves 34.57%, 52.00% and 72.59% pose estimation accuracy in images with complex texture after rotation transformation under the threshold conditions of 5 °, 10 ° and 20 °, and achieves an average accuracy of 82.25%. The implementation of this algorithm provides a new improved method for image matching.

    参考文献
    相似文献
    引证文献
引用本文
分享
相关视频

文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2026-02-02
  • 最后修改日期:2026-02-02
  • 录用日期:2026-03-06
  • 在线发布日期:
  • 出版日期:
文章二维码