基于深度学习SuperGlue的图像拼接与实现  

Imagemosaic and implementation based on deep learning SuperGlue

作　　者：王康涛 WANG Kangtao(Information Engineering College,Wenzhou Business College,Wenzhou 325000,China)

机构地区：[1]温州商学院信息工程学院,浙江温州325000

出　　处：《山东理工大学学报（自然科学版）》2025年第2期10-15,21,共7页Journal of Shandong University of Technology:Natural Science Edition

摘　　要：针对传统算法在特征匹配模块准确度较低、在纹理差的区域很难识别特征点等问题,提出一种基于深度学习的SuperPoint特征点提取算法与SuperGlue匹配算法相结合、用于图像配准的拼接算法。首先将图像用共享的编码器处理以提取图像深层特征,再分别经过特征点和描述子的两个解码器以提取特征点和对应的描述子;然后传入SuperGlue网络中,将提取的特征点和描述子使用SuperGlue算法建立良好的匹配,再通过变换矩阵使待匹配图像变换到统一的坐标下,实现图像重叠区域的对齐;最后通过渐入渐出的加权融合算法对图像进行融合,得到更宽视角和高分辨率的图像。实验证明在特征点提取模块,本文算法比对比SIFT算法、ORB算法和SIFT-FREAK算法的匹配正确率更高,同时得到的拼接效果图质量更优。To address the issues of low accuracy of the feature matching module and difficulties in identifying the feature points in the area with poor texture,this study proposes a novel approach that combines a deep learning-based SuperPoint feature point extraction algorithm and SuperGlue matching algorithm for image mosaic research in image registration.Initially,the image is processed through a shared encoder to extract the deep features of the image.Then,the feature points and corresponding descriptors can be extracted through the two decoders of the feature points and descriptors.These extracted components are then input into the SuperGlue network and well matched by SuperGlue algorithm.Once the fine matching is established,the transformation matrix is applied to transform the images into a unified coordinate to realize the proper alignment of the overlapping areas.Finally,the image is fused by the weighted fusion algorithm of gradual in and gradual out,and a wider perspective and high resolution image can be obtained.Experiments demonstrate that the proposed algorithm in this study achieves higher matching accuracy in the feature point extraction module compared to SIFT,ORB and SIFT-FREAK algorithms.Additionally,the quality of the mosaic effect image obtained by this method is better.

关 键 词：超级点算法 超级胶水算法 图像融合 全景图 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]