Application and implementation of SuperGlue algorithm based on deep learning in image stitching

In response to the low accuracy of traditional algorithms in feature matching modules and their difficulty in identifying feature points in sparse texture areas, this paper proposes an image registration and stitching algorithm based on deep learning, integrating SuperPoint feature point extraction algorithms and SuperGlue matching algorithms. The algorithm first extracts deep features from images through shared encoders, then uses feature point and descriptor decoders to extract feature points and their corresponding descriptors; subsequently, the extracted feature points and descriptors are fed into the SuperGlue network, where the SuperGlue algorithm establishes precise matches, and transformation matrices are used to convert the images to be matched into a unified coordinate system, achieving alignment of overlapping regions; finally, a weighted fusion algorithm with gradual in-out transitions is employed to fuse the images, generating wide-angle and high-resolution stitched images. Experimental results show that in the feature point extraction and matching module, the proposed algorithm significantly improves matching accuracy compared to SIFT, ORB, and SIFT-FREAK algorithms, while also producing stitched images with superior quality.