Research on Natural Image Stitching based on Extended Point-line Feature
Qi Liu, Xiyu Tang, Ju Huo · 2023
Generating high-quality stitched images is a very challenging task in computer vision. In this paper, we propose a single-perspective natural image stitching method based on extended point-line features, which preserves the matching line features of image overlapping areas to the greatest extent and preserves image edge information as much as possible through the method of regional line segment connection. The matched line features are then extended to generate complementary matched point features, which optimizes image pre-alignment and constraints image geometry features. Finally, the image stitching is realized by parametric warping and grid deformation, and the point-line features of the stitching image are quantitatively evaluated by the root mean squared error (RMSE) and root mean squared error distance (RMSEdis) respectively. Experiments show that our method has obvious advantages for complex image stitching with obvious geometric shapes. Compared with the existing methods, RMSE has a maximum increase of 28.8%, and RMSEdis has a maximum increase of 63.2%.