A Robust Image Stitching Based on Solution Set Smoothing
Yan Qing Hu, Rui Zhou · 2024
The importance of a smooth warp field in image stitching is on the rise. Existing techniques like as-projective-as-possible (APAP) and adaptive as-natural-as-possible (AANAP), which rely on Direct Linear Transformation, leverage smooth projection fields to enhance alignment and naturalness in image stitching. Nonetheless, experiments reveal their lackluster computational efficiency. In this study, we propose a robust as-fast-as-possible image stitching based on solution set smoothing, which focuses on both accuracy and computational efficiency of image stitching. We take the matched feature point pairs as the initial seed for the warping field. We first compute the base set of precise warping parameters for each seed. Then, we filter outliers in the solution space and weight all precise solutions to obtain the locally optimal solution, and finally obtain the overall warping field by inverse distance interpolation. The experimental results show that our algorithm outperforms the state-of-the-art algorithms in terms of computational performance, and the quality and naturalness of the alignment reach the same level.