Image Stitching Based on Multi-Scale Meshes

Generating high-quality stitching images with a natural structure is a challenging task in computer vision. Recent image stitching methods based on warps failed to suppress the distortion of the images. They often bend the salient lines in the image, which is inconsistent with human perception. In this paper, we succeed in proposing a novelty model called multi-perspective warps for natural image stitching which is related to the density of feature points in the images. With it we can get more precise matching results. Three new energy terms are developed to stitch quality to specify and balance the expected for aligning the vertices of the multi-scale mesh, which can constrain the transformation of the mesh. We also explore and introduce three feature point reconstruction algorithms to enrich the features in the images. Extensive experiments demonstrate that the proposed method outperforms most state-of-the-arts by effectively preserving the linear structure in the image and improving the robustness.