Image stitching with positional relationship constraints of feature points and lines

Abstract

The elimination of parallax and the processing of natural issue in complex scenes are challenging tasks for image stitching. In this paper, an image stitching method with positional relationship constraints of feature points and lines, which can accomplish accurate alignment and reduce projection distortion, is proposed. At first, to reduce the computational cost and the number of outliers on subsequent feature matching, we combine the template matching to propose a quick way for detecting overlapping regions. Then, the appropriate reference image is determined to mitigate the projection distortion that the image warping is in the cases of non-planar geometry of the scenes. Furthermore, a local mesh model based on dual features is established to guide the mesh deformation. And an energy function is designed to refine alignment. In addition to alignment error terms, a novel positional relationship constraint term is proposed to improve quality of naturalness of final stitching results. Finally, experimental results demonstrate that our approach is superior to the existing image stitching algorithms in improving quality of the stitched image naturalness.