Optimal local warp model for image stitching
Tao Yuan, Xiaoming Li · 2016
Geometrical warp model for alignment is a key issue in image stitching. For robustness, most of the commercial tools for stitching use global parametric warps to bring image into alignment, which is suitable for ideal imaging conditions that the images to be stitched differ purely by rotation or the imaged scene is purely planar. When such conditions are violated alignment artifacts or “ghosting” will appear in the final result which leads to unconvincing visual effect. For this reason, J. Zaragoza et al[1]proposed a local homography based registration warp model which reduces the alignment error and improve the stitching quality greatly. But we find that this model is in general not optimal mainly due to the involved parameters is same for all positions in the stitched images. In this paper, we propose an optimal local warp model in which the parameters for computing the local warp model are location dependent. Experiment results verified that our model obtained a more accurate alignment; hence the final stitching quality is improved.