Structure-Aware Image Compositing

Abstract

The classic task of image compositing is complicated by the fact that the source and target images need to be carefully aligned and adjusted. Otherwise, it is not possible to achieve convincing results. Visual artifacts are caused by image intensity mismatch, image distortion or structure misalignment even if the images have been globally aligned. In this paper we extend classic Poisson blending by a constrained structure deformation and propagation method. This approach can solve the above-mentioned problems and proves useful for a variety of applications, e.g. in de-ghosting of mosaic images, classic image compositing or other applications such as superresolution from image databases. Our method is based on the following basic steps. First, an optimal partitioning boundary is computed between the input images. Then, features along this boundary are robustly aligned and deformation vectors are computed. Starting at these features, salient edges are traced and aligned, serving as additional constraints for the smooth deformation field, which is propagated robustly and smoothly into the interior of the target image. If very different images are to be stitched, we propose to base the deformation constraints on the curvature of the salient edges for C1-continuity of the structures between the images. We present results that show the robustness of our method on a number of image stitching and compositing tasks.