Inpainting basé motif d'images et de vidéos appliqué aux données stéréoscopiques avec carte de profondeur

摘要

We focus on the study and the enhancement of greedy pattern-based image processing algorithmsfor the specific purpose of inpainting, i.e., the automatic completion of missing data in digitalimages and videos. We first review the state of the art methods in this field and analyze the important steps of prominent greedy algorithms in the literature. Then, we propose a set of changesthat significantly enhance the global geometric coherence of images reconstructed with this kindof algorithms. We also focus on the reduction of the visual bloc artifacts classically appearing inthe reconstruction results. For this purpose, we define a tensor-inspired formalism for fast anisotropic patch blending, guided by the geometry of the local image structures and by the automaticdetection of the artifact locations. We illustrate the improvement of the visual quality brought byour contributions with many examples, and show that we are generic enough to perform similaradaptations to other existing pattern-based inpainting algorithms. Finally, we extend and applyour reconstruction algorithms to stereoscopic image and video data, synthesized with respect tonew virtual camera viewpoints. We incorporate the estimated depth information (available fromthe original stereo pairs) in our inpainting and patch blending formalisms to propose a visuallysatisfactory solution to the non-trivial problem of automatic disocclusion of real resynthesizedstereoscopic scenes.