Image structures: From augmented reality to image stylization. (Structure d'une image : de la réalité augmentée à la stylisation d'images)

In this thesis we consider in general image structures and more specifically, image gradients and contours. They have been proven useful in recent years for various computer graphics applications, such as Augmented Reality (AR), image and video stylization. The goal of analyzing image structures is to describe a high level understanding of image contents and to provide a powerful support to improve the quality of applications, such as visual legibility, accuracy, spatial and temporal coherence. We first demonstrate the important role of image structures in Focus+Context compositing. For Focus+Context rendering in AR, a technique dedicated to the visualization of hidden scenes in video streams, the use of screen segmentation and feature lines significantly emphasizes the depth cue of occluded scenes, and reveals the correct occluding order. We further extend Focus+Context rendering to importance-driven image synthesis, where image gradient and saliency map are used to composite multiple rendering styles in a coherent manner. In the second part, we thus introduce a new approach to estimate a continuous gradient field without oversmoothing original details contained in an image. For this purpose, we develop a new and higher-order local approximation method for discrete non-oriented gradient fields based on a moving least square (MLS) formalism. We show that our isotropic linear approximation outperforms classical structure tensor: image details are better preserved and instabilities are significantly reduced. We demonstrate how our non-oriented MLS gradient field benefits to various image stylization approaches. Finally, we demonstrate that the use of a feature profile analysis for image line extraction via fitting techniques permits to distinguish sharp and smooth features. Profile parameters are then mapped to stylistic parameters such as brush orientation, size or opacity to give rise to a wide range of line-based styles. Key-Words : Image structure, Augmented reality, image and video stylization, gradient field, contours and lines.