Fast Example-based Surface Texture Synthesis

Example-based texture synthesis has received a lot of attention in computer graphics recently, with applications ranging from image editing, photo-realistic rendering, visualization and geometric modeling. Current texture synthesis technique has three main directions: (1) manipulating synthesis process by user interaction and performing visualization and data editing via texture synthesis; (2) combining texture synthesis with real-time rendering and generating photo-realistic scene details in real-time; and (3) dealing with more general data such as geometry, motion etc. and taking texture synthesis as a general modeling tool. In each of these scenarios, synthesis speed is important. However, most existing texture synthesis techniques are too slow for interactive or real-time applications. Consequently it is necessary to study acceleration techniques for texture synthesis.This thesis mainly focuses on fast texture synthesis techniques for 3D mesh surfaces. Several acceleration algorithms are proposed for different types of textures. The main contributions are as follows:1. A fast texture synthesis technique for general image textures is proposed. Computational speed is improved by exploiting parallelism of the synthesis algorithm via an efficient GPU implementation. The technique extends texture optimization algorithm by using a discrete solver based on k-coherence. The solver allows a fast, quality neighborhood search while avoiding the blurry blending problem in previous work. The algorithm is further extended to 3D mesh surfaces by defining a new texture energy function, allowing interactive flow texture animation.2. A fast synthesis technique for globally varying textures is presented. Synthesis speed is improved by reducing the data size of sample textures. Traditional acceleration techniques for texture synthesis are not applicable to globally varying textures due to their large size. An inverse texture synthesis technique is proposed to automatically produce a small compaction from a large globally varying texture via energy optimization. This small compaction encodes all the regions in the large original texture and can be used to reconstruct the original texture or to re-synthesize new textures under user-supplied control maps. More important, it allows real-time synthesis of globally varying textures on GPU. Besides, the technique can automatically compute orientation fields for anisotropic textures that contains both low- and high-frequency regions.3. A fast synthesis technique for geometric textures is proposed. A set of texture tiles is precomputed to accelerate the synthesis process. Acceleration techniques for image texture synthesis are not applicable to geometric textures due to their difference in representation. A technique termed geometric texture Wang Tiles is proposed to address this problem. A set of Wang Tiles is pre-computed from a specific geometric texture and used to generate new geometric textures on different 3D mesh surfaces. The proposed technique consumes much less storage and computation cost than existing methods.