Image Synthesis by Numbers of Texture Patches

Abstract

In this paper we investigate two kinds of prevalent applications including the texture synthesis and texture transfer. The texture sythesis technique takes an input sample and generates synthetic texture of unlimited size, which perceived by human being to be very similar to the given source. Initially, we adopt patch-based sampling algorithm to complete the goal, which dervies from the Markov Random Field model and performs faster than the pixel-based counterpart. Secondly, rather than wxhaustive search we explore theaccelerative thchnique of nearest neighbors search to improve the performance in the synthesis process of best match. Thirdly, to find the minimum error path the dynamic programming technique is used for tackling the problem of consistent transition between two overlapped patches. In particular, our automatic system framework is amenable for sythesizing the tiling and constrained texture. The usage of tiling texture is very conventional in texture-mapped geometric models and webpage images, and the constrained texture synthesis is applicable to a wide category of applications such as hole filling and object removing. In addition, we propose a non-iterative method for texture transfer algorithm, which renders the target image by transferring source tecture patches very simular to the texture synthesis algorithm. The method runs more than an order of magnitude faster than multi-pass counterpart, and takes into account two match principles of target fidelity and neighbor coherence. Moreover, er succeed in extending the technique for the label controlled texture transfer, and demonstrate its employment on landscape gardening.