Enhanced Texture Synthesis Based On Structure Matching And It's Applications

Texture synthesis is a very hot research area in both computer graphics and computer vision. Recently a new sample-based algorithm is proposed. First it divides texture into many small patches, which could be described with a 2-D Markov model, and the patch to be synthesized is determined by the information of its neighbors which are already synthesized. The best-match patch is searched as the current patch by the overlapped region of adjacent patches in the sample.The traditional matching algorithms which are only dependent on color information cannot capture the structural difference of two images, thus often give bad results. Through a lot of experiments, we find that the sample-based texture synthesis algorithm turns to an under-constrained problem at most of the time. For a given patch, we can always find several patches have similar matching score but quite different content information, so the selected patch from sample is usually unreasonable for it could not keep good coherence with existing patches. In this paper, we present a novel algorithm to solve this problem. It takes into account of the color-difference and structure-difference information at the same time. First, the structure feature is extracted. For the traditional point-by-point matching algorithm takes exponential complexity, it's difficult to be used directly. This problem is overcome by a new two-level matching algorithm also presented in this paper. Thus the matching complexity is greatly reduced. First, the boundary points are connected into curves. After the curves are matched, the point-to-point matching is employed between matched curves. The structure-difference is regarded as the sum of the difference between the matched points. Experiments show that the presented method produces more reasonable and credible synthesis results for the improved matching algorithm.Furthermore, we find that the matching algorithm is also applicable in image inpainting and super-resolution sampling and produces good results.