Graph cut-based fast texture synthesis with CUDA

This paper presents a graph cut-based fast texture synthesis algorithm which utilizes CUDA to greatly accelerate the synthesis speed. The whole sample serves as a patch and is laid into the synthesized image repeatedly to build the final texture, and graph cut is used to determine the optimal patch boundaries. To speed up the synthesis speed, the patch locations are only selected based on a small set, which consists of some optimal relative offsets between two patches and is constructed in the preprocessing step. The matching errors during patch selection are also estimated without much computation from the information provided by this step. The graph cut algorithm is implemented every time a patch is laid, resulting in its becoming the bottleneck of the synthesis process. This problem is solved by virtue of the CUDA implementation of the maxflow/mincut algorithm. Graph construction and the preprocessing step are also accelerated with CUDA. The speed of the synthesis is several to dozens of times faster than the traditional algorithm without GPU acceleration.