On pixel-based texture synthesis by non-parametric sampling

In this paper, we propose a pixel-based method for texture synthesis with non-parametric sampling. On top of the general framework of pixel-based approaches, our method has three distinguishing features: window size estimation, seed point planting, and iterative refinement. The size of a window is estimated to capture the structural components of the dominant scale embedded in the texture sample. To guide the pixel sampling process at the initial iteration, a grid of seed points are sampled from the example texture. Finally, an iterative refinement scheme is adopted to diffuse the non-stationarity artifact over the entire texture. Our objective is to enhance texture quality as much as possible with a minor sacrifice in efficiency in order to support our conjecture that the pixel-based approach would yield high quality images.