User Constrained Multiscale MRF Model for Texture Mixture Synthesis and its Application to Texture Replacement

The original multiscale MRF texture model proposed by Paget (IEEE Transactions on Image Processing, 1998, page 925-931) can be used to synthesize a broad range of textures but is limited to taking a single input texture and outputting a homogeneous texture similar to the input. This is insufficient for textures that have combined visual characteristics from several different sources. To address this problem, this paper presents a new method, called User Constrained multiscale MRF model, for synthesizing a new texture mixture from multiple input textures. Since the Gibbs sampler and exhaustive search are used in the original multiscale MRF model, a brute force implementation of the algorithm is slow. To overcome this problem, an existing fast neighborhood search technique is adapted for our model, and the run time is decreased by a factor of 5001000. We also demonstrate that our method can be used in texture replacement. The experimental results show that our algorithm performs well in the quality of results.