Patch-based multi-resolution real-time texture synthesis in wavelet domain

Texture synthesis has been studied for many years, but it is still difficult to achieve real-time implementation and high quality simultaneously. We present a discrete-wavelet-transform (DWT) based multi-resolution technique for synthesizing high quality textures in real-time. In the wavelet domain, our algorithm first generates randomly an initial texture patch from the input texture picture and subsequently all other texture patches are assigned by measuring the similarity in a defined overlapped region. A final texture is synthesized by the inverse DWT. We demonstrate that the developed algorithm works well for a wide range of texture images (including regular to stochastic), especially for natural texture images that include quasi-repeating patterns consisting of small similar objects of different sizes.