CPU synthesis method results

I have implemented the CPU method (most naive one, single solution, full search mode) and have some synthesis result as following:

And the results seems to be pretty nice. The noise like pixels on the first several rows and columns are caused by the noise initialization.

Also, I got some bad results and analyzed why they look bad.

 

This one looks terrible because I chose a wrong size of L-shape neighborhood when I was synthesizing. The proper size should be chosen according to the features of the texture. Take the bad result as an example, the distance between to centers of the circles is about 26 pixels, and the neighbor size I chose is 15, at times it chose a pure black neighborhood as the best matched one, and it never converges. In the "good results" figures, I chose the neighborhood size as 30, and it gave me a satisfying result.

This one looks disgusting, and what's surprising me is it used the same parameters as the one in the "good results" figures. I figured it's influenced by the initialize noise, some noise make it not converge, maybe I can fix this by using another initialize mode, I will do it in parallel with implementing the GPU methods.