forum.alglib.net

Hello,

when I use the ALGBLIB rbf-ml algorithm to approximate a 128x128 pixel image, I get more non-zero elements than there are pixels.
I thought that the hierarchical model tried to fit the mimimum amount of required basis functions. Is it possible to use ALGLIB to perform this kind of image compression?


PS: If I manually select a random subset of the pixels (say 10%) as initial centers, the reconstruction is far worse than with methods like Residual Subsampling. (http://www.sciencedirect.com/science/article/pii/S0898122107000454)

Thanks,
Jeroen