This repository is changing SIHR codes MATLAB/GNU Octave to python.

At this moment, I implement Akashi2016 algorithm using numpy and jax.

This implementation is much faster than original.

The other algorithms are coming soon.

The table below is the time of each model to process 1820x1024x3 size single image.
And I marked sec per iteration for fair comparison in table.
Because Akashi algorithm has random variables in code, so each execution has different iterations.

This repository is largely based on the SIHR repository by vitorsr.

Y. Akashi and T. Okatani, “Separation of reflection components by sparse non-negative matrix factorization,” Computer Vision and Image Understanding, vol. 146, pp. 77–85, May 2016, http://dx.doi.org/10.1016/j.cviu.2015.09.001

Highlight removal, diffuse, specular