Enhancement and acceleration of the Retinex algorithm

Post date: Mar 3, 2010 10:28:44 PM

This was my personal project for the Biometrics class in 2009 as part of the M.Sc. in Artificial Intelligence, Pattern Recognition and Digital Image. One year before, I was doing some experiments on multigrid anisotropic diffusion with GPUBase as a testing problem for our new GeForce 9800 GX2 GPU. Back then I didn't know the Retinex algorithm, which was quite close to what I was already doing. I though it was a good testing case, but I didn't find any real application of its results.

As I learned about the Retinex algorithm for face preprocessing one year later, I started testing some modifications and adaptations of my old source code to the new problem. The results were quite astonishing: faces that hardly could be seen were quite clear and only in a fraction of a second, nearly 300 times faster than its CPU version.

I have not published yet the source code, as I want to perform some additional research and, maybe, publish a paper about this. However, I want to show some of the results. These images come from the Yale Face Database B:

The algorithm implementation works in RGB mode though all these examples come from greyscale images. Further research has to be done about combining the color channels to enhance the results in color images.

Even if source and algorithm details have not been published yet, please don't hesitate to ask any questions if you're interested on this.