We obtained an image with known ground truth separation into CMY and K, and applied three types of degradations to this input:

1. Adding Gaussian noise
2. Raising the gray level of black lines
3. Applying low frequency deformation

We cranked each type to amounts way beyond what would be expected in practice to test the breaking points of our algorithm. The links below show a quantitative evaluations the individual stages of our algorithm.

1. Overview and graphs
2. Visual results on a small crop