I wanted to see if texture generation could be used to remove the red text from
this image (I later paid the $30 registration fee).
Efros and Leung [1] is one common method for generating homogeneous textures,
but this problem is different because the textures are different in different
parts of the image and we only need to generate a small texture area.
So, instead of matching similar neighborhoods as was done in by Efros and Leung,
I assumed that the pixels were locally similar.
To generate the texture, I first did a sweep through the image and for
each pixel p that was red to within a
hand-tuned threshold using Euclidean distance, I sampled from the neighborhood of p
using a Gaussian and replaced p with the sample. If that sampled value was again red, I resampled.
I then used a uniform smoothing mask of size 9 to smooth each pixel that had been
replaced.
The result was this image
I also tried converting the RGB values to the uniform color space CIE LAB to see
if a better image could be produced. Contrary to expectations, the RGB image
was still the best. When I used L, a, and b, and the Euclidean distance with a hand-tuned
cutoff threshold then the best image produced was
When I only used only a and b, and the Euclidean distance with a hand-tuned cutoff threshold then
the best image produced was
[1] A.A. Efros and T.K. Leung, "Texture synthesis by non-parametric sampling", In Proceedings of the Seventh International Conference on Computer Vision, Corfu, Greece, 1999.