Non Local Means Image denoising

This is a python implementation of A non-local algorithm for image denoising paper. Here is the colab link for this code.

Config Files Parameters and assumptions

• Patch size = 7. This is mainly because it was somewhat neither too large nor too small. Too large a size would mean very few patches are similar to the current patch and too small would make a huge number of patches similar. (paper has also taken 7)
• Search Size = 21x21 window around the current pixel. This is mainly from a time complexity point of view.
• h = 0.1, std of gaussian kernel(a)=1.5. This is because that was when the image was denoised the best. The function from scikit image library also suggests the same
• Sigma = 0.1. The standard deviation of the noise added.

Results

Qualitative

Quantitative

Peak Signal to Noise Ratio results for 10 different images and comparision between gaussian denoising method and NL means image denoising method is as shown below