search

yu4u / noise2noise

An unofficial and partial Keras implementation of "Noise2Noise: Learning Image Restoration without Clean Data"
MIT License
1.08k stars 234 forks source link

rational for picking loss function #19

Open matteo411 opened 6 years ago

matteo411 commented 6 years ago

Is there some reasoning you can share about why you used a different loss function for each type of noise generation? gaussian : mse text : mae impulse random : L0

yu4u commented 6 years ago

Please refer to the original paper. Theoretically these noise and loss pairs are optimal. (mean, median, or mode of noisy images is a clean image if a noise is Gaussian, text, or impulse noise respectively)