Can we learn to denoise images with noisy images only?
How they solve it?
The authors motivate their work with a temperature predictor example. If we are given the temperature of last few days then using an L2 loss predictor will lead to a model that predicts the mean of given samples as the temperature on the next day. Similarly, if a model is given noisy images as targets ( where the input is also a noisy image that has to be denoised ), then the model will learn to predict the mean as its impossible to predict random noise. They then go on experimenting with different noises.
For the experiments, they use a recent State-of-art model RED30 [1]. They then explain how this model is trained with noisy images in a slightly odd manner. Though what they broadly mean is that there are three training cases.
Case I The traditional case. Say the dataset has 100 clean images and 19 noisy ones for each. Now we have 100 * (19+1) * 1 training pairs ( input = noisy or clean; output = clean ).
Case II Here the model is trained with target also being possibly noisy. Hence we have 100 * (19+1) * (19) choices. This is because the same image can’t be both input and output.
Case III Noise2Noise case. Here the input and output both are noisy images only while having only two noisy images for each image and 1000 image samples. This means there are 1000* 2 training samples.
They report that Case III performs the best, then case II and then case I. The reason is that in case 3 there are more real images on the same budget of 2000 images. While case 2 has much more training samples than any of them.
They then go on to experiment with different noise types showing that it is possible to denoise images with only noisy samples.
[1] Image restoration using convolutional auto-encoders with symmetric skip connections