Early approaches to learning-based denoising where supervised, i.e. they required pairs of clean and noisy data for training. While theoretically applicable to many types of data and noise, in practice the requirement for paired greatly limited their applicability.
Later, the community developed self-supervised methods, requiring only unpaired noisy data for training. Such methods are based on making assumptions about the nature of the noise, which are often (but not always) correct.
Finally, unsupervised methods, which train deep generative models for the data and noise. Like self-supervised methods, they can be trained using only noisy data. instead of outputting a single denoised image, these methods allow us to account for the uncertainty in the denoising process by producing samples of possible solutions.
The talk will give an overview over these three types of methods as well as their strengths and limitations.