Abstract: Deep neural networks have shown promising performance in image restoration such as denoising, compressive sensing and so on. One of the most important components for them is how to train them - it does not only influence on the performance itself, it also enables to overcome constraints such as ground truth images. In this talk, I will introduce a number of training ways to enable self-supervised learning without ground truth via Stein's unbiased risk estimator for diverse image restoration tasks, to correct for scatter contamination in nuclear medicine via weakly supervised learning, and to progressively restore blurred images with a lightweight architecture in a supervised manner. If time permits, I will also talk about our recent works for multiple image restoration tasks in a single (pre-trained) model.