Deep learning for inverse problems in imaging

Deep learning has significantly advanced several areas of imaging, from image analysis tasks such as classification and segmentation to image restoration tasks. Besides being able to produce photorealistic images, generative architectures such as GANs and diffusion models have turned out to be very useful for solving inverse problems, providing powerful priors that can be used even with lack of ground truth data, or lack of training data at all. This talk will focus on recent advances in deep models for restoration, touching on aspects such as perception, complexity and priors. I will cover models for supervised, self-supervised, unsupervised and one-shot restoration, in the single- and multi-image case, as well as the generation of multiple ""good"" solutions that are consistent with the observed data. I will show examples of models learning non-local relationships in the data, along with applications to a variety of data types including optical/radar satellite images and point clouds. I will discuss the implications of these methods on accuracy, perceptual quality and complexity, and I will discuss multimodality as a way to further improve image restoration accuracy.

Перевод пока недоступен