Deep learning for inverse problems in imaging

Abstract

Inverse problems have been widely studied in image processing, with applications in areas such as image denoising, blind/non-blind deblurring, super-resolution and compressive sensing. Lately deep learning techniques and architectures have made significant impact in the solution of various inverse problems, surpassing the performance of classical variational optimization algorithms.In this talk, we will review state-of-the-art deep architectures for inverse problems in imaging. We will compare the data-driven solutions of deep learning with standard iterative methods in terms of performance, speed and practicality. We will discuss adversarial learning, generative adversarial networks (GANs) and denoising auto -encoders (DAEs) that are used to learn the distribution of the data in the context of inverse problems. We will then provide a unified framework for the application of deep learning to the solution of various inverse problems, including motion deblurring, single image super-resolution, compressive sensing and sparse recovery. The tutorial will finish by summarizing the recent trends in literature to develop general, model-independent solution to inverse problems using novel deep architectures and learning strategies.