Photographic Image Priors in the Era of Machine Learning

Explore the intersection of machine learning and image processing in this 42-minute lecture by Prof. Eero Simoncelli from New York University and The Flatiron Institute. Delve into the evolution of solving inverse problems in image processing and computer vision, from traditional prior probability densities to the revolutionary impact of artificial neural networks. Examine the properties of priors implicitly embedded in denoising networks and learn about methods for drawing samples from them. Discover how these sampling methods can be extended to solve deterministic linear inverse problems without additional training, broadening the applications of supervised learning for denoising. Gain insights into the minimal assumptions required, the robust convergence across various parameter choices, and the state-of-the-art performance achieved in deblurring, super-resolution, and compressive sensing. This talk, presented at the Institut des Hautes Etudes Scientifiques (IHES), offers a comprehensive look at photographic image priors in the context of modern machine learning techniques.