Défloutage non-aveugle d'images dans des situations réelles
Thomas Eboli · HAL (Le Centre pour la Communication Scientifique Directe) · 2021
Sharpness is an important criterion for shooting acceptable photographs. Several factors such as the camera settings, motion or defocus may decrease image sharpness and lead to blur, resulting in the loss of details. Typical image process- ing techniques rely either on optimization algorithms using a handcrafted image prior or machine learning approaches leveraging supervisory pairs of sharp and synthetic blurry images. In this thesis, we follow a recent trend that combines the two sorts of techniques previously detailed and achieving the state-of-the-art image deblurring results. We first present a parametric function for RGB image deblurring, embedding preconditioned Richardson fixed-point iterations to replace the classical fast Fourier transform (FFT) algorithm prone to ringing artifacts, especially at the image boundaries. In a second contribution, we propose a model for joint non-blind deblurring and demosaicking of raw images.