Une exploration du déflouage d’images et vidéos : les détails qui font la différence

This thesis studies the problem of image and video blur and its removal.In the first part we focus on the restoration of image bursts, in particular for deblurring and super-resolution.First we study Fourier Burst Accumulation, which efficiently fuses the frames temporally by a weighted average in the Fourier domain.Then we show that the recent advances in satellite design allow to increase the spatial resolution using multi-frame super-resolution algorithms.We propose a method based on a spline interpolation model and quantify the gain of resolution. We apply the method with success on raw Skysat image bursts lent by Planet.In the second part we focus on the non-blind deconvolution problem.While most methods assume an over-simplistic image formation model, we propose to explicitly handle saturation, quantization, and gamma correction, with considerable improvement of the results.In the third part we tackle the difficult problem of blind deblurring, where the blur kernel is not known.First we propose an anatomy of the Goldstein and Fattal method which models statistical irregularities in the power spectrum of blurred natural images in order to estimate a blur kernel.Then we analyze a blur kernel estimation method that uses an L0 prior on the image gradients.While the method performs well on ideal settings, we show that its performance degrades rapidly under high noise conditions.To cope with this issue, we propose improvements of the method in order to handle high noise levels while maintaining its efficiency.The proposed approach yields results that are equivalent to those obtained with computationally far more demanding methods.Finally we propose to quantify the sharpness of images from the PlanetScope constellation to discard low quality images or deconvolve blurry ones.