This dissertation contributes to two fundamental problems of digital image formation: the modeling and estimation of the blur introduced by an optical digital camera and the fast generation of realistic synthetic images. The accurate estimation of the camera's intrinsic blur is a longstanding problem in image processing. Recent technological advances have significantly impacted on image quality. Thus improving the accuracy of calibration procedures is imperative to further push this development. The first part of this thesis presents a mathematical theory that models the physical acquisition of digital cameras. Based on this modeling, two fully automatic algorithms to estimate the intrinsic camera blur are introduced. For the first one, the estimation is performed from a photograph of a specially designed calibration pattern. One of the main contributions of this dissertation is the proof that a pattern with white noise characteristics is near optimal for the estimation purpose. The second algorithm circumvents the tedious process of using a calibration pattern. Indeed, we prove that two photographs of a textured planar scene, taken at two different distances with the same camera configuration, are enough to produce an accurate estimation. In the second part of this thesis, we propose an algorithm to accelerate realistic image synthesis. Several hours or even days may be necessary to produce high-quality images. In a typical renderer, image pixels are formed by averaging the contribution of stochastic rays cast from a virtual camera. The simple yet powerful acceleration principle consists of detecting similar pixels by comparing their ray histograms and letting them share their rays. Results show a significant acceleration while preserving image quality.
