On the generation of high dynamic range images: theory and practice from a statistical perspective

Abstract

This dissertation studies the problem of high dynamic range (HDR) image generation from a statistical perspective. A thorough analysis of the camera acquisition process leads to a simplified yet realistic statistical model describing raw pixel values. The analysis and methods then proposed are based on this model. First, the theoretical performance bound of the problem is computed for the static case, where the acquisition conditions are controlled. Furthermore, a new method is proposed that, unlike previous methods, improves the reconstructed HDR image by taking into account the information carried by saturated samples. From a more practical perspective, two methods are proposed to generate HDR images in the more realistic and complex case where both objects and camera may exhibit motion. The first one is a multi-image, patch-based method, that simultaneously estimates and denoises the HDR image. The other is a single image approach that makes use of a general restoration method to generate the HDR image. This general restoration method, applicable to a wide range of problems, constitutes the last contribution of this dissertation.