This thesis is about the design of new algorithm tools that improve the compression ratio of current video coding standards, such as H.264/AVC. To reach this goal, a preliminary study on a set of image restoration methods identified two distinct lines of research. The first is based on methods of texture analysis and synthesis. This kind of method, also known as template matching, is commonly used in video coding contexts to predict a portion of an image texture from an analysis of its neighborhood. We tried to improve the prediction model by taking into account the specificities of video encoders such as H.264/AVC. In particular, the rate-distortion function used in video coding standards is commonly based on an objective measure. This mechanism is inherently incompatible with the concept of texture synthesis, whose effectiveness is usually measured perceptually. It was this contradiction that motivated this first line of research. The second is inspired by image regularization methods based on total variation minimization. These methods were originally developed in order to improve the quality of an image according to prior knowledge of its damage. Starting from this work, we designed a predictive model of transformed coefficients obtained from a natural image, which is integrated into a conventional video encoding scheme.
