Section: Overall Objectives

Representation and compression of visual data

The objective is to develop algorithmic tools for constructing low-dimensional representations of various imaging modalities (2D images and videos, multi-view, light fields, ...). Our approach goes from the design of specific algorithmic tools to the development of complete compression algorithms. The algorithmic problems that we address include data dimensionality reduction, the design of compact representations using overcomplete dictionaries, transforms on graphs, or autoencoders based on deep learning architectures. Low rank and sparse models are the essence of transform coding and of many other processing methods (e.g., denoising, classification, registration, super-resolution, inpainting). Developing complete compression algorithms necessarily requires tackling topics beyond the issues of sparse data representation and dimensionality reduction. For example, the problem of spatial, inter-view or temporal prediction using deep learning techniques is also addressed. Finally, rate-distortion models for constructing rate-efficient representations with various features of scalability or low dynamic range compatibility in the case of high dynamic range content are also studied.