Computer vision is one of the most challenging research domains in engineering sciences. The aim is to reproduce human visual perception through intelligent processing of visual data. CVN aims at proposing innovative techniques towards automatic structuring, recovering, interpreting, and modeling big (visual) data. CVN is associated with Inria Paris-Saclay through a joint research project-team (OPIS). Researchers of CVN are also members of the Fédération de Mathématiques de CentraleSupélec (FR CNRS 3487). Our primary objective is to remain a reference center of international scientific excellence and to contribute significantly to the theory and practice in the field of computer vision, artificial intelligence, and (bio)medical imaging.
Variational problems requiring the estimation of a very large number of variables have now to be tackled, especially in the field of inverse problems (e.g., ≥ 109 variables in 3D imaging). In addition to the curse of dimensionality, another difficulty to overcome is that the cost function usually reads as the sum of several loss/regularization terms, possibly composed with large-size linear operators. These terms can be nonsmooth and/or nonconvex, as they may serve to promote the sparsity of the sought solution in some suitable representation or to fulfill some physical constraints. In such a challenging context, we develop advanced deterministic and stochastic optimization methods based on fixed point iterations, proximal techniques, majoration-minimisation (MM) approaches, and distributed/parallel implementations.
Machine learning methods have led to impressive results in various domains of Data Science. Nevertheless, the fundamental reasons for their excellent performance are often still poorly understood. We are developing robust, explainable, and efficient inference methods. Deep neural networks are the current state-of-the-art methods for solving a wide range of diverse tasks in signal/image classification or regression. We are working on reliable design and learning techniques for neural networks. We address challenges raised by partially annotated data and weakly supervised learning. Machine learning on graphs is also an important task with a plethora of practical applications. Our goal is to develop a systematic framework for large-scale data mining and representation learning on graphs.
Imaging devices provide a huge amount of information at various scales thanks to a wide range of modalities (MRI, Xray, PET, ultrasound, microscopy, …). These images can be multispectral, volumetric or correspond to sequences of 2D or 3D fields. Our group has developed a long-term expertise in image analysis, segmentation, denoising, restoration, and reconstruction. Advanced image models are built thanks to convex or nonconvex variational approaches. Bayesian methods are also employed, as well as techniques based on mathematical morphology and graphs.