Search this site
Embedded Files
Luca Calatroni
  • Home
  • Short CV
  • Papers
  • Team
  • Codes
  • Talks & visits
  • Projects
    • ANR JCJC TASKABILE
  • Organising
  • Teaching
Luca Calatroni
  • Home
  • Short CV
  • Papers
  • Team
  • Codes
  • Talks & visits
  • Projects
    • ANR JCJC TASKABILE
  • Organising
  • Teaching
  • More
    • Home
    • Short CV
    • Papers
    • Team
    • Codes
    • Talks & visits
    • Projects
      • ANR JCJC TASKABILE
    • Organising
    • Teaching

ANR JCJC 2022 - TASKABILE

Task-adapted bilevel learning of flexible statistical models for imaging and vision (ANR-22-CE48-0010)

Context and objectives

The project TASKABILE is positioned at the interface between three different areas: inverse problems, optimisation and learning. It focuses on the development of reliable bilevel optimisation methods for the optimal and task-adapted modelling of problems arising in imaging and vision. It is structured around three main questions.

  • Task-adapted model assessment: The quantitative evaluation of how well a hand-crafted imaging model performs when used to solve a specific problem is, intrinsically, a very subjective and debatable question: there is no universal quality measure tailored to quantify the performance of a model. To control the error committed in the reconstruction process w.r.t. to a (often unavailable) ground-truth reference image, global error-based metrics are typically used, while other quality metrics more inspired to the actual visual processing of image contents happening in our brain can be used for the computation of visually satisfying results. The functional expression of these metrics is often very hard to manipulate as structural/colour image features are often modelled in a highly non-linear way. Furthermore, they often rely on empirical psycho-physical evaluations.

  • Content-adapted modelling: The description of meaningful image contents in terms of relevant features is a challenging problem in mathematical image processing and computer vision. Exemplar salient features classically used to describe geometric (i.e. non-textured) image contents are, for instance, edges and orientations, while scale and frequency information are typically used to model textured components. Due to the high-heterogeneity of image contents observed in many applications (e.g., biomedical imaging) as well as in the processing of natural images, it is often worth considering a flexible modelling where such contents are represented by a variety of local  descriptors whose functional form is derived from statistical, analytical and geometrical considerations.

  • Fast and reliable hyper-parameter tuning: All existing model- and data-driven approaches for image processing and analysis require careful hyper-parameter tuning to get meaningful results. Coming up with a problem-specific and mathematically-grounded parameter selection strategy is often the most time-consuming and challenging problem for the validation of a method. In statistics, this is typically performed by maximum-likelihood or conditional-mean estimators; in the context of sparse regularisation, this is typically done by exploiting prior knowledge of the degradation process (noise, blur) observed in the data while in machine- and deep-learning approaches this is achieved by the design of suitable architectures. Bilevel learning approaches provide a natural and interpretable strategy to effectively optimise parameter models tailored to the task at hand.

TASKABILE aims to make the framework of bilevel learning a paradigm for the reliable and task-dependent estimation of adaptive feature-dependent models for imaging and vision. Differently from deep-learning black boxes, the approaches described in TASKABILE are theoretically grounded, interpretable and provide a flexible tool combining statistical/variational modelling with vision-inspired optimisation in a unified way.

Timeline, hosting institution and granted amount

The duration of the project is 48 months for the period March 1 2023 - February 28 2027 (ended in October 2024 due to change of affiliation of PI). The project will be carried out in the I3S laboratory of Sophia-Antipolis, France. The allocated project funding is 272 350€.

Team members

  • PI: Luca Calatroni (CNRS, FR)

  • Alessandro Lanza (Università di Bologna, IT)

  • Thomas Moreau (Inria Saclay, FR)

Collaborators

  • Xavier Descombes (Inria SAM, FR)

  • Kostas Papafitsoros (QMU, UK)

Project publications (HAL repository): complete list

Google Sites
Report abuse
Google Sites
Report abuse