Vous êtes ici : GIPSA-lab > Formation > Thèses soutenues
PICONE Daniele

Model Based Signal Processing Techniques for Non-conventional Optical Imaging Systems


Directeur de thèse :     Mauro DALLA MURA

Co-directeur de thèse :     Laurent CONDAT

École doctorale : Electronique, electrotechnique, automatique, traitement du signal (EEATS)

Spécialité : Signal, image, parole, télécoms

Structure de rattachement : Université Grenoble Alpes

Établissement d'origine : Univervité de Salerne - Italie

Financement(s) : contrat à durée déterminée ; Sans financement ; Sans financement


Date d'entrée en thèse : 01/10/2017

Date de soutenance : 25/11/2021


Composition du jury :
Andrés Almansa, DR au CNRS, MAP5, CNRS, Université Paris Descartes, Paris, France
- Magnús Örn Úlfarsson, Professor University of Iceland, Reykjavík, Iceland
- Valérie Perrier, Professeur des universités, Université Grenoble Alpes, CNRS, Grenoble INP, LJK, Grenoble, France
- Enrico Magli, Professore ordinario Politecnico di Torino, Turin, Italy
- Etienne le Coarer, Ingénieur de recherche, IPAG/UGA-CNRS, Université Grenoble Alpes, Grenoble, France
- Mauro Dalla Mura, Maître de conférences, Université Grenoble Alpes, CNRS, Grenoble INP, GIPSA-lab, Grenoble, France; Institut Universitaire de France, Paris, France; Tokyo Tech WRHI, School of Computing, Tokyo Institute of Technology, Tokyo, Japan
- Laurent Condat, Research sci


Résumé : There is an increasing demand for images with higher spectral and spatial resolution for applications in several domains such as health, environment, quality checking and natural disasters monitoring. Hyperspectral imagery provides the necessary spectral diversity to recover the composition of materials on site for applications such as the detection of fires, anomalies, chemical agents, targets and changes in the scene. The requirement for cheaper and more compact devices (e.g. to be embarked on low cost satellites and airborne platform) which are capable of capturing this information has led to the development of nonconventional innovative design concepts to overcome the technological limitations of traditional cameras. Data acquired by such novel imaging devices following the computational imaging paradigm are typically not readily exploitable for the final application. A computational phase is hence needed for extracting useful information from the raw acquisitions. This thesis addresses this issue by setting up an inversion problem. The general approach is to characterize the data fidelity term with a physical model, describing the underlying optical transformations performed by the device. The challenge is then shifted on the regularization step to properly characterizes the features of the quantities of interest and improve the accuracy of the estimation, which can be tackled with variational techniques. The analysis is applied to two novel concepts for nonconventional optical devices. The first one is a novel compressed acquisition imaging system based on color filter arrays, which embeds information from sensors with different spatial and spectral characteristics into a single mosaiced product. As opposed to existing compressed sensing based devices, the goal is not to recover the original uncompressed mul- tiresolution sources, but instead to directly recover a synthetic fused image with both high spatial and spectral resolution. The proposed solution relies on the to- tal variation regularization and is the subject of a detailed analysis, comparing its compressive power with straightforward software alternatives, evaluating its performances as the amount of channels changes, and validating its efficiency in comparison to state of the art methods when applied to classical fusion or mosaicing algorithms separately. The second class of devices is based on the ImSPOC patent, a design concept for a low finesse snapshot imaging spectrometer based on the interferometry of Fabry-Pérot. Its ideal behaviour follows the principle of the Fourier Transform Spectroscopy, as its acquisition can be interpreted as a sampled version of an interferogram, arranged across different sub-images distributed on the same focal plane. After defining a physical model based on optical geometry, its validity is evaluated over real acquisitions by setting up a Bayesian inference problem to determine its parameters, with approaches based on maximum likelihood estimators, regular-grid searches and nonlinear regression. A variety of preliminary tests are then carried out on the inversion method, with approaches based on singular value decomposition and sparse-inducing regularizers, accompanied by a analysis of their robustness to model mismatches.

GIPSA-lab, 11 rue des Mathématiques, Grenoble Campus BP46, F-38402 SAINT MARTIN D'HERES CEDEX - 33 (0)4 76 82 71 31