PhD Thesis Subject 2020-2022
Blind restoration of hyperspectral images acquired by UAVs
IETR – UMR 6164 CNRS / SHINE Research Group - Lannion
Inverse problems, modelling, estimation, filtering, restoration, deconvolution, optimization, multi-criteria, regularization, bigdata, UAVs.
Master M2 with strong skills in Signal and Image Processing, Statistics and Applied Mathematics; Python, Matlab, C/C++ skills; fluency in written/oral Scientific English.
FOR MORE information, PLEASE CONTACT CO-SUPERVISORS:
BACKGROUND AND DESCRIPTION
The new generation of airborne hyperspectral imaging sensors, embedded on small aircraft or UAVs, is of great economic, technical and scientific interest. These datasets allow retrieving valuable information on the nature (content) and the spatiotemporal evolution of over flown areas. However, their analysis and interpretation remain difficult in practice, when the acquired data are distorted by several sources of degradation related to the acquisition system and/or its environment.
To interpret the content of such data in an optimal way (so as to reveal, for instance, the accurate spectral signature of in-situ minerals and vegetable species imaged on the whole available spectrum), a preliminary stage of restoration (including denoising and deconvolution) must be introduced to compensate for the different sources of degradation, either depending on the sensor and/or the acquired scene.
Some solutions have already been developed in the literature, but they are limited in their exploitation in terms of accuracy, robustness and automatic implementation. In most cases, these methods are semi-blind and require a priori knowledge related to the degradation function and/or the image to be restored. In addition, they are often sensitive to the settings of the values of the regularization parameters involved in the cost functions used.
To solve this complex problem, it is necessary to develop a restoration approach introducing a minimum of a priori knowledge and relying on a joint exploitation of local spatial and spectral information in the acquired images. We propose in this thesis to develop an original multi-criteria restoration approach both taking into account the heterogeneity of involved environments and being adaptive to the acquisition conditions and the content of over flown areas. This approach will address three issues together:
•The first issue focuses on the analysis and the estimation of the characteristics of a signal dependent observation noise (and therefore not stationary), especially for images acquired with the last generation of hyperspectral sensors.
•The second issue relates to the deconvolution problem. It requires first an advanced modelling of the point spread function (PSF) of the whole imaging system. The objective is to estimate both the dimension and the variability of the PSF spatial and spectral spreads by integrating the information of all the spectral bands of the acquired image.
•The third issue aims to formalize the impact on the efficiency of the subsequent processings of uncertainty achieved on the estimated characteristics of the PSF and the observation noise. The goal is essentially to adapt these correction procedures so as to make them robust to uncertainties of the observed models and to their estimated parameters.
The methodological contribution of the restoration approach proposed in this thesis will be directly evaluated on targeted topics conducted in collaboration with our academic and/or industrial partners, in particular within the framework of the Centre Technologique Drone Ouest and the Anticipa Lannion technopole.
This project follows the work of a PhD thesis defended in our laboratory (December 2018).
(c) GdR 720 ISIS - CNRS - 2011-2020.