Robust Multispectral data analysis for target detection and IDentification – MultID

Multispectral imaging is an observation modality that already plays an essential role in a very large range of scientific fields ranging from remote sensing, geoscience, astronomy to bioimaging to only name a few. A very appealing aspect of multispectral is the possibility to highly discriminate between physical contributions based on their spatial distribution and more importantly their spectral behavior. These potentially high discrimination capabilities have motivated the use of multispectral images in scientific domains as different as astronomy, bioimaging or remote sensing. In the scope of astronomy, most recent space missions are now equipped with multispectral imagers (e.g. ESA’s missions Mars Express, Herschel, Planck or NASA’s satellite Fermi are just a few examples of missions embarking multispectral imagers). Multispectral imagers have recently emerged as a new promising tool in passive surveillance imaging, in civil and military contexts, where enhanced anomaly detection capacities are expected based on the discrimination ability of multispectral images.
These seemingly different domains of applications share the same data processing problems: how to perform a coherent processing to extract useful features. These features can be either physically meaningful components in astronomy or anomalies in surveillance imaging. In the context of surveillance imaging, target detection can be viewed as a target/background. For that purpose, blind source separation has been widely applied in astronomy but has not been adequately and efficiently extended to target detection problems. In this context their adaptivity to the data should provide better detection rate and more robustness to the variability of the data (scene variability, weather conditions …etc) thus departing from the methods classically used in this field. Therefore the objective of this proposal is to extend the source separation framework to tackle the problem of target detection in surveillance imaging.
In general target detection/identification problems and beyond the standard source separation framework, a priori information are very often known on either the targets or the background. Taking this a priori knowledge into account in source separation is rarely done if never when the number of targets exceeds the number of observations. Exploring novel extensions of the standard source separation framework to take spectral prior information into account will be part of the proposal. This will be carried out by getting inspiration from recently sparsity-based source separation methods and sparse decomposition techniques; these techniques being well-suited for large-scale problems (i.e. high number of potential targets). These methods are also well known to provide robustness to instrumental noise and mismodeling which are highly desirable properties for robust anomaly detection.
Underlying the duality of this proposal, we would like to point out that the two targeted applications of this project will highly make profit of the joint development of source separation methods. The first application will focus on astronomical component detection and extraction, more specifically in the scope of the Mars Express (for evaluation purposes) and the Planck project. We will put a strong emphasis on evaluating the performances of source separation methods and the proposed extensions in the context of surveillance imaging for anomaly detection purposes with our industrial partner SAGEM.