Ricardo Borsoi

CNRS researcher
University of
Lorraine


Scholar profile
Github projects

Research

In the following you can find quick description of the research problems I've been working on.

Spectral Unmixing

Hyperspectral devices sample the reclectance spectrum of a material at hundreds of contiguous spectral bands. This high spectral resolution allows for a precise characterization of a material based on its spectral characteristics. However, due to physical limitations hyperspectral cameras have very limited spatial resolution, which means that most pixels in the image are actually composed of a mixture of several pure materials, which are called endmembers. Unmixing attempts to recover the spectra of the pure materials (i.e., the endmembers) as well as the propoetions to which they contribute to each pixel of the image (which are called the abundances).

Several challenges underlie the unmixing problem:

  • Variability of the endmember spectra

  • Nonlinear interactions between light and the materials in the scene

  • Multitemporal data

In my work, I've addressing these problems by developing algorithms based on sparse regression, multiscale image decomposition, lor-rank tensor models, and deep learning.

Image Fusion and Super Resolution

Most imaging modalities (such as, e.g., hyperpectral or multispectral/RGB images) have limitations and trade-off in the statial, spectral or temporal resolution of the acquisitions. Multimodal image fusion or super resolution aims at overcoming the physical limitations of the imaging sensors to explore complementary strenghts of each of them to obtain high-resolution image data. Similarly, super-resolution aims to improve the resolution of images acquired from a single modality. Nonetheless, this problem can be ill-posed and requires a proper framework to obtain high quality solutions. I investigated this problem in multiple frameworks, for RGB videos and hyperspectral and multispectral image fusion, developing methods that are computationally efficient and can represent the nonidealities of the image acquisition process to obtain higher quality recontructions.

Graph Identification and change detection

Data in many real world applications are supported in irregular domains or networks, which can be represented as graphs. This is the case of, e.g., brain networks and power grids. Graph signal processing (GSP) aims at extending classical signal processing tools to process data that is supported on graphs. However, several challenges exist. First, in many applications the connectivity of the graph is not known in advance and must be estimated from measurements. Second, applications of GSP are often large-scale problems, which require efficient methods. In particular, two important problems are to investigate the identification of network connectivity from nonlinear measurements, and the detection of anomalie or abrupt changes in temporal data streams, which often occur in groups of datapoints that are highly connected accordingn to the graph topology. A current research direction consists of developing efficient online algorithms for these task and analysing their performance.