Object detection, characterization and reconstructionfrom faint signals in images: applications in astronomy and microscopy
| When |
Oct 30, 2020
from 10:30 to 11:30 |
|---|---|
| Where | R 116 (GN1 Monod) |
| Attendees |
Olivier Flasseur |
| Add event to calendar |
vCal
iCal
|
[Click here to participate in visioconference]
Astronomy is a field of study in which optical advances have driven the design of new generations of instruments always more efficient and dedicated to specific tasks. In particular, the detection of exoplanets and their characterization by direct imaging from the Earth is a hot topic. Beyond the detection of exoplanets, the reconstruction of circumstellar disks made of gas and dust is of primary astrophysical interest since exoplanets could form inside such structures by accretion.
Microscopy is another field of study in which recent advances in terms of resolution and sensitivity have opened the door to new medical diagnoses. Among the large variety of imaging modalities, digital holography appears to be a cost-effective method of choice for characterizing microscopic objects.
For both application fields, the detection, characterization and reconstruction of the objects of interest are very challenging due to the underlying low signal-to-noise ratio regime, thus requiring a fine processing of the data by dedicated and versatile algorithms. In this seminar, we will present some of the processing algorithms we have proposed in the context of high-contrast direct imaging, in astronomy, and of digital holography, in microscopy, see Figure1. The underlying imaging challenges are formalized within an inverse problems framework. The main focus is put on the use of statistical and/or physics-based approaches to derive reliable and quantitative estimates characterizing the detected objects.Information redundancies (e.g.,temporal,multi-spectral) are also exploited.Robust processing strategies are also considered to improve their systematic deployment on data often corrupted by outliers. All the developed algorithms are totally unsupervised: weighting and/or regularization parameters are estimated in a data-driven fashion making the methods efficient for the processing of real data of uneven quality.
The related publications can be found on my personal webpage: https://olivier-flasseur.github.io