ANR research project · 2027–2030
FLUO
Donaldson Matrix Reconstruction from Spectral Imaging and Physics-Informed Neural Networks
Fluorescent materials re-emit part of the light they absorb at longer wavelengths. Measuring this behaviour precisely today requires expensive laboratory instruments. FLUO aims to make it accessible: the project combines affordable multispectral cameras with physics-informed neural networks to estimate the complete optical fingerprint of a material, and tests a common simplifying assumption about how fluorescent light is emitted.
- 48 months · January 2027 – December 2030
- ANR single-team project (PRME), CE42 Sensors, imagers and instrumentation
- ImViA laboratory, CORES team, Université Bourgogne Europe, Dijon
Context
The Donaldson matrix D(λex, λem) is the optical fingerprint of a material (Donaldson 1954). For each excitation wavelength λex, it gives the light re-emitted at each emission wavelength λem:
- its diagonal is the reflected radiance factor, i.e. the ordinary colour of the material;
- its lower part (λem > λex, the Stokes shift) describes the fluorescence.
Scientific barriers
- The reference instrument, the bispectrometer, is costly, confined to the laboratory and slow (point-by-point scanning).
- Separating a weak fluorescence from a dominant reflectance is a severely ill-posed inverse problem.
- Low-cost snapshot sensors have a low spectral resolution and a low signal-to-noise ratio.
- Reference D-matrices are scarce, so fully supervised deep learning is not viable.
- Fluorescent emission is usually assumed to be Lambertian (identical in all directions), but this is rarely verified.
Hypotheses
- Computational hypothesis
- A few illuminants and spectral bands are enough to estimate the D-matrix, provided they are coupled with physical constraints and deep unmixing.
- Physical hypothesis
- The Lambertian assumption of fluorescent emission is an oversimplification for complex real-world materials.
Objectives
- Compact and affordable acquisition hardware: a snapshot multispectral camera with multi-LED illumination.
- Physics-informed neural networks (PINNs), semi-supervised and robust to low signal-to-noise ratio.
- An open database of raw spectral images and reference D-matrices.
- An empirical test of the Lambertian assumption using Reflectance Transformation Imaging (RTI) and spectro-goniometry.
Methodology
- WP1 — Acquisition and open dataset (2027–2028)
- A reference bispectral setup provides ground-truth D-matrices; a low-cost snapshot multispectral camera with multi-LED illumination is designed and optimised. Both feed an open, standardised database.
- WP2 — Physics-informed neural networks (2028–2030)
- Analytical baselines and deep backbones, then a semi-supervised PINN built on a differentiable optical model y = DE and a physics-informed loss (Cuomo et al. 2022).
- WP3 — Testing the Lambertian assumption (2028–2030)
- A spectro-goniometer and a multispectral RTI dome measure fluorescence under many angles, to quantify non-Lambertian emission and derive hardware recommendations and physical priors.
The materials studied range from reference samples (dyes, security inks) to industrial materials (paints, plastics, papers, minerals) and biological specimens (leaves, ex vivo tissues).
Expected impact
- Science
- First empirical models of the directionality of fluorescence; PINNs for spectral unmixing at low signal-to-noise ratio.
- Open science
- An open D-matrix database, open-source code and models, and a benchmark challenge.
- Applications
- Anti-counterfeiting, quality control of coatings, paints and plastics, precision agriculture (chlorophyll fluorescence), photorealistic rendering.
Team
FLUO is carried out by the CORES team of the ImViA laboratory (UR 7535), Université Bourgogne Europe, Dijon.
- Yannick Benezeth
Project coordinator - Wahabou Abdou
- Thomas Boussey
- Stéphanie Bricq
- Hermine Chatoux
WP2 leader - Pierre Gouton
- Alamin Mansouri
WP3 leader - Franck Marzani
- Matthieu Rossé
- Jean-Baptiste Thomas
WP1 leader
Open positions. The project will recruit a postdoctoral researcher in optical metrology and two PhD students (computer vision / AI, and computational imaging / optics). Announcements will be published here.
