Deep Learning Techniques for HDR Modulo Imaging
Brayan Monroy (Universidad Industrial de Santander, Colombia)
| Quand ? |
Le 16/09/2025, de 13:00 à 14:00 |
|---|---|
| Où ? | M7 101 |
| Participants |
Brayan Monroy |
| Ajouter un événement au calendrier |
vCal
iCal
|
Nathan Buskulic BRAYAN MONROY
Title: Deep Learning Techniques for HDR Modulo Imaging
Abstract: High dynamic range (HDR) imaging faces inherent limitations in conventional CCD and CMOS sensors, which quickly saturate under strong illumination, leading to loss of detail in bright regions. Modulo imaging, provide an innovative solution by cyclically resetting pixel intensities once they exceed a predefined threshold. This sensing strategy enables the theoretical acquisition of unbounded dynamic range, but introduces discontinuities and noise that make the reconstruction process challenging. Recovering an HDR representation from such wrapped measurements requires inversion algorithms capable of distinguishing true scene structures from artificial wrap artifacts while mitigating the presence of noise. Recent developments present deep learning strategies, optimization-driven formulations, and hybrid frameworks that combine the advantages of both. These advances demonstrate the ability to achieve robustness against noise, improved perceptual quality, and generalization to out-of-distribution lighting conditions. Beyond HDR reconstruction, HDR modulo imaging holds strong promise for real-world applications where extreme lighting is common, such as autonomous driving and aerial robotics, where reliable perception must be maintained across challenging scenarios from bright headlights to deep shadows. Together, these developments position HDR modulo imaging as a promising direction in computational imaging.
Website: https://bemc22.github.io/
In Room M7 101, 1st floor, Monod campus, ENSL.