Bayesian Deep Image Priors for denoising XRF maps
Abstract
Abstract Low-Energy X-ray Fluorescence (XRF) mapping at synchrotrons is limited by acquisition time and dose, especially for sensitive samples. Sparse scanning and traditional denoising methods may struggle with high noise and low photon counts. While supervised deep learning can hallucinate “nice looking” features when ground truth is unavailable, Deep Image Prior (DIP) methods avoid pre-training by optimising a network on a single measured image. In this work, we extend DIP with Bayesian reasoning and a Poisson likelihood tailored to XRF maps. We test our method on noisy synchrotron XRF data, demonstrating reliable image reconstruction.
Fields of Study & Subjects
Advanced X-ray Imaging Techniques
Advanced X-ray and CT Imaging
Machine Learning in Materials Science
Publication details
- Journal / source
- Journal of Instrumentation
- Volume
- 21
- Issue
- 01
- Pages
- C01006–C01006
- Publication date
- 1 January 2026
- Language
- EN