Speaker
Description
Synchrotron light source facilities are evolving into the fourth generation with extreme spatial, temporal and energy resolving capabilities, which pushes the transition of experiment modes into high resolution, multiscale, ultra-fast, and in-situ characterization with dynamic loading or under operando conditions. Such transition raises challenges to balance acquisition efficiency and data quality, where denoising algorithms play a crucial role in signal-to-noise improvement and physical information retrieval. Herein, we develop two efficient denoising algorithms by integrating physical prior knowledge of diffraction/scattering images into deep learning methods.
The first algorithm [1] is based on a small, yet efficient machine learning model designed specifically for SAXS/WAXD experimental image denoising. This model allows for the preservation of physical information and signal-to-noise ratio even when exposure time or dose is significantly reduced, providing a customized solution compared to traditional denoising models designed on natural images. In particular, the proposed model demonstrates superior performance in processing highly textured SAXS/WAXD images compared to the mainstream denoising algorithms. Additionally, the versatility of the proposed model enables wide application in other synchrotron imaging experiments, particularly when data volume and image complexity is concerned.
The second algorithm [2] leverages the intrinsic physical symmetry of X-ray patterns, achieving excellent blind denoising and physical information recovery capabilities without high signal-to-noise ratio reference data. This method is more efficient and effective than the deep learning approaches without physical symmetry considered. It can effectively recover physical information from spatially and temporally resolved data acquired in X-ray diffraction/scattering and Pair Distribution Function experiments, while also maintaining high tolerance on asymmetric distribution of experimental patterns. As a self-supervised denoising approach, the proposed algorithm benefits and facilitates photon-hungry as well as time-resolved in-situ experiments with dynamic loading.
This talk will introduce our most recent discoveries on a systematic denoising solution comprising both supervised and self-supervised denoising methods. The presented material is suitable and contributes to one of the covered topics of NOBUGS 2024: AI/ML applications. We are looking forward to having further discussions with the audience interested in such an intriguing topic.
[1] Zhou, Z., Li, C., Bi, X. et al. A machine learning model for textured X-ray scattering and diffraction image denoising. npj Comput Mater 9, 58 (2023).
[2] Zhou, Z., Li, C., Fan, L. et al. Denoise X-ray Image by Exploring the Power of Its Physical Symmetry. J. Appl. Cryst., Accepted (2024).
| Abstract publication | I agree that the abstract will be published on the web site |
|---|
