Impulsive Noise Reduction for Self-Emitted X-Ray Images From a Scintillator-Based Imaging System

In this article, we present a novel method for image signal recovery from a scintillator-based X-ray imaging system inherently corrupted by impulsive noise. This approach combines a physically based image synthesis process with deep neural networks to effectively reduce neutron-induced impulsive noise. For the synthesis, experimental data are meticulously decomposed into constituent components based on different physical processes: the clean signal from self-emitted X-rays, shot noise induced by photons, impulsive noise from neutrons, and electronic noise from digital devices, each analyzed and modeled independently. A Legendre function is employed to model the clean hot spot signal, emphasizing its asymmetry. The noise power spectrum (NPS), derived from experimental data, is utilized for shot noise simulation, while the statistical characteristics of impulsive noise are analyzed and integrated into the synthesis process. A deep neural network, leveraging the U-Net architecture, is trained on these synthetic images and subsequently applied to real experimental data. This method significantly enhances the signal-to-noise ratio (SNR) of experimental data, improving from $\sim -30$ to −20 dB and $\sim ~20$ to 30 dB, and maintains high fidelity in clean signal representation. Our approach demonstrates significant potential to improve image quality and diagnostic accuracy in experimental settings and could be adapted for similar applications in other research contexts involving imaging systems in radiation environments.