Neural network denoising of high-energy-density x-ray images

Noise is a significant problem for x-ray images of High-Energy-Density (HED) experiments which reduces our ability to infer quantitative physical properties about the system. In these images we measure the spatially varying x-ray transmission across the system. We typically image over very short (10-100 ps) time scales, so these images are subject to significant statistical (Poisson) noise, and harder x-rays coupled with the camera response produces an additive noise. Denoising is necessary for improving the quality of these images for analysis, but the choice of filter also affects the underlying information in the image. Our goal is to measure small fluctuations in density from these images, so we want a denoiser which retains as much real information as possible. Spatially varying Poisson noise is not typically encountered in common optical images, however, so denoising methods are not usually tested for this kind of noise. We train a Deep Convolutional Neural Network (DnCNN) to denoise our experimental images, which generally performs better than others.