Self-supervised denoising method for single neutron image based on the S2S-NR network

Abstract

Fast neutron radiography technology has unique application advantages in the field of non-destructive testing. However, during the imaging process, the imaging system is inevitably affected by various factors, leading to significant noise contamination in the resulting neutron images, which affects subsequent processing and analysis. In recent years, self-supervised learning has become a powerful tool for single image denoising. We propose a self-supervised denoising method based on the Self2Self-Neutron Radiography (S2S-NR) network to remove noise from fast neutron images. We train the network using a single noisy fast neutron image, employ gated convolution for feature extraction, and perform dropout training on Bernoulli sampling instances of the neutron image. The results are estimated by averaging the predictions from various instances of the network with dropout. Furthermore, we incorporate no-reference image quality assessment metrics into the loss function to optimize the training process. Experimental results show that the method achieves state-of-the-art denoising performance on both simulated and real fast neutron images, demonstrating the effectiveness and practicality of this method as a potential solution for the denoising task in fast neutron imaging.