Research on two-stage image denoising algorithm based on Feature Fusion for neutron imaging

Neutron imaging suffers from severe white spot noise characterized by high intensity, random spatial distribution, and resistance to conventional filtering, significantly degrading image quality and obscuring critical structural details. To address this challenge, we propose DFGU-Net (Dual-stage Feature-Gated U-Net), a novel self-supervised denoising framework that leverages intrinsic statistics from single noisy observations. The method integrates three key components: harmonic mean filtering for initial high-intensity noise suppression, a cascaded coarse-to-fine architecture for progressive structural restoration, and a gated feature fusion module for adaptive multi-level feature integration. Comprehensive experiments on the SIXray dataset demonstrate superior performance, achieving PSNR and SSIM improvements of 1.57 dB and 0.036 over S2S-NR, respectively. On real thermal neutron images, the method reduces BIQI and NIQE scores by 4.43 and 0.013, respectively, outperforming classical approaches (BM3D) and state-of-the-art deep learning methods (DnCNN, SUNet, GAN-based models) in both quantitative metrics and perceptual quality. The proposed framework strikes an effective balance between computational efficiency and restoration accuracy, making it suitable for practical deployment in neutron-based non-destructive testing applications.