Deep learning-based white-spot suppression in neutron radiography using real gamma-induced artifact modeling

Neutron radiography is widely used for nondestructive inspection of advanced equipment. However, under strong radiation fields, radiographic images are often contaminated by high-energy particles-induced white spots with extreme intensity, which obscure critical internal details. Existing deep learning-based suppression methods rely heavily on manually synthesized white-spot data, leading to incomplete removal or excessive smoothing due to insufficient physical realism. To address these limitations, we propose a real white-spot modeling guided SAG-Restormer for neutron radiography. First, a physically realistic dataset construction strategy based on real white-spot extraction and transfer is developed. Second, a spatial anomaly gating (SAG) module is introduced into the Restormer backbone to alleviate attention degradation in global self-attention. Additionally, an asymmetric edge-preserving Charbonnier loss (AEPCL) is designed for high-fidelity reconstruction. Experiments demonstrate that the proposed method achieves superior performance on multiple no-reference image quality assessment (NR-IQA) metrics, effectively suppressing white spots while preserving intrinsic microstructural details.