Deep image prior-based self-supervised framework for noise reduction with structural preservation in neutron imaging

초록

Neutron radiography often suffers from severe noise, multiple scattering, and system-induced blur, which collectively degrade structural visibility and quantitative reliability. To address these challenges, we propose a self-supervised denoising framework that integrates a deep image prior (DIP) network with gradient-Laplacian structural consistency and frequency-guided early stopping mechanism. The proposed method operates solely on a single noisy image unlike supervised approaches that require large paired datasets, effectively preserving edges and high-frequency features. The full-width-at-half-maximum value of the neutron image processed using the proposed DIP method was similar to 1.53, which demonstrates its superior ability to preserve edge information compared to those of compared noise-reduction techniques. The contrast-to-noise ratio and coefficient of variation results showed notable improvements when using the DIP method, yielding average increases of similar to 164% and 91%, respectively, compared with the noisy input across two neutron-image cases. A comparative analysis confirms that the framework mitigates over-smoothing common to conventional methods and avoids noise overfitting associated with DIP optimization. This method provides a generalizable and data-efficient solution for neutron imaging restoration with potential applicability across radiography, tomography, and time-resolved measurements.

키워드