Super-Resolution Reconstruction for Neutron Radiography Using Improved Real-ESRGAN

Neutron radiography is a well-established nondestructive testing technique, but it remains insufficient to meet the demands of high-resolution (HR) imaging, particularly when based on a compact neutron source. Real-ESRGAN, a powerful image super-resolution (SR) method based on a generative adversarial network (GAN), is extended in this study to practical HR restoration applications in neutron radiography. The complex degradation factors in neutron radiography (e.g., blur, noise, and gamma white spots) are considered to enhance the generalization ability of the SR network for neutron image data. Thereafter, the U-Net discriminator is integrated with attention gates (AGs) to enable the SR network to focus on important features and suppress irrelevant information. To balance the cost and reliability of dataset construction, X-ray images are used as data drivers for model training. The experimental results on both synthetic and real data demonstrate that the proposed method reconstructs HR neutron radiographs with sharper texture details and less noise than state-of-the-art SR techniques, in terms of both visual quality and quantity.