Spatial-frequency attention-based neutron radiographic image restoration and quality assessment method

Neutron radiographic images are often degraded by insufficient beam intensity, exposure time, scattering, detector noise, and collimator size limitations, resulting in low contrast, complex noise, spatial blur, and white spots. Current studies in neutron radiography often treat image restoration and quality assessment as isolated tasks. To achieve a unified framework, we developed a two-step improvement scheme by utilizing the same spatial–frequency attention module and consistent datasets on both the restoration and assessment networks. For the proprietary dataset construction, an updated simulation method for multiple degradations is employed to mitigate the discrepancy between simulated and real-world images. For the network design, centered on the same spatial-frequency attention module, our approach integrates the U-Net-based restoration network with a no-reference image quality assessment network specifically designed for seed screening tasks. Experimental results show that the proposed restoration-assessment scheme significantly enhances the visual quality of authentic neutron radiographic images, thereby providing a potential image-preprocessing method for the subsequent industrial defect detection task.