Enhancing the spatial resolution of neutron radiography with generative adversarial networks

Abstract Neutron imaging inevitably suffers from image degradation such as image information loss and image blurring. Many experimental methods and device optimization schemes have been proposed to further improve the resolution of neutron imaging, this often sacrifices time resolution or incurs high retrofitting costs. This study proposes a flexible and effective super-resolution method based on neutron imaging experiment and deep learning algorithm. In the experiment, a large number of neutron images were measured using the reactor neutron imaging facility. The experiment measured multiple samples under various conditions to ensure that the network has good generalization ability. In the network structure, a more complex image super-resolution generative adversarial network was introduced. In the network, residual module, skip technology and dense connection were used to enhance the interaction performance of multi-scale feature information. More high-frequency details were reconstructed by introducing perceptual loss instead of pixel loss. The perceptual loss was calculated by extracting deep features of the image through an independent VGG network. Finally, the model was verified through real neutron imaging, with clarity, richness of information, spatial resolution, and subjective visual effect as the evaluation metrics. The results show that this method can achieve higher quality super-resolution reconstruction. The super-resolution image has better definition, greater information recovery, and improved spatial resolution level. The deep learning super-resolution method can express the intrinsic laws of neutron images, and reduce the optical blur caused by the imaging device acting as a frequency filter.