Enhancement of neutron radiography resolution through residual hybrid attention network

Neutron radiography is widely used in aviation, nuclear energy, metallurgy, geology and other fields. Improving the image quality and spatial resolution of neutron radiography is crucial to promoting the development of non-destructive testing. This paper proposes a novel solution based on deep learning to improve the quality of neutron radiography. In order to train a more reliable super-resolution network, a real neutron image dataset was constructed using the reactor neutron radiography facility. This dataset contains neutron images accumulated experimentally over the past decade, which amalgamates numerous imaging conditions and various sample types. Based on deep learning, this paper proposes a hybrid attention network that can efficiently extract multi-scale information. The multi-scale branch, residual-in-residual, channel attention and spatial attention can effectively enhance the expressive ability of the network. After two stages of transfer learning, the network demonstrated excellent performance. Both subjective vision and objective scores showed that the network has outstanding abilities to diminish blur artefacts, restore texture details, and enhance resolution. The solution based on deep learning reduces the point diffusion effect in the neutron space detection mechanism, enhances the resolution of neutron radiography, and further improves the accuracy of non-destructive testing.