Research On Neutron Image Restoration Based On Deep Learning

Neutron imaging is a non-destructive testing technique that utilizes the interaction between neutrons and atomic nuclei of materials.It has wide applications in fields such as material science,archaeological research,etc.Neutron imaging can achieve tasks that are difficult for X-ray imaging techniques,such as high-sensitivity imaging of low atomic number elements(such as light elements,hydrogen,water,carbon,etc.)or deep structure analysis through heavy elements(such as lead,titanium,etc.).Neutron imaging has a similar principle to X-ray imaging techniques,but it faces many limitations in practical applications.On one hand,neutron imaging systems require high-intensity neutron sources,which are costly and hard to popularize.On the other hand,images obtained from small neutron generators are often of poor quality with noise,blur and other problems.Therefore,neutron image restoration methods have significant importance for improving the quality and usability of neutron images.Traditional neutron image restoration methods have achieved remarkable results in many image denoising and restoration tasks.However,these methods are usually designed for specific tasks and lack flexibility and generalization when dealing with multiple tasks.In recent years,deep learning methods have attracted wide attention and achieved good results in natural scene images,remote sensing,medical image processing,etc.This paper explores the unique properties of neutron images and proposes a deep learning-based neutron image restoration method that can flexibly handle various neutron image restoration tasks on a single model.The work and innovation of this paper are as follows:(1)A novel Layered Input Gradi Net for Image Denoising(LIGN)is proposed for neutron image restoration tasks.Layered input can effectively separate and extract different structural information in neutron images;gradient-based methods effectively enhance the restoration of edge texture structures in neutron images;the proposed plug-and-play sharpening loss function greatly enhances the visual quality of restored neutron images.(2)To solve the problem of lack of real neutron image datasets,this paper considers that neural networks are data-driven,based on neutron imaging principles and characteristics of neutron images,proposes an effective X-ray imaging dataset for network training and learning,which effectively improves training speed and experimental accuracy.This article selects representative,high-performance deep learning restoration networks in the field of natural images,based on different methods,and trains and tests them on neutron images to obtain experimental comparison data.To ensure the fairness of the experiments,the same hardware environment,training parameters,and evaluation metrics(PSNR and SSIM)are used for quantitative analysis and experimental comparison.Through analysis of the quantitative indicators,the superiority of the proposed method in terms of lightweight and applicability is demonstrated.The experimental results show that the proposed method achieves the best performance and excellent visual quality on both simulated noise neutron images and real noisy neutron images.