Search

• A neutron image denoising method based on multi-branch network is proposed. • Each branch is separately targeted for denoising one type of noise. • It would be more effective to use a combination of pixel-level and feature-level fusion to synthesise the noise residuals. • Multi-branch network structure is more flexible and effective in mixed noise denoising compared to a single model. Neutron imaging technology is a non-destructive testing method that uses the interaction characteristics between neutrons and matter to obtain images. In the process of neutron imaging, it will be affected by many noises. To address Gaussian, Poisson and Gamma white spot noise in neutron images, a new denoising method based on multi-branch network is proposed. The basic idea is to use different network branches to deal with different types of noise, resulting in more accurate and effective denoising. Finally, the weight fusion module network was used to synthesize the noise residual output of each branch to generate the final total residual image. Experimental results show that the proposed method is significantly more effective in dealing with mixed noise in neutron images compared to several other single-model networks, proving the superiority of the multi-branch network in the field of neutron image denoising. [ABSTRACT FROM AUTHOR]

Neutron radiography has been widely employed in nondestructive investigations. The noise and blur that are inevitably generated in the process of neutron radiography seriously decrease the image quality and lead to the loss of image information. In particular, white spot noise, which has the characteristics of nonuniform distribution, large size, and high magnitude, is difficult to remove. To solve this issue, we apply a generative adversarial network-based to convert a degraded neutron image into a clean image. The basic idea is to integrate attention mechanisms into our generation and discrimination networks. A fusion block with a visual attention mechanism is proposed, which can extract more potential features from the images and preserve the image texture as much as possible. In addition, due to the lack of neutron image datasets, we propose a neutron image degradation model to simulate the noise and blur in real neutron images. The results of the experiments show that our method can effectively eliminate noise and blur from neutron images while retaining texture information well. • Proposed a deep learning-based method to remove hybrid noise and blur from neutron images. • A fusion block combines a dense residual block and a spatial channel attention mechanism. • Perceptual loss is used to make our network efficiently learn the high-level information of the image. • Visual evaluation and quantitative analysis illustrate the effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]