Npj Comput. Mater.: 纵然藏她千百度，暮回首，AI把她全揭露！

海归学者发起的公益学术平台

分享信息，整合资源

交流学术，偶尔风月

然而，目前为提高采集效率所做出的努力，大多需要引入新的硬件设备，将降低整个实验系统的可靠性；而且单纯地去提升采集效率，往往还会降低数据的信噪比，最终降低数据质量。那么，到底有没有同步提升采集效率和数据质量的策略呢？那就是从算法的角度出发，对实验图像进行降噪，直接提升其数据质量，从而可以在实验中采用更快的采集速率，进而间接提高实验效率，最终获得实验效率和数据质量的同步提升。随着近年来人工智能技术的突飞猛进，基于机器学习和深度学习的降噪网络模型则是一个大有潜力的突破口。

来自中国科学院高能物理研究所的董宇辉研究员课题组提出了一种新型的、由多个Encoder和Decoder块作为主体架构的人工智能降噪模型：SEDCNN（图1）。这一模型以监督学习的方法，研究了SAXS/WAXD实验中低信噪比图像的信号恢复问题（图2）。模型的稳定性和降噪效果优于多数轻量化SOTA网络，如PMRID、REDCNN等。该模型的优势还在于其小型、轻量、易于训练的特性，未来可以被自然地部署在为新一代光源打造的实验过程控制与数据采集软件框架中，实现SAXS/WAXD实验图像的自动化和在线化降噪。此外，该工作对SAXS/WAXD物理图像的特征进行了充分考量，使用图像预处理方法对光斑中心高强度散射信号进行了屏蔽；使用方位角积分计算物理结果差异，形成了全新的降噪效果评价统计指标体系（图2），是一种为SAXS/WAXD实验定制化的人工智能降噪解决方案。

图1：Overall architecture of SEDCNN

图2：Visually, the details of the diffraction patterns have been successfully recovered after applying either radial integration or line profile in ALBULA; Numerically, various performance metrics show improved image quality after performing radial integration.

研究成果在其他有类似降噪需求的同步辐射光源实验站上，同样具有较高的应用价值。该文近期发表于npj Computational Materials 9：58(2023)，英文标题与摘要如下，点击左下角“阅读原文”可以自由获取论文PDF。

图3: Line profile and Q-value integration on experimental WAXD images.

图4: Line profile and Q-value integration visually compare the denoising effect on experimental WAXD images using our SEDCNN (WAXD, batch size equaling 12 and 16) and PMRID.

A machine learning model for textured X-ray scattering and diffraction image denoising

Zhongzheng Zhou, Chun Li, Xiaoxue Bi, Chenglong Zhang, Yingke Huang, Jian Zhuang, Wenqiang Hua, Zheng Dong, Lina Zhao, Yi Zhang & Yuhui Dong

With the advancements in instrumentations of next-generation synchrotron light sources, methodologies for small-angle X-ray scattering (SAXS)/wide-angle X-ray diffraction (WAXD) experiments have dramatically evolved. Such experiments have developed into dynamic and multiscale in situ characterizations, leaving prolonged exposure time as well as radiation-induced damage a serious concern. However, reduction on exposure time or dose may result in noisier images with a lower signal-to-noise ratio, requiring powerful denoising mechanisms for physical information retrieval. Here, we tackle the problem from an algorithmic perspective by proposing a small yet effective machine-learning model for experimental SAXS/WAXD image denoising, allowing more redundancy for exposure time or dose reduction. Compared with classic models developed for natural image scenarios, our model provides a bespoke denoising solution, demonstrating superior performance on highly textured SAXS/WAXD images. The model is versatile and can be applied to denoising in other synchrotron imaging experiments when data volume and image complexity is concerned.