扫描透射电子显微镜（STEM）可以在原子皮米级别上对物体材质进行电子成像，并利用图像进行原子结构解读。但是，要获得高质量的原子尺度STEM图像需要高端的STEM设备以及熟练的操作者，各种环境因素都会在STEM的成像过程中引入难以预计的非均匀噪声，从而严重影响图像质量，进而影响原子结构分析结果。基于深度神经网络的预测模型可以通过去噪或数据拟合来减少噪声的影响，但存在过拟合问题。该文将材料结构条件先验建模到深度神经网络模型中，设计了一种基于材料结构条件先验的高噪声STEM图像原子结构分割方法。该方法通过对比学习方式将材料结构条件先验建模成分割网络的注意力（包括自注意力和交叉注意力）并加以计算，不仅使得分割网络能够自适应地关注图像中的关键区域，还能自适应地关注来自结构坐标向量模态的控制信息。在仿真测试集中，该方法相比AtomAI Segmentor方法，在倒角距离、Jaccard分数和F1分数上分别提升175%、49.7%和42.7%；相比作者课题组早期提出的多尺度方法，在倒角距离、Jaccard分数和F1分数上分别提升167%、28%和23.9%。在实验室样本测试集中，该方法相比AtomAI Segmentor方法，在倒角距离、Jaccard分数和F1分数上分别提升63%、9.3%和7.4%；相比作者课题组早期提出的多尺度方法，在倒角距离上提升12.8%，在Jaccard分数和F1分数上性能持平。材料结构条件先验的引入，使得分割网络模型能更准确地分割高噪声STEM图像中的原子结构，并预测被噪声或顶层遮挡的次级结构信息。

Scanning transmission electron microscopy (STEM) can perform electron imaging of material properties at the atomic picometer level and interpret the atomic structure using the obtained images. However, obtaining high-quality atomic-scale STEM images requires advanced STEM equipment and skilled operators. Various environmental factors can introduce unpredictable non-uniform noise during the STEM imaging process, thereby significantly affecting image quality and consequently influencing the results of atomic structure analysis. The prediction model based on deep neural networks can reduce the impact of noise through denoising or data fitting, but there exists a problem of overfitting. This paper introduces materials structure conditions as priors in the deep neural network model and designs a method for atomic structure segmentation of high-noise STEM images based on materials structural priors. In this method, the materials structural priors are modelled as the attention (including self-attention and cross-attention) of the segmentation network and are calculated, which not only enables the segmentation network to adaptively focus on the key regions of the image but also to adaptively focus on the control information from the structural coordinate vector modalities. In the simulation test set, as compared with AtomAI Segmentor method, the proposed method improves the chamfer distance, Jaccard and F1 metrics by 175%, 49.7% and 42.7%, respectively; as compared with the early multi-scale method proposed by the research group, it improves the chamfer distance, Jaccard and F1 metrics by 167%, 28% and 23.9%, respectively. In the laboratory sample test set, as compared with AtomAI Segmentor method, the proposed method improves the chamfer distance, Jaccard and F1 metrics by 63%, 9.3% and 7.4%, respectively; as compared with the early multi-scale method proposed by the research group, it improves the chamfer distance by 12.8%, and the Jaccard and F1 metrics remain largely unchanged. The introduction of materials structural priors enables the segmentation network model to more accurately segment the atomic structure in high-noise STEM images and predict the secondary structure information that is affected by noise or top-level occlusion.