基于卷积神经网络的滑动窗口算法在高分辨电镜图像中的应用

基于卷积神经网络的滑动窗口算法在高分辨电镜图像中的应用

南 虎，路 璐，麻晓晶，侯振铎，张 辰，汪 洁，刘卫华，王大威^*

（1.西安交通大学，电子与信息学部微电子学院，陕西西安710049；2.中国科学院西安光学精密机械研究所，陕西西安710119；3.杭州电子科技大学，电子信息学院，浙江杭州310018）

摘  要 原子峰位置的精确定位、原子像绝对强度的标定是利用原子分辨电子显微像确定材料结构的重要前提。近年来快速发展的深度学习技术在计算机视觉和目标检测领域都取得了极大的成功。本文建立了针对单原子图像的卷积神经网络，并构建了含有大量单原子图像的训练数据集，用于这一卷积神经网络的训练，实现了对单原子图像目标的可靠检测。利用训练完成的卷积神经网络改进滑动窗口算法，可以发现这一方法在原子位置检测效果上取得了明显提升。

关键词 高分辨电镜图像；深度学习；卷积神经网络；滑动窗口算法

中图分类号：O766⁺.1；TP183；TP319  文献标识码：A   doi:10.3969/j.issn.1000-6281.2021.03.005

Application of sliding window algorithm with convolutional neural network in high-resolution electron microscope image

NAN Hu¹，LU Lu¹，MA Xiao-jing¹，HOU Zhen-duo¹，ZHANG Chen²，WANG Jie³，LIU Wei-hua¹，WANG Da-wei^1*

(1. School of Microelectronics, Faculty of Electronics and Information Engineering，Xi’an Jiaotong University，Xi’an Shaanxi 710049；2.Xi'an Institute of Optics and Precision Mechanics, Chinese Academy of Sciences, Shaanxi Xi'an 710119；3. School of Electronics and Information Engineering，Hangzhou Dianzi University，Hangzhou Zhejiang 310018，China)

Abstract   The precise detecting of atomic peak positions and the calibration of the atomic image intensities are important prerequisites for the investigation of material structures. Using computers to automatically complete these tasks can greatly accelerate the image analysis process, making it possible to quickly process many high-resolution electron microscope images. In recent years, the rapid development of deep learning technology has achieved great success in the field of computer vision and target detection. Here, a convolutional neural network for single-atom images is built, and trained with a training data set containing many single-atom images to achieve reliable detection of single-atom image targets. Using the trained convolutional neural network along with the sliding window algorithm, this method achieves a significant improvement in determining atomic positions.

Keywords  high-resolution electron microscopy image；deep learning；convolutional neural network；sliding window algorithm

“全文下载请到同方知网，万方数据库或重庆维普等数据库中下载！”

[1] HONKALA K，HELLMAN A，REMEDIAKIS I N，et al. Ammonia synthesis from first-principles calculations[J]. Science，2005，307(5709): 555-562.

[2] URBAN K W. Studying atomic structures by aberration-corrected transmission electron microscopy[J]. Science，2008，321(5888): 506-510.

[3] JIA C L，MI S B，BARTHEL J，et al.  Determination of the 3D shape of a nanoscale crystal with atomic resolution from a single image[J]. Nature Materials，2014，13(11)：1044-1049.

[4] 邹敏杰，唐云龙，冯燕朋，等. 铁电薄膜中180°荷电畴壁的亚埃尺度结构特性[J]. 电子显微学报，2018，37:468-473.

[5] 明文全，陈江华，沈若涵，等. 成像参数对ADF STEM三维重构成像精确性的影响[J].电子显微学报，2019，38(5)：502-511.

[6] GALINDO P L，KRET S，SANCHEZ A M，et al. The Peak Pairs algorithm for strain mapping from HRTEM images[J]. Ultramicroscopy，2007，107(12): 1186-1193.[S1]

[7] GONG Y J，LIU Z，LUPINI A R，et al.  Band gap engineering and layer-by-layer mapping of selenium-doped molybdenum disulfide[J]. Nano Letters，2014，14(2): 442-449.

[8] 南虎，卢江波，刘明，等.  高分辨原子像中原子峰位置的精确自动判定与电极化畴的快速可视化[J]. 电子显微学报，2016，35(3)：191-200.

[9] LU L，NAHAS Y，LIU M，et al. Topological defects with distinct dipole configurations in PbTiO₃/SrTiO₃ multilayer films[J].  Physical Review Letters，2018，120(17):177601.

[10] LECUN Y, BENGIO Y, HINTON G. Deep learning[J].  Nature, 2015，521(7553): 436-444.

[11] ABADI M，AGARWAL A，BARHAM P，et al. TensorFlow: Large-scale machine learning on heterogeneous systems，2015. Software available from tensorflow.org.