Statistical analysis of helium bubbles in transmission electron microscopy images based on machine learning method
Statistical analysis of helium bubbles in transmission electron microscopy images based on machine learning method
复制标题
DOI：
10.1007/s41365-021-00886-y
复制
发表时间：
2021-05
影响因子：
2.8
通讯作者：
Zhong-Hang Wu;Juju Bai;Di-Da Zhang;Gang Huang;Tian-Bao Zhu;Xi-Jiang Chang;Ren-Duo Liu;Jun Lin;Jiu-Ai Sun
中科院分区：
文献类型：
--
作者：
Zhong-Hang Wu;Juju Bai;Di-Da Zhang;Gang Huang;Tian-Bao Zhu;Xi-Jiang Chang;Ren-Duo Liu;Jun Lin;Jiu-Ai Sun
氦泡是在金属或合金中观察到的典型辐射微结构，通常使用透射电子显微镜（TEM）进行研究。然而，调查需要人工输入来定位和标记所获取的TEM图像中的气泡，使得这项任务费力且容易出错。本文提出了一种能够自动识别和分析氦气泡TEM图像的机器学习方法，从而提高了研究的效率和可靠性。在该方法中，氦气泡簇首先通过基于密度的空间聚类的应用程序与噪声算法确定后，去除背景和噪声像素。对于每个氦气泡簇，氦气泡的数量基于取决于特定图像分辨率的簇大小来确定。最后，使用高斯混合模型对氦泡团进行了分析，得到了氦泡的位置和大小信息。与需要使用大量注释图像进行训练以建立准确分类器的其他方法相比，使用少量TEM图像来确定所建立模型中使用的参数。通过人工标定的氦泡图像验证了该方法的有效性，获得了较高的F1值。此外，所建立的模型可以识别人类无法轻易识别的气泡状物体。这种计算效率高的方法实现了对材料结构识别的对象识别，这可能有利于科学工作。
Helium bubbles, which are typical radiation microstructures observed in metals or alloys, are usually investigated using transmission electron microscopy (TEM). However, the investigation requires human inputs to locate and mark the bubbles in the acquired TEM images, rendering this task laborious and prone to error. In this paper, a machine learning method capable of automatically identifying and analyzing TEM images of helium bubbles is proposed, thereby improving the efficiency and reliability of the investigation. In the proposed technique, helium bubble clusters are first determined via the density-based spatial clustering of applications with noise algorithm after removing the background and noise pixels. For each helium bubble cluster, the number of helium bubbles is determined based on the cluster size depending on the specific image resolution. Finally, the helium bubble clusters are analyzed using a Gaussian mixture model, yielding the location and size information on the helium bubbles. In contrast to other approaches that require training using numerous annotated images to establish an accurate classifier, the parameters used in the established model are determined using a small number of TEM images. The results of the model formulated according to the proposed approach achieved a higherF1 score validated through some helium bubble images manually marked. Furthermore, the established model can identify bubble-like objects that humans cannot facilely identify. This computationally efficient method achieves object recognition for material structure identification that may be advantageous to scientific work.