npj: 机器学习添视觉—材料缺陷快分析

2018 年 8 月 18 日 知社学术圈

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

电子显微镜已广泛应用于各种材料的晶界、杂质等缺陷的分析研究，可为其提供微观结构和性能的详细信息，因而成为材料科学研究的基石。然而，这需要大量的图像才能提取出统计上显著的信息，并需手动完成，既耗时又会因人而异导致结果不一致。若为电子显微镜创建一个强大而灵活的缺陷自动识别和分类平台，将可在图像记录后甚至在图像采集过程中，快速地完成分析任务。

现在，美国威斯康星大学和橡树岭国家实验室的研究团队将机器学习、计算机视觉和图像分析技术相结合，开发了一种自动识别工具；依次应用了级联对象检测器、卷积神经网络和局部图像分析方法，以获得有关缺陷尺寸和缺陷类型的信息。与人工分析相比，自动分析有可能显著提高分析的效率、准确性和可重复性，还可通过日渐显要的自动数据生成方法进行扩展。他们的结果证明，该自动化工具在回溯和精确度方面的表现与人工手动水平相当或更好，在图像/缺陷定量分析指标方面与人类平均水平接近，适用于不同对比度、不同亮度和不同放大倍数的图像。他们的设计有望检测多种缺陷类型，并可能定位、分类、定量测量一系列缺陷类型、多种材料和各种电子显微技术的特征，也值得进一步改进，以便对大数据集作实时分析。

该文近期发表于npj Computational Materials 4:36 (2018)，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。

Automated defect analysis in electron microscopic images

Wei Li, Kevin G. Field & Dane Morgan

Electron microscopy and defect analysis are a cornerstone of materials science, as they offer detailed insights on the microstructure and performance of a wide range of materials and material systems. Building a robust and flexible platform for automated defect recognition and classification in electron microscopy will result in the completion of analysis orders of magnitude faster after images are recorded, or even online during image acquisition. Automated analysis has the potential to be significantly more efficient, accurate, and repeatable than human analysis, and it can scale with the increasingly important methods of automated data generation. Herein, an automated recognition tool is developed based on a computer vison–based approach; it sequentially applies a cascade object detector, convolutional neural network, and local image analysis methods. We demonstrate that the automated tool performs as well as or better than manual human detection in terms of recall and precision and achieves quantitative image/defect analysis metrics close to the human average. The proposed approach works for images of varying contrast, brightness, and magnification. These promising results suggest that this and similar approaches are worth exploring for detecting multiple defect types and have the potential to locate, classify, and measure quantitative features for a range of defect types, materials, and electron microscopic techniques.