School of Instrumentation and Opto-Electronic Engineering, Beihang University, Beijing, China.
Key Laboratory of Precision Opto-Mechatronics, Technology of Education Ministry, Beihang University, Beijing, China.
J Microsc. 2021 Aug;283(2):117-126. doi: 10.1111/jmi.13013. Epub 2021 Apr 27.
Through-focus scanning optical microscopy (TSOM) is an economical, non-contact and nondestructive method for rapid measurement of three-dimensional nanostructures. There are two methods using TSOM image to measure the dimensions of one sample, including the library-matching method and the machine-learning regression method. The first has the defects of small measurement range and strict environmental requirements; the other has the disadvantages of feature extraction method greatly influenced by human subjectivity and low measurement accuracy. To solve the problems above, a TSOM dimensional measurement method based on deep-learning classification model is proposed. TSOM images are used to train the ResNet50 and DenseNet121 classification model respectively in this paper, and the test images are used to test the model, the classification result of which is taken as the measurement value. The test results showed that with the number of training linewidths increasing, the mean square error (MSE) of the test images is 21.05 nm² for DenseNet121 model and 31.84 nm² for ResNet50 model, both far lower than machine-learning regression method, and the measurement accuracy is significantly improved. The feasibility of using deep-learning classification model, instead of machine-learning regression model, for dimensional measurement is verified, providing a theoretical basis for further improvement on the accuracy of dimensional measurement.
基于聚焦扫描光学显微镜（TSOM）的三维形貌测量技术具有经济、非接触和无损等优点。目前，基于 TSOM 图像的三维形貌测量方法主要有两种，分别为基于模板匹配的方法和基于机器学习回归的方法。基于模板匹配的方法存在测量范围小、环境要求苛刻等问题；基于机器学习回归的方法存在特征提取方法易受人为因素影响、测量精度低等问题。针对以上问题，本文提出了一种基于深度学习分类模型的 TSOM 三维形貌测量方法。该方法分别使用 ResNet50 和 DenseNet121 分类模型对 TSOM 图像进行训练，并用测试图像对模型进行测试，将模型的分类结果作为测量值。实验结果表明，随着训练线宽数量的增加，DenseNet121 模型和 ResNet50 模型的测试图像的均方误差（MSE）分别为 21.05nm²和 31.84nm²，均远低于机器学习回归方法，测量精度得到了显著提高。验证了使用深度学习分类模型代替机器学习回归模型进行尺寸测量的可行性，为进一步提高尺寸测量精度提供了理论依据。
