Machine-learning models for analyzing TSOM images of nanostructures

Through-focus scanning optical microscopy (TSOM) is an economical and nondestructive method for measuring three-dimensional nanostructures. After obtaining a TSOM image, a library-matching method is typically used to interpret optical intensity information and determine the dimensions of a measurement target. To further improve dimensional measurement accuracy, this paper proposes a machine learning method that extracts texture information from TSOM images. The method extracts feature vectors of TSOM images in terms of the Gray-level Co-occurrence Matrix (GLCM), Local Binary Pattern (LBP), and Histogram of Oriented Gradient (HOG). We tested models trained with these vectors in isolation, in pairs, and a combination of all three to test seven possible feature vectors. Once normalized, these feature vectors were then used to train and test three machine-learning regression models: random forest, GBDT, and AdaBoost. Compared with the results of the library-matching method, the measurement accuracy of the machine learning method is considerably higher. When detecting dimensional features that fall into a wide range of sizes, the AdaBoost model used with the combined LBP and HOG feature vectors performs better than the others. For detecting dimensional features within a narrower range of sizes, the AdaBoost model combined with HOG feature extraction algorithm performs better.