Edge detection system and its use for machine learning

TLDR: In this paper, the authors proposed a method to remove noise from roughness measurements to determine roughness of a feature in a pattern structure using an inverse linescan model, and evaluated a high-frequency portion of the biased power spectral density (PSD) dataset to determine a noise model for predicting noise over all frequencies, and subtracting the noise predicted by the determined noise model from a biased roughness measure.

Abstract: Systems and methods are disclosed that remove noise from roughness measurements to determine roughness of a feature in a pattern structure. In one embodiment, a method includes generating, using an imaging device, a set of one or more images, each including an instance of a feature within a respective pattern structure. The method also includes detecting edges of the features within the pattern structure of each image using an inverse linescan model, generating a biased power spectral density (PSD) dataset representing feature geometry information corresponding to the edge detection measurements, evaluating a high-frequency portion of the biased PSD dataset to determine a noise model for predicting noise over all frequencies of the biased PSD dataset, and subtracting the noise predicted by the determined noise model from a biased roughness measure to obtain an unbiased roughness measure provided as part of a training data set to a machine learning model.