Abstract View
Comparing Multiple Types of Machine Learning for Characterizing TEM Images of Soot
TIMOTHY SIPKENS, Hamed Nikookar, Max Frei, Frank Einar Kruis, Steven Rogak, University of British Columbia
Abstract Number: 452
Working Group: Instrumentation and Methods
Abstract
TEM imaging is a mainstay in the characterization of aerosols, like soot. One of the largest barriers to population-wide statistics is the often-labor-intensive manual approach to image analysis. This work examines a series of machine learning approaches to distinguishing soot from the background in TEM images, including (1) unsupervised k-means clustering, (2) a convolutional neural network, and (3) trainable Weka segmentation, via the Fiji image analysis package. Classifications are compared to a largely-manual, in-house method where a user-specified threshold is applied in local regions of the image. Methods show a marked improvement over existing classifiers across a range of images, with over 99% classified pixel accuracy and only a small number of falsely identified aggregates. We also present sensitivities of a range of automated primary particle sizing techniques to the chosen classifier, which is a step towards fully automated analysis of soot TEM images. We conclude by examining the relationship between primary particle size and projected-area equivalent diameter of soot.