Technical note: Using k-means to identify soot aggregates in transmission electron microscopy images. (February 2021)

• Record Type: Journal Article
• Authors: Sipkens, T.A.; Rogak, S.N.
• Is Part Of: Journal of aerosol science. Volume 152(2021)
• Publication Date: 2021-02
• Subjects: TEM -- Soot -- Aggregate morphology -- Particle size distribution k-means -- Machine learning; Aerosols -- Periodicals; Aerosols -- Periodicals; Aérosols -- Périodiques; 541.34515
• DOI: 10.1016/j.jaerosci.2020.105699
• Languages: English
• ISSNs: 0021-8502

Abstract

Abstract: This note describes a new approach to identifying soot aggregates in transmission electron microscopy images using k-means clustering. As k-means generally performs poorly for images with noise or gradients, segmenting the image first requires image pre- and post-processing methods, including background removal, denoising, a measure of image texture, and an adjusted threshold binary image. Three pre-processed versions of the image are compiled into feature layers prior to segmentation using k-means. A rolling ball transform is applied post hoc to improve the accuracy of the resultant segmentations. Results are compared to those using a largely-manual method, wherein the threshold is adjusted with a slider in a GUI; automatic Otsu thresholding with a rolling ball transformation; and trainable WEKA segmentation via Fiji. Graphical abstract: Image 1 Highlights: We present a new k -means classifier for segmenting TEM images of soot. The classifier uses three pre-processed feature layers as input to k -means. Results are compared to a manual classification, as well as Otsu and Fiji outputs. The method is shown to be reasonably robust on a set of over 200 images.