Detection of nano particles in TEM images using an ensemble learning algorithm

Abstract

Transmission electron microscopy (TEM) has an ability to depict material structures on nanoscales (~0.1 nm). High resolution TEM has found applications in a wide range of domains such as the studies of biological tissues, reactive chemical compounds and product defect inspection. For the past decade, Nano-research has generated a large number of TEM images, each containing immense amount of information that cannot be processed and interpreted manually. The combination of image processing and big data mining becomes the only viable solution. This thesis investigates the feasibility of using a Cascade AdaBoost algorithm to detect and count nanoparticles automatically. Experiments with cube-shaped objects have yielded very promising results with high detection rate (true positive rate) and low false alarm rate (false positive rate). The impacts of labeling variation, sample size and feature size on the detection accuracy were also discussed.