|About this Abstract
||2020 TMS Annual Meeting & Exhibition
||Algorithm Development in Materials Science and Engineering
||L-1 (Digital): Machine Learning and Computer Vision on Classification of Carbon Nanotube and Nanofiber Structures for TEM Dataset
||Qixiang Luo, Elizabeth A. Holm
|On-Site Speaker (Planned)
We introduce a convolutional neural networks (CNNs) based machine learning and computer vision method to classify CNTs/CNFs and other airborne particles from a TEM image dataset. The model starts with a transfer learning pipeline achieved by hypercolumn extractions on CNN fully-connected layers from both precise edges at shallow and abstract semantics at deep, followed by K-means visual dictionary for creating feature clusters. It then uses VLAD encoder to mitigate the ambiguity between feature clusters and local feature descriptors based on image retrieval, with boosting step for classifiers. The unsupervised t-SNE method was used to visualize classification performance for data cluster analysis. The method achieves 85% accuracy on full dataset and over 90% on pure nano-carbon dataset for 5-fold cross-validation. We also discuss image pre-processing with data augmentation techniques for imbalanced dataset and hierarchical recognition, to utilize a tree-like learning structure to explore attributes from generalized to specific.
||Planned: Supplemental Proceedings volume