Effectiveness of Data Augmentation of SEM Images on a Small Database Based on Deep-Learning Intelligence

This work has been published in Braz. J. Phys, 52, 59 (2022) 

The NFFA Europe Projects in Nanoscience Foundries & Fine Analysis require a way to train scanning electron microscope (SEM) images for classification. However, the cost of preparing a large SEM dataset is prohibitively high. In order to assist the NFFA Projects, we are looking for a computational method that can classify SEM images using a small dataset.

Our goal is to identify the composition of nanowire-fiber-mixtures images, specifically focusing on optimizing the performance of image classification between nanowires, fibers, and tips due to their geometric similarities. To achieve this, we employ deep-learning techniques and compare the validation accuracies of 11 convolutional neural network (CNN) models. By increasing the diversity of data through approaches such as reflection, translation, and scale factor variations, we achieve the highest validation accuracy of 97.1% in recognizing nanowires, fibers, and tips.

Furthermore, we extend our classification capabilities to determine the level of porosity in anodized aluminum oxide for self-assisted nanowire growth. We are able to optimize the validation accuracy at 93% for this task. Our software provides scientists with a tool to accurately count the percentage of fibers in any nanowire-fiber composite and automatically design a porous substrate for embedding nanowires of different sizes.

In summary, our software offers a solution for training SEM images using a small dataset, enabling scientists to classify nanowires, fibers, and tips with high accuracy, as well as determine the level of porosity in anodized aluminum oxide for nanowire growth.

Fig 1: Classifying various types of nanowires with high similarity

Fig 2: Estimate the porosity level