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

The software development in Nanoscience Foundries & Fine Analysis (NFFA) Europe Projects needs to train numerous scanning electron microscope (SEM) images. However, the price of preparing a large SEM dataset is always very high. To assist the NFFA Projects, we search for a suitable computational method to classify the SEM images on a small dataset. To prepare for identifying the composition of nanowire-fiber-mixtures images, we optimize the performance of image classification between nanowires, fibers, and tips due to their geometric similarities. The SEM images are analyzed by deep-learning techniques where the validation accuracies of 11 convolutional neural network (CNN) models are compared. By increasing the diversity of data such as reflection, translation, and scale factor approaches, the highest validation accuracy of recognizing nanowires, fibers, and tips is 97.1%. We proceed to classify the level of porosity in anodized aluminum oxide for the self-assisted nanowire growth where the validation accuracy can be optimized at 93%. Our software presents a path for scientists to count the percentage of fibers in any nanowire-fiber composite and design the porous substrate for embedding different sizes of nanowires automatically.