Deep Learning-Based Classification of Nanoscale SEM Features Using Inception

Troughs of mark are well known for the role it plays on scanning electron microscope images (SEM) that are used for imaging, which allows users to study the morphological and structural characteristics of several nanomaterials. Regardless of its significance, the classification of images taken from a scanning electron microscope SEM is often subjective and manual, making it tedious and time-consuming. In this research, we propose an automated classification of SEM images on a nanoscale using Inception V3, which is a deep learning model part of convolutional neural networks (CNN). A set of 10 distinct classes was provided, which included particles, nanowires, porous sponges, MEMS devices, and coated surfaces, all containing features at the nanoscale. The dataset was prepared with increased variability by augmenting, resizing, and normalizing the features to improve generalization of the model. The InceptionV3 model was trained with Adam optimizer using categorical cross-entropy loss function to enhance the model precision. The model showed remarkable class prediction capabilities with a testing accuracy of 90.18%. Analysis with the confusion matrix and report classification revealed strong precision and recall scores in most categories save for some, where slight misclassification of some poorly defined objects with similar structure took place. The findings indicated the use of deep learning could greatly increase the automation in classifying images from SEM, streamlining the workflow during nanomaterials characterization in both scientific and industrial fields. Further development will focus on dataset resolution, self-supervised learning methods, and explainability to achieve better performance of the model.