Deep learning for nanoscience scanning electron microscope image classification
TL;DR
A multi-class classification framework is proposed for SEM images of nanostructures, categorized into nanowires, fibers, and tips, and the experimental results demonstrate that both SMOTE strategies effectively address class imbalance, while the pixel-level approach achieved the highest classification performance.
Abstract
Materials science investigates the relationships between a material’s structure, properties, and fabrication processes. Recently, artificial intelligence (AI) and deep learning (DL) techniques have significantly enhanced the ability to analyze complex visual data, particularly in scanning electron microscopy (SEM) images. These approaches enable the detection of subtle morphological patterns that may not be easily identified through manual inspection, facilitating more accurate characterization of nanomaterials. In this study, a multi-class classification framework is proposed for SEM images of nanostructures, categorized into nanowires, fibers, and tips (NFT). To address the issue of class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was incorporated using two complementary strategies: pixel-level and feature-level representations. In the pixel-level approach, synthetic samples were generated directly from flattened image data, while in the feature-level approach, deep features were first extracted using a pre-trained ResNet50 model before applying SMOTE. For classification, transfer learning was employed using three convolutional neural network architectures: SqueezeNet, ShuffleNet, and GoogLeNet. In addition, a Multi-Layer Perceptron (MLP) classifier was used for feature-level representations. The experimental results demonstrate that both SMOTE strategies effectively address class imbalance, while the pixel-level approach achieved the highest classification performance. The pixel-level approach achieved a classification accuracy of up to 98.35% using GoogLeNet and ShuffleNet, whereas the feature-level approach achieved an accuracy of 97.44% while offering a computationally efficient alternative. These findings highlight the effectiveness of combining SMOTE with transfer learning for handling imbalanced SEM datasets and illustrate the trade-off between classification performance and computational efficiency.