Convolutional neural networks applied to classification of nanoparticles and nanotubes images

Abstract

Nanotechnology is a promising research area involving the manipulation of nanomaterials and it has a wide variety of applications. To study these nanomaterials, researchers often use scanning electron microscopes (SEM) to generate and store images. However, different users tend to label the images using subjective criteria, which generates a large amount of data that is difficult to manage. To address this problem, we propose a convolutional neural network for image classification of nanomaterials. This work compares two convolutional neural networks, a custom network and a VGG16 network, and implements grid search to determine the best parameters for both networks. The results show that the custom network is more effective than VGG16 network, with a mean accuracy of 77.1% and F1 score of 76% versus 66.6% and 0.56%, respectively. Based on these results, we recommend using a custom network for image classification of nanomaterials with the following parameters: learning rate of 0.001, minibatch of 16, number of epochs 20, number of layers 3, and rmsprop optimizer.