Performance Comparison of Data Classification based on Modern Convolutional Neural Network Architectures

Abstract

In recent years, convolutional neural network (CNN) has become the most powerful tool to deal with some computer vision applications, for example, classification, object detection, semantic segmentation, and so on. This benefits from the development of computer hardware and theory of deep learning. Many state-of-the-art CNN architectures have been proposed and they produce better performance than previous conventional methods. However, the effects of these modern CNN architectures are usually demonstrated based on some public datasets. The performance of these architectures on unpublished datasets is still not clear. To solve this problem, we compare classification performance of six different CNN architectures based on our own Lead isotope dataset. Six architectures are ZFNet, VGG16, ResNet18, ResNet101, GoogLeNet, DenseNet, respectively. Through experiments, we find that all CNN architectures can produce satisfactory classification results, and the performance of some architectures in our dataset is different from that in some public datasets.