A New Convolutional Neural Network Architecture for Mars Landmarks Classification Based on MobileNetV2

Convolutional neural networks have achieved outstanding state-of-the-art performance in many mainstream application scenarios. However, its application in niche but critical fields such as planetary science requires more exploration. Among those specific application scenarios, the Mars landmark clas...

Full description

Saved in:
Bibliographic Details
Published in2023 IEEE 3rd International Conference on Software Engineering and Artificial Intelligence (SEAI) pp. 68 - 74
Main Authors Xia, Runpeng, Du, Sirui, Liu, Jiahang, Li, Chenjia
Format Conference Proceeding
LanguageEnglish
Published IEEE 16.06.2023
Subjects
Architecture
Computer architecture
deep learning
image classification
Mars
mars landmarks
Network architecture
neural networks
Smoothing methods
Space vehicles
Transfer learning
Online AccessGet full text

Cover

More Information
Summary:Convolutional neural networks have achieved outstanding state-of-the-art performance in many mainstream application scenarios. However, its application in niche but critical fields such as planetary science requires more exploration. Among those specific application scenarios, the Mars landmark classification is one of the ongoing tasks. While, due to issues with some in-use neural network architectures such as inadequate accuracy and excessively large size, this task has yet to be effectively fulfilled. To address these challenges, we proposed a new network architecture based on MobileNetV2 by introducing non-local operations and DropBlock. Also, with the help of label smoothing and transfer learning, the accuracy of our new proposed network architecture can achieve 93.64%, which outperforms several in-use models including VGG-16, original MobileNetV2, ResNet-34 and HiRISENet.
DOI:10.1109/SEAI59139.2023.10217391