This paper is dedicated to the topic of terrain recognition on Mars using advanced techniques based on the convolutional neural networks (CNN). The work on the project was conducted based on the set of 18K images collected by the Curiosity, Opportunity and Spirit rovers. The data were later processed by the model operating in a Python environment, utilizing Keras and Tensorflow repositories. The model benefits from the pretrained backbones trained for analysis of the RGB images. The project achieves an accuracy of 83.5% when extending the scope of classification to unknown objects and 94.2% when omitting unknown results. The results were compared with related projects of Zooniverse and NASA's Jet Propulsion Laboratory scientific group. From amongst the evaluated configurations, the best results and resource utilization were achieved by applying the UNet architecture with resnext_50 backbone and Adam optimizer.
Authors
Additional information
- DOI
- Digital Object Identifier link open in new tab 10.62036/isd.2024
- Category
- Aktywność konferencyjna
- Type
- publikacja w wydawnictwie zbiorowym recenzowanym (także w materiałach konferencyjnych)
- Language
- angielski
- Publication year
- 2024
