1. Transfer learning for real-time crater detection on asteroids using a Fully Convolutional Neural Network.
- Author
-
Latorre, F., Spiller, D., Sasidharan, S.T., Basheer, S., and Curti, F.
- Subjects
- *
CONVOLUTIONAL neural networks , *ASTEROID detection , *METEORITE craters , *LUNAR craters , *ASTEROIDS , *DIGITAL elevation models , *DATA augmentation - Abstract
This paper proposes the use of Transfer Learning from the Moon to Ceres for real-time autonomous crater detection operations on asteroids. This approach, based on the use of a U-Net for crater segmentation, starts with the network training on a Moon Digital Elevation Model; the trained model is then refined using two different strategies, one based on data augmentation and one based on fine tuning with small sets of Ceres images. Data from the DAWN mission was used, while the reference crater catalogue for Ceres is the one created by Zeilnhofer from LAMO data, containing 44,594 craters with diameter greater than 1 km in a latitude range of 84.66°S–89.62°N, full longitude. The loss function chosen during network training is the Focal Tversky Loss, which takes into consideration the imbalanced distribution of class pixels in the ground truth (pixels corresponding to crater rims represent approximately only 3% over the whole dataset). For both the Moon and Ceres, the U-Net output is post-processed with a template matching algorithm. The detected craters are counted and compared to the crater catalogue locations in order to compute metrics performances (precision, recall, F1 score) in the crater domain. Post-processing on the Moon reached performance of 84.12% F 1 score on the baseline model. The baseline and augmented models gave a lower performance on Ceres data, although comparable with the outcomes found in other works, with a best F 1 score of 70.17%. However, when the model is completely retrained starting from the baseline, it ends up with a best F 1 score of 79.96%. Finally, a strategy for the actual on-board implementation is proposed considering the NVIDIA Jetson TX2 accelerator. We demonstrated that the inference time of the model on the accelerator is on the order of few milliseconds without loss of performances, thus proving the possibility to consider this approach for future real-time on-board navigation systems. • U-Net training on Moon DEM data and TL on Ceres for autonomous crater detection. • Loss-based imbalanced class training for segmentation performance optimization. • Application of U-Net for asteroid crater detection for the first time. • On-board algorithm implementation in real-time on the Nvidia GPU TX2. [ABSTRACT FROM AUTHOR]
- Published
- 2023
- Full Text
- View/download PDF