1. AiTARs-Net: A novel network for detecting arbitrary-oriented transverse aeolian ridges from Tianwen-1 HiRIC images.
- Author
-
Cao, Zhen, Kang, Zhizhong, Hu, Teng, Yang, Ze, Chen, Dong, Ren, Xiaolan, Meng, Qingyu, and Wang, Dong
- Subjects
- *
CONVOLUTIONAL neural networks , *GAUSSIAN function - Abstract
Transverse aeolian ridges (TARs) are enigmatic landforms found across Mars, whose formation mechanism remains largely unknown. China's Tianwen-1 mission, which landed on Mars in 2021, provided extensive data aiding in-depth investigations of TARs. However, manually identifying TARs across large regions of Mars is time-consuming and labor-intensive, making it impossible to complete the entire region's TAR identification. To solve this issue, we propose the AiTARs-Net, an automatic arbitrary-oriented TRA detection network. This model begins by extracting TAR features using an enhanced dimension-aware global-local attention module, which focuses on interactions between spatial and channel features to capture discriminative features of TARs. After that, we employ an anchor-free proposal generation network to produce TAR candidates with arbitrary orientations. The proposal generation network uses a nonaxis-aligned two-variable Gaussian function to model the target as an oriented center heatmap. Then, the oriented bounding box and category information are predicted at the corresponding center position. Finally, we introduce the rotated region-based convolutional neural network to refine the proposals to obtain more accurate TARs' locations and orientations. To assess the efficacy of our proposed method, we built the Martian TARs dataset (M-TARset), an compilation of TARs labeled in six different topographical and morphological types, containing various shapes and illumination scales, to facilitate training and prediction of potential TARs. The experimental results obtained on a Martian TARs dataset and a large-scale TARs extraction at the Zhurong landing site confirm that the proposed framework outperforms the leading generic object extraction methods in accuracy, demonstrating its strong generalization abilities for large-scale TAR detection. The source code and M-TARset are available at https://github.com/PlanetaryScience3510/M-TARset. [Display omitted] • We proposed a novel AiTARs-Net framework for large-scale TARs automatic extraction. • We proposed enhanced global-local feature extraction module for capture TAR features. • We introduced anchor-free network for generating TAR proposal. • We created a Martian TARs dataset to train and test TAR detection model. [ABSTRACT FROM AUTHOR]
- Published
- 2024
- Full Text
- View/download PDF