Multi-scale feature pyramid-based crater detection on lunar surface

Impact craters, prominent geomorphological aspects of planetary surfaces, provide valuable insights into terrain development and the evolutionary timeline of the solar system. However, detecting Crater objects is exciting but demanding in lunar surface exploration. At the same time, the recent popular object recognition techniques perform well overall. As lunar images are disturbed by illumination conditions and the object's scale varies depending on the craters' size, the performance of detecting lunar objects is imperfect. Therefore, this study designed a multi-scale feature structure that effectively integrates information at all levels with the help of ResNet-50 as the framework, with an augmented feature pyramid network (Aug-FPN) combined with a consistent super-vision module. Also, this study used a residual feature augmentation module to boost the extraction of unchangeable proportional contextual data and reduced data loss at the uppermost level of the feature map in the pyramid network. Finally, Soft-RoI selection is applied to acquire more effective RoI features from various pyramid levels and generate an improved RoI feature following position enhancement and classification. The study evaluated the proposed model using standard metrics on the fusion of a global dataset, which gives precision, recall and F1-Score nearly equal to 92%, 89.2%, and 90.5%, respectively, compared to other object detector models. The proposed model efficiently enhanced the recognition accuracy of multi-scale objects for craters, such as varying scales of craters on the Moon's surface in between [0.1-5] km as small, nearly greater than [5-10] km as medium, and > 10 km as large craters. The results found from the ablation study demonstrated that the proposed method greatly enhanced the detection accuracy with the low false positive rate (FPR) with the scale mentioned, particularly in complex scenarios such as noisy, overlapping, and illumination conditions without post-processing techniques.