宇航学报 ›› 2021, Vol. 42 ›› Issue (11): 1439-1445.doi: 10.3873/j.issn.1000-1328.2021.11.010

• 制导、导航、控制与电子 • 上一篇    下一篇

暗弱环境下小天体陨石坑智能检测算法

邵巍,郗洪良,王光泽,肖扬,马广飞,姚文龙   

  1. 青岛科技大学自动化与电子工程学院,青岛 266061
  • 收稿日期:2020-12-28 修回日期:2021-03-23 出版日期:2021-11-15 发布日期:2021-11-15
  • 基金资助:
    国家重点研发计划(2019YFA0706500);国家自然科学基金(61773227,61971253)

An Intelligent Detection Algorithm for Small Body Craters in Faint Environment

SHAO Wei, XI Hong liang, WANG Guang ze, XIAO Yang, MA Guang fei, YAO Wen long   

  1. College of Automation and Electronic Engineering, Qingdao University of Science and Technology, Qingdao 266061, China
  • Received:2020-12-28 Revised:2021-03-23 Online:2021-11-15 Published:2021-11-15

摘要: 针对小天体环境暗弱、存在自旋情况下,陨石坑难以检测的问题,提出了一种陨石坑智能检测算法。通过局部方差均衡算法增强暗弱环境下图像中的陨石坑特征,然后利用多视角重投影模拟小天体自旋及相机视角变化,并进行数据增强。对于陨石坑漏标或标注错误问题,采用YOLOv4结合自学习的网络结构更新数据集标注信息,有效提高标注率。同时,对于小陨石坑漏检问题,将预测图像切分成若干个有重叠区域的子图像并合并预测结果。实验结果表明,该算法与现有主流检测网络相比具有更好的效果。


关键词: 深度学习, 小天体陨石坑检测, 多视角重投影, 数据增强, 自学习

Abstract: An intelligent crater detection algorithm is proposed for the problem that the craters are difficult to detect in the faint environment and the presence of spin of small bodies. The crater features in the image in the faint environment are enhanced by a local variance equalization algorithm. Then a multi view reprojection method is used to simulate the small bodies spin and camera view variation, and the data augmentation is performed. For the crater miss labeling or mislabeling problem, YOLOv4 combined with a self learning network structure is used to update the dataset labeling information and effectively improve the labeling rate. Meanwhile, for the small crater miss detection problem, the predicted image is sliced into several sub images with overlapping regions and the prediction results are combined. The experimental results show that the algorithm has better results compared with the existing mainstream detection networks.


Key words: Deep learning, Small body craters detection, Multi view reprojection, Data augmentation, Self learning

中图分类号: