宇航学报

宇航学报 ›› 2022, Vol. 43 ›› Issue (6): 830-838.doi: 10.3873/j.issn.1000-1328.2022.06.014

• 空间科学 • 上一篇    

月表陨坑检测轻量化深度学习方法

高艾，周永军，王俊伟，兀泽朝   

  1. 1. 北京理工大学宇航学院，北京 100081；2. 深空自主导航与控制工业和信息化部重点实验室，北京 100081；3. 飞行器动力学与控制教育部重点实验室，北京 100081
  • 收稿日期:2021-12-21 修回日期:2022-01-14 出版日期:2022-06-15 发布日期:2022-06-15
  • 基金资助:
    国家自然科学基金（11872110）

Lightweight Deep Learning Method for Lunar Surface Crater Detection

GAO Ai, ZHOU Yongjun, WANG Junwei, WU Zezhao   

  1. 1. School of Aerospace Engineering，Beijing Institute of Technology，Beijing 100081, China; 2. Key Laboratory of Autonomous Navigation and Control for Deep Space Exploration，Ministry of Industry and I nformation Technology，Beijing 100081, China;3. Key Laboratory of Dynamics and Control of Flight Vehicle, Ministry of Education, Beijing 100081, China
  • Received:2021-12-21 Revised:2022-01-14 Online:2022-06-15 Published:2022-06-15

PDF

145

摘要： 针对目前基于深度学习的陨坑检测方法存在的模型参数量大和检测速度慢的问题，提出了一种轻量化的深度学习陨坑检测方法。首先，采用通道剪枝方法删减卷积神经网络中冗余的卷积核，得到结构紧凑高效的陨坑检测模型。然后，使用轻量化的深度可分离卷积操作替换基础陨坑检测模型中的标准卷积操作，进一步降低了模型的复杂度。仿真实验结果表明，所提出的轻量化陨坑检测模型能够保证较高的像素预测精度，并且能够适应亮度、图像噪声等干扰因素的影响。同时，与轻量化处理前的模型相比，参数量减少了99.2%，检测速度提升了94%。

关键词: 月球着陆探测, 陨坑检测, 深度学习, 卷积神经网络, 轻量化处理

Abstract: A lightweight deep learning crater detection method is proposed to address the problems of large number of model parameters and slow detection of the current deep learning crater detection methods. Firstly, the channel pruning method is used to delete the redundant convolution kernel in convolution neural network to obtain a compact and efficient crater detection model. Then, the lightweight depthwise separable convolution operation is used to replace the standard convolution operation in the basic crater detection model, which further reduces the complexity of the model. The simulation results show that the proposed lightweight crater detection model can ensure high pixel prediction accuracy, and can adapt to the influence of interference factors such as brightness and image noise. Moreover, compared with the model before lightweight processing, the amount of parameters is reduced by 99.2% and the detection speed is improved by 94%.

Key words: Lunar landing exploration, Crater detection, Deep learning, Convolutional neural networks, Lightweight processing

中图分类号: