Spacecraft Components Detection Based on a Lightweight YOLOv3 Model

Man Yuan,Guhong Zhang,Zhuoqun Yu,Yufan Wu,Zhonghe Jin

DOI: https://doi.org/10.1109/itaic54216.2022.9836634

2022-01-01

Abstract:With the rapid development of the global aerospace industry, vision-based space target detection has become one of the hotspots of space technology. Traditional detection algorithms are mainly based on handcrafted features which have poor generalization performance. While algorithms based on deep learning, especially CNN models, can be assumed as a candidate to deal with this problem. However, it is challenging to directly deploy large neural networks on embedded devices of spacecraft due to the limited capacity. To address this problem, a lightweight YOLOv3 network structure is proposed in this paper, which reduces the number of parameters of the original by a factor of 42×, from 61M to 1.4M. And then Knowledge Distillation strategy could be used to improve the performance of the lightweight model when its accuracy drops a lot. In addition, because of the shortage of publicly available spacecraft datasets, a simulated large-scale dataset for spacecraft components detection based on a virtual environment is constructed. The experimental results show that, compared with the baseline model, the lightweight YOLOv3 model performs well on our dataset which has a smaller size and higher speed. So the lightweight model we proposed is more suitable for intelligent applications of spacecraft embedded devices.