.png){kind=icon}
.png){kind=icon}
Return
Spacecraft component recognition based on frequency-spatial awareness and comparative auxiliary training strategy
.png){kind=icon}
PRE
AI
DOI：10.1016/j.asr.2025.09.077.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Spacecraft components recognition is a crucial technology which is of great significance for on-orbit maintenance, space docking, space target monitoring and other on-orbit missions. However, the adverse imaging conditions in space make it difficult to recognize spacecraft components. Especially under uneven lighting and shadow occlusion conditions, existing methods often suffer from missed detections, false detections, and even failures. To tackle these problems, we propose two novel methods under the existing Encoder-Decoder segmentation architecture, namely multistage frequency-space dual domain attention(MFSA) module and comparative auxiliary training strategy(CATS). The MFSA module comprises two branches, the frequency-domain branch adaptively retains high-frequency information and enhances low-frequency information according to the stage of network which enables the network to achieve improved segmentation efficiency with more refined boundary segmentation results. The spatial-domain branch effectively suppresses irrelevant background. CATS introduces an additional branch during the training stage to input images with random contrast and brightness transformations, and guides the network to alleviate the effects of lighting changes through auxiliary training loss. Furthermore, we construct a spacecraft component semantic segmentation dataset(SRSD) based on laboratory-acquired and ray-tracing algorithms which includes the measured data collected by our designed data acquisition platform considers various lighting conditions and the simulated data generated by ray-tracing considers the orbit relationship between satellites. Meanwhile, We test the proposed methods on the SRSD and a large-scale public spacecraft component segmentation dataset(UESD). Experimental results indicate that our proposed methods achieve excellent performance and could maintain a certain performance even under harsh illumination conditions. The partial source code and dataset will be available at https://github.com/yemu1138178251/SRSD_dataset.
Journal
IF:
2.8
Papers:
1.3K
Citations:
2.0W