Sate-DDRNet: A Lightweight Dual-Branch Semantic Segmentation Network for Spacecraft Inspection

Deep learning segmentation methods have shown large application potential in space missions such as Active Debris Removal (ADR) and On-Orbit Servicing (OOS). However, the impressive performance of contemporary models typically comes with the trade-off of high computational complexity and extended inference latency, making them impractical for spacecraft inspection missions. Meanwhile, there is currently little research on semantic segmentation of spacecraft components. Therefore, this paper focuses on the design of light-weight semantic segmentation (LWSS) algorithms, proposes an efficient method, which introduces a new pyramid pooling module (PPM) with an attention mechanism. Furthermore, we comprehensively evaluated 12 modern LWSS algorithms with 4 metrics on the Adelaide dataset. Our method achieves a state-of-the-art balance of accuracy (76.58% mIoU) and inference speed (79.2 FPS). Requiring only 1/10 of the computational complexity of ordinary algorithms with the same level of precision.