SCI-Mamba: Unsupervised AI Clears Up Low-Light Spacecraft Images

A team of researchers (Yiyong Sun et al.) has introduced SCI-Mamba, an unsupervised low-light image enhancement framework designed specifically for non-cooperative spacecraft visual perception. The method addresses a critical bottleneck in on-orbit servicing missions: spaceborne cameras suffer severe low-light degradation, yet paired normal/low-light space samples are extremely scarce, limiting supervised approaches. SCI-Mamba unites self-calibrated unsupervised learning, a linear-complexity VMamba architecture (instead of quadratic self-attention), and Retinex-based physical priors. The result is a lightweight pipeline suitable for resource-constrained space hardware, capable of enhancing images without any ground-truth pairs.

To train and evaluate the model, the authors constructed Space Dark-1.0, a dedicated low-light spacecraft dataset combining real orbital footage, darkroom hardware-in-the-loop measurements, and physically constrained synthetic data covering diverse illumination, motion, and attitude conditions. Comprehensive comparisons against CNN, Transformer, and other Mamba-based approaches demonstrate SCI-Mamba's advantages in visual authenticity, color fidelity, and inference speed. The code has been released on GitHub. This work provides a practical solution for close-proximity space operations like autonomous rendezvous, component detection, and robotic capture, where reliable vision in darkness is essential.