To address the failure of conventional 3D reconstruction methods in complex space environments characterized by weak textures, strong symmetry, and highly reflective surfaces, this paper proposes a lightweight framework integrating adaptive deep feature learning and robust geometric matching. First, a novel low-light enhancement algorithm is developed by combining Fast Guided Filtering with an adaptive gamma correction strategy. This algorithm dynamically balances illumination distribution and suppresses overexposed regions through pixel-wise intensity adjustment, significantly improving the signal-to-noise ratio of low-light images. Subsequently, the ALIKED network is employed to detect keypoints and generate descriptors. By integrating deformable convolutions and sparse sampling strategies, ALIKED enhances geometric invariance under complex surface deformations while reducing computational redundancy. Local features are further aggregated into global descriptors via NetVLAD, which clusters and encodes residuals to strengthen spatial context awareness. Furthermore, the LightGlue network leverages rotary positional encoding to embed geometric relationships and employs adaptive attention layers with dynamic computation mechanisms. This design enables efficient bidirectional cross-attention matching, effectively resolving ambiguities caused by repetitive structures and illumination variations. Finally, an incremental Structure-from-Motion pipeline is adopted to fuse optimized matches into dense point clouds through bundle adjustment and multi-view stereo. The experimental results demonstrate that the proposed algorithm achieves a point cloud coverage rate of 90.5% in satellite simulation model reconstruction, exhibiting significant improvements over traditional algorithms and other deep learning methods. This advancement provides critical insights for enhancing Three-dimensional reconstruction accuracy of non-cooperative targets.