Advanced Instance Segmentation for Non-Cooperative Spacecraft Components in On-Orbit Servicing: Dataset Synthesis and Hybrid Transformer-CNN Network

To address the challenges of high-precision instance segmentation in on-orbit servicing missions, this study proposes an innovative instance segmentation framework that integrates Transformer and CNN, and constructs a high - fidelity hybrid dataset. Departing from traditional purely synthetic data approaches, we adopt a low-cost dataset construction strategy: 1) Physical rendering of spacecraft models using 3ds Max to simulate multi-illumination conditions (e.g., solar flares, shadow occlusion) and 6-DOF motion states; 2) Integration of limited real images to fine-tune data distributions, reducing reliance on complex GAN-generated data. Adaptive geometric augmentation (random rotation, scaling) and noise injection are further applied to enhance data authenticity. At the algorithmic level, we propose TransMask, a hybrid network integrating Swin Transformer modules for global context modeling and CNN-based feature pyramids (FPN) for multi-scale feature extraction. To tackle small component detection, a Dual Attention Module (DAM) is embedded, fusing spatial-channel attention with deformable convolution to focus on critical regions. A lightweight knowledge-distilled mask head is introduced to maintain accuracy while reducing computational overhead. Additionally, self-supervised pre-training with unlabeled on-orbit images enhances the model’s generalization capability for unknown component states. Experiments on self-built dataset demonstrate that TransMask significantly outperforms existing mainstream methods in segmentation accuracy and robustness, while exhibiting efficient real-time inference capabilities on embedded platforms, validating its feasibility for on-orbit deployment. This study bridges the gap between synthetic training and real-space applications, providing a reliable solution for autonomous on-orbit servicing.