基于虚拟合成数据的卫星装配状态视觉检测方法

To address the high cost and inefficiency of acquiring annotated data for deep learning-based visual inspection in complex scenarios，a method integrating virtual synthetic data with deep learning is proposed for satellite assembly inspection. First，a high-fidelity synthetic training set is constructed by automatically generating multi-lighting， multi-viewpoint virtual images with pixel-level annotations based on a satellite assembly model library and a physical simulation engine. Then，a multi-task detection framework is designed for satellite compartment assembly features， decomposing anomaly detection into sub-tasks to guide model training with synthetic data. The trained model is deployed to real inspection scenarios，using hybrid training and style transfer to reduce the virtual-real data gap. Finally，the effectiveness of the proposed method is verified through experiments on a satellite Z-board assembly task，which achieve 96. 1% accuracy for missing part detection and 81. 2% for foreign bolt/nut detection with minimal real samples. A scalable path is provided by this approach for quality inspection in low-sample precision manufacturing.