Spacecraft components recognition is essential for space on-orbit service missions such as space docking, on-orbit maintenance, and other applications. Compared with the expensive recognition methods based on the Light Detection and Ranging (LIDAR) system, the optical camera provides an economical alternative. However, owing to the adverse illumination in space, spacecraft images often suffer from shadow and poor visibility, which seriously damages the recognition accuracy. To overcome the image quality limitation, we proposed a novel concept “Multi-Illumination Angles Image Fusion” (MIAF), which aims to fuse the information of single spacecraft in different illumination angles and reconstruct the shadow-free spacecraft image. In this study, a new dataset containing spacecraft images under multiple illumination angles is created. Then a novel image fusion model for spacecraft image shadow removal is proposed. Different from existing image fusion methods that use handcrafted fusion rules, a weight subnet block is designed to learn the optimal fusion strategy automatically. Additionally, a novel loss function combining supervision and self-supervision is adopted to train the weight subnet. To verify the advantage, the proposed method is evaluated and compared with 8 competitive image fusion methods on the released dataset. The proposed method achieves the state-of-the-art performance in both qualitative and quantitative experiment results. The dataset and MIAF implementation code are open-sourced, which can be found in (https://github.com/xiang-ao-data/Spacefuse-shadow-removal/tree/master).