PFU-Net: A Spacecraft Pose Estimation Utilizing Feature Fusion and Dense Residuals Network

With the continuous progress of space technology, space in-orbit service has become an important means to ensure the stable operation of spacecraft in complex space environments. However, the quality of non-cooperative spacecraft ISAR images acquired is low, the image edge is not clear, the resolution is poor, and the data provided is also limited, and the data is sparse. To solve this problem, a deep neural network architecture based on improved U-Net is proposed: PFU-Net. The close-connected structure can effectively extract in-orbit spacecraft features from images captured by inverse synthetic aperture radar (ISAR) and be used for pose estimation of non-cooperative in-orbit spacecraft. Specifically, the network first uses transfer learning to alleviate the problem of data sparsity through pre-training models. Secondly, aiming at the sparse features of spacecraft ISAR imaging, a dense residual structure combining channel attention mechanism is proposed in the coding and decoding part to achieve adaptive feature thinning, thereby reducing feature loss in the process of downsampling. In addition, the spectral gated network and multi-head self-attention mechanism are integrated to introduce additional frequency domain information, which is combined with spatial features to capture more global information. The experimental results show that compared with other methods, the improved network structure reduces the spacecraft Angle prediction error from the original 2.7° to 0.55°~0.58°, which is better than other advanced spacecraft attitude estimation models, and has a good performance in practicality and applicability.