Abstract: Non-cooperative spacecraft targets, such as spacecrafts engaged in autonomous rendezvous and docking, spacecrafts that have failed during missions, and spacecrafts performing on-orbit maintenance tasks, require rapid and accurate 6D relative pose estimation for subsequent guidance and control of service spacecraft. Previous methods often utilize CNNs (convolutional neural networks) and transformers for long-range feature interactions. However, the operations like downsampling tend to neglect high-frequency features such as edges, resulting in feature distortion. Additionally, researchers have designed multi-head loss functions or incorporated adversarial learning to perform transfer learning for pose estimation, but the tasks are complicated and time-consuming. In comparison, this paper proposes a wavelet feature enhancement approach to achieve accurate pose estimation for cross-domain spacecraft. Firstly, the extracted multi-scale features are decomposed into high-frequency and low-frequency bands. An effective transformer is used to perform spatial and contextual attention operations in the frequency domain, thus improving feature representation. Secondly, wavelet decomposition is performed on the processed features to enhance the high-frequency features of low-resolution images. Finally, the model trained in the source domain is directly applied to pose prediction in the target domain, and iterative self-training is conducted based on the obtained pseudo-labels. Comprehensive experiments on the SwissCube, SPEED, and SPEED+ pose estimation datasets demonstrate that the proposed method can effectively improve pose estimation accuracy and achieve competitive inference speed.