Wavelet-Enhanced Feature Pyramid Network for Pose Estimation of Non-Cooperative Spacecraft

Monocular six-degree-of-freedom pose estimation of non-cooperative spacecraft is critical for a variety of on-orbit missions. However, it remains a challenging task due to extreme illumination variations, pronounced scale differences, complex rotational motions, and the sparse geometric characteristics of spacecraft. While deep learning has advanced this field, existing methods relying on standard convolution and pooling operations lose high-frequency details and spatial resolution, thus limiting pose estimation accuracy. To overcome these limitations, this paper introduces a novel wavelet-enhanced feature pyramid network for spacecraft monocular pose estimation, which enables multiscale feature fusion and representation enhancement across frequency and spatial domains together with cross-sample feature interaction to improve estimation accuracy and robustness. For large-scale features, a wavelet-enhanced frequency-spatial downsampling module integrating Haar wavelet transform and channel attention is designed to strengthen pose-sensitive features and suppress illumination noise interference while retaining geometric structural details. For small-scale features, a cross-sample feature enhancement downsampling module with a dual-path architecture is proposed to strengthen high-level semantic representation through cross-sample feature interaction against complex spacecraft rotations. Extensive offline validations on the SPEED, SPEED[Formula: see text], and SwissCube datasets demonstrate that our method achieves competitive pose estimation accuracy while striking a better balance between accuracy and inference speed.