Star image motion deblurring via angular velocity assisted U-Net with feature fusion and CBAM

Abstract

For star sensors operating in space, star images are susceptible to blur degradation caused by complicated motions, leading to difficulties in star identification and reduced accuracy in star spots extraction. Although deep learning methods have been applied to mitigate these issues, their performance is often limited by the neglect of physical motion information. To address this, we embed angular velocity data to guide the network in establishing the relationship between angular velocity and blur parameters. Furthermore, considering the sparse nature of star spot features, we introduce the Convolutional Block Attention Module (CBAM) to enhance the feature saliency of star regions and their motion trajectories. A dedicated feature fusion module is also designed to effectively integrate inertial data and visual representations, facilitating joint learning of physical motion patterns and image features. Simulation experiments indicate that the proposed method significantly improves image restoration quality compared to the baseline U-Net, achieving a 16.84 % increase in PSNR and reducing the centroid error to 0.0836 pixels. Experiments based on in-orbit observation data validate the effectiveness of our method in real-world scenarios.