A zero-shot image denoising pipeline for space target detection

Optical imaging of space targets using ground-based or space-based telescopes is typically affected by complex noise. Due to the sparse features and limited data, the denoising performance of star images is often suboptimal. In this paper, we propose a lightweight zero-shot star image denoising framework featuring an improved 3-layer U-Net backbone, which can efficiently complete the denoising task without a complete dataset. This network extracts feature information through two pair of down-sampling and up-sampling layers, as well as several convolution modules. The spatial attention module is employed to focus on attention regions, enhancing the model efficiency and generalization ability. In the experiments conducted with real star images, the denoising pipeline primarily consists of three steps: image preprocessing, network training and inference. The results demonstrate that our method effectively removes noise from star images and outperforms existing techniques, facilitating the accurate detection and extraction of space targets in subsequent research.