CDN-Net: Faint Celestial Target Detection Based on Densely Nested Hierarchical Network

Abstract

The detection of celestial objects in ground-based wide-field optical telescope images serves as the foundational step for subsequent celestial analysis tasks. Existing methods for astronomical target detection have not addressed the challenges posed by a high dynamic range, faintness of targets, and an inaccurate supervision map. This paper presents a faint celestial target detection framework named the Celestial Densely Nested Network (CDN-Net). First, a hierarchical bit-depth decomposition strategy is designed to address high dynamic range astronomical FITS images, ensuring effective representation of faint targets. Second, a densely nested hierarchical network is introduced to extract high-resolution features of these faint astronomical targets. Lastly, a soft segmentation map, along with the corresponding loss, is proposed to guide the network's focus toward faint targets. Experiments were conducted on both simulated and real data sets, separately comprising 2560 images and 24,087 images, respectively, to evaluate the performance of CDN-Net. Compared to six existing methods, CDN-Net achieves superior precision, recall, and F1 score, especially for faint targets with signal-to-noise ratios below 3. Additionally, comparisons with star catalogs validate the effectiveness of CDN-Net. The code for this work is available at https://github.com/AeroFirefly/CDN-Net.