ATD-DL: A Deep Learning Framework for Faint Astronomical Target Detection

Junyao He^1,2Hao Luo^1*Wei Xiao¹Shuwei Liu¹Zhaoxiang Qi¹

• ¹Shanghai Astronomical Observatory, Chinese Academy of Sciences (CAS), Shanghai, China
• ²University of Chinese Academy of Sciences, Beijing, China

Astronomical imaging data frequently exhibit low signal-to-noise ratios (SNRs), especially for observations obtained from small-aperture, wide-field survey instruments, in which the detected signals are inherently faint and dominated by noise. Such characteristics pose a substantial technical challenge for subsequent target detection and quantitative measurement tasks. This challenge is particularly pronounced for faint astronomical targets with SNRs ranging from 1 to 10. When the SNR decreases below approximately 5, the useful signal approaches or falls beneath the detection threshold imposed by background noise, leading to a pronounced degradation in the performance of traditional threshold-based detection algorithms, such as SExtractor. Furthermore, astronomical imaging data are typically characterized by a high 16-bit dynamic range. This wide dynamic range results in the intensities of faint targets being compressed into a narrow interval of low pixel values. Standard global normalization strategies employed in deep learning models further compress this narrow intensity band, thereby suppressing and obscuring discriminative target features. To address these challenges, we propose ATD-DL, a deep learning–based framework specifically designed for faint astronomical target detection. The core of the proposed framework is an enhanced U-Net–based segmentation architecture. This architecture is integrated with a multi-stage image preprocessing pipeline, target separation, and centroid extraction modules to enable efficient and robust detection of astronomical objects. Experimental results demonstrate that the proposed method achieves excellent performance in detecting extremely faint targets with SNRs in the range 2 ≤ SNR < 5. Compared with traditional approaches, including SExtractor and DAOPHOT, the proposed framework exhibits a markedly superior detection capability under low-SNR conditions near the detection limit.