DNRFA-Net: dense nested network based on receptive field attention for faint celestial object detection

Detecting faint celestial bodies is crucial for space surveillance and planetary defense. However, it faces significant challenges due to the ultra-low signal-to-noise ratio, the fact that targets often occupy only a few pixels, interference from cosmic backgrounds such as atmospheric turbulence, and the scarcity of information caused by indistinguishable shapes. To overcome these limitations, proposed Dense Nested Receptive Field Attention Network (DNRFA-Net), an integrated segmentation detection framework with three key innovations: (1) A dense nested U-Net with extended cross level connections that preserves faint target details by enhancing gradient flow in low contrast regions; (2) An Adaptive Receptive Field Attention Module (ARFAM) that dynamically adjusts the receptive field through a Scale Estimation Unit (SEU) to improve attention to faint targets and enhance the ability to distinguish between targets and backgrounds; (3) A Cross Attention Feature Fusion Module (CAFFM) that refines multi-scale feature fusion through interaction across spatial and channel dimensions. By assigning importance weights to features in each dimension, multi-scale features are fused, preserving the saliency information of targets in each dimension and improving the quality of feature map fusion. After evaluating on real ground optical images, DNRFA-Net achieved excellent detection results, achieving a recall of 93.978%, a precision of 97.460%, and a F1 score of 95.687%. It demonstrated excellent robustness in detecting faint celestial targets and promoted the development of deep learning-based space object detection. The code in this article will be published on https://github.com/qinqinmie/xlljjjjj-DNRFA-Net.