To address the challenges of low signal-to-noise ratio and poor localization in detecting faint celestial objects from ground-based optical images, we propose CRFusion-Det, a plug-and-play probabilistic detection head. It introduces innovations at both the feature representation and inference levels. First, dilated convolutions and the CBAM attention module are integrated into the heatmap and width-height regression branches to enhance multi-scale contextual perception. Second, for offset estimation, the keypoint coordinate regression is innovatively reformulated as a probability distribution modeling problem. This is achieved via a learnable Prospect/Background Probability Estimation Module (PBPEM) and a Spatial-Appearance message Transmission Module (SATM), which explicitly capture inter-target geometric constraints and appearance consistency. A mean-field iterative algorithm is employed for structured inference, enabling progressive distribution refinement and sub-pixel localization. Extensive experiments on real-world datasets demonstrate CRFusion-Det’s effectiveness and generalization. When integrated into five different baseline networks, it consistently improved the recall by 1.68%–5.24% and reduced the normalized mean error to as low as 0.05 (0.73 pixels). The proposed CRFusion-Det significantly enhances the detection and localization accuracy of baseline models for faint targets, validating its superiority as a solution for astronomical image processing.
扫码关注我们
求助内容:
应助结果提醒方式:
