Answers are grounded in this paper and cite their sources.
Target detection in low-light environments has wide applications, but it faces challenges due to extremely weak photon responses, severe sensor noise, and broken target geometry. We propose a target detection architecture based on feature-level physical prior decoupling–Illumination-Guided and Reliability-aware YOLO (IGU-YOLO). First, we design a reliability-aware illumination gating module (UIG-Block) in the shallow backbone layers. It estimates illumination priors and learns a feature-reliability map for confidence-guided suppression of unreliable responses. To further separate illumination from noise, we design a self-supervised illumination consistency loss (SIC Loss). Together, these two designs encourage the model to learn physically consistent illumination distributions without real illumination labels and to reduce unreliable feature injection through confidence gating. Second, to address broken target contours in extremely dark regions, we deploy an illumination-guided deformation refinement module (IGD-R) in the backbone. This module uses shallow illumination priors to guide deformable convolution for spatial structure reconstruction and introduces a frequency-domain mechanism to suppress secondary semantic artifacts. Experiments on the extremely low-light benchmark dataset ExDark show that IGU-YOLO improves \(mAP₅₀\) , \(mAP₇₅\) , and \(mAP50:95\) by 2.6%, 2.7%, and 2.1%, respectively, over the baseline, demonstrating its effectiveness under low-light conditions.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: