Detection Utility, Not Visual Quality: Degradation-Aware Dual-Branch Fusion for Low-Light Object Detection
Enhanced visibility does not necessarily translate into improved object detection under low-light conditions. On ExDark, eight enhanced variants evaluated with the same detector produce [email protected] scores ranging from 76.4\% to 77.6\%, with five falling below the un-enhanced baseline of 77.0\%, indicating that enhancement quality and detection utility can be misaligned. This observation motivates a key question: how can the detector assess the reliability of enhanced representations and adapt their contribution to the detection task? To address this problem, we propose DADF-Det, a degradation-aware dual-branch framework for low-light object detection. A Global Degradation Discriminator (GDD) estimates the detection utility of the enhanced image from the original--enhanced image pair and provides a global reliability prior for feature fusion. Guided by this prior, the Enhanced Texture Branch (ETB) extracts semantic texture representations from the enhanced image, while the Low-Light Edge Branch (LEB) preserves structural and boundary cues from the original image. A Dynamic Discrepancy Calibration Fusion (DDCF) module further calibrates local feature discrepancies according to the estimated global reliability, while a boundary consistency loss constrains the fused representations to preserve localization-sensitive structures. Experiments on ExDark, DarkFace, and ACDC-Night show that DADF-Det improves [email protected] on ExDark by 1.1 percentage points over the strongest publicly reported method while achieving competitive performance on the remaining benchmarks.
Authors
- Hui Zong (ORCID: https://orcid.org/0000-0003-1345-3749)
- Xin Cheng (ORCID: https://orcid.org/0000-0003-1738-3522)
- Chenbin Wang
- Chunpeng Wang
- Haoran Liu
- Pengfei Xu
Institutions
- Tongji University (CN)
- Qilu University of Technology (CN)
- Huaiyin Institute of Technology (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-18
- DOI
- https://doi.org/10.21203/rs.3.rs-11076264/v1
- Primary Topic
- Advanced Neural Network Applications
- Type
- preprint