LLE-YOLO: A Lightweight Network for Low-Light Object Detection
Abstract
Real-time object detection under low illumination remains challenging because weak contrast, sensor noise, and feature degradation can reduce target discriminability, while practical deployment also requires careful control of model complexity and computational cost. Existing enhancement-then-detect pipelines can improve image visibility but introduce additional computational and latency overhead, whereas direct detection methods must balance weak-feature representation against model efficiency. To address these challenges, this paper proposes LLE-YOLO, a lightweight end-to-end detector for low-light object detection. LLE-YOLO integrates three coordinated components. First, the Global Context Feature Enhancement (GCFE) module and the Dilated Receptive-field Attention Module (DRAM) combine lightweight global-context aggregation, multi-scale dilated convolution, and content-dependent feature reweighting to improve feature representation under low illumination. Second, the Multi-Scale Adaptive Fusion (MSAF) mechanism adaptively integrates aligned multi-scale features using learnable spatial weights to alleviate cross-scale feature inconsistency. Third, the Intersection Auxiliary over Union (IAoU) loss is adopted as a geometry-aware regression objective, using enclosing-region and corner-alignment constraints to provide additional geometric information for bounding-box regression, particularly under limited-overlap conditions. Experiments on the DroneVehicle nighttime subset and the Night-Det dataset demonstrate competitive detection performance with moderate model complexity. On DroneVehicle, LLE-YOLO achieves 76.3% <inline-formula> <tex-math notation="LaTeX">$\mathrm {mAP}_{50}$ </tex-math></inline-formula> and 50.3% <inline-formula> <tex-math notation="LaTeX">$\mathrm {mAP}_{50:95}$ </tex-math></inline-formula>, with 19.4 M parameters, 32.0 G FLOPs, and a throughput of 176.1 FPS on an NVIDIA A16 GPU. On Night-Det, it achieves 75.5% <inline-formula> <tex-math notation="LaTeX">$\mathrm {mAP}_{50}$ </tex-math></inline-formula> and 44.8% <inline-formula> <tex-math notation="LaTeX">$\mathrm {mAP}_{50:95}$ </tex-math></inline-formula>. These results indicate that LLE-YOLO provides a favourable balance between detection accuracy and computational efficiency for low-light object detection.