LE-YOLO: task-driven curve enhancement with attention-guided ghostnet for low-light object detection

Object detectors often suffer substantial performance degradation under low-light conditions due to the combined effects of a low signal-to-noise ratio, reduced contrast, and loss of fine-grained details. This limitation is especially critical in safety-sensitive scenarios. We propose Light Enhancement YOLO (LE-YOLO), an end-to-end low-light object detection framework that integrates a task-driven dynamic curve enhancement module with a GhostNet-based detection head on top of YOLOv10n. The enhancement module, inspired by Zero-DCE++, is embedded as a learnable “zeroth layer” whose parameters are optimized solely by the final detection loss. In a single forward pass, the module predicts pixel-wise high-order curve parameters and implicitly learns representations that are more conducive to detection rather than human visual perception. To partially offset the additional computational overhead of this front-end module, we redesign the neck and head of YOLOv10n using C3Ghost blocks derived from GhostNet, which reduce parameters and floating-point operations (FLOPs) while preserving detection accuracy. Furthermore, we incorporate the Convolutional Block Attention Module (CBAM) into the neck to refine feature responses along both the channel and spatial dimensions, thereby improving robustness to complex low-light backgrounds. We evaluate LE-YOLO on the ExDark dataset. Compared with the YOLOv10n baseline, our model improves mean Average Precision (mAP)@0.5 from 64.9% to 75.0% and mAP@0.5:0.95 from 41.4% to 52.8%, while recall increases from 56.8% to 66.5%. Experiments on the DarkFace dataset further demonstrate that LE-YOLO generalizes well to unseen low-light domains, achieving a 7.1 percentage-point gain in mAP@0.5 over YOLOv10n. Taken together, these results show that the proposed task-driven enhancement and GhostNet-based head provide a strong accuracy-efficiency trade-off and robust accuracy gains for low-light object detection. Therefore, we position LE-YOLO as an accuracy-oriented detector with computational compensation.