DRE-YOLO: A dual-path retinex enhancement network for low-light object detection

Existing object detectors often suffer from degraded accuracy and limited real-time performance under low-light and uneven illumination. To overcome these limitations, we present DRE-YOLO, an efficient end-to-end framework integrating a Dual-path Retinex Enhancement (DRE) network with YOLOv3. Grounded in retinex theory, our method decomposes input images into reflectance and illumination components. We then introduce two specialized modules: the Reflectance Feature Boosting (RFB) module, which enhances semantic features through multi-scale fusion and cross-attention; and the Illumination Adaptive Balance (IAB) module, which regulates lighting via dual-branch attention channels. The two refined components are recombined to generate enhanced images for detection. Trained with joint reconstruction and detection losses, we evaluate our method on the Exdark and Dark Face datasets. Experimental results demonstrate that our DRE-YOLO achieves superior performance and surpasses existing low-light object detection methods.