<p>Deep learning-based object detection methods face persistent challenges in complex lighting environments. Existing approaches often rely on serially connecting image enhancement models with detection networks to improve visual quality, thereby indirectly boosting detection performance. However, this strategy is not always effective due to the differing optimization objectives of enhancement and detection networks. To overcome this limitation, we propose a dual-path fusion framework and a novel Illumination-Guided Feature Fusion (IGFF) network. By integrating enhancement and detection networks through a hybrid serial-parallel strategy, our method enhances detection performance via illumination-guided, multi-level, and multi-scale deep fusion. Specifically, the IGFF framework leverages a pre-trained enhancement model to guide the detection network in adapting to varying lighting conditions. It achieves deep fusion by capturing and comparing spatial information before and after enhancement, thereby facilitating efficient information integration across both channel and spatial dimensions. The framework is highly adaptable, requiring no architectural modifications to existing enhancement or detection networks. Experimental results demonstrate that our method improves detection performance under complex lighting conditions and further demonstrates strong generalization capability in foggy scenarios.</p>