An Improved YOLOv11 for Nighttime Drone-View Object Detection

To address the dual challenges of image quality degradation and detection performance limitations in UAV visual perception systems under nighttime conditions, this study proposes a collaborative optimization framework integrating image quality enhancement and target feature reinforcement. Specifically, issues such as brightness attenuation, noise interference, and detail loss caused by nighttime environments severely constrain the feature extraction capabilities of traditional detection algorithms. To overcome this, the research team innovatively developed a Dual-Domain Real-time Guided nighttime Image Enhancement Network (DDNet): on one hand, it enhances edge details through Log operator-based feature extraction, while on the other hand, employs convolutional layers and residual connections to preliminarily boost brightness and suppress noise, forming a complementary image enhancement solution. Building upon this foundation, the YOLOv11 detection network was implemented using the VisDrone2019 dataset. The framework incorporates a Super-Resolution Spatially-Adaptive Feature Modulation (SRSAFM) module and a Lightweight Cross Attention (LCA) mechanism, achieving efficient image superresolution and complementary brightness-color information interaction. Through a cascaded architecture that deeply integrates enhancement modules with improved detection models, the detection network optimizes feature weight allocation via attention mechanisms and enhances small-target recognition accuracy using high-resolution feature maps. Experimental validation demonstrates that this joint optimization strategy preserves image detail textures while improving the average detection accuracy in nighttime scenarios by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{2. 8 \%}$</tex>.