YOLO-LVSOD: A Multi-Scale Dilated Attention Network for Foggy Small Object Detection

UAV aerial imagery is widely used in deep learning, but most methods target medium-to-large objects in clear conditions, with limited research on small road object detection in fog. This paper proposes YOLO-LVSOD (Low-Visibility Small Object Detector), a fog-adapted small road object detector based on YOLOv11s. The backbone integrates a Multi-Scale Feature Augmentation (MSFA) module using dilated convolutions and attention mechanisms to enhance feature representation. Additionally, a Lightweight Deep Difference Enhancement Convolutional Detection head (LDDECD) is introduced to improve small object detection accuracy. Experiments are conducted on a fog-simulated version of the VisDrone2019 dataset containing diverse small road objects. Performance is evaluated by using precision, recall, mAP and parameters, demonstrating advantages over state-of-the-art models.