Research on Small Target Detection and Recognition Algorithm Based on Improved YOLOv8 for Aerial Imagery

Unmanned Aerial Vehicles (UAVs) encounter various challenges in flight and imaging in everyday scenarios. These challenges include the limited representation of ground objects in UAV-captured images, the dense distribution of ground objects, and the characteristics of small targets such as their small size, high speed, and maneuverability. These factors often result in small pixel points in images, making it challenging to accurately detect them using existing UAV detection methods. The study introduces a novel dynamic head framework, named Dynamic Head (DyHead), which integrates target detection head and attention mechanisms to emphasize the fusion of shallow features for efficient information interaction and fusion. Furthermore, the study enhances the feature weights of dense small targets and incorporates a multi-branch fusion module called Diverse Branch Block (DBB) in C2f to enhance the feature extraction capability of small objects. In the experiment of the VisDrone2019 dataset, the enhanced PDD-YOLOv8 model demonstrates significant performance improvements compared to the standardized YOLOv8s model without increasing parameters. Notably, the mean average precision (mAP) increases by 8.5 and 5.9 percentage points at IoU thresholds of 0.5 and 0.5:0.95, respectively, demonstrating strong specificity and performance.