UAV Small Target Detection in Complex Scenes Based on Improved YOLOv8s

Aiming at the phenomenon of misdetection and omission caused by the question of tiny target size and difficulty in capturing details when UAVs perform target detection tasks in intricate scenes. An modified target detection algorithm YOLOv8s-SME with YOLOv8s as the base algorithm is proposed. Firstly, the improved idea of Alpha-IOU is introduced into SIOU to construct a new exponential SIOU loss function (E- SIOU), which introduces more nonlinear properties in the target detection task, suppresses the error propagation and accelerates the model convergence, and makes it more sensitive in calculating the bounding box overlap degree; Secondly, the An attention mechanism that operates through multiple dimensions is introduced(MCA), which takes into account To model the importance of multi-dimensional cooperation in a novel way, improving performance while incurring little additional computational overhead; Subsequently, the EMA attention mechanism is used to construct a new C2f-E module in the feature extraction network, which groups the channel dimensions into multiple sub-features and reshapes some of the channel dimensions into batch dimensions to maximise the feature information of the small targets for the purpose of improving the detection accuracy. Finally validated on the VisDrone-2019 dataset, the experimental results show that the algorithm in this paper has a 3% improvement in mAP@0.50 compared to YOLOv8s algorithm and a 1.9% improvement in mAP@0.50:0.95.