Object Detection Model of YOLOv8-CSD for UAV Images

In order to address the challenges faced in detecting objects in UAV aerial images, we propose an enhanced YOLOv8-CSD object detection algorithm. Our aim is to overcome issues such as low accuracy, complex model structure, limited generalization ability, small size of detected objects, variations in scale and occlusion due to background elements. These factors often lead to missed detections and false alarms. To improve the feature extraction capability and enhance detection accuracy, we introduce the SimAM attention mechanism into the network's neck section. Additionally, we leverage the benefits offered by deformable convolutional networks and incorporate a C2fDCN module into the YOLOv8 neck network for improved feature fusion ability specifically targeting small targets. Furthermore, we replace the original detection head with DyHead for unified scale consideration along with spatial and task awareness. Experimental results on VisDrone2019 dataset consisting of aerial photography demonstrate that our improved algorithm achieves a mean Average Precision (mAP) of 41.9%, which is 3.4% higher than that achieved by the original YOLOv8s algorithm. This showcases excellent performance while validating its effectiveness.