EAMS-YOLOv8: An Object Detection Algorithm for Drone Aerial Images

Object detection in drone aerial images encounters numerous challenges, notably the presence of small objects occupying minimal pixel areas, as well as occlusions and overlaps among objects, leading to a reduction in detection accuracy. To tackle these issues, we present EAMS-YOLOv8, an enhanced model derived from the YOLOv8n. The innovations of EAMS-YOLOv8 are fourfold: First, the Efficient-C2F block is incorporated into the backbone network, replacing the C2F block. This enhancement bolsters the capability of feature extraction and facilitates the extraction of multi-scale features. Second, the SPPF block is replaced by the Adaptive-SPPF block, which employs diverse adaptive pooling to capture global and local information and decreases information loss during pooling. Furthermore, a layer of up-sampling and concatenation operations is added to change the connection of the neck network to enrich the feature information of the small object layer. Finally, the SEAM attention mechanism is introduced before the detection head to alleviate the occlusion problem among objects and enhance the recognition and localization capability of EAMS-YOLOv8. Experiments show that EAMS-YOLOv8 improves mAP by approximately 7% on the VisDrone2019 dataset compared to YOLOv8n, effectively increasing detection accuracy in complex and dense scenes, and providing a practical solution object detection area.