CSD-YOLOv8: A Small-Object-Detection Model Based on Improved YOLOv8 for UAV Aerial Images

In view of the challenges of many small-featured and informative objects in UAV aerial photography scenes, and the difficulty of distinguishing adjacent objects, a new model called CSD-YOLOv8 is proposed. To evolve the backbone network in this regard, Coordinate Attention (CA) is introduced by this model, which enables the model to pay attention to important features of small-sized objects. It also contains a shallow small object-detection-specialized layer (SOL) that grabs shallow features. This makes it combine shallow and deep features well and greatly lowers the small object missing rate. The original detection heads are replaced by the Dynamic Head (DH), which is more adaptive to the input data, making the model better at detecting small objects. Experimental results on the dataset VisDrone2019 validate that CSD-YOLOv8 outperformed the standard YOLOv8 by a margin of 8.5% and 5.9% for mAP50 and mAP50:95, respectively. CSD-YOLOv8 outperforms previous works with the improved ability to detect small objects tremendously.