DAT-YOLO: Small Object Detection Model from the Perspective of Drones

A DAT-YOLO model incorporating deformable attention is proposed to address the challenges of dense target distribution, small pixel size, and uneven sample size in object detection from the perspective of unmanned aerial vehicles. This model is based on YOLOv8s and optimizes the Backbone and Neck structures, allowing the model network to be shallower in depth while being more focused on small targets; Using RepNCSPELAN4 structure instead of C2f structure to better adapt the model to different computing and hardware requirements; The deformable attention mechanism DAT is integrated into the spatial pyramid pooling layer, allowing the network to capture more effective features while expanding the receptive field; The use of inner-ShapeIoU loss function improves the algorithm's localization and classification performance for small targets in complex environments. The experimental results of DAT-YOLO on the VisDrone dataset show that with a 57.5% reduction in parameter count, the mAP@50 and the mAP@50:95 has also increased by 6.7% and 5.0% respectively, making it suitable for drone target detection tasks.