An improved YOLO11 UAV toy models target detection model

The challenges of Unmanned Aerial Vehicle (UAV) military target detection mainly come from the lack of training datasets and the requirements of high real-time and high precision. Due to the particularity and confidentiality of military-related data, publicly available military target datasets are very scarce. In this study, ArmoredCar, FixedWing, and Tank toy models were photographed in different environments and manually annotated to construct the Toy-3 dataset. Then, a multi-scene image enhancement method was designed based on the Toy-3 dataset. By simulating the common degradation factors in the UAV image acquisition process, more complex and diverse training data were generated, and an extended dataset Toy-3-Enhanced, which is closer to the actual mission scene, was constructed. Subsequently, the improved network Toy-Efficient-YOLO11 was proposed with YOLO11 as the benchmark model. The experimental results on Toy-3-Enhanced show that the improved model has a mAP50 of 0.982, a mAP50:95 of 0.805, and a Frames Per Second (FPS) of 333.33 frames/second, which is superior to the YOLO11n baseline model in terms of detection accuracy and real-time performance, providing a good technical solution for the UAV military target detection task.