中圖分類號：TP391.4 文獻(xiàn)標(biāo)志碼：A DOI: 10.16157/j.issn.0258-7998.256845 中文引用格式： 譚勛瓊，王穎林. YOLO-PDS:基于改進(jìn)的YOLOv11的無人機(jī)小目標(biāo)檢測算法[J]. 電子技術(shù)應(yīng)用，2025，51(12)：96-102. 英文引用格式： Tan Xunqiong，Wang Yinglin. YOLO-PDS: a small object detection algorithm for drones based on the improved YOLOv11[J]. Application of Electronic Technique，2025，51(12)：96-102.
YOLO-PDS： a small object detection algorithm for drones based on the improved YOLOv11
Tan Xunqiong，Wang Yinglin
School of Physics and Electronics， Changsha University of Science and Technology
Abstract： Object detection has broad application prospects in the field of remote sensing. Although object detection algorithms have made significant progress in natural images, these methods still face numerous challenges when directly applied to remote sensing images. The background of remote sensing images is often complex, and the objects are relatively small, which leads to an extremely imbalanced distribution of foreground and background information. To address the issues of small targets and object occlusion in drone images, this paper proposes an improved drone small object detection algorithm based on PinwheelConv. To enhance the model's performance in detecting small objects, the PinwheelConv is used in place of regular convolution in the backbone network, which better adapts to the extraction of small target features. Additionally, a C2f-PC module based on the windmill convolution idea is designed to replace the C3k2 module in the backbone. To address the severe occlusion problem in drone images, this paper innovatively introduces the C2f-PDWR module to replace the C3k2 module in the neck network, enhancing the model's feature fusion capability. Moreover, a Spatially Enhanced Attention Module (SEAM) is incorporated to improve the model's detection of occluded objects. Finally, this paper proposes a more efficient small object detection model, YOLO-PDS, based on YOLOv11. The proposed method improves the mAP50 by over 3.7% and the recall rate by more than 2.2% compared to the baseline YOLOv11 detection method on the VisDrone2019 dataset.
Key words : object detection；YOLOv11；Pinwheel Convolution；multidimensional attention mechanism
