无人机视角轻量化红外与可见光图像检测方法  

Lightweight infrared and visible image detection methods for UAV perspectives

作　　者：蒋行国 王峣 林国军 孙潇 刁豪杰 黎明 JIANG Xing-guo;WANG Yao;LIN Guo-jun;SUN Xiao;DIAO Hao-jie;LI Ming(School of Automation and Information Engineering,Sichuan University of Science and Engineering,Yibin 644000,China;Artificial Intelligence of Key Laboratory of Sichuan Province,Yibin 644001,China)

机构地区：[1]四川轻化工大学自动化与信息工程学院,四川宜宾644000 [2]人工智能四川省重点实验室,四川宜宾644001

出　　处：《激光与红外》2025年第3期452-460,共9页Laser & Infrared

基　　金：四川轻化工大学人才引进项目(No.2019RC12);四川省科技计划项目(No.2024YFHZ0026)资助。

摘　　要：针对无人机航拍视角目标检测空间尺度变化大、物体像素占比少,算法部署边缘计算平台存储空间占比大等问题。本文基于YOLOv8n网络结构,提出一种改进航拍视角轻量化小目标检测方法DSF-YOLO-P算法。首先,主干网络C2f模块融合FasterNet,组成Faster-C2f轻量化模块,以保证模型在不影响检测精确度的同时实现网络轻量化与提升检测速度。然后,新增160×160预测头并重构网络通道,以提升模型对小型目标检测的准确度及鲁棒性。改进后的DSF-YOLO算法在可见光数据集VisDrone2019与红外数据集HIT-UAV上准确度分别提升2.5%与0.6%,参数量降低10%。最后,将DSF-YOLO算法进行依赖图剪枝操作,减少模型冗余参数又不影响模型性能。剪枝后的DSF-YOLO-P算法在VisDrone2019数据集上相较于DSF-YOLO算法,精确度不变,计算量与参数量分别降低45%与26%。实验结果充分证明了DSF-YOLO-P算法在无人机航拍视角小目标检测中的有效性。Aiming at the UAV aerial photography viewpoint target detection spatial scale change is large,the object pixels account for a small proportion,and the algorithm deployment edge computing platform storage space occupies a large proportion of the problem.In this paper,based on the YOLOv8n network structure,an improved aerial photography viewpoint lightweight small target detection method DSF-YOLO-P algorithm is proposed.Firstly,the backbone network C2f module is integrated with FasterNet to form the Faster-C2f lightweight module to ensure that the model achieves network lightweighting and improves the detection speed without affecting the detection accuracy.Then,a new 160×160 prediction head is added and the network channels are reconfigured to improve the accuracy and robustness of the model for small target detection.The improved DSF-YOLO algorithm improves the accuracy by 2.5%and 0.6%on the visible dataset VisDrone2019 and infrared dataset HIT-UAV,respectively,and reduces the number of parameters by 10%.Finally,the DSF-YOLO algorithm is subjected to the dependency graph pruning operation to reduce the redundant parameters of the model without affecting the model performance.The pruned DSF-YOLO-P algorithm achieves the same accuracy and reduces the computational effort and number of parameters by 45%and 26%,respectively,compared with the DSF-YOLO algorithm on the VisDrone2019 dataset.The experimental results fully demonstrate the effectiveness of the DSF-YOLO-P algorithm in detecting small targets in the aerial view of UAVs.

关 键 词：边缘计算平台 无人机 FasterNet 预测头 剪枝 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] V279[自动化与计算机技术—计算机科学与技术]