摘要
无人机在军事情报、航拍检测等领域能够提供目标相关的图像信息,为处理任务提供目标信息。针对无人机图像背景复杂、检测目标小、可提取特征少等问题,提出基于YOLOv5s的改进无人机图像识别算法。首先,结合CotNet模块对网络结构进行优化,提升模型自学习能力并增强识别精度;其次,对颈部网络进行改进,通过跨层链接和提高特征图分辨率更好地利用浅层特征图中包含的丰富信息来定位目标,并且在检测头部分采用解耦检测头,减少预测过程中定位与分类任务对于特征信息的冲突;最后,为了提高收敛速度和模型精度,在CIoU和EIoU损失函数的基础上对损失函数的宽高纵横比进行优化。在公开数据集VisDrone测试集上进行测试,所提算法相比原始YOLOv5s算法的mAP_(50)与mAP_(50∶95)分别提升了6.1与2.9个百分点,实验结果表明,所提模型能够有效提升无人机图像识别的准确率。
UAVs can provide target-related image information in such fields as military intelligence and aerial photography detection providing target information for image processing tasks.An improved UAV image recognition algorithm based on YOLOv5s is proposed to address the issues of complex background small detection targets and limited extractable features in UAV images.Firstly the network structure is optimized by using CotNet module to enhance the model s self-learning ability and improve recognition accuracy.Secondly the Neck network is improved to better utilize the rich information contained in shallow feature maps to locate targets through cross-layer linking and an improvement to feature map resolution.In the detection head section,decoupled detection heads are used to reduce the conflict between localization and classification tasks over feature information utilization in the prediction process.Finally in order to improve the convergence rate and model accuracy the aspect ratio of the width to the height of the loss function is optimized based on CIoU and EIoU loss functions.Tests are conducted on the test set of public dataset VisDrone.The proposed algorithm improves mAP 50 and mAP 50︰95 by 6.1 and 2.9 percentage points respectively in comparison with the original algorithm.The experimental results show that the proposed model can effectively improve the accuracy of UAV image recognition.
作者
李杰
王峰
马晨
吴国瑞
赵伟
康智强
LI Jie;WANG Feng;MA Chen;WU Guorui;ZHAO Wei;KANG Zhiqiang(College of Electronic Information and Optical Engineering Taiyuan University of Technology, Jinzhong 030000 China;33rd Institute of China Electronics Technology Group Corporation, Taiyuan 030000 China)
出处
《电光与控制》
CSCD
北大核心
2024年第4期22-27,91,共7页
Electronics Optics & Control
基金
山西省重点研发计划(202102150101008)
山西省留学人员科技活动择优资助项目(20230063)
山西省水利科学技术研究与推广项目(2022GM002)。
作者简介
李杰(1997-),男,山西太原人,硕士生。;通讯作者:王峰(1975-),男,山西太原人,博士,教授,硕导。
