摘要
针对无人机图像中由于目标微小且相互遮挡、特征信息少导致检测精度低的问题,提出一种基于改进YOLOv7的无人机图像目标检测算法。在颈部和检测头中加入了坐标卷积,能更好地感受特征图中目标的位置信息;增加P2检测层,减少小目标特征丢失、提高小目标检测能力;提出多信息流融合注意力机制——Spatial and Channel Attention Mechanism(SCA),动态调整注意力对空间信息流和语义信息流的关注,获得更丰富的特征信息以提高捕获目标的能力;更换损失函数为SIoU,加快模型收敛速度。在公开数据集VisDrone2019上进行对比实验,改进后算法的mAP50值相比YOLOv7提高了4%,达到了52.4%,FPS为37,消融实验验证了每个模块均提升了检测精度。实验表明,改进后的算法能较好地检测无人机图像中的目标。
To solve the problems of small targets,mutual occlusion,and less feature information in UAV images,which lead to low detection accuracy,an improved YOLOv7 UAV image target detection algorithm is proposed.CoordConv is added to the neck and detection head,which can better sense the position information of the target in the feature map;the P2 detection layer is added to reduce the loss of small target features and improve the detection ability of small targets;multiple information flow fusion attention—Spatial and Channel Attention Mechanism(SCA)is proposed to dynamically adjusts the focus on spatial information flow and semantic information flow to obtain more comprehensive feature information to improve the ability to capture targets;the loss function is replaced with SIoU to speed up model convergence.A comparison experiment is conducted on the public dataset VisDrone2019.The mAP50 value of the proposed algorithm is 4%higher than that of YOLOv7,reaching 52.4%,and the FPS is 37.The ablation experiments verify that each module improves the detection accuracy.Experiments show that the improved algorithm can better detect objects in UAV images.
作者
梁秀满
贾梓涵
于海峰
刘振东
LIANG Xiuman;JIA Zihan;YU Haifeng;LIU Zhendong(College of Electrical Engineering,North China University of Science and Technology,Tangshan 063210,China)
出处
《无线电工程》
2024年第4期937-946,共10页
Radio Engineering
基金
河北省自然科学基金(F2018209289)。
