基于改进的Yolov5的无人机图像小目标检测  

UAV image small-target detection based on improved Yolov5

作　　者：何宇豪 易明发 周先存 王冠凌 HE Yuhao;YI Mingfa;ZHOU Xiancun;WANG Guanling(College of Electrical Engineering,Anhui Polytechnic University,Wuhu 241000,China;The Department of Electronic and Information Engineering,West Anhui University,Lu’an 237012,China)

机构地区：[1]安徽工程大学电气工程学院,安徽芜湖241000 [2]皖西学院电子与信息工程学院,安徽六安237012

出　　处：《智能系统学报》2024年第3期635-645,共11页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金项目(61572366);安徽高校自然科学研究重大项目(J2021ZD0116);皖江高端装备制造协同创新中心开放基金项目(GCKJ2018013).

摘　　要：为了解决无人机航拍图像小目标检测算法检测速度与精度无法兼顾的问题,在Yolov5的基础上,提出了针对于无人机图像小目标检测的Yolov5_GBCS算法。在新的算法中,添加一个额外的检测头,以便增强对小目标的特征融合效果;在主干网络中分别采用GhostConv卷积模块、GhostBottleneckC3模块替换部分Conv模块和C3模块用以提取丰富特征和冗余特征以提高模型效率;引入加权双向特征金字塔网络(bidirectional feature pyramid network,BiFPN)结构,用以提高对小目标的检测精度;在主干网络和颈部网络中引入轻量化的卷积块注意力模块(convolutional block attention module,CBAM),关注重要特征并抑制不必要的特征,增强小目标特征表达能力;使用Soft-NMS算法来替换NMS,因此降低了小目标在密集场景下的漏检率。通过在VisDrone2019数据集上的实验结果表明,集成了所有改进的方法后的Yolov5_GBCS算法,不仅提高了检测精度,而且有效地提高了检测速度,模型的mAP从38.5%提高到43.2%,检测速度也从53 f/s提高到59 f/s。Yolov5_GBCS算法可以有效地实现无人机航拍图像中小目标识别。The detection speed and accuracy of small targets captured by UAV aerial photography cannot be considered at the same time.To address this problem,a new algorithm based on the Yolov5 algorithm,namely Yolov5_GBCS,is proposed for small-target detection of UAV-captured images.In the new algorithm,an additional detector head is added to enhance the feature fusion effect of small targets.In the backbone network,the GhostConv convolution module and GhostBottleneckC3 module are used to replace some original Conv modules and C3 modules for extracting rich and redundant features,respectively,to improve the model efficiency.The weighted bidirectional feature pyramid network structure is introduced to enhance the detection accuracy of small targets.The lightweight convolutional block attention module is introduced into the backbone and neck networks to focus on important features and suppress unnecessary features to boost the ability of small-target feature expression.The Soft-NMS algorithm is used to replace the NMS for reducing the miss detection rate of small targets in dense scenes.Experimental results on the VisDrone 2019 dataset show that the Yolov5_GBCS algorithm integrating all improved methods enhances the detection accuracy and effectively improves the detection speed.The mAP of the subject model has been increased from 38.5%to 43.2%,and the detection speed from 53 f/s to 59 f/s.Therefore,the Yolov5_GBCS algorithm can effectively recognize small targets in the image captured by UAV aerial photography.

关 键 词：图像处理 GhostConv卷积模块 双向特征金字塔网络 卷积块注意力模块 Soft双向特征金字塔网络 轻量化模型 小目标检测 VisDrone数据集 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] V19[自动化与计算机技术—计算机科学与技术]