摘要
针对无人机航拍图像存在一些较小目标、目标相互遮挡等问题,提出一种基于注意特征重构的无人机航拍检测算法。设计密集目标双向特征金字塔网络,将不同尺度的特征进行融合,获得更丰富的语义信息和空间上下文,使目标特征信息利用得更充分。设计最大坐标注意力,其采用最大池化,不仅保留了目标的主要特征,而且有助于增大网络的感受野,构建SiLU激活函数,可以减少梯度消失问题,使算法更准确地定位目标。设计高效分组卷积,采用分组卷积将输入通道分成多组,组内独立进行卷积操作,并且使用更小的卷积核尺寸,从而减少了参数量和计算量。采用WIoUv3作为回归损失函数,引入动态非单调机制以及使用合理的梯度增益分配策略,提高了算法的目标定位能力。在DOTA和VisDrone2019数据集上的实验结果显示,该算法的mAP分别为59.70%和43.07%,相比YOLOv8分别提高了4.25%和4.87%,满足实时检测的要求。
In order to solve the problems of small objects and mutual occlusion in UAV aerial images,a new detection algorithm based on attention feature reconstruction is proposed.A dense bidirectional feature pyramid network is designed to integrate features of different scales to obtain richer semantic information and spatial context,so as to make full use of target feature information.The design of maximum coordinate attention,which adopts maximum pooling,not only retains the main features of the target,but also helps to increase the receptive field of the network.The construction of SiLU activation function can reduce the problem of gradient disappearance and enable the algorithm to locate the target more accurately.Efficient grouping convolution is designed.The input channels are divided into multiple groups by grouping convolution,and the convolution operation is carried out independently within the groups,and the convolution kernel size is smaller,thus reducing the number of parameters and the calculation amount.WIoUv3 is used as regression loss function,dynamic non-monotonic mechanism is introduced and a reasonable gradient gain allocation strategy is used to improve the target location ability of the algorithm.Experimental results on DOTA and VisDrone2019 data sets show that the mAP of the algorithm is 59.70%and 43.07%,respectively,which is 4.25%and 4.87%higher than that of YOLOv8,meeting the requirements of real-time detection.
作者
李莉
夏晨博
王文鼎
武芳芳
LI Li;XIA Chenbo;WANG Wending;WU Fangfang(School of Information and Electrical Engineering,Hebei University of Engineering,Handan 056038,China;Department of Computer Science,Handan Vocational College of Science and Technology,Handan 056000,China)
出处
《兵器装备工程学报》
北大核心
2025年第8期309-318,共10页
Journal of Ordnance Equipment Engineering
基金
河北省教育厅科学研究基金项目(SQ2023096)
邯郸市科学技术研究与发展计划基金项目(21422031289)。
作者简介
李莉(1984-),女,博士,讲师,硕士生导师,E-mail:hdlili@126.com;通信作者:夏晨博(1998-),男,硕士研究生,E-mail:xiachenbo16@163.com。
