一种融合注意力机制与边缘计算的遥感影像车辆检测算法

A Vehicle Detection Algorithm Based on Remote Sensing Image Combining Attention Mechanism and Edge Computing

针对遥感影像目标检测任务中存在的背景复杂、小目标检测困难、大尺寸影像导致的检测效率低等问题,提出了一种融合注意力机制与边缘计算的遥感影像车辆检测算法。首先在YOLOv5的主干网络中引入多维协作注意机制(Multidimensional Collaborative Attention,MCA),然后提出多分支卷积模块替换特征融合网络中的CBL模块,最后将训练的网络权重部署到Atlas 200开发者套件(Developer Kit,DK)边缘计算设备。在UCAS-AOD数据集上进行模型验证与对比,实验结果表明,该网络模型在复杂环境下对车辆的检测精度达到94.9%,较YOLOv5提升8.2%;网络部署到边缘设备后速度较GPU设备提升258%。同时,该方法也可以扩展应用到其他遥感影像目标检测任务中,为相关领域提供了一种高效、准确的检测方法。

To address the challenges posed by complex backgrounds,difficulty in detecting small targets,and low detection efficiency due to large-size images in the task of remote sensing image target detection,a vehicle detection algorithm for remote sensing images that integrates the attention mechanism and edge computing is proposed.Firstly,the multidimensional collaborative attention mechanism(MCA)is introduced into the backbone network of YOLOv5.Subsequently,a multi-branch convolutional module is proposed to replace the CBL module in the feature fusion network.Finally,the trained network weights are deployed to the Atlas 200 Developer Kit(DK)edge computing device.The proposed model is validated and compared against the UCAS-AOD dataset.The experimental results show that the network model achieves 94.9% vehicle detection accuracy in complex environments,an improvement of 8.2% over YOLOv5;the speed of the network when deployed to edge devices is 258% higher than that of GPU devices.Importantly,this approach can be extended to address other remote sensing image target detection tasks,thereby offering an efficient and accurate method for related fields.