动态聚焦多维注意力遥感弱小目标检测  

Dynamic Focusing Multidimensional Attention Remote Sensing Weak Target Detection

作　　者：张朝阳 张上 胡益民 张岳 熊偌炎 ZHANG Zhaoyang;ZHANG Shang;HU Yimin;ZHANG Yue;XIONG Ruoyan(College of Electrical Engineering&New Energy,China Three Gorges University,Yichang 443002,China;Hubei Province Engineering Technology Research Center for Construction Quality Testing Equipments,Yichang 443002,China;Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydroelectric Engineering,Yichang 443002,China;College of Computer and Information Technology,China Three Gorges University,Yichang 443002,China)

机构地区：[1]三峡大学电气与新能源学院,湖北宜昌443002 [2]湖北省建筑质量检测装备工程技术研究中心,湖北宜昌443002 [3]水电工程智能视觉监测湖北省重点实验室,湖北宜昌443002 [4]三峡大学计算机与信息学院,湖北宜昌443002

出　　处：《无线电通信技术》2025年第1期196-209,共14页Radio Communications Technology

基　　金：国家级大学生创新创业训练计划(202111075012,202011075013)。

摘　　要：遥感图像的目标检测在各领域有着广泛的应用场景,由于遥感图像中检测目标存在形态多变、弱小目标较多以及背景复杂等原因,导致遥感图像在目标检测方面存在检测精度识别不高、模型参数过大等问题。为提升算法对遥感图像目标的检测准确率以及缩减算法模型量,提出了动态聚焦多维注意力检测算法——YOLO-WiseGOD。在YOLOv8n基线网络中使用WIoU(Wise-IoU)构建动态聚焦机制的边界框损失,弱化因几何因素导致的梯度增益泛化能力不足的问题,在协调高低质量锚框竞争力的同时,使之适用于聚焦普通锚框,提高网络模型检测的定位能力。在网络末端融合改进的L-ODConv(Leaky ReLU-Omni-Dimensional Dynamic Convolution)多维注意力机制,避免梯度锯齿问题,在减少模型参数的同时,优化输出特征和卷积权值的调制,提升网络加权特征融合。在主干网络中引入轻量化注意力模块C2FGhostV2,在更好地捕捉输入特征图中的多尺度特征和全局上下文信息的同时,保持较低的参数量和计算复杂度,更好地平衡训练精度和模型量之间的关系。所提算法在遥感数据集NWPU VHR-10(Northwestern Polytechnical University Very High Resolution-10)和RSOD(Remote Sensing Object Detection)上进行实验验证,对比当前主流算法模型YOLOv8n,其平均检测准确率(mean Average Precision,mAP)分别提高了2.0%和2.3%,模型参数量减少4.5%,计算量减少10.9%,能有效提高遥感图像中微小目标的检测精度和实现一定的模型轻量化。The identification of targets in remote sensing pictures has a wide range of possible applications in various fields.The existence of variable morphology in remote sensing images,coupled with the existence of numerous faint targets and complex background,leads to challenges such as reduced detection accuracy and an excessive number of model parameters in target detection.This paper presents the YOLO-WiseGOD technique,which seeks to improve the precision of target detection in remote sensing photos while reducing the reliance on several computational models.The technology accomplishes this by utilizing a dynamically focused multidimensional attention strategy.The WIoU(Wise-IoU)approach is utilized to generate the bounding box loss for the dynamic focusing mechanism in the YOLOv8n baseline network.This approach tackles the problem of limited ability to apply general rules due to geometric considerations,and finds a balance between the significance of high-and low-quality anchor frames.Additionally,it improves the network model s capacity to detect and locate objects by specifically emphasizing regular anchor frames.Afterwards,the improved Leaky ReLU-Omni-Dimensional Dynamic Convolution(L-ODConv)multidimensional attention mechanism is incorporated at the end of the network to address the problem of gradient irregularity and optimize the adjustment of output characteristics and convolutional weights.This enhances the combination of weighted network features while reducing the number of model parameters.The incorporation of C2FGhostV2,a compact attention module,into the primary network improves its capacity to capture features at various scales and global contextual information inside the input feature maps.This is accomplished by simultaneously minimizing the number of parameters and computational complexity,while efficiently managing the trade-off between training accuracy and model size.The algorithm described in this research is evaluated using two remote sensing datasets,Northwestern Polytechnical University Very

关 键 词：遥感图像 弱小目标检测 YOLOv8 动态聚焦机制 多维注意力 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]