摘要
针对遥感图像目标检测中存在图像分辨率低、小目标特征信息不足以及检测难度大等问题,提出一种基于改进YOLO11的遥感图像目标检测算法SO-YOLO。首先,设计一个新的卷积神经网络(CNN)构建块SDOD-Conv,由空间到深度转换层和全维动态卷积组成,代替主干网络中的跨步卷积和池化层,加强特征提取,在特征提取过程中避免细粒度信息损失;其次,在颈部网络中引入空间和通道重建卷积(SCConv),压缩特征之间的空间和通道冗余,减少冗余计算并促进代表性特征学习;最后,采用Inner-IoU损失函数作为回归损失,通过引入比例因子的辅助边界框计算IoU损失,获得更快、更准确的回归结果。在HRSC2016数据集和DOTA数据集上的实验结果表明,相较于YOLO11,改进后算法的平均精度均值分别提高了2.3%和2.0%,表明改进算法具有良好的检测性能。
To solve the problems of low image resolution and insufficient feature information of small targets and difficulty of target detection in remote sensing image,a remote sensing image detection method based on SO-YOLO is proposed on the basis of YOLO11 model.First,a new convolutional neural network(CNN)building block SDOD-Conv is designed,which consists of a spaceto-depth layer and a full-dimensional dynamic convolution,replacing each stride convolution layer and pooling layer in the backbone network,to enhance the feature extraction and avoid the loss of fine-grained information during the feature extraction process.Second,Introducing Spatial and Channel Reconstruction Convolution(SCConv)into the neck network to compress spatial and channel redundancies in features,redundant computation is reduced and discriminative feature learning is enhanced.Finally,the Inner-IoU loss function is used as the regression loss,and the IoU loss is computed by introducing an auxiliary bounding box for the scale factor to obtain faster and more efficient regression results.The improved algorithm is verified on the HRSC2016 data set and DOTA data set,which are improved by 2.1% and 1.5% respectively,compared to the average accuracy of YOLO11,and demonstrates good detection performance.
作者
王佳
李芳
WANG Jia;LI Fang(Shenyang Ligong University,Shenyang 110159,China)
出处
《沈阳理工大学学报》
2026年第2期47-52,60,共7页
Journal of Shenyang Ligong University
基金
国家自然科学基金项目(62102272)。
关键词
遥感图像
目标检测
YOLO11
卷积模块优化
remote sensing images
target detection
YOLO11
convolution module optimization
