摘要
遥感图像的语义分割在土地资源利用、城市规划和军事侦察等很多领域都有广泛的应用。文中针对遥感图像数据集样本分布不均衡、小目标物体难以识别的问题,提出一种结合边缘检测特征细化的轻量级双分支结构语义分割方法。该方法主体对空间细节分支和语义分支建立动态特征关联,实现双分支特征的协同增强与自适应聚合;设计边缘特征细化分支,通过残差跨层连接和特征细化方法的添加,在保持语义一致性的同时,进一步提升分割边界的锐度。对比实验结果有效证明了所提方法的分割精度有明显提高,小目标物体的识别度更高。提出的模型在数据集Vaihingen上的平均像素精度(mPA)和平均交并比(mIoU)两个指标分别提高3.83%和2.74%,证明了所提方法在遥感图像语义分割领域的适用性。
Semantic segmentation of remote sensing images has a wide range of applications in fields such as land resource utilization,urban planning,and military reconnaissance.This paper proposes a lightweight dual⁃branch structure semantic segmentation method combined with edge detection feature refinement.This method strives to avoid imbalanced sample distribution in remote sensing image datasets and cope with the difficulties in identifying small objects.The main body of the method is responsible for establishing dynamic feature associations between the spatial detail branch and the semantic branch,achieving collaborative enhancement and adaptive aggregation of dual⁃branch features.An edge feature refinement branch is designed,and the residual cross⁃layer connections and feature refinement modules are designed,which enhances the sharpness of segmentation boundaries while maintaining semantic consistency.The results of comparative experiments effectively demonstrate that the proposed method can improve segmentation accuracy and recognizability of small objects significantly.The proposed model achieves improvements of 3.83%in mean pixel accuracy(mPA)and 2.74%in mean intersection over union(mIoU)on the Vaihingen dataset,which proves the applicability of the proposed method in the field of semantic segmentation of remote sensing images.
作者
李秀娟
刘中胜
LI Xiujuan;LIU Zhongsheng(School of Information,Xi’an University of Finance and Economics,Xi’an 710100,China;School of Computer Science and Engineering,Xi’an University of Technology,Xi’an 710048,China)
出处
《现代电子技术》
北大核心
2025年第19期25-30,共6页
Modern Electronics Technique
基金
全国统计科学研究项目(2024LY095)
陕西省自然科学基础研究计划资助项目(2024JC-YBQN-0261)。
关键词
遥感图像
双分支结构
语义分割
动态特征关联
交叉注意力
特征融合
remote sensing image
dual⁃branch structure
semantic segmentation
dynamic feature association
cross attention
feature fusion
