基于递归特征与边缘解耦的遥感图像语义分割  

Semantic Segmentation of Remote Sensing Images Based on Recursive Features and Edge Decoupling

作　　者：王学伟 WANG Xue-wei(Datang Northeast Electric Power Test Research Institute,Changchun 130102,China)

机构地区：[1]大唐东北电力试验研究院有限公司,吉林长春130102

出　　处：《计算机技术与发展》2025年第3期40-48,共9页Computer Technology and Development

基　　金：吉林省科技发展计划项目(20230203177SF)。

摘　　要：图像语义分割是实现遥感图像智能解译的关键技术。然而,在面对复杂的遥感图像时,传统图像语义分割方法对于弱特征目标的分割仍然存在一定局限性,尤其受到遥感图像目标边缘混叠的影响,导致对边缘细节的处理相对粗糙。因此,该文提出了一种基于递归特征与边缘解耦的遥感图像语义分割模型。首先,根据特征复用和跨层连接思想,设计了一个编解码结构的层级递归特征网络,以增强对弱特征的提取能力。其次,结合多尺度融合预测和边缘解耦方式,通过融合低、高级特征图,深化对细节的处理,并引入目标本体与边缘的联系形成边缘监督,从而实现对边缘细节的精细化处理。最后,在ISPRS提供的Vaihingen和Potsdam两个数据集上进行了消融和对比实验。实验结果表明,该语义分割模型能够较好地保持目标内部区域的一致性,并在分割效果上实现了对边缘细节的精细化处理,有效提高了遥感图像语义分割的精度。Semantic segmentation of images is a crucial technology for intelligent interpretation of remote sensing images.However,traditional semantic segmentation methods still face limitations in segmenting weak feature targets,especially when dealing with complex remote sensing images where the edges of targets are often mixed,resulting in relatively coarse handling of edge details.Therefore,we propose a remote sensing image semantic segmentation model based on recursive features and edge decoupling.Firstly,according to the ideas of feature reuse and cross-layer connections,we design a hierarchical recursive feature network in an encoder-decoder structure,aiming to enhance the extraction capability of weak features.Secondly,by combining multi-scale fusion prediction and edge decoupling,the model merges low-and high-level feature maps,deepening the processing of details.It introduces edge supervision by establishing a connection between the target body and the edge,achieving refined handling of edge details.Finally,we conduct ablation and comparative experiments on the Vaihingen and Potsdam datasets provided by ISPRS.The experimental results demonstrate that the proposed semantic segmentation model effectively maintains the internal consistency of targets and achieves refined processing of edge details in segmentation,which significantly improves the accuracy of remote sensing image semantic segmentation.

关 键 词：语义分割 遥感图像 递归特征 多尺度特征融合 边缘解耦结构 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]