拉普拉斯卷积的双路径特征融合遥感图像智能解译方法  

Dual-path feature fusion remote sensing image intelligent interpretation method by Laplacian convolution

作　　者：曾军英[1] 顾亚谨 曹路 秦传波[1] 邓森耀 翟懿奎[1] 甘俊英[1] 谢梓源 ZENG Junying;GU Yajin;CAO Lu;QIN Chuanbo;DENG Senyao;ZHAI Yikui;GAN Junying;XIE Ziyuan(School of Electronics and Information Engineering,Wuyi University,Jiangmen 529020,China)

机构地区：[1]五邑大学电子与信息工程学院,广东江门529020

出　　处：《现代电子技术》2024年第17期65-72,共8页Modern Electronics Technique

基　　金：国家自然科学基金项目(61771347);广东普通高校人工智能重点领域项目(2019KZDZX1017);广东普通高校重点领域专项(2020ZDZX3031);广东省基础与应用基础研究项目(2021A1515011576);江门市基础与理论科学研究类科技计划项目(江科[2021]87号2021030101900004391,江科[2022]110号2022JC01022);广东省科技创新战略专项资金(攀登计划专项资金)(pdjh2022b0530)。

摘　　要：由于遥感图像存在多尺度变化和目标边缘模糊等问题,对其进行智能解译仍然是一项极具挑战性的工作。传统的语义分割方法在处理这些问题时存在局限性,难以有效捕捉全局和局部信息。针对上述问题,文中提出一种双路径特征融合分割方法 DFNet。首先,使用Swin Transformer作为主干提取全局语义特征,以处理像素之间的长距离依赖关系,从而促进对图像中不同区域相关性的理解;其次,将拉普拉斯卷积嵌入到空间分支,以捕获局部细节信息,加强目标地物边缘信息表达;最后,引入多尺度双向特征融合模块,充分利用图像中的全局和局部信息,以增强多尺度信息的获取能力。在实验中,使用了三个公开的高分辨率遥感图像数据集进行验证,并通过消融实验验证了所提模型不同模块的作用。实验结果表明,所提方法在Uavid数据集、Potsdam数据集、LoveDA数据集的mIoU达到了71.32%、85.58%、54.01%,提高了语义分割的性能,使分割结果更为精细。Due to the multi-scale variations of images and blurred target edges,the intelligent interpretation of remote sensing images is still a challenging task.The traditional semantic segmentation methods have limitations in dealing with these problems and fail to capture global and local information effectively.Therefore,a dual-path feature fusion segmentation method DFNet is proposed.The Swin Transformer is used as the backbone to extract global semantic features,so as to facilitate addressing long-range dependencies between pixels,thereby enhancing the comprehension of correlations in different regions of the image.The Laplacian convolution is incorporated into the spatial branch to capture local detail information,so as to strengthen the expression of edge information for target features.A multi-scale bidirectional feature fusion module is introduced to leverage both global and local information within the image,so as to enhance the capability to capture multi-scale information.In the experiments,three publicly available high-resolution remote sensing image datasets are used for validation.The ablation experiments are utilized to validate the roles of different modules of the proposed model.Experiment results show that the mIoU(mean intersection over union)of the proposed method reaches 71.32%,85.58%,and 54.01%in dataset Uavid,dataset Potsdam and dataset LoveDA,respectively.To sum up,the proposed method can improve the performance of semantic segmentation and make the segmentation result more refined.

关 键 词：语义分割 遥感图像 多尺度信息 拉普拉斯卷积 边缘信息 双路径 特征融合 智能解译 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP751[电子电信—信息与通信工程]