摘要
为了解决遥感图像语义分割任务中上下文依赖关系提取不足、空间细节信息损失导致分割精度下降等问题,提出了一种结合上下文与类别感知特征融合的语义分割方法。该方法首先以ResNet-50作为特征提取的主干网络,并在下采样中采用注意力模块,以增强特征表示和上下文依赖关系的提取;然后在跳跃连接上构建大尺寸的感受野块,提取丰富的多尺度上下文信息,以减少目标之间尺度变化的影响;其后并联场景特征关联融合模块,以全局特征来引导局部特征融合;最后在解码器部分构建类别预测模块和类别感知特征融合模块,准确融合底层的高级语义信息与高层的细节信息。将所提方法在Potsdam和Vaihingen数据集上验证可行性,并与DeepLabv3+,BuildFormer等6种常用方法进行对比实验,以验证其先进性。实验结果表明,所提方法在Recall, F1-score和Accuracy指标上均优于其他方法,尤其是对建筑物分割的交并比(intersection over union, IoU)在2个数据集上分别达到90.44%和86.74%,较次优网络DeepLabv3+和A2FPN分别提升了1.55%和2.41%。
To address the accuracy reduction in the semantic segmentation of remote sensing images due to insufficient extraction of contextual dependencies and loss of spatial details,this study proposed a semantic segmentation method based on context-and class-aware feature fusion.With ResNet-50 as the backbone network for feature extraction,the proposed method incorporates the attention module during downsampling to enhance feature representation and contextual dependency extraction.It constructs a large receptive field block on skip connections to extract rich multiscale contextual information,thereby mitigating the impacts of scale variations between targets.Furthermore,it connects a scene feature association and fusion module in parallel behind the block to guide local feature fusion based on global features.Finally,it constructs a class prediction module and a class-aware feature fusion module in the decoder part to accurately fuse the low-level advanced semantic information with high-level detailed information.The proposed method was validated on the Potsdam and Vaihingen datasets and compared with six commonly used methods,including DeepLabv3+and BuildFormer,to verify its effectiveness.Experimental results demonstrate that the proposed method outperformed other methods in terms of recall,F1-score,and accuracy.Particularly,it yielded intersection over union(IoU)values of 90.44%and 86.74%for building segmentation,achieving improvements of 1.55%and 2.41%,respectively,compared to suboptimal networks DeepLabv3+and A2FPN.
作者
何晓军
罗杰
HE Xiaojun;LUO Jie(College of Software,Liaoning Technical University,Huludao 125105,China)
出处
《自然资源遥感》
北大核心
2025年第2期1-10,共10页
Remote Sensing for Natural Resources
基金
辽宁省教育厅科学研究经费项目“基于智能多主体的并行化海量遥感影像分割方法研究”(编号:LJKZ0350)资助。
关键词
类别感知
语义分割
遥感图像
上下文信息
特征融合
class-aware
semantic segmentation
remote sensing image
contextual information
feature fusion
作者简介
第一作者:何晓军(1975-),男,博士,副教授,主要从事遥感影像处理、人工智能、大数据处理等方面的研究。Email:hexiaojun@lntu.edu.cn;通信作者:罗杰(1995-),男,硕士研究生,研究方向为遥感图像处理。Email:1349876941@qq.com。
