用于遥感图像语义分割的多重预测网络  

Multi-Prediction Network for Semantic Segmentation of Remote Sensing Images

作　　者：倪家辉 周激流[2] Ni Jiahui;Zhou Jiliu(College of Electronics and Information Engineering,Sichuan University,Chengdu 610065;College of Computer Science,Sichuan University,Chengdu 610065)

机构地区：[1]四川大学电子信息学院,成都610065 [2]四川大学计算机学院,成都610065

出　　处：《现代计算机》2022年第7期87-91,共5页Modern Computer

摘　　要：现如今,越来越多的遥感卫星广泛地用于监测地质环境、气候环境以及人类活动环境的变化,对遥感图像的处理和分析也越来越重要,其中的一种重要研究方向便是遥感图像的语义分割。然而,由于遥感图像多波段及地貌特征的复杂性,基于传统方法的语义分割方法并不能有效地处理遥感图像。随着近年来深度学习的发展,为遥感图像的处理分析带来新的思路,通过训练端到端的卷积神经网络模型能够快速地对遥感图像的地貌类别进行分类。本文提出了一种基于卷积神经网络的多重亲和图监督并预测的网络,利用骨干网络提取遥感图像不同深度的语义特征,并对相应的特征进行亲和图预监督,最后整合不同深度的预测结果,能够增强长短距离像素的特征表示,实现有效的语义分割。此外,使用不同数据集进行实验,对比了常见的深度学习语义分割方法。实验结果表明,本文提出的方法能够有效提升遥感图像语义分割的精度。Nowadays,more and more remote sensing satellites are widely used to monitor changes in the geological environ⁃ment,climate environment,and human activity environment.The processing and analysis of remote sensing images are becoming more and more important.One of the important research directions is Semantic segmentation of remote sensing images.However,due to the complexity of remote sensing images with multiple bands and topographic features,semantic segmentation methods based on traditional methods cannot effectively process remote sensing images.With the development of deep learning in recent years,new ideas have been brought to the processing and analysis of remote sensing images.The end-to-end convolutional neural network model can be trained to quickly classify the topography categories of remote sensing images.This paper proposes a multiaffinity map supervision and prediction network based on a convolutional neural network.The backbone network is used to extract the semantic features of different depths of remote sensing images,and the corresponding features are pre-supervised by affinity maps,and finally the integration of different depths The prediction result can enhance the feature representation of long and short distance pixels and realize effective semantic segmentation.In addition,experiments with different data sets are used to compare common deep learning semantic segmentation methods.Experimental results show that the method proposed in this paper can effec⁃tively improve the accuracy of remote sensing image semantic segmentation.

关 键 词：深度学习 遥感图像 语义分割 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]