基于改进pix2pix的遥感图像语义分割  

Remote Sensing Image Semantic Segmentation Based on Improved Pix2pix

作　　者：黄聪 周坤 安学刚 HUANG Cong;ZHOU Kun;AN Xuegang(School of Electronic and Infornation Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China;Guohua Satelite Data Technology Co.,Ltd.,Lanzhou 730050,China)

机构地区：[1]兰州交通大学电子与信息工程学院,兰州730070 [2]国华卫星数据科技有限公司,兰州730050

出　　处：《兰州交通大学学报》2023年第5期57-65,共9页Journal of Lanzhou Jiaotong University

摘　　要：针对高分辨率遥感图像语义分割面临的边界模糊、错分和漏分割问题,提出一种改进的pix2pix模型用于遥感图像语义分割。首先,在分割网络编码中,通过引入空洞空间金字塔池化模块来提取不同尺度的图像上下文特征;然后,在跳跃连接阶段引入空间注意力机制对地物边缘细节信息进行增强,以提高网络模型对地物的分割能力。将所提出的方法在Vaihingen和Gaofen Image Dataset(GID)数据集上进行语义分割仿真实验,mIoU分别达到82.69%和81.27%。仿真实验结果表明:所提出的方法充分地利用了上下文信息,有效减少了错误分割,使分割边界更清晰,优于其它经典语义分割方法。An improved pix2pix model is proposed to address the issues of boundary blur,misclassification,and missed segmentation in high-resolution remote sensing image semantic segmentation.Firstly,in the segmentation network coding,the hollow space pyramid pooling module is introduced to extract contextual features of images at different scales;Then,in the skip connection stage,a spatial attention mechanism is introduced to enhance the edge detail information of ground objects,in order to improve the network model's segmentation ability for ground objects.The proposed method was applied to semantic segmentation simulation experiments on Vaihingen and Gaofen Image Dataset(GID)datasets,and the mIoU reached 82.69%and 81.27%,respectively.The simulation experimental results show that the proposed method fully utilizes contextual information,effectively reduces erroneous segmentation,makes the segmentation boundary clearer,and is superior to other classical semantic segmentation methods.

关 键 词：高分辨率遥感图像 语义分割 pix2pix 空间注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]