一种改进Unet网络的遥感影像分割算法  

Algorithm for segmentation of remote sensing imagery using the improved Unet

作　　者：李娇娇 刘志强 宋锐[1] 李云松[1] LI Jiaojiao;LIU Zhiqiang;SONG Rui;LI Yunsong(The State Key Laboratory of Integrated Services Networks,Xidian University,Xi’an 710071,China)

机构地区：[1]西安电子科技大学综合业务网理论及关键技术国家重点实验室,陕西西安710071

出　　处：《西安电子科技大学学报》2022年第6期67-75,128,共10页Journal of Xidian University

基　　金：国家自然科学基金青年基金(61901343);地理信息工程国家重点实验室基金(SKLGIE2020-M-3-1);中国博士后特别资助科学基金(2018T111019);空间智能控制实验室科学与技术基金(ZDSYS-2019-03);高等学校学科创新引智计划(111计划)(B08038);芜湖-西电产学研合作基金(XWYCXY-012021002-HT)。

摘　　要：现有的遥感影像分割算法将边缘信息和语义信息进行简单的结合,往往不能确保语义建模的整体改进。为解决此类问题,提出了基于改进Unet网络的遥感影像分割算法。该算法在基础编码器的基础上添加了边缘提取模块,此模块融合骨干网络所提取的语义特征信息以及由输入图像经过Canny算子和膨胀数学形态学操作获得的边缘特征信息,更好地学习了遥感影像的边缘。为了进一步获取遥感影像全局信息以提高分割精度,提出了边缘引导上下文聚合模块。该模块通过捕获边缘区域的像素和物体内部像素之间的长距离依赖关系,通过聚合上下文信息而加强类内一致性。在"天智杯"人工智能挑战赛数据集的测试下,改进后的模型总体准确度达到84.5%,平均交并比达到68.6%,精度与经典Unet模型相比分别提高了5.3%和9.2%。改进后的模型在ISPRS Vaihingen和Potsdam基准数据集上总体准确度分别达到了91.2%和91.6%,更适于精确的遥感影像分割。Existing remote sensing image segmentation algorithms simply combine edge information with semantic information, they often fail to ensure an overall improvement in semantic modelling.To solve such problems, an improved remote sensing image segmentation algorithm using Unet networks is proposed.The improved algorithm adds an edge extraction module to the base encoder module.The module fuses the semantic feature information of the backbone network and the boundary feature information obtained from the input image by Canny operator and dilated mathematical morphological operations to learn the edges of remote sensing image.To further acquire global information of remote sensing images for improving segmentation accuracy, an edge-guided context aggregation module is proposed.This module enhances the intra-class consistency by capturing the long-distance dependencies between pixels in the boundary region and pixels inside the object, and then aggregates the contextual information.Under the test of the "Tianzhi Cup" AI Challenge dataset, the overall accuracy of the improved model reached about 84.5% and the average intersection ratio reached about 68.6%,with an accuracy improvement of 5.3% and 9.2% respectively compared with the Unet model.The improved model achieved an overall accuracy of 91.2% and 91.6% on the ISPRS Vaihingen and Potsdam benchmark datasets respectively, making it more suitable for accurate remote sensing image segmentation.

关 键 词：语义分割 CANNY算子 数学形态学操作 深度学习 图像处理 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]