一种结合语义分割模型和图割的街景影像变化检测方法  

A street view image change detection method combining semantic segmentation model and graph cuts

作　　者：李文国 黄亮[1,2] 左小清 王译著[1] LI Wenguo;HUANG Liang;ZUO Xiaoqing;WANG Yizhu(Faculty of Land Resource Engineering,Kunming University of Science and Technology,Kunming 650093,China;Surveying and Mapping Geo-Informatics Technology Research Center on Plateau Mountains of Yunnan Higher Education,Kunming 650093,China)

机构地区：[1]昆明理工大学国土资源工程学院,昆明650093 [2]云南省高校高原山区空间信息测绘技术应用工程研究中心,昆明650093

出　　处：《全球定位系统》2021年第1期98-104,共7页Gnss World of China

基　　金：国家自然科学基金(41961039,41961053);云南省应用基础研究计划面上项目(2018FB078);云南省高校工程中心建设计划资助课题。

摘　　要：由于街景影像具有地物尺度多样化、地物界限不明确、地物光谱信息复杂等问题,造成应用统计方法、机器学习等方法对复杂度高的街景影像变化检测性能欠佳.因此提出一种结合语义分割模型和图割(GC)的街景影像变化检测方法.该方法首先采用Camvid数据集训练DeeplabV3+网络得到的迁移学习模型对两个时期的街景影像进行语义分割;然后采用GC方法实现消除天空和植被等对街景影像的影响;接着采用变化向量分析(CVA)获取差异影像,最后对差异影像进行二值化和精度评价.研究结果表明,提出的方法总体精度优于大津法(OTSU)、K均值法、Segnet网络迁移学习模型方法和DeeplabV3+网络迁移学习模型方法,是一种可行的街景影像变化检测方法.Due to the problems of the diversity of the scale,the unclear boundary and the complex spectral information of the ground objects,the performance of the statistical method and machine learning method for the change detection of the high complexity street view images is poor.Therefore,a street view images change detection method combining semantic segmentation model and graph cuts(GC)is proposed in this paper.Firstly,the DeeplabV3+semantic segmentation model combined with migration learning is used to pre-train the Camvid data set to obtain the pre-trained model in this method;Then a small number of annotated samples from the data set of this paper were used to Fine Tune the pre-training model,which was respectively used for semantic segmentation of street view images in two periods.Then GC method is used to remove the sky,roads,vegetation and other factors,which is impacting on the street view.Finally,change vector analysis(CVA)is used to obtain the difference images,and binarization and accuracy evaluation were carried out for the difference images.The results show that the overall accuracy of the proposed method is better than the Otsu method(OTSU),K-means method,Segnet network migration learning method and DeeplabV3+network migration learning method,it is a feasible method for detecting changes in street view images.

关 键 词：DeeplabV3+网络 图割(GC) 变化检测 迁移学习 变化向量分析(CVA) 

分 类 号：P232[天文地球—摄影测量与遥感]