基于微调语义分割模型的街景影像变化检测方法  

A Change Detection Method of Street View Image Based on Fine Tuning Semantic Segmentation Model

作　　者：李文国 黄亮[1,2] 左小清 王译著[1] LI Wen-guo;HUANG Liang;ZUO Xiao-qing;WANG Yi-zhu(Faculty of Land Resource Engineering,Kunming University of Science and Technology;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 [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,云南昆明650093

出　　处：《软件导刊》2021年第2期200-205,共6页Software Guide

基　　金：国家自然科学基金项目(41961039,41961053)。

摘　　要：利用传统统计方法与机器学习方法对街景影像进行变化检测时,存在精度低、鲁棒性差和泛化能力弱等问题。因此提出一种微调的DeeplabV3+网络街景影像变化检测方法。首先将街景数据集和Camvid数据集合为联合数据集,用于微调的DeeplabV3+网络训练;然后将训练得到的模型用于街景影像分类,并采用变化向量分析获取差异影像;最后对差异影像进行二值化和精度评价。研究结果表明,该方法正确率比大津法、K均值法提升25%以上,同时具有较强的鲁棒性和泛化能力,是一种可行的街景影像变化检测方法。Traditional statistical methods and machine learning methods for street view image change detection are prone to problems such as low accuracy,poor robustness and weak generalization ability.To solve these problems,a method is proposed to solve the problem based on fine-tuned DeeplabV3+network for street view image change detection.The method first combines the street view data set and the Camvid data set to form a joint data set,and uses the data set for the fine-tuned DeeplabV3+network training,then the trained model is used to classify street view images,the change vector analysis is used to obtain the difference images.Finally,the difference image is binarized and the accuracy is evaluated.The research results show that the accuracy of this method is more than 25%higher than that of Otsu method and K-means method,and it has strong robustness and generalization ability.It is a feasible image change detection method for street view.

关 键 词：街景 变化检测 DeeplabV3+ 

分 类 号：TP317.4[自动化与计算机技术—计算机软件与理论]