摘要
针对现有很多深度学习的建筑物变化检测方法未考虑图像的结构特征,导致建筑物边缘像素分割精度低的问题,本文提出了一种增强边缘信息的遥感影像建筑物变化检测模型。首先采用Canny算法和概率霍夫变换算法提取双时相影像中建筑物的直线边缘特征图作为图像结构特征;然后将双时相影像及其对应的边缘特征图输入到增强边缘信息的全卷积神经网络(FCN)中;最后采用骰子损失和交叉熵损失加权组合函数衡量网络模型。试验表明,增强边缘信息的FCN网络在精度评价和视觉分析上具有一定的优越性。
For many existing deep learning building change detection methods,it is difficult to obtain image structure features,which leads to the problem of low segmentation accuracy of building edge pixels.In this paper,a building change detection model based on enhanced edge information in remote sensing images is proposed.Firstly,the Canny algorithm and the probabilistic Hough transform algorithm are used to extract the linear edge feature map of the building in the bitemporal image as the image structure feature.Then the bi-temporal images and their corresponding edge feature maps are input into a fully convolutional neural network(FCN)that enhances edge information.Finally,the weighted combination function of Dice Loss and CrossEntropy Loss is used to measure the network model.Experiments show that the FCN network with enhanced edge information has certain advantages in accuracy evaluation and visual analysis.
作者
陈婕
刘纪平
徐胜华
CHEN Jie;LIU Jiping;XU Shenghua(Chinese Academy of Surveying&Mapping,Beijing 100036,China)
出处
《测绘通报》
CSCD
北大核心
2023年第6期61-67,共7页
Bulletin of Surveying and Mapping
基金
国家重点研发计划(2020YFC1511704)。
