面向高分辨率建筑物影像的图像分割方法研究

Research on Image Segmentation Methods for High Resolution Building Images

为解决城市近景影像中建筑物的分割问题,有效提高图像分割效率,针对不同纹理、不同摄影距离的建筑物,开展了分割方法的确立及精度分析工作。利用传统方法及基于深度学习的图像分割方法对多组建筑物影像进行了图像分割实验,并进行了可视化及分割准确度分析。结果表明,传统图像分割方法计算量较小,处理速度较快,对于门窗、管道、墙面3类标志的分割平均准确率分别为36.8%、33.3%、53.0%,但受噪声影响较大;基于深度学习的方法对3类目标的分割准确率分别为93.3%、76.7%、80.8%,分割效果及准确率优于传统方法,但需要训练大量的参数及较强的计算资源支持。

To solve the segmentation problem of buildings in urban close range images and effectively improve image segmentation efficiency,the establishment of segmentation methods and accuracy analysis were carried out for buildings with different textures and different photographic distances.Conducted image segmentation experiments on multiple sets of building images using traditional methods and deep learning based image segmentation methods,and conducted visualization and segmentation accuracy analysis.The results have show that traditional image segmentation methods have lower computational complexity and faster processing speed.The average accuracy of segmentation for three types of signs is doors,windows,pipes,and walls is 36.8%,33.3%,and 53.0%,respectively.However,they are greatly affected by noise.The segmentation accuracy of deep learning based methods for three types of targets is 93.3%,76.7%,and 80.8%,respectively.The segmentation effect and accuracy are better than traditional methods,but it requires training a large number of parameters and strong computational resource support.