多尺度特征融合的建筑物及损毁语义分割

Semantic segmentation of building and damage based on multi-scale feature fusion

遥感影像分割作为影像信息提取过程中的关键步骤,近年来基于深度学习的影像语义分割模型已经成为影像分割的主要研究导向。文中提出一种基于深度学习的多尺度特征融合语义分割网络,用来分割遥感影像中建筑物和损毁建筑物,该网络充分利用不同尺度特征图的信息,获得更精确的分割边缘。同时探究了不同样本数量和不同网络深度对于训练得到模型分割性能的影响,对深度学习网络应用于遥感影像参数选择提供了一定经验指导。

Remote sensing image segmentation is a key step in the process of image information extraction.In recent years,image semantic segmentation model based on depth learning has become the main research direction of image segmentation.A multi-scale feature fusion semantic segmentation network based on deep learning is proposed to segment buildings and damaged buildings in remote sensing images in this paper.The network can obtain more accurate segmentation edges by making full use of the information of different scale feature maps.At the same time,the influence of different sample sizes and different network depths on the segmentation performance of the training model is explored.It provides some experience guidance for the deep learning network to be used in the processing of remote sensing images.