改进SegNet与迁移学习的遥感建筑物分割方法

A building segmentation method for remote sensing with improved SegNet and transfer learning

针对传统SegNet应用于遥感影像建筑物分割出现分割不连续的问题,该文提出了一种改进的SegNet模型,并引入迁移学习方法,以提高遥感影像建筑物分割精度。以SegNet为基础,加入能够提取多尺度特征的改进空洞空间卷积池化金字塔模块,并引入跳层连接使分割结果更为精细。选取了FCN、SegNet、载入ImageNet预训练权重参数的SegNet作为对比算法,对遥感建筑物分割数据集Inria Aerial Image Labeling Dataset进行训练和测试。实验结果表明,在有限的迭代次数及实验区域内,该文算法拥有更好的分割效果和更强的泛化能力。

Aiming at the problem of discontinuous segmentation results when SegNet is applied to building segmentation of remote sensing image, an improved SegNet was proposed and a transfer learning method was introduced to improve the segmentation accuracy of building in remote sensing image. Based on SegNet, an improved atrous spatial pyramid pooling module was added to extract multi-scale features, and the skip-connected was introduced to refine the result of the building segmentation. In this paper, FCN,SegNet and the SegNet loaded ImageNet pre-trained weight parameter were selected as comparison algorithms to train and test the Inria Aerial Image Labeling Dataset of remote sensing buildings segmentation dataset. The experimental results showed that the proposed method had better segmentation effect and stronger generalization ability in the limited number of iterations and the experimental area.