基于多尺度特征融合模型的遥感图像建筑物分割

Building Segmentation of Remote Sensing Images based on Multiscale- feature Fusion Model

摘要:

针对传统深度网络模型难以精确提取建筑物边缘轮廓特征及对不同尺寸建筑物无法自适应提取的问题,提出一种膨胀卷积特征提取的多尺度特征融合深度神经网络模型(Multiscale-feature fusion Deep Neural Networks with dilated convolution,MDNNet)对遥感图像建筑物自动分割的方法。首先在ResNet101模型中引入膨胀卷积扩大提取视野保留更多特征图像分辨率；其次利用多尺度特征融合模块获取多个尺度的建筑物特征并将不同尺度的特征融合；最终利用特征解码模块将特征图恢复到原始输入图像尺寸,实现遥感图像建筑物精确分割。在WHU遥感图像数据集的实验结果表明,提出模型有效克服道路、树木和阴影等因素影响,分割结果有效保留建筑物边界细节信息,有效提升分割精度,像素准确率PA达到0.864,平均交并比mIoU达到0.815,召回率Recall达到0.862。

Abstract:

Towards how to solve the problem that traditional deep neural networks model is difficult to accurately extract the edge contour features of buildings and cannot adaptively extract buildings of different sizes, a method of automatic segmentation of remote sensing image buildings on account of Multiscale-feature fusion Deep Neural Networks with dilated convolution (MDNNet) is proposed. To begin with, expansion convolution is introduced into ResNet101 model to expand the extraction field and preserve more feature image resolution. Secondly, multiscale feature fusion module is used to obtain building features of multiple scales and fuse features of different scales. Eventually, the feature decoding module is used to restore the feature image to the original input image size, thus realizing accurate segmentation of remote sensing image buildings. The experimental results on WHU Building change detection dataset show that the proposed model effectively overcomes the influence of road, trees and shadows, and the segmentation results effectively retain the detailed information of building boundaries and improve the segmentation accuracy. The pixel accuracy PA comes to 0.864, the mean Intersection over Union mIoU comes to 0.815 and the Recall rate comes to 0.862.