基于改进U-Net的建筑物集群识别研究

Research on building cluster identification based on improved U-Net

针对U-Net在高分影像建筑物提取中部分建筑物边缘特征易模糊或丢失的问题,提出一种对高分影像建筑物边缘增强,同时对U-Net部分卷积过程进行改进的优化的建筑物提取方法。首先利用域变化递归滤波的方式对建筑物边缘进行增强,将增强后影像输入U-Net神经网络中进行训练;其次为充分利用建筑物在高分影像上丰富的细节特征,尝试在原U-Net结构基础上,从训练图像和标签中提取成对的补丁以增加训练数据,这些补丁进一步加强了正反向深度学习中建筑物高维特征的获取;最后在影像上实现建筑物提取。对辽宁省盘锦市邻接渤海湾地区2017年9月29日高分二号影像建筑物提取实验结果表明,对于包含阴影区域干扰较多的非理想样本数据,用U-Net识别建筑物得到的整体分类精度为75.99%,而改进方法最高整体分类精度可达83.12%,较原U-Net网络精度提高7.13百分点,证明该方法行之有效。

Aiming at tackling the problem that some edge features of buildings are easily blurred or lost in the extraction of buildings with high resolution image by U-Net,this paper proposes an optimized building extraction method,which firstly enhances the edge of buildings with high resolution image and simultaneously improves the partial convolution process of U-Net.Specific process is as follows:Firstly,the domain change recursive filtering method is used to enhance the edge of the building,and the enhanced image is input into U-Net neural network results for training.To make full use of the rich details characteristics of the buildings on the GF-2 images,the authors tried to extract pairs from training images and label patch on the basis of the original U-Net structure and in the process of coding decoding,so as to increase the training data.These patches further strengthened the positive and negative deep learning of high-dimensional feature for buildings,thus successfully realizing building image segmentation.In this paper,the experimental results of the extraction of GF-2 image buildings in Panjin City of Liaoning Province adjacent to Bohai Bay on September 29,2017 show that the overall classification accuracy of the buildings detected by U-Net is 75.99%for the shaded and unsatisfied area sample data,and the maximum overall classification accuracy of this method can reach 83.12%,which is 7.13 percentage higher than that of the original U-Net network.It is proved that the U-NET model combined with domain change recursive filtering is effective.