深层特征提取与多输出增强融合的建筑物分割

Building Segmentation Based on Deep Feature Extraction andMulti Output Enhancement Fusion

针对高分辨率遥感影像建筑物语义分割中存在单体建筑部分漏分与纹理相近的非建筑物错分的问题,提出深度特征提取与多输出增强建筑物语义分割网络。首先,在编码与解码交替处设计并行连续空洞空间注意力金字塔模块,对建筑物高维特征实现深层提取;然后,在网络解码阶段设计多输出增强融合模块,提升不同尺度建筑物特征对输出结果的有效参与度。选取U-Net、DeeplabV3+、MA-FCN、BRRNet同类算法进行对比,在Massachusetts和WHU公共数据集上测试,OA、precision、recall、IoU、F1指标分别达到98.87%、94.53%、95.40%、90.41%、94.96%,与同类算法相比,除precision之外其他4项指标更高,与U-Net相比,依次高出0.25%、1.12%、1.14%、2.02%、1.12%。

This paper proposes a deep feature extraction and multi output enhanced building semantic segmentation network to address the problem of missing individual building parts and similar texture non building misclassification in high-resolution remote sensing image building semantic segmentation.Firstly,a parallel continuous void spatial attention pyramid module is designed at the alternation of encoding and decoding to achieve deep extraction of high-dimensional features of buildings.Then,in the network decoding stage,a multi-output enhanced fusion module is designed to improve the effective participation of building features at different scales in the output results.Selecting the similar algorithms of U-Net,DeeplabV3+,MA-FCN and BRRNet for comparison and testing on the public datasets of Massachusetts and WHU,the OA,precision,recall,F1 and IoU indicators can reach 98.87%,94.53%,95.40%,90.41%,and 94.96%,respectively.OA,F1,recall and IoU accuracy are higher than that of the other four indicators.The OA,precision,recall,IoU and F1 of the proposed method are higher than that of U-Net by 0.25%,1.12%,1.14%,2.02%,and 1.12%,respectively.