基于FPN的高分辨率建筑物图像的语义分割

Building Extraction from High-resolution Imagery Based on FPN

针对目前的语义分割网络仅使用部分卷积层产生最终输出而导致的建筑物的边缘精度较低、预测图像质量不高等问题,提出了一种基于特征金字塔结构的BuildingNet网络,并引入了空洞空间金字塔池化模块,使得该网络具有针对不同分辨率图像的良好特征提取性能。另外,设计了一种改进的Lovász损失函数训练所提出的网络,有效提高了提取结果的图像质量。实验表明,对于高分辨率建筑物图像的语义分割,所提出的BuildingNet网络在两个实验数据集上的F1得分分别为94.58%和92.15%,优于常用的三种语义分割网络(SegNet、U-Net、Deeplabv3+),证明了其有效性。

Aiming at the problem that the currently widely used semantic segmentation network only uses part of the convolutional layer to generate the final output,resulting in low edge accuracy of the building and low quality of the predicted image,this paper proposes a feature pyramid net based structured BuildingNet network,and introduces the atrous spatial pyramid pooling module,making the network have good feature extraction performance for images of different resolutions.Moreover,it designs an improved Lovász loss function to train the proposed network,which effectively improves the image quality of the extracted results.Experimental results show that for the semantic segmentation of high-resolution building images,the F1 scores of the BuildingNet network proposed in this paper on the two datasets are 94.58%and 92.15%,respectively,which are better than that of the other three semantic segmentation networks(SegNet,U-Net,and Deeplabv3+),which proves the effectiveness of the method proposed in this paper.