基于跨层注意力Unet网络的高分辨率遥感影像建筑物提取研究

Research on Building Extraction from High Resolution Remote Sensing Image Based on Cross Layer Attention Unet

使用全卷积神经网络提取高分辨率遥感影像中的建筑物对城市规划、土地资源管理等应用具有重要意义。本文提出一种全卷积神经网络SC-Unet,以Unet架构为基础,采用SELU激活函数,放弃批标准化;使用空间金字塔池化替换最后一个卷积模块,使得SC-Unet的参数量仅为Unet的50%;使用轻量级跨层注意力模块让高层语义指导低层语义。实验在WHU建筑物数据集上进行,结果显示:SC-Unet在测试集上的IOU达到88.1,比Unet高3.2,且推理速度SC-Unet是Unet的2倍。

It is of great significance to extract buildings from high-resolution remote sensing images by using full convolution neural network for urban planning,land resource management and other applications.In this paper,a full convolution neural network SCUnet is proposed,which is based on Unet architecture.SC-Unet adopts SELU activation function and abandons batch normalization.Using spatial pyramid pooling to replace the last convolution module,the number of parameters of SC-Unet is only 50%of that of Unet.Use a lightweight cross-layer attention module to let higher-level semantics guide lower-level semantics.The experiment was carried out on the WHU building dataset.The experiment results show that the IOU of SC-Unet on the test set is 88.1,which is 3.2 higher than that of Unet,and the inference speed of SC-Unet is two times the inference speed of Unet.