一种融合注意力机制的建筑物变化检测模型

A model for detecting building changes incorporating attention mechanisms

针对城市建筑物变化检测问题,该文基于U-net深度学习语义分割模型,提出了一种融合残差结构和注意力机制的遥感影像建筑物变化检测模型,以U-net模型为基础,引入ResNet50的残差结构用来代替编码阶段中的卷积层,在加深网络深度的同时解决梯度消失的问题;在解码阶段横向连接结构中引入注意力机制,加强网络对变化建筑物特征的学习。实验表明,在U-net结构的基础上加入残差结构和注意力模块后,建筑物变化检测的精确率、召回率、F1值分别提升了6.28%、6.02%、5.88%。

In order to address the problem of deterrence of urban building variations,this paper proposed a remote sensing image building variation detection model based on the U-net deep learning semantic segmentation model incorporating residual structure and attention mechanism.The model was based on the U-net model jointly with the residual structure of ResNet50to replace the convolutional layer in the coding stage.Meanwhile,the problem of gradient disappearance was solved when deepening the network depth.An attention mechanism was introduced in the lateral connectivity structure of the decoding stage to be used to enhance the network’s learning of changing building characteristics.The experiment results showed that after adding the residual structure and attention module to the U-net structure,the accuracy,recall and F1 values of building change detection were improved by 6.28%,6.02%and 5.88%respectively.