基于注意力门残差网络的遥感影像道路提取

Road extraction from remote sensing images based on attention gate residual network

针对遥感影像道路提取出现的无关噪声多,道路不连续问题,本文通过改进U-Net提出了基于注意力门残差网络的道路提取算法。首先,编码器部分引入残差块传递原始特征,在保证网络深度的同时,使梯度能够有效传递;其次,在连接层使用多尺度空洞卷积特征提取模块,来充分挖掘图像中的多尺度特征信息;最后,用注意力门将浅层网络信息和反卷积信息融合实现解码,以抑制浅层噪声特征。使用的数据集包括Massachusetts Roads Dataset数据集和CVPR DeepGlobe 2018道路提取挑战赛数据集。实验结果表明,该算法可以有效提升道路分割的效果。

Aiming at the problem that there are many independent noises and discontinuity in road extraction from remote sensing images, a semantic segmentation algorithm of Residual Attention U-Net is proposed by improving U-Net.Firstly, the encoder introduces the original character of residual block transfer, which can guarantee the depth of the network and make the gradient transfer efficiently. Then, multiscale dilated convolution extraction module is used in the connection layer to fully mine the feature information in the image.Finally, the decoder uses the attention gate to fuse the shallow network information with the deconvolution information to suppress the shallow noise characteristics.The used datasets include the Massachusetts Roads Dataset and the CVPR DeepGlobe 2018 Road Extraction Challenge dataset.The experimental results show that the algorithm can effectively improve the effect of road segmentation.