基于GLNet和HRNet的高分辨率遥感影像语义分割

High-resolution Remote Sensing Image Semantic Segmentation Based on GLNet and HRNet

在GLNet(Global-Local Network)中,全局分支采用ResNet(Residual Network)作为主干网络,其侧边输出的特征图分辨率较低,而且表征能力不足,局部分支融合全局分支中未充分学习的特征图,造成分割准确率欠佳。针对上述问题,提出了一种基于GLNet和HRNet(High-Resolution Network)的改进网络用于高分辨率遥感影像语义分割。首先,利用HRNet取代全局分支中原有的ResNet主干,获取表征能力更强,分辨率更高的特征图。然后,采用多级损失函数对网络进行优化,使输出结果与人工标记更为相似。最后,独立训练局部分支,以消除全局分支中特征图所带来的混淆。在高分辨率遥感影像数据集上,对所提出的改进网络进行训练和测试,实验结果表明,改进网络在全局分支和局部分支上的平均绝对误差(Mean Absolute Error,MAE)分别为0.0630和0.0479,在分割准确率和平均绝对误差方面均优于GLNet。

The backbone of a convolutional neural network global branch,a residual network(ResNet),obtains low-resolution feature maps at side outputs that lack feature representation.The local branch aggregates the feature maps in the global branch,which are not fully learned,resulting in a negative impact on image segmentation.To solve these problems in GLNet(Global-Local Network),a new semantic segmentation network based on GLNet and High-Resolution Network(HRNet)is proposed.First,we replaced the original backbone of the global branch with HRNet to obtain high-level feature maps with stronger representation.Second,the loss calculation method was modified using a multi-loss function,causing the outputs of the global branch to become more similar to the ground truth.Finally,the local branch was trained independently to eliminate the confusion produced by the global branch.The improved network was trained and tested on the remote sensing image dataset.The results show that the mean absolute errors of the global and local branches are 0.0630 and 0.0479,respectively,and the improved network outperforms GLNet in terms of segmentation accuracy and mean absolute errors.