基于改进HRNet的遥感图像地物覆盖语义分割研究

Research on Semantic Segmentation Method for Remote Sensing Image Land Cover Based on Improved HRNet

摘要: 随着科技的发展，利用深度学习的方法帮助遥感图像地物覆盖语义分割工作取得了很大的进展。然而，针对于遥感语义分割数据集中存在的像素分布不平衡问题，我们提出了基于注意力的HRNet (Attention-based HRNet, AbHRNet)结构。首先，针对于各个类别目标之间的像素数量不平衡问题，本文在特征提取网络中引入了卷积注意力模块，使得网络对于我们感兴趣的目标特征尤其是数量较少的目标特征赋予了更多的关注，并减小了由复杂的背景信息带来的干扰；其次，针对于目标和目标、目标和背景之间像素数量不平衡的问题，在基准网络交叉熵损失的基础上又引入了二元交叉熵损失和Dice Loss，以实现对背景样本的有效监督，并解决由于像素数量不平衡带来的模型难以优化的问题。在LoveDA数据集上的实验结果表明，我们提出的AbHRNet的平均交并比达到了51.14%，相较于基准HRNet模型提升了1.97%，尤其是帮助分割效果很差的荒地类别的精度提升了一倍。

Abstract: With the development of science and technology, great progress has been made in the remote sensing image land cover semantic segmentation task by using deep learning methods. However, in view of the unbalanced pixel distribution of remote sensing semantic segmentation dataset, we propose the Attention-based HRNet (AbHRNet) structure. First, in view of the imbalance problem of the number of pixels between various category targets, Convolutional Block Attention Module (CBAM) is introduced into the feature extraction network, so that the network gives more attention to the target features that we are interested in, especially the target features with a small number, and reduces the interference caused by complex background information; second, in view of the problem of the unbalanced pixels number between target and target, target and background, bina-ry cross entropy loss and Dice Loss are introduced on the basis of cross entropy loss of baseline network to achieve effective supervision of background samples and to solve the problem is difficult to be optimized due to the unbalanced pixels number. The experimental results on the LoveDA dataset show that the mean intersection over union (mIoU) of our proposed AbHRNet reaches 51.14%, which has a 1.97% improvement compared to the baseline HRNet model, especially helping the barren category with poor segmentation effect to double the accuracy.