基于改进Upernet的遥感影像语义分割算法

A semantic segmentation algorithm for remote sensing image based on improved Upernet

针对现有遥感图像语义分割算法中存在的分割精度低、浅层特征利用不充分的问题,提出了一种基于改进Upernet的遥感影像语义分割算法.首先引入Resnest的分裂注意力网络连接结构来重构原有的骨干网络,并在其中集成可变形卷积以提高网络对不同尺度遥感影像的特征提取能力;然后在Upernet的下采样路径中设计融合高效通道注意力机制模块重构特征融合模块,改善特征表示能力;最后引入交叉熵损失函数和Dice Loss联合训练样本来改善训练中存在的样本不平衡的问题并加速模型收敛.实验结果显示:改进后的网络在遥感影像ISPRS的Potsdam数据集和Vaihingen数据集上分别达到了79.38%和74.70%的MIoU.相比DeepLabV3Plus、Pspnet、FCN以及经典Upernet算法,该算法性能在Potsdam数据集上获得了平均2.16%的提升,在Vaihingen数据集上亦获得了平均2.21%的提升.

To address the issues of low segmentation accuracy and insufficient utilization of shallow features in existing remote sensing image semantic segmentation algorithms,a remote sensing image semantic segmentation algorithm based on improved Upernet is proposed.Initially,it introduces a split-attention network connection structure from Resnest to reconstruct the original backbone network and integrates deformable convolution to enhance the feature extraction capability for remote sensing images of different scales.Then,it designs a feature fusion module in Upernet's downsampling path,which incorporates an efficient channel attention mechanism to improve feature representation ability.Finally,it adopts a combination of cross entropy loss function and Dice Loss for training to address sample imbalance problem during training and to accelerate model convergence.Experimental results demonstrate that the improved network achieved MIoU scores of 79.38%and 74.70%on the ISPRS Potsdam and Vaihingen datasets,respectively.Compared to DeepLabV3Plus,Pspnet,FCN,and the classic Upernet algorithm,the proposed algorithm shows average improvement of 2.16%on the Potsdam dataset and 2.21%on the Vaihingen dataset.