遥感图像语义分割的多特征注意力融合网络

MULTI-FEATURE ATTENTION FUSION NETWORK FORSEMANTIC SEGMENTATION OF REMOTE SENSING IMAGES

针对高分辨率遥感图像中存在背景复杂、目标大小不一、类间具有相似性的问题,提出一种用于遥感图像语义分割的多特征注意力融合网络(Multi-feature Attention Fusion, MAFNet)。MAFNet基于编码和解码结构,在编码阶段,采用空间金字塔池化获取多尺度的上下文信息,同时融合特征通道之间的关联信息,提高特征图的语义表征能力;在解码阶段,基于注意力机制将高层特征与低层特征自适应地融合,逐级恢复目标的细节特征。在公开的数据集Potsdam和Vaihingen上设计了对比实验,PA值分别达到了89.6%和89.1%,验证了该方法的有效性。

Aimed at the problems of complex background,different target sizes,and similarity between classes in high-resolution remote sensing images,a multi-feature attention fusion network(MAFNet)for semantic segmentation of remote sensing images is proposed.MAFNet was based on the encoding and decoding structure.In the encoding stage,spatial pyramid pooling was used to obtain multi-scale context information,and at the same time,the associated information between feature channels was merged to improve the semantic representation ability of the feature map.In the decoding stage,the high-level features and low-level features were adaptively fused to restore the detailed features of the target level by level.Comparative experiments were carried out on the public data sets Potsdam and Vaihingen.The PA values reached 89.6%and 89.1%respectively,which verified the effectiveness of the method.