融合注意力机制的建筑物提取方法

A Building Extraction Method Combined Attention Mechanism

针对高分辨率遥感影像中因细节信息繁多导致的建筑物特征提取困难问题,提出了一种融合注意力机制的建筑物提取方法RSDANet,进一步从复杂的特征中增强有效特征,抑制非有效特征。该方法采用了Encoder-Decoder结构。在Encoder中,设计了一个多级的残差连接块RSD Block和一个双向的FPN结构用于捕获多尺度特征信息,接着使用注意力机制进一步处理多尺度特征,以全局上下文信息为约束条件抑制或增强特征;在Decoder中,使用Skip-Connection把低层特征逐级融合起来,低层特征含有比较丰富的空间信息,更利于提高分割精度。实验研究表明,该方法取得了较好的提取结果,相比于Deeplabv3+方法,MIoU提高了2.24%。

Aiming at the problem of the difficulty in building feature extraction in high-resolution remote sensing imagery due to the large amount of detailed information,this paper proposes a building extraction method RSDANet combined with attention mechanism to further enhance effective features from complex features and suppress ineffective features.This method uses the Encoder-Decoder structure.In the Encoder,a multi-level residual block RSD Block and a bilateral FPN structure are designed to capture multi-scale feature information.It then uses an attention mechanism to further process multi-scale features,suppressing or enhancing features with global context information as constraints.In Decoder,skip-connection is used to fuse the lower feature layer step by step.The lower feature layer contains richer spatial information,which is more conducive to improving the segmentation accuracy.The experimental results show that this method achieves better extraction results.Compared with the Deeplabv3+method,the MIoU of this method is increased by 2.24%.