基于Res-UNet高分辨率遥感影像建筑提取方法

Building Extraction Method for High-Resolution Remote Sensing Images Based on Res-UNet

摘要: 针对现有语义分割技术在处理复杂遥感影像时细节特征识别能力不足、信息丢失等问题，本研究提出了一种融合注意力机制的遥感影像语义分割网络模型。该模型以编码器–解码器架构的U-Net模型为基础，通过将残差结构嵌入主干网络以缓解梯度消失和网络退化问题。此外，模型还整合了通道和空间注意力模块，以兼顾影像的细节特征和提高模型的鲁棒性。在ISPRS Vaihingen数据集上的分析验证表明，引入CBAM模块的Res-UNet在去除“噪声”、地物边缘“平滑”、细窄地物“连续性”以及细小目标分割等方面，其语义分割精度显著优于传统网络模型。

Abstract: In view of problems of the existing semantic segmentation technology in processing complex remote sensing images, such as lack of detail feature recognition ability and information loss, this paper proposes a network segmentation model of remote sensing images that integrates attention mechanism. The model is based on the U-Net model of the encoder-decoder architecture, which alleviates the gradient vanishing and network degradation problems by embedding the residual structure into the backbone network. In addition, the model integrates the channel and spatial attention modules to balance the detailed characteristics of the image and improve the robustness of the model. Analysis and validation on the ISPRS Vaihingen dataset show that Res-UNet introduced into CBAM module is significantly better than traditional network models in removing “noise”, “smoothing”, “ground” “continuity”, and fine target segmentation.