融入CBAM的Res-UNet高分辨率遥感影像语义分割模型

Res-UNet High-resolution Remote Sensing Image Semantic Segmentation Model Integrated into CBAM

针对现有语义分割方法处理复杂遥感影像细节特征识别能力差、信息丢失等问题,提出一种融合注意力机制的遥感影像语义分割网络模型。模型主干网络采用编码器-解码器架构的U-Net模型,为了缓解梯度和网络退化问题,将残差结构嵌入到主干网络中;同时融入通道、空间注意力模块,兼顾影像的细节特征和模型鲁棒性。在ISPRS Potsdam数据集上进行分析验证,实验结果表明,在去除“噪声”、地物边缘“平滑”、细窄地物“连续”、细小目标分割等方面,融入CBAM模块的ResUNet语义分割精度要优于传统网络模型。

Aiming at the problems of poor recognition ability and information loss of the existing semantic segmentation methods for complex re-mote sensing images,we proposed a semantic segmentation network model of remote sensing images based on attention mechanism.We adopted the U-NET model with encoder-decoder architecture in the model backbone network.In order to alleviate the gradient and network degradation,we embedded the residual structure into the backbone network,and integrated the channel and spatial attention modules to take into account the detail features of image and model robustness.We analyzed and verified this model on ISPRS Potsdam dataset.The experimental results show that the accuracy of Res-UNet semantic segmentation with CBAM module is better than that of traditional network model in terms of removing noise,smoothing edge of ground objects,continuing narrow ground objects and small object segmentation.