基于双分支融合策略的遥感建筑物提取轻量化网络

Lightweight network for remote sensing building extractionbased on dual-branch fusion strategy

针对高分辨率遥感影像建筑物提取时带来的庞大参数与计算量的问题,文章提出了双分支策略融合的轻量化语义分割网络。网络包含全局特征信息提取分支和空间信息提取分支。双分支共享图像底层特征,全局特征信息提取分支连续利用mobileNetV2中的反向瓶颈残差块结构提取特征,并利用空洞空间金字塔池化模块融合多尺度特征;空间信息提取分支添加卷积块注意力模块,更关注待分割区域的细节特征,实现精细化特征融合。最后将双分支特征融合,利用双线性插值上采样恢复图像大小。将文章模型在WHU建筑物影像数据集上进行实验,正确提取建筑物准确度可达92.09%,IoU达86.28%,模型参数大小仅为1.19 M,保证了建筑物提取精度的同时达到了轻量化预期。

In order to solve the problem of large parameters and computation when extracting buildings from high-resolution remote sensing images,a lightweight semantic segmentation network based on two-branch strategy fusion was proposed.The network includes global feature information extraction branch and spatial information extraction branch.Two branches share the underlying features of the image,and the global feature information extraction branch continuously uses the reverse bottleneck residual block structure of mobileNetV2 to extract features,and uses the pyramid pool module of empty space to fuse multi-scale features.Convolution block attention module is added in the spatial information extraction branch to pay more attention to the detailed features of the region to be segmented to achieve fine feature fusion.Finally,the two-branch features are fused and the image size is restored by up-sampling with bilinear interpolation.The proposed model was tested on the WHU building image data set,and the accuracy of correct building extraction was up to 92.09%,the IoU was up to 86.28%,and the model parameter size was only 1.19M,which ensured the building extraction accuracy and reached the lightweight expectation.