基于分支合并策略的轻量级遥感图像分割

Lightweight remote sensing image segmentation based on branch merging strategy

为有效解决现有遥感图像分割网络在提升分割精度时带来大量参数和计算量的问题,提出基于分支合并策略的轻量级语义分割网络。网络的两个分支分别为全局语境分支和空间信息分支。全局语境分支采用轻量级网络CGNet(context guided network)的思想作为基网络提取像素级别和分割级别的上下文纹理特征,空间信息分支采用三层卷积提取图像的空间特征,通过特征融合模块将两种特征融合,设计加权的多尺度交叉熵损失函数增强小目标分割效果。在开源遥感数据集上实验,其结果验证了该网络在同等条件下利用较少的资源能达到较高的精度。

To solve the problem that the existing remote sensing image segmentation network brings a lot of parameters and computation while improving the segmentation accuracy,a lightweight semantic segmentation network based on branch merging strategy was proposed.The two branches of network were global context branch and spatial information branch.The global context branch used the idea of lightweight network CGNet(context guided network)as the base network to extract the context texture features at pixel level and segmentation level.The spatial information branch used three-layer convolution to extract the spatial features of the image.The two features were fused through the feature fusion module,and the weighted multi-scale cross entropy loss function was designed to enhance the small target segmentation.Experimental results on open source remote sensing datasets show that the network can achieve high accuracy with less resources under the same conditions.