LRSS-Net:轻量级遥感地物分割网络

LRSS-Net:Lightweight Remote Sensing Segmentation Network

针对目前基于深度学习的高分辨率遥感图像分割模型由于参数量大、计算复杂而导致高延迟、低响应的问题,提出了一种轻量级遥感地物分割方法,较好的平衡了速度和精度.该方法使用MobileNetV2进行特征粗提取,通过构建空间信息嵌入分支实现不同尺度的特征细提取,不同层次之间引入密集连接以获取密集的上下文信息.解码端设计特征融合优化策略逐层融合不同尺度的特征增加对细粒度特征的感知,同时以反卷积与双线性插值交替的上采样方式减少图像边缘信息丢失.最后采用交叉熵损失与Dice损失结合的方式加快网络收敛速度.为了验证所提方法的有效性,与几种常用的语义分割方法进行了对比实验.实验结果表明,所提算法的分割准确率为93.7%,MIoU为88.01%,可以实现地物的有效分割.

The current high-resolution remote sensing image segmentation model based on deep learning has the problems of high delay and low response caused by a large number of parameters and complex calculations.Considering the problems,this study proposes a lightweight remote sensing feature segmentation method,which can better balance speed and accuracy.This method uses MobileNetV2 for rough feature extraction,constructs spatial information embedding branches to achieve fine feature extraction on different scales,and introduces dense connections between different levels to obtain dense contextual information.The decoding end designs the feature fusion optimization strategy to fuse the features of different scales layer by layer to increase the perception of fine-grained features.Meanwhile,upsampling with alternating deconvolution and bilinear interpolation is employed to reduce the image edge information loss.Finally,the cross-entropy loss is combined with the Dice loss to accelerate network convergence.Comparative experiments are carried out with several commonly used semantic segmentation methods to verify the effectiveness of the proposed method.The experimental results show that the segmentation accuracy of the proposed algorithm is 93.7%,and the MIoU is 88.01%,which can achieve effective segmentation of ground objects.