基于改进U-Net的遥感图像道路提取算法

Road Extraction from High Resolution Remote Sensing Images Based on Improved U-Net

针对基于U-Net模型对遥感图像道路特征提取能力不足、分割结果不清晰等问题,文章提出了一种改进的U-Net算法:首先在编码器中引入级联的空洞空间金字塔模块充分利用图像全局上下文信息从而改善分割结果模糊的问题;再通过在通道中嵌入坐标注意力机制模块加强对道路特征信息的提取,最后在解码器部分引入空间注意力机制旨在提高道路分割边缘的清晰度。实验表明:在马赛诸塞州数据集下改进后的U-Net模型比原始U-Net网络模型在Recall、F1-S cores和Io U三个指标下分别了提高了0.085、0.038、0.045,提取的道路结构更完整且相互连通,证明了算法优化的有效性。

Aiming at the problems of insufficient road feature extraction ability and unclear segmentation results of remote sen-sing images based on U-Net model,this paper proposes an improved U-Net algorithm:firstly,a cascading void space pyramid module is introduced into the encoder to make full use of the global context information of the image to improve the problem of blurring of segmentation results;Then,by embedding the coordinate attention mechanism module in the channel,the extraction of road feature information is strengthened,and finally the spatial attention mechanism is introduced in the decoder part,aiming to improve the clarity of the road segmentation edge.Experiments show that the improved U-Net model under the Massachusetts dataset is improved by 0.085,0.038 and 0.045 respectively compared with the original U-Net network model under the three in-dicators of Recall,F1-Scores and IoU,respectively,and the extracted road structure is more complete and interconnected,which proves the effectiveness of algorithm optimization.