基于改进的U-Net的遥感图像语义分割

SEMANTIC SEGMENTATION OF REMOTE SENSING IMAGE BASED ON IMPROVED U-NET

遥感图像语义分割是根据土地覆盖类型对图像中每个像素进行分类,是遥感图像处理领域的一个重要研究方向,由于在研究的过程出现了相似地物导致分割不准确,为了解决这个问题,提出一种基于U-Net和残差网络的遥感图像语义分割网络DeepResU-Net,对传统的U-Net语义分割网络进行改进,以U-Net为骨架网络,采用残差卷积单元替换原始U-Net的编码层和解码层中的卷积层,防止网络梯度消失,网络中还含有丰富的跳跃连接可以促进信息传播。在遥感(ISPRS)Vaihingen数据集上的实验表明,该方法比FCN-8s、SegNet、U-Net、ResU-Net的分割准确度更高。

Semantic segmentation of remote sensing images is to classify each pixel in the image according to the type of land cover.It is an important research direction in the field of remote sensing image processing.The segmentation is inaccurate due to similar features in the research process.In order to solve this problem,DeepResU-Net,a remote sensing image semantic segmentation network based on U-Net and residual network is proposed.It improved the traditional U-Net semantic segmentation network,used U-Net as the skeleton network,and used residual convolution unit to replace the convolutional layer in the coding layer and decoding layer of the original U-Net,so as to prevent the network gradient from disappearing.The network contained rich jump connections that could promote information dissemination.Experiments on the remote sensing(ISPRS)Vaihingen dataset show that the results obtained by this method are more accurate than FCN-8s,SegNet,U-Net,and ResU-Net.