基于改进的SegNet分割网络的遥感图像分割

Remote Sensing Image Segmentation Based on Improved SegNet Segmentation Network

摘要: 城市遥感图像的建筑物分割技术，对城市的土地规划、资源利用、灾害防治具有广泛的应用价值，对此提出了一种基于改进的SegNet分割网络的城市遥感图像分割方法。1) 为了解决所需数据量过大、网络层数过深等问题，模型在SegNet网络结构的基础上，删减了部分卷积层，使网络结构更加简洁，运行速度更快。2) 针对SegNet网络保留特征信息不足、图像分割边缘粗糙的缺点，针对这一问题，我们设计了一个新的多尺度特征提取模块，通过三个不同尺度的卷积核提取目标信息。以SegNet为基线，通过用该模块替换跳过连接，我们提出了一种多尺度特征提取SegNet。该方法可以对跳过连接中的浅层特征信息进行二次特征提取，细化细节信息，缩小低级特征与高级特征之间的语义差距。它不仅可以提高网络提取多尺度特征信息的能力，从更大的范围到更多的层次来提取遥感图像中建筑物的边缘细节信息，而且可以增加跳过连接的数量，以减少网络过度拟合。多尺度模块实验结果表明，所提出的方法与现有的FCN、U-Net、SegNet网络相对比，准确率和交并比有明显的提升，并且对遥感图像的城市建筑物分割边缘有了较好的改善。

Abstract: The building segmentation technology of urban remote sensing image has a wide application value for urban land planning, resource utilization and disaster prevention and control. This paper proposes a segmentation method of urban remote sensing image based on improved SegNet seg-mentation network. 1) In order to solve the problems such as too large amount of data required and too deep number of network layers, the model has deleted some convolution layers based on the SegNet network structure, making the network structure simpler and faster. 2) In view of the shortcomings of SegNet network, such as insufficient feature information and rough edge of image segmentation, we designed a new multi-scale feature extraction module to extract target infor-mation through three convolution kernels of different scales. Taking SegNet as the baseline, we propose a multi-scale feature extraction SegNet by replacing the skip connection with this module. This method can extract secondary features from shallow feature information in the skip link, refine the details, and narrow the semantic gap between low-level features and high-level features. It can not only improve the ability of the network to extract multi-scale feature information, extract the edge details of buildings in remote sensing images from a wider range to more levels, but also increase the number of skipped connections to reduce network over fitting. The experimental results of multi-scale modules show that the proposed method has significantly improved the ac-curacy and intersection/merge ratio compared with the existing FCN, U-Net and SegNet networks, and has better improved the urban building segmentation edge of remote sensing images.