多深度神经网络集成的道路提取

Road Extraction Based on Integration of Multiple Deep Neural Networks

高分辨率遥感图像道路提取是遥感信息分析领域中的一项重要工作。尽管深度学习的高分辨率遥感影像道路提取方法已经获得了先进的性能,但大多数深度学习网络具有严重的数据依赖性,缺少一种普遍适用的网络模型。针对以上问题,分别探讨了深度神经网络的宽度及结构对高分辨率遥感图像道路提取的影响。选取3种经典的端到端网络模型(SegNet、U-Net和FRRN)进行试验,针对每种网络结构依次设置3×3、5×5、7×7的卷积核。最后,剖析卷积核的尺度对道路提取的影响,提出了一种多深度学习网络集成的道路提取方法。结果表明,对于SegNet、FRRN、U-Net网络,3×3的卷积核普遍获得了较好的道路提取结果,SegNet方法能够较好地平衡道路提取的完整性和正确性。

Road extraction from high-resolution remote sensing images data is an essential work in the field of remote sensing information analysis.The high-resolution imagery extraction method based on deep learning has achieved state-of-the-art performance.However,most deep learning networks have serious data dependence and lack a universally applicable network model.In view of the above problems,this study discusses the influence of the width of the deep neural network and the model structure on the road extraction of high-resolution imagery.Specifically,three classic end-to-end network structure models (SegNet,U-Net,and FRRN) are selected for testing,and 3×3,5×5,and 7×7convolution kernels are set for each network structure in order to explore the impact of different scale convolution kernels on road extraction.Finally,the influence of the scale of the convolution kernel on the road extraction of the network model is analyzed,and a road extraction method for multi-scale deep learning network integration is proposed.The results show that for SegNet,FRRN and u-net networks,road extraction results by the 3×3 convolution kernel are all good and for SegNet,the completeness and correctness of road extraction can be achieved.