高分辨率遥感图像语义分割在航空图像分析领域具有重要的理论和应用价值。但由于高分辨率遥感图像中楼房语义的丰富性和图像背景的复杂性,以往的分割方法往往容易产生边缘模糊、细节信息丢失和分辨率低等缺点。为了解决高分辨率卫星图像语义分割边界模糊和信息丢失的问题,本文提出一种端到端的卷积神经网络Dilated-UNet(D-UNet)。首先通过改进U-Net网络结构,采用Dilation技术拓展四通道的多尺度空洞卷积模块,每个通道采用不同的卷积扩张率来识别多尺度语义信息,从而提取更丰富的细节信息。其次设计了一种交叉熵和Dice系数的联合损失函数,更好的训练模型以达到预期分割效果。最后,在Inria航空图像数据集上进行综合评估与检验,实验结果表明提出的遥感图像分割方法能够有效地从高分辨率遥感图像中进行像素级城市建筑物分割,分割精度更高,优于其它方法,具有较高的实际应用价值。

Objective: Semantic segmentation of high-resolution remote sensing image has important theoretical and practical value in the field of aerial image analysis. However, due to the richness of building semantics and the complexity of image background in high-resolution remote sensing images, the traditional segmentation methods are prone to edge blur, loss of detail information and low resolution. Method: To solve the problem of fuzzy boundary and information loss in high-resolution satellite image semantic segmentation, an end-to-end convolutional neural network named Dilated-UNet (D-UNet) has been proposed. Firstly, the U-Net network structure is improved and the multi-scale dilated convolution module of four channels is expanded by using the division technology. Each channel uses different convolution expansion rate to identify the multi-scale semantic information, so as to extract richer detailed information. Secondly, a joint loss function of cross entropy and Dice coefficient is designed to achieve the desired segmentation effect. Results: The model is comprehensively evaluated and tested on the Inria aerial image dataset. The experimental results show that the proposed remote sensing image segmentation method can effectively segment urban buildings at pixel level from high-resolution remote sensing images, and the segmentation accuracy is higher, which is better than other methods. Conclusions: We conclude that our proposed D-UNet has the potential to deliver automatic building segmentation from high-resolution remote sensing images at an accuracy that makes it a useful tool for practical application scenarios.