U型卷积神经网络的ZY-3影像道路提取方法

Road extraction from ZY-3 remote sensing image based on U-Net like convolution architecture

针对经典全卷积神经网络在池化和上采样过程中造成图像分辨率不断下降以及对各个像素进行分类时忽略了像素之间的关系,导致提取道路比较模糊和平滑的问题。该文提出一种基于U型卷积网络的ZY-3道路提取方法。首先,参考医学图像分割领域表现突出的U-Net模型,采用对称式网络结构将低级细节信息与高级语义信息相结合,提高道路的初提取精度;其次考虑到卷积神经网络对百万量级的参数优化程度相对不足,采用集成学习的方法,通过变更权重获得若干个模型进行融合,进一步提升了道路提取的精度;最后,通过使用形态学开运算完成孔洞的去除等工作。实验结果表明,该文方法的提取结果在不同实验区域中平均准确度达到了95%以上,显著优于基于经典全卷积网络模型、基于纹理与形状特征提取道路的方法。

Aiming at the classical full convolutional neural network(FCN) causes the image resolution to decrease continuously during the pooling and upsampling layers,and the relationship between image pixels is neglected when classifying each pixel,which leads to the blurring and smoothing of the extracted road. This paper proposed a road extraction method from ZY-3 image based on U-NET. Firstly,referring to the prominent U-Net model in the field of medical image segmentation,a symmetrical network structure was used to combine low-level detail information with high-level semantic information to improve segmentation accuracy. Secondly,considering the convolutional neural network(CNN) is relatively insufficient for the parameter optimization of millions on degrees,ensemble learning method was adopted to obtain several models for fusion by changing weights,which further improves the accuracy of road extraction. Finally,mathematical morphology opening operation was applied to complete holes removal. Experimental results showed that average accuracy of the proposed method was more than 95% in different experimental areas,which is significantly better than classical FCN model and the method based on texture and shape feature extraction.