基于生成对抗网络的遥感图像居民区分割

Remote sensing image residential area segmentation based on generative adversarial network

针对传统遥感图像居民区分割存在的孔洞和边缘不准确的问题,在研究传统滤波方法和全卷积网络方法的基础上,提出了一种新的生成式对抗网络(GAN)分割方法。构造全卷积网络作为生成网络(G),用于学习遥感图像居民区分布规律,生成居民区分割结果图(fake);设计一个卷积神经网络分类器,作为判别网络(D),用于区分生成网络分割结果图(fake)和人工标注结果图(label)。G试图欺骗D网络,D尽可能区分fake图,传播误差训练G网络,通过这种对抗式训练,使得G生成的fake图与label图尽可能相似,从而获得更好的分割结果。实验结果表明:基于GAN网络方法进行遥感图像居民区分割与Gabor滤波方法和FCN方法相比具有更好的鲁棒性。

Aiming at the problem of holes and inaccurate edges in traditional remote sensing image residential area segmentation,a new generative adversarial network(GAN)segmentation method is proposed based on traditional filtering method and full convolutional network method.Firstly,a full convolutional network is constructed as the generation network(G)for learning the distribution rule of the remote sensing image residential area,and generate the residential area segmentation result map(fake);then,a convolution neural network classifier is designed as a discriminant network(D)to distinguish the generated network segmentation result map(fake)from the manual result map(label).G tries to deceive the D network.D distinguishes the fake image as much as possible and propagates error to train G network.Through adversarial training,the fake map generated by G is as similar as possible to the label map.Thereby it can obtain better segmentation results.The experimental results show that the method based on GAN network has better robustness in the remote sensing image residential area segmentation than Gabor filtering method and fcn method.