基于生成对抗网络的遥感图像超分辨率重建  

Remote Sensing Image Superresolution Reconstruction Based on GAN

作　　者：李昂[1] 宋晓莹 LI Ang;SONG Xiao-ying(The 4808 Armament Repair Factory of PLA,Qingdao 266000,China;Higher Education Key Lab for Measuring&Control Technology and Instrumentations of Heilongjiang,Harbin University of Science and Technology,Harbin 150080,China)

机构地区：[1]中国人民解放军第四八零八工厂军械修理厂,山东青岛266000 [2]哈尔滨理工大学测控技术与通信工程学院,黑龙江哈尔滨150080

出　　处：《光学与光电技术》2019年第6期39-44,共6页Optics & Optoelectronic Technology

摘　　要：生成对抗网络模型可以用来生成服从原始真实图像分布规律的高频细节信息。为了进一步提高重建图像的视觉质量,对生成对抗网络的生成网络、判别网络及感知损失三个方面进行了改进。首先移除了生成网络中的BN层,同时在残差块中采用密集连接的方式,增加网络模型的容量,降低了计算复杂性,增强了网络训练的稳定性。然后采用迁移学习技术来促进深度模型的训练,解决了遥感数据不足的问题。实验结果表明提出的算法通过对遥感图像超分辨率重建算法进行改进,可以获得更好的主观视觉效果,PSNR和SSIM均有显著提高。Generative adversarial network model can be used to generate high-frequency detail information which obeys the distribution of original real image. In order to further improve the visual quality of the reconstructed image,the generation network,discrimination network and perceptual loss of the generated confrontation network are improved in this paper. Firstly,the BN layer in the generated network is removed. Meanwhile,dense connections are used in the residual blocks to increase the capacity of the network model,reduce the computational complexity and enhance the stability of network training. Then the transfer learning technology is used to promote the training of depth model,which solves the problem of insufficient remote sensing data. The experimental results show that the proposed algorithm improves the super-resolution reconstruction algorithm of remote sensing images and can obtain better subjective visual effects. PSNR and SSIM are improved visibly.

关 键 词：生成对抗网络 遥感图像超分辨率重建 残差密集块 迁移学习 感知损失 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]