一种改进的SRGAN红外图像超分辨率重建算法  

An Improved SRGAN Infrared Image Super-Resolution Reconstruction Algorithm

作　　者：胡蕾[1] 王足根 陈田 张永梅[3] Hu Lei;Wang Zugen;Chen Tian;Zhang Yongmei(School of Computer Information Engineering,Jiangxi Normal University,Nanchang 330022,China;Electric Power Research Institute,State Grid Jiangxi Electric Power Co.,Ltd,Nanchang 330096,China;School of Information,North China University of Technology,Beijing 100144,China)

机构地区：[1]江西师范大学计算机信息工程学院,江西南昌330022 [2]国网江西省电力有限公司电力科学研究院,江西南昌330096 [3]北方工业大学信息学院,北京100144

出　　处：《系统仿真学报》2021年第9期2109-2118,共10页Journal of System Simulation

基　　金：国家自然科学基金(61662033,61262036)。

摘　　要：针对红外图像分辨率偏低的问题,设计了一种改进的超分辨率生成对抗网络(Super-Resolution Using a Generative Adversarial Network,SRGAN)算法。在生成网络中,提出应用残差密集网络获取各网络层提取的图像特征以保留图像更多的高频信息,并采用渐进式上采样方式以提升大缩放因子下超分辨率重建效果。在损失函数方面采用更符合人类感官的感知损失,使生成图像在感官和内容上与真实高分辨率图像更加接近。实验结果表明:所提方法重建的超分辨率红外图像质量在主观及客观评价中均要优于当前具有代表性的方法。Aiming at the low resolution of infrared images,an improved SRGAN super-resolution reconstruction algorithm is designed.In the generative network,the method of applying the residual dense network to obtain the image features extracted from each network layer so as to retain more high-frequency information of the image,and adopting a progressive upsampling method to improve the super-resolution reconstruction effect under a large scaling factor.In terms of the loss function,the perceptual loss that is more in line with human senses is adopted to make the generated image being closer to the real high-resolution image of senses and content.Experimental results show that the quality of reconstructed infrared image is better than that of the current representative methods in the subjective and objective evaluation.

关 键 词：红外图像 超分辨率重建 生成式对抗网络 残差密集网络 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]