基于DDR-CycleGAN的红外图像数据增强  

Infrared image data enhancement based on DDR-CycleGAN

作　　者：张浩[1,2,3,4] 杨坚华[1,2,3,4] 李启航 花海洋 ZHANG Hao;YANG Jian-hua;LI Qi-hang;HUA Hai-yang(Key Laboratory of Opto-Electronic Information Processing,Chinese Academy of Sciences,Shenyang 110016,China;Shenyang Institute of Automation,Chinese Academy of Sciences,Shenyang 110016,China;Institutes for Robotics and Intelligent Manufacturing,Chinese Academy of Sciences,Shenyang 110169,China;University of Chinese Academy of Sciences,Beijing 100049,China;School of Information Science and Technology,Northwest University,Xi ′an 710127,China)

机构地区：[1]中国科学院光电信息处理重点实验室,辽宁沈阳110016 [2]中国科学院沈阳自动化研究所,辽宁沈阳110016 [3]中国科学院机器人与智能制造创新研究院,辽宁沈阳110169 [4]中国科学院大学,北京100049 [5]西北大学信息科学与技术学院,陕西西安710127

出　　处：《激光与红外》2022年第4期600-606,共7页Laser & Infrared

基　　金：中科院创新基金项目(No.E01Z040101)资助。

摘　　要：提出了一种基于双鉴别器相对循环一致性生成对抗网络(DDR-CycleGAN)的红外图像数据生成方法。针对双鉴别器监督机制易出现的过度优化造成的性能下降问题,该方法在双鉴别器循环一致性生成对抗网络(DD-CycleGAN)中加入了相对概率的思想,采用鉴别器鉴别图像的相对真实概率替代绝对真实概率,使得生成图片更加接近真实图片。本文方法采用FLIR数据集进行训练和测试,实验结果表明本文方法相比DD-CycleGAN在可见光图像生成红外图像的图像质量上,峰值信噪比提高了3.91%,FID(Frechet Inception Distance score)降低了3.81%。An infrared image data generation method based on DDR-CycleGAN is proposed.The idea of relative probability is added on the basis of DD-CycleGAN,and the relative real probability of discriminator is used to replace the absolute real probability thus to solve the problem of performance degradation caused by over optimization.The experimental results show that compared with DD-CycleGAN,the peak signal-to-noise ratio(PSNR)of this method is improved by 3.91%,and the FID(Frechet perception distance score)is reduced by 3.81%.

关 键 词：循环一致性生成对抗网络 双监督机制 相对概率 红外图像数据增强 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]