基于循环生成对抗网络的光照补偿方法  

Illumination compensation method based on cycle generative adversarial networks

作　　者：赵亮 张鸿[1,2] ZHAO Liang;ZHANG Hong(College of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China;Hubei Province Key Laboratory of Intelligent Information Processing and Real-Time Industrial System,Wuhan University of Science and Technology,Wuhan 430065,China)

机构地区：[1]武汉科技大学计算机科学与技术学院,湖北武汉430065 [2]武汉科技大学智能信息处理与实时工业系统湖北省重点实验室,湖北武汉430065

出　　处：《计算机工程与设计》2020年第9期2566-2573,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(61373109)。

摘　　要：为以较好的效果、较高的鲁棒性解决一些图片中存在的光照相关问题,对图片进行光照补偿,设计基于循环生成对抗网络的光照补偿方法。使用循环生成对抗网络对光照条件不佳的图片进行光照补偿,以分别处在正常光照和复杂光照下的同类事物或场景的图片作为对抗训练集训练CycleGAN,使用Switchable Normalization代替原CycleGAN使用的Instance Normalization,解决光照问题。通过实验中的图片对比和亮度、对比度、峰值信噪比的比较,该方法在过暗图片和强反光图片上的光照补偿效果均高于传统算法和未经更改的CycleGAN网络,验证了算法的有效性和鲁棒性。To solve the illumination-related problems in some pictures with better effect and higher robustness,the illumination compensation method based on the cycle generative adversarial networks was designed.CycleGAN was used to counter the network to compensate the photos with poor lighting conditions,the model was trained by a confrontation training set with pictures of similar things or scenes under normal illumination and complex illumination,and Switchable Normalization was used instead of Instance Normalization used by original CycleGAN to solve the lighting problem.Through the comparison of image contrast and brightness,contrast and peak signal-to-noise ratio in the experiment,the illumination compensation effect of the method on the over dark picture and the strong reflection picture is higher than that of the traditional algorithm and the unmodified CycleGAN network.The effectiveness and robustness of the proposed algorithm are identified.

关 键 词：生成对抗网络 光照补偿 计算机视觉 深度学习 归一化 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]