一种基于条件生成对抗网络的去雾方法.pdf

信息与电脑 算法语言 China ComputerCommunication 2018年第9期 一种基于条件生成对抗网络的去雾方法 贾绪仲 文志强 （湖南工业大学 计算机学院，湖南 株洲 412007） 摘 要： 雾霾环境下所捕获的照片中，主要信息往往被雾所覆盖。从带雾图像中恢复有效信息一直以来都是计算机 视觉领域研究的热门，如何快速行之有效地从带雾图像中恢复出高质量的无雾图像一直困扰着人们。笔者受到生成对抗 网络应用于图像风格转换的启发，设计出基于条件生成对抗网络的去雾模型，为了能达到更好的去雾效果，对损失函数 进行重新设计。在损失函数中加入 l1 loss 解决了生成图像的模糊问题，并针对所生成的无雾图像边界信息缺失问题， 引入了感知损失函数来解决。通过本文方法所训练的去雾模型，无论是在整体视觉效果、多层结构相似度、均方差等方 面都要优于传统方法。 关键词：图像去雾；条件生成对抗网络；深度学习 中图分类号：TP391.41    文献标识码：A    文章编号：1003-9767(2018)09-060-04 Image Dehazing Using Conditional Generative Adversarial Nets Jia Xuzhong, Wen Zhiqiang (School of Computer Science, Hunan University of Technology, Zhuzhou Hunan 412007, China) Abstract: Photos which were captured in the haze environment, the main information is often covered by fog. Restoring effective information from images with fog has always been a hot topic in the field of computer vision. How to recover high-quality fog-free images from fog images has been plaguing us all the time. The author is inspired by the generation of confrontation networks to apply image style conversion, and designs a defogging model based on conditional generation against network. In order to achieve a better defogging effect, the loss function is redesigned. Adding l1 loss to the loss function solves the blurring problem of the generated image, and solves the problem of lack of boundary information of the generated fog-free image by introducing a perceptual loss function. The defogging model trained by this method is superior to the traditional method in terms of overall visual effect, similarity of multi-layer structure, and mean square error. Key words: image dehazing; condition generative adversarial nets; deep learning 1 引言