作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2022, Vol. 48 ›› Issue (6): 207-212. doi: 10.19678/j.issn.1000-3428.0061823

• 图形图像处理 • 上一篇    下一篇

一种用于图像去雾的改进生成对抗网络

丁泳钧, 黄山   

  1. 四川大学 电气工程学院, 成都 610065
  • 收稿日期:2021-06-03 修回日期:2021-08-04 发布日期:2021-08-16
  • 作者简介:丁泳钧(1997—),男,硕士研究生,主研方向为计算机视觉、图像处理;黄山,教授、博士。
  • 基金资助:
    教育部产学合作协同育人项目(202002109040)。

An Improved Generative Adversarial Network for Image Dehazing

DING Yongjun, HUANG Shan   

  1. College of Electrical Engineering, Sichuan University, Chengdu 610065, China
  • Received:2021-06-03 Revised:2021-08-04 Published:2021-08-16

摘要: 雾霾天气会使计算机视觉相关系统获取到的图像质量下降并影响系统的正常工作。传统图像去雾算法通过分析大量图像建立模型,并找出清晰图像与模型之间的映射关系,但该类算法大多存在颜色失真和图像失真的问题,且在某些特定场景下可能失效。基于卷积神经网络的去雾算法得到的图像质量相对较好,但是该类算法对数据集要求较高,普遍需要成对数据进行训练,而获取同一时刻和场景下的有雾图像与无雾图像难度较高。提出一种基于循环生成对抗网络(CycleGAN)的改进图像去雾算法,该算法无需使用成对的数据进行训练。通过优化生成器生成的无雾图像与真实无雾图像之间的颜色损失,使得生成器能够生成与无雾图像具有相同颜色分布的图像。同时,通过向2个生成器中分别添加对应目标域图像的输入并引入特征损失函数,以解决经典CycleGAN在图像转换时存在的图像失真问题,从而更好地还原原始图像的细节特征。实验结果表明,相较DCP、CycleGAN、AOD-Net、Cycle-dehaze等算法,该算法的结构相似度提高4.3%~23.0%,峰值信噪比提高2.3%~36.9%,其能取得更好的去雾效果。

关键词: 图像去雾, 图像复原, 图像增强, 深度学习, 生成对抗网络

Abstract: Hazy weather degrades the quality of images obtained by computer vision related systems and affects the normal operation of the system.However, most traditional algorithms establish the relationship between the clear image model and image distortion, and most can determine the relationship between the specific image model and image distortion.The image quality obtained by the dehazing algorithm based on Convolutional Neural Network(CNN) is relatively high, but this type of algorithm has high dataset requirements and typically requires training on the data in pairs.However, it is difficult to obtain the hazy image and the haze-free image of the same scene at the same time.An improved image dehazing algorithm based on Cycle Generative Adversarial Network(CycleGAN) is proposed.The algorithm does not need to use paired data for training.By optimizing the color loss between the haze-free image generated by the generator and the real haze-free image, the generator can generate an image with the same color distribution as the haze-free image.Moreover, by adding the input of the corresponding target domain image to the two generators and introducing the feature loss function, the image distortion problem in the image conversion of classical CycleGAN is solved, improving the restoration of the detailed features of the original image.The results show that the structural similarity of the algorithm improved by 4.3%~23.0% compared with DCP, CycleGAN, AOD-Net, Cycle-dehaze, and other algorithms, and the Peak Signal-to-Noise Ratio(PSNR) improved by 2.3%~36.9%, thereby achieving an improved dehazing effect.

Key words: image dehazing, image restoration, image enhancement, deep learning, Generative Adversarial Network(GAN)

中图分类号: