Computer Engineering and Applications ›› 2023, Vol. 59 ›› Issue (9): 225-236.DOI: 10.3778/j.issn.1002-8331.2201-0105

• Graphics and Image Processing • Previous Articles     Next Articles

Hybrid Samples Image Dehazing via Latent Space Translation

ZHENG Yutong, SUN Haoying, SONG Wei   

  1. 1.School of Information Engineering, Minzu University of China, Beijing 100081, China
    2.Key Laboratory of Marine Environmental Survey Technology and Application, Ministry of Natural Resource, Guangzhou 510310, China
    3.National Language Resource Monitoring & Research Center of Minority Languages, Beijing 100081, China
  • Online:2023-05-01 Published:2023-05-01

隐空间转换的混合样本图像去雾

郑玉彤,孙昊英,宋伟   

  1. 1.中央民族大学 信息工程学院, 北京 100081
    2.自然资源部海洋环境探测技术与应用重点实验室,广州 510310
    3.国家语言资源监测与研究少数民族语言中心,北京 100081

Abstract: Deep learning learns the inherent laws of samples from datasets which determine the performance of the model to a certain extent. However, it may be lack of paired real data, or difficult to synthesize paired data to simulate the real environment to train in single image dehazing dataset. This problem may cause the trained model doesn’t perform well in real hazy image. This paper proposes hybrid samples learning problem, and the hybrid samples learning algorithm based on latent space translation, aiming to make full use of paired data and unpaired data(hybrid samples). VAE-GAN(variational auto-encoder, generative adversarial networks) is used to encode hybrid samples into latent space, and then the adversarial loss is used to align real data with synthesis data. The mixup of feature adaptive fusion(MFF) module included in mapping net is used to learn the translation between paired data. So that, a dehazing data path from the real hazy image to the clear image is established. The experimental results show that proposed model performs well in real hazy images compared with other algorithms, and has outstanding effect on thick hazy images, and the peak signal to noise ratio of the proposed algorithm is higher than that of comparison algorithms.

Key words: single image dehazing, latent space translation, hybrid sample, variational autoencoder(VAE), generative adversarial network(GAN)

摘要: 深度学习从数据集中学习样本的内在规律,数据集的质量一定程度上决定了模型的表现。在去雾任务的公开数据集中,由于缺少成对真实数据,合成的成对数据难以模拟真实环境等问题,可能导致训练出的模型在实际环境中表现不佳。为此,提出混合样本学习问题,利用合成的成对数据和真实数据(混合样本)同时训练模型,通过隐空间的转换实现混合样本间的转换。算法利用变分自编码器和生成对抗网络(VAE-GAN)将混合样本分别编码到隐空间,利用对抗损失将真实数据的隐空间向合成雾图的隐空间对齐,利用含特征自适应融合(MFF)模块的映射网络学习成对数据隐空间之间的转换,从而建立起从真实雾图域到清晰图像域之间的去雾数据通路。实验结果表明,该算法相比其他去雾算法在真实雾图上的去雾结果更加清晰,对于较厚的雾图也有突出的效果,且该算法的峰值信噪比高于对比算法。

关键词: 单幅图像去雾, 隐空间转换, 混合样本, 变分自编码器(VAE), 生成对抗网络(GAN)