0%

(AAAI 2018) Towards Perceptual Image Dehazing by Physics-based Disentanglement and Adversarial Training

Yang X, Xu Z, Luo J. Towards perceptual image dehazing by physics-based disentanglement and adversarial training[C]//Thirty-second AAAI conference on artificial intelligence. 2018.



1. Overview


现有的de-hazing方法

  • on synthetic dataset
  • hand-designed priors
  • supervised training

因此,论文提出Disentangled Dehazing Network

  • weakly-supervised
  • GAN, multi-scale discriminator
  • physical-model based disentanglement
  • reconstruction
  • collect HazyCity dataset
  • DehazeNet
  • MSCNN
  • AOD-Net
  • CycleGAN
  • dualGAN
  • AIGN
  • WaterGAN
  • discoGAN
  • UNIT

1.2. 模型



1.2.1. Generator

  • G_J. 生成clean image
  • G_A.生成atmosphere light
  • G_t. 生成transmission map

1.2.2. Discriminator

multi-scale结构

  • local discriminator (感知域70x70). Model high-frequency structure (texture/style recognition)
  • global discriminator (感知域256x256). Global information, alleviate artifacts


1.3. Reconstruction



  • Reconstruction Loss



  • Adversarial Loss



  • Regularization Loss
    the smoothness of the medium transmission map.



1.4. Recovering Method of Haze-free Image

  • The output of G_J
  • Generate from A and t


  • Combine


1.5. 数据集

  • D-HAZY (synthetic)
    • β=1, A=255.
    • NYU-Depth (23 images).
    • Middlebury (1449 images).
  • HazyCity (real)
    • natural, outdoor. Build on PM25 dataset.
    • hazy (845), haze-free (1891)
    • crawled from tourist website and photos of various attraction sites and street scenes in Beijing. 三个标注者,选取标注一致的图片。

1.6. Future Work

  • de-raining
  • image matting



2. Experiments


2.1. Baseline

2.1.1. prior-based

  • DCP
  • CAP
  • NCP

2.1.2. learning-based

  • DehazeNet
  • MSCNN
  • CycleGAN

2.2. Metric

  • PSNR
  • SSIM
  • CIEDE2000. measure color difference

2.3. 实验结果




2.4. Ablation Study