0%

(CVPR 2018) Gated fusion network for single image dehazing

Posted on In , ,

Ren W, Ma L, Zhang J, et al. Gated fusion network for single image dehazing[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018: 3253-3261.


1. Overview

1.1. Motivation

Most existing methods follow the atmosphere scattering model.

In this paper, it proposed Gated Fusion Network (GFN)

  • directly restore a clear image from a hazy image
  • encoder-decoder
  • fusion-based strategy + confidence maps. White Balance, Contrast Enhancing and Gamma Correlation
  • multi-scale to avoid halo artifacts
  • GAN loss


1.2. Major Factors of Hazy Image

  • color cast introduced by atmosphere light
  • lack of visibility due to attenuation
  • solution
    • WB. eliminating chromatic casts
    • CE. better global visibility in thick hazy region but too dark in light hazy region
    • GC. recover the light hazy region

1.3. Contribution

  • Network not follow the atmosphere scattering model
  • demonstrate the utility and effectiveness of GFN
  • multi-scale approch to eliminate halo artifacts

1.4.1. Multiple-image aggregation

1.4.2. Hand-crafted Priors Based Methods

  • maximize the contrast
  • dark channel
  • color-lines
  • non-local
  • fusion luminance, chromatic and saliency maps

1.4.3. Data-driven Methods

  • combine four feature with Random Forest
  • color attenuation prior
  • deep learning

1.5. Dataset

A∈(0.8, 1.0), β∈[0.5, 1.5]. random sample 7 groups, add 1% Gaussian noise.

  • Train. 1400 images x7
  • Test. remained 49 images x7
  • RESIDE dataset. benchmark


2. Network



2.1. White Balance Input

  • recover the latent color of the scene and eliminate chromatic cast
  • but still present low contrast

2.1.1. Gary World Assumpation

the average value of the R,G,B components should average out to a common gray value. (Link)



2.2. Contrast Enhanced Input

  • subtracting the average luminance value
  • dark image region tend to black


2.3. Gamma Corrected Input

  • overcome the dark limitation of CE



2.4. Network

  • dilation Conv
  • Relu
  • 3 Conv + 3 DeConv, stride 1
  • first layer 5x5, other 3x3x32


2.5. Multi-Scale Refinement



  • vary the image resolution (x2) to preserve halo artifacts
  • loss function


  • motivation
    the human visual system is sensitive to local changes (e.g., edges) over a wide range of scales. As a merit, the multi-scale approach provides a convenient way to incorporate local image details over varying resolutions.

2.6. GAN Loss

apply to finest image.



2.7. Total Loss




3. Experiments

3.1. Details

  • 128x128 patches, batch size 10, 240,000 iteration
  • Adam, 0.0001 LR, LR decay
  • weight decay 0.00001
  • train 35 hours on K80

3.2. Comparison



light, medium and heavy (β=0.8, 1.0, 1.2).




3.3. Ablation Study

  • Multi-scale



  • Gated Fusion



3.4. Limitation



  • can not handle very large fog