0%

(AAAI 2018) Densely connected pyramid dehazing network

Posted on In , ,

Zhang H, Patel V M. Densely connected pyramid dehazing network[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 3194-3203.


1. Overview

1.1. Motivation

Most existing method

  • inaccuracies in the estimation of transmission map
  • not leverage end-to-end training, unable to capture the inherent relation among transmission map, atmospheric light and dehazed image

In this paper, it proposed Densely Connected Pyramid Dehazing Network (DCPDN)

  • jointly learn transmission map, atmospheric light and dehazed image
  • multi-level pyramid pooling for the estimation of transmission map
  • edge-preserving loss function
  • joint-discriminator
  • stage-wise learning

1.2. Contribution

  • jointly network
  • jointly discriminator
  • edge-preserving loss, edge-preserving pyramid densely connected encoder-decoder network
  • experiments, ablation study

1.3.1. Prior-based

  • dark-channel
  • contrast color-lines
  • haze-line

    1.3.2. Learning-based

    1.3.3. GAN

1.4. Dataset

A∈[0.5, 1]， β∈[0.4, 1.6]. random sample 4 groups.

  • TrainA. 1000 images from NYU-depth2, get 4000 paris
  • TestA. 100 images from NYU-depth2, get 400 paris
  • TestB. 200 paris fromMiddlebury and Sun3D


2. Network



2.1. Pyramid Densely Connected Transmission Map Estimation Net



dense: maximize information flow.

  • Encoder. 5 (dense block + down-sample transition block)
  • Decoder. 5 (dense block + up-sample transition block)
  • Multi-level pyramid pooling module. deal with lacking of global structure information

2.2. Atmosphere Light Estimation Net

  • U-net. 4 (Conv-BN-Relu) + 4 (DeConv-BN-Relu)

2.3. Joint Discriminator



2.4. Edge-preserving Loss



  • (set 1) L2 loss of transmission map
  • (set 0.8) two-directional gradient loss

edge corresponds to the discontinuities in the image intensities.



  • (0.8) feature edge loss

low-level features (edges and contour) can be captured in the shallow layers.



relu1-1 and relu2-1 of VGG16.

2.5. Loss Function



  • L^t. edge-preserving loss
  • L^a. L2 of atmospheric loss
  • L^d. L2 of dehazed loss
  • (0.25) L^j. joint Discriminator loss

Stage-wise training and then fine-tuned.


3. Experiments

3.1. Details

  • LR. 0.002, Adam
  • batch size 1, 512x512, 400000 iteration

3.2. Ablation Study



  • multi-level pooling. better preserve the global structural for obj with relatively larger scale
  • edge-preserving loss. better refined the edge of transmission map
  • joint Discriminator. enhance the transmission map

3.3. Comparison