《NTIRE 2020 Challenge on NonHomogeneous Dehazing》阅读笔记


NTIRE 2020 Challenge on NonHomogeneous Dehazing

Dataset

  • NH-Haze: 55对不同的室外有不均质雾图(5456×3632 with 24 bit depth)
  • DenseHaze,O-Haze

评价指标

  • PSNR
  • SSIM
  • LPIPS
  • PI

Architectures

  • U-Net, ResNet, DenseNet and Inception

Modths

Trident Dehazing Network (TDN)☑️

问题:

  • 对非均质雾没有鲁棒性
  • 浓雾区域的信息难以估计

主要思想:

技术路线:

==HDMGN子网的本质为求T图==

ECNU-KT☑️

dehaze sneaker

Spider

NTU Dehazing

VICLAB-DoNET

iPAL-NonLocal☑️

Team JJ

iPAL-EDN☑️

NTUEE LINLAB

​ The method is based on a encoder-decoder generator model with a multi-scale kernel encoder in the front (size is 3, 5, and 7). It is trained with part of the densenet-161. Next, the authors used BicycleGAN to enhance the generator. The hazy image and the difference of the hazy image and ground truth are used as the input when training the cVAE-GAN Encoder.

NTUST-merg

​ The proposed method is based on the At − DH Network as the backbone network. The authors used use DenseNet as the pretrained model in the encoder network. On the other hand, they employed the similar structure of the DenseNet with additional residual block in the decoder network. They used two decoders to estimate A (atmosphere light) and t(transmission map). Moreover, L2 loss and perceptual loss were used as loss functions. In the loss term, they both only calculate the estimated haze-free image and ground truth loss.

SIAT

neptuneai

Neuro-avengers☑️

NITREXZ

he proposed model contains a standard generative adversarial network.

AISAIL

DeBlurGAN

ICAIS dehaze

RETINA

This approach, named spatio-temporal retinex-inspired by an averaging of stochastic samples (STRASS) , is based on the spatio-temporal envelope retinex-inspired with a stochastic sampling framework (STRESS) [32] and also from the random spray retinex (RSR) [46]. In this work,

the authors used the idea of the relation developed in [19] replacing the envelope structure of the samples used in [32] by an average of these samples. Due to the local properties of the algorithm, this modified computation in the frame- work also impacted regions of the image far from the cam- era.

hazefreeworld


文章作者: Amber Ye
版权声明: 本博客所有文章除特別声明外,均采用 CC BY 4.0 许可协议。转载请注明来源 Amber Ye !
  目录