Self-augmented Unpaired Image Dehazing via Density and Depth Decomposition program running record

Thesis title:

"Self-augmented Unpaired Image Dehazing via Density and Depth Decomposition"

Author List:

Yang Yang (Tianjin University), Wang Chaoyue (Sydney University), Liu Risheng (Dalian University of Technology), Zhang Lin (Tongji University), Guo Xiaojie (Tianjin University), Tao Dacheng (Sydney University, JD Discovery Research Institute)

Paper abstract:

To overcome the overfitting problem of dehazing models trained on synthetic datasets, many recent approaches attempt to use unpaired data for training to improve the generalization ability of the models.However, most of these methods simply follow the idea of ​​CycleGAN to construct the dehazing cycle and the fogging cycle, but ignore the physical characteristics of the haze environment in the real world, that is, the influence of haze on the visibility of objects varies with depth and fog density.In this paper, we propose a self-enhancing image dehazing framework, called D4 (Dehazing via Decomposing transmission map into Density and Depth), for image dehazing and fog generation.Rather than simply estimating transmission maps or clear images, our proposed framework focuses on exploring scattering coefficients and depth information in hazy and clear images.With the estimated scene depth, our method is able to re-render foggy images with different thicknesses of fog and improve the performance of the dehazing network as a self-data augmentation mechanism.Notably, the entire training process relies only on unpaired hazy and clear images, successfully recovering scattering coefficients, depth maps, and clear images from a single hazy image.Comprehensive experiments show that our method outperforms state-of-the-art non-pairwise dehazing methods with fewer parameters and FLOPs.

Paper Information:

[1] Yang Yang, Chaoyue Wang, Risheng Liu, Lin Zhang, Xiaojie Guo, Dacheng Tao. Self-augmented Unpaired Image Dehazing via Density and Depth Decomposition. CVPR 2022. 

Paper link:

https://www.aliyundrive.com/s/WqWvoBkrzBi

Code link:

https://github.com/YaN9-Y/D4

Related instructional videos:

https://www.bilibili.com/video/BV1YF411L72F?share_source=copy_web&vd_source=7e63a4719cd8caa0c6c43bdcc04828df

Code debugging:

1. Required packages

cv2

pip install opencv-python

yaml

pip install pyyaml

kornia

pip install kornia

matplotlib

pip install matplotlib

2. Some errors and their solutions

(1) The yaml.load function is missing the Loader parameter

Change the corresponding line in the corresponding config.py file to

self._dict = yaml.load(self._yaml, Loader=yaml.FullLoader)

(2) No response to the Internet

Load local model instead, src/blockLine 57 in .py

def _make_pretrained_efficientnet_lite3(use_pretrained, exportable=False):#efficientnet = torch.hub.load(# "rwightman/gen-efficientnet-pytorch",# "tf_efficientnet_lite3",# pretrained=use_pretrained,# exportable=exportable,# ) It needs to be connected to the Internet, so change it to the following writing and load the local fileefficientnet = torch.hub.load("/home/yanhaorui/.cache/torch/hub/rwightman_gen-efficientnet-pytorch_master","tf_efficientnet_lite3",pretrained=use_pretrained,exportable=exportable,source='local')return _make_efficientnet_backbone(efficientnet)

3. The method in kornia is wrong

The corresponding line in D4.py is changed to

depth = kornia.filters.median_blur(depth,(9,9))

4. save() parameter problem

The corresponding function of D4.py is changed to

def save(self, save_best=False, psnr=None, iteration=None):self.model.save()