DADM

Density-Aware Diffusion Model for Efficient Image Dehazing

1. Our Result comparison with other methods.

Guided by the haze density, our DADM can handle images with dense haze and complex environments.

2. Overview of our DADM.

For image dehazing, the haze image can be considered as noisy data, while the clear haze-free image is the target data. By training a diffusion model, we can learn a mapping from haze images to haze-free images. The diffusion model contains a forward diffusion process and a reverse diffusion process.
In the reverse diffusion process of our DADM, we introduce a density-aware dehazing network (DADNet) to estimate the noise in the input image and recover a haze-free image from a haze image (considered as a noisy state).

3. The haze density of the mage.

The dark channel value of noise areas is also very low, resulting in the inability to extract accurate haze density information. We introduce a cross-feature density extraction module (CDEModule) to optimize the dark channel map and obtain the accurate haze density for the image.

4. Analysis of the test sampling process on four datasets.

The dehazing results have achieved an optimum at some intermediate time point, while the image quality may instead degrade as the dehazing process continues.

5. Statistics on Datasets.

We evaluate all the sampling results at various time steps using the PSNR values against the ground truth. Based on the mean value $t_2$ and the variance $\sigma$ of the distribution, we can plot the corresponding normal distribution curves $c \sim \mathcal{N}(t_2, \sigma^2)$. Next, we set a criterion that the area under the corresponding normal distribution curve from $t_1$ to $t=0$ is 90% of the total area.

6. Sampling process for testing.

We use $t_1$ as the dividing time point. t > $t_1$ is the first stage, while t ≤ $t_1$ is the second stage. These two stages use different sampling methods.

7. Analysis of xt on four datasets.

We evaluate the PSNR vaules between $x_t$ and ground-truth for each time step on four datasets.

8. Visual Comparison with IR-SDE.

We compared our methods with IR-SDE.

Dependenices Version

OS: Ubuntu 20.04.6
nvdia: 11.6
python: 3.7.1
pytorch: 1.13.0

How to train?

datasets preparation and parameters configuration

You can directly configure the hazy and GT paths of the dataset and other parameters in the corresponding files under the Option folder. We have configured four different types of processing methods for different types of data sets hazy image and gt image formats.

indoor
outdoor
NH_Haze
NH-Haze2

train

bash train.sh

test

bash test.sh

DADM

Density-Aware Diffusion Model for Efficient Image Dehazing

1. Our Result comparison with other methods.

Guided by the haze density, our DADM can handle images with dense haze and complex environments.

2. Overview of our DADM.

For image dehazing, the haze image can be considered as noisy data, while the clear haze-free image is the target data. By training a diffusion model, we can learn a mapping from haze images to haze-free images. The diffusion model contains a forward diffusion process and a reverse diffusion process.
In the reverse diffusion process of our DADM, we introduce a density-aware dehazing network (DADNet) to estimate the noise in the input image and recover a haze-free image from a haze image (considered as a noisy state).

3. The haze density of the mage.

The dark channel value of noise areas is also very low, resulting in the inability to extract accurate haze density information. We introduce a cross-feature density extraction module (CDEModule) to optimize the dark channel map and obtain the accurate haze density for the image.

4. Analysis of the test sampling process on four datasets.

The dehazing results have achieved an optimum at some intermediate time point, while the image quality may instead degrade as the dehazing process continues.

5. Statistics on Datasets.

We evaluate all the sampling results at various time steps using the PSNR values against the ground truth. Based on the mean value $t_2$ and the variance $\sigma$ of the distribution, we can plot the corresponding normal distribution curves $c \sim \mathcal{N}(t_2, \sigma^2)$. Next, we set a criterion that the area under the corresponding normal distribution curve from $t_1$ to $t=0$ is 90% of the total area.

6. Sampling process for testing.

We use $t_1$ as the dividing time point. t > $t_1$ is the first stage, while t ≤ $t_1$ is the second stage. These two stages use different sampling methods.

7. Analysis of xt on four datasets.

We evaluate the PSNR vaules between $x_t$ and ground-truth for each time step on four datasets.

8. Visual Comparison with IR-SDE.

We compared our methods with IR-SDE.

Dependenices Version

OS: Ubuntu 20.04.6
nvdia: 11.6
python: 3.7.1
pytorch: 1.13.0

How to train?

datasets preparation and parameters configuration

You can directly configure the hazy and GT paths of the dataset and other parameters in the corresponding files under the Option folder. We have configured four different types of processing methods for different types of data sets hazy image and gt image formats.

indoor
outdoor
NH_Haze
NH-Haze2

train

bash train.sh

test

bash test.sh

DADM

Density-Aware Diffusion Model for Efficient Image Dehazing

1. Our Result comparison with other methods.

Guided by the haze density, our DADM can handle images with dense haze and complex environments.

2. Overview of our DADM.

For image dehazing, the haze image can be considered as noisy data, while the clear haze-free image is the target data. By training a diffusion model, we can learn a mapping from haze images to haze-free images. The diffusion model contains a forward diffusion process and a reverse diffusion process.
In the reverse diffusion process of our DADM, we introduce a density-aware dehazing network (DADNet) to estimate the noise in the input image and recover a haze-free image from a haze image (considered as a noisy state).

3. The haze density of the mage.

The dark channel value of noise areas is also very low, resulting in the inability to extract accurate haze density information. We introduce a cross-feature density extraction module (CDEModule) to optimize the dark channel map and obtain the accurate haze density for the image.

4. Analysis of the test sampling process on four datasets.

The dehazing results have achieved an optimum at some intermediate time point, while the image quality may instead degrade as the dehazing process continues.

5. Statistics on Datasets.

We evaluate all the sampling results at various time steps using the PSNR values against the ground truth. Based on the mean value $t_2$ and the variance $\sigma$ of the distribution, we can plot the corresponding normal distribution curves $c \sim \mathcal{N}(t_2, \sigma^2)$. Next, we set a criterion that the area under the corresponding normal distribution curve from $t_1$ to $t=0$ is 90% of the total area.

6. Sampling process for testing.

We use $t_1$ as the dividing time point. t > $t_1$ is the first stage, while t ≤ $t_1$ is the second stage. These two stages use different sampling methods.

7. Analysis of xt on four datasets.

We evaluate the PSNR vaules between $x_t$ and ground-truth for each time step on four datasets.

8. Visual Comparison with IR-SDE.

We compared our methods with IR-SDE.

Dependenices Version

OS: Ubuntu 20.04.6
nvdia: 11.6
python: 3.7.1
pytorch: 1.13.0

How to train?

datasets preparation and parameters configuration

You can directly configure the hazy and GT paths of the dataset and other parameters in the corresponding files under the Option folder. We have configured four different types of processing methods for different types of data sets hazy image and gt image formats.

indoor
outdoor
NH_Haze
NH-Haze2

train

bash train.sh

test

bash test.sh

DADM

Density-Aware Diffusion Model for Efficient Image Dehazing

1. Our Result comparison with other methods.

Guided by the haze density, our DADM can handle images with dense haze and complex environments.

2. Overview of our DADM.

For image dehazing, the haze image can be considered as noisy data, while the clear haze-free image is the target data. By training a diffusion model, we can learn a mapping from haze images to haze-free images. The diffusion model contains a forward diffusion process and a reverse diffusion process.
In the reverse diffusion process of our DADM, we introduce a density-aware dehazing network (DADNet) to estimate the noise in the input image and recover a haze-free image from a haze image (considered as a noisy state).

3. The haze density of the mage.

The dark channel value of noise areas is also very low, resulting in the inability to extract accurate haze density information. We introduce a cross-feature density extraction module (CDEModule) to optimize the dark channel map and obtain the accurate haze density for the image.

4. Analysis of the test sampling process on four datasets.

The dehazing results have achieved an optimum at some intermediate time point, while the image quality may instead degrade as the dehazing process continues.

5. Statistics on Datasets.

We evaluate all the sampling results at various time steps using the PSNR values against the ground truth. Based on the mean value $t_2$ and the variance $\sigma$ of the distribution, we can plot the corresponding normal distribution curves $c \sim \mathcal{N}(t_2, \sigma^2)$. Next, we set a criterion that the area under the corresponding normal distribution curve from $t_1$ to $t=0$ is 90% of the total area.

6. Sampling process for testing.

We use $t_1$ as the dividing time point. t > $t_1$ is the first stage, while t ≤ $t_1$ is the second stage. These two stages use different sampling methods.

7. Analysis of xt on four datasets.

We evaluate the PSNR vaules between $x_t$ and ground-truth for each time step on four datasets.

8. Visual Comparison with IR-SDE.

We compared our methods with IR-SDE.

Dependenices Version

OS: Ubuntu 20.04.6
nvdia: 11.6
python: 3.7.1
pytorch: 1.13.0

How to train?

datasets preparation and parameters configuration

You can directly configure the hazy and GT paths of the dataset and other parameters in the corresponding files under the Option folder. We have configured four different types of processing methods for different types of data sets hazy image and gt image formats.

indoor
outdoor
NH_Haze
NH-Haze2

train

bash train.sh

test

bash test.sh

DADM

Density-Aware Diffusion Model for Efficient Image Dehazing

1. Our Result comparison with other methods.

Guided by the haze density, our DADM can handle images with dense haze and complex environments.

2. Overview of our DADM.

For image dehazing, the haze image can be considered as noisy data, while the clear haze-free image is the target data. By training a diffusion model, we can learn a mapping from haze images to haze-free images. The diffusion model contains a forward diffusion process and a reverse diffusion process.
In the reverse diffusion process of our DADM, we introduce a density-aware dehazing network (DADNet) to estimate the noise in the input image and recover a haze-free image from a haze image (considered as a noisy state).

3. The haze density of the mage.

The dark channel value of noise areas is also very low, resulting in the inability to extract accurate haze density information. We introduce a cross-feature density extraction module (CDEModule) to optimize the dark channel map and obtain the accurate haze density for the image.

4. Analysis of the test sampling process on four datasets.

The dehazing results have achieved an optimum at some intermediate time point, while the image quality may instead degrade as the dehazing process continues.

5. Statistics on Datasets.

We evaluate all the sampling results at various time steps using the PSNR values against the ground truth. Based on the mean value $t_2$ and the variance $\sigma$ of the distribution, we can plot the corresponding normal distribution curves $c \sim \mathcal{N}(t_2, \sigma^2)$. Next, we set a criterion that the area under the corresponding normal distribution curve from $t_1$ to $t=0$ is 90% of the total area.

6. Sampling process for testing.

We use $t_1$ as the dividing time point. t > $t_1$ is the first stage, while t ≤ $t_1$ is the second stage. These two stages use different sampling methods.

7. Analysis of xt on four datasets.

We evaluate the PSNR vaules between $x_t$ and ground-truth for each time step on four datasets.

8. Visual Comparison with IR-SDE.

We compared our methods with IR-SDE.

Dependenices Version

OS: Ubuntu 20.04.6
nvdia: 11.6
python: 3.7.1
pytorch: 1.13.0

How to train?

datasets preparation and parameters configuration

You can directly configure the hazy and GT paths of the dataset and other parameters in the corresponding files under the Option folder. We have configured four different types of processing methods for different types of data sets hazy image and gt image formats.

indoor
outdoor
NH_Haze
NH-Haze2

train

bash train.sh

test

bash test.sh

DADM

Density-Aware Diffusion Model for Efficient Image Dehazing

1. Our Result comparison with other methods.

Guided by the haze density, our DADM can handle images with dense haze and complex environments.

2. Overview of our DADM.

For image dehazing, the haze image can be considered as noisy data, while the clear haze-free image is the target data. By training a diffusion model, we can learn a mapping from haze images to haze-free images. The diffusion model contains a forward diffusion process and a reverse diffusion process.
In the reverse diffusion process of our DADM, we introduce a density-aware dehazing network (DADNet) to estimate the noise in the input image and recover a haze-free image from a haze image (considered as a noisy state).

3. The haze density of the mage.

The dark channel value of noise areas is also very low, resulting in the inability to extract accurate haze density information. We introduce a cross-feature density extraction module (CDEModule) to optimize the dark channel map and obtain the accurate haze density for the image.

4. Analysis of the test sampling process on four datasets.

The dehazing results have achieved an optimum at some intermediate time point, while the image quality may instead degrade as the dehazing process continues.

5. Statistics on Datasets.

We evaluate all the sampling results at various time steps using the PSNR values against the ground truth. Based on the mean value $t_2$ and the variance $\sigma$ of the distribution, we can plot the corresponding normal distribution curves $c \sim \mathcal{N}(t_2, \sigma^2)$. Next, we set a criterion that the area under the corresponding normal distribution curve from $t_1$ to $t=0$ is 90% of the total area.

6. Sampling process for testing.

We use $t_1$ as the dividing time point. t > $t_1$ is the first stage, while t ≤ $t_1$ is the second stage. These two stages use different sampling methods.

7. Analysis of xt on four datasets.

We evaluate the PSNR vaules between $x_t$ and ground-truth for each time step on four datasets.

8. Visual Comparison with IR-SDE.

We compared our methods with IR-SDE.

Dependenices Version

OS: Ubuntu 20.04.6
nvdia: 11.6
python: 3.7.1
pytorch: 1.13.0

How to train?

datasets preparation and parameters configuration

You can directly configure the hazy and GT paths of the dataset and other parameters in the corresponding files under the Option folder. We have configured four different types of processing methods for different types of data sets hazy image and gt image formats.

indoor
outdoor
NH_Haze
NH-Haze2

train

bash train.sh

test

bash test.sh

DADM

Density-Aware Diffusion Model for Efficient Image Dehazing

1. Our Result comparison with other methods.

Guided by the haze density, our DADM can handle images with dense haze and complex environments.

2. Overview of our DADM.

For image dehazing, the haze image can be considered as noisy data, while the clear haze-free image is the target data. By training a diffusion model, we can learn a mapping from haze images to haze-free images. The diffusion model contains a forward diffusion process and a reverse diffusion process.
In the reverse diffusion process of our DADM, we introduce a density-aware dehazing network (DADNet) to estimate the noise in the input image and recover a haze-free image from a haze image (considered as a noisy state).

3. The haze density of the mage.

The dark channel value of noise areas is also very low, resulting in the inability to extract accurate haze density information. We introduce a cross-feature density extraction module (CDEModule) to optimize the dark channel map and obtain the accurate haze density for the image.

4. Analysis of the test sampling process on four datasets.

The dehazing results have achieved an optimum at some intermediate time point, while the image quality may instead degrade as the dehazing process continues.

5. Statistics on Datasets.

We evaluate all the sampling results at various time steps using the PSNR values against the ground truth. Based on the mean value $t_2$ and the variance $\sigma$ of the distribution, we can plot the corresponding normal distribution curves $c \sim \mathcal{N}(t_2, \sigma^2)$. Next, we set a criterion that the area under the corresponding normal distribution curve from $t_1$ to $t=0$ is 90% of the total area.

6. Sampling process for testing.

We use $t_1$ as the dividing time point. t > $t_1$ is the first stage, while t ≤ $t_1$ is the second stage. These two stages use different sampling methods.

7. Analysis of xt on four datasets.

We evaluate the PSNR vaules between $x_t$ and ground-truth for each time step on four datasets.

8. Visual Comparison with IR-SDE.

We compared our methods with IR-SDE.

Dependenices Version

OS: Ubuntu 20.04.6
nvdia: 11.6
python: 3.7.1
pytorch: 1.13.0

How to train?

datasets preparation and parameters configuration

You can directly configure the hazy and GT paths of the dataset and other parameters in the corresponding files under the Option folder. We have configured four different types of processing methods for different types of data sets hazy image and gt image formats.

indoor
outdoor
NH_Haze
NH-Haze2

train

bash train.sh

test

bash test.sh

DADM

Density-Aware Diffusion Model for Efficient Image Dehazing

1. Our Result comparison with other methods.

Guided by the haze density, our DADM can handle images with dense haze and complex environments.

2. Overview of our DADM.

For image dehazing, the haze image can be considered as noisy data, while the clear haze-free image is the target data. By training a diffusion model, we can learn a mapping from haze images to haze-free images. The diffusion model contains a forward diffusion process and a reverse diffusion process.
In the reverse diffusion process of our DADM, we introduce a density-aware dehazing network (DADNet) to estimate the noise in the input image and recover a haze-free image from a haze image (considered as a noisy state).

3. The haze density of the mage.

The dark channel value of noise areas is also very low, resulting in the inability to extract accurate haze density information. We introduce a cross-feature density extraction module (CDEModule) to optimize the dark channel map and obtain the accurate haze density for the image.

4. Analysis of the test sampling process on four datasets.

The dehazing results have achieved an optimum at some intermediate time point, while the image quality may instead degrade as the dehazing process continues.

5. Statistics on Datasets.

We evaluate all the sampling results at various time steps using the PSNR values against the ground truth. Based on the mean value $t_2$ and the variance $\sigma$ of the distribution, we can plot the corresponding normal distribution curves $c \sim \mathcal{N}(t_2, \sigma^2)$. Next, we set a criterion that the area under the corresponding normal distribution curve from $t_1$ to $t=0$ is 90% of the total area.

6. Sampling process for testing.

We use $t_1$ as the dividing time point. t > $t_1$ is the first stage, while t ≤ $t_1$ is the second stage. These two stages use different sampling methods.

7. Analysis of xt on four datasets.

We evaluate the PSNR vaules between $x_t$ and ground-truth for each time step on four datasets.

8. Visual Comparison with IR-SDE.

We compared our methods with IR-SDE.

Dependenices Version

OS: Ubuntu 20.04.6
nvdia: 11.6
python: 3.7.1
pytorch: 1.13.0

How to train?

datasets preparation and parameters configuration

You can directly configure the hazy and GT paths of the dataset and other parameters in the corresponding files under the Option folder. We have configured four different types of processing methods for different types of data sets hazy image and gt image formats.

indoor
outdoor
NH_Haze
NH-Haze2

train

bash train.sh

test

bash test.sh

Name		Name	Last commit message	Last commit date
datasets		datasets
figs		figs
models		models
nets		nets
options		options
utils		utils
README.md		README.md
test.py		test.py
test.sh		test.sh
train.py		train.py
train.sh		train.sh

Name		Name	Last commit message	Last commit date
datasets		datasets
figs		figs
models		models
nets		nets
options		options
utils		utils
README.md		README.md
test.py		test.py
test.sh		test.sh
train.py		train.py
train.sh		train.sh

Name		Name	Last commit message	Last commit date
datasets		datasets
figs		figs
models		models
nets		nets
options		options
utils		utils
README.md		README.md
test.py		test.py
test.sh		test.sh
train.py		train.py
train.sh		train.sh

Name		Name	Last commit message	Last commit date
datasets		datasets
figs		figs
models		models
nets		nets
options		options
utils		utils
README.md		README.md
test.py		test.py
test.sh		test.sh
train.py		train.py
train.sh		train.sh

Name		Name	Last commit message	Last commit date
datasets		datasets
figs		figs
models		models
nets		nets
options		options
utils		utils
README.md		README.md
test.py		test.py
test.sh		test.sh
train.py		train.py
train.sh		train.sh

Latest commit

History

Folders and files

Repository files navigation

DADM

Density-Aware Diffusion Model for Efficient Image Dehazing

1. Our Result comparison with other methods.

2. Overview of our DADM.

3. The haze density of the mage.

4. Analysis of the test sampling process on four datasets.

5. Statistics on Datasets.

6. Sampling process for testing.

7. Analysis of xt on four datasets.

8. Visual Comparison with IR-SDE.

Dependenices Version

How to train?

datasets preparation and parameters configuration

train

test

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

DADM

Density-Aware Diffusion Model for Efficient Image Dehazing

1. Our Result comparison with other methods.

2. Overview of our DADM.

3. The haze density of the mage.

4. Analysis of the test sampling process on four datasets.

5. Statistics on Datasets.

6. Sampling process for testing.

7. Analysis of xt on four datasets.

8. Visual Comparison with IR-SDE.

Dependenices Version

How to train?

datasets preparation and parameters configuration

train

test

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

DADM

Density-Aware Diffusion Model for Efficient Image Dehazing

1. Our Result comparison with other methods.

2. Overview of our DADM.

3. The haze density of the mage.

4. Analysis of the test sampling process on four datasets.

5. Statistics on Datasets.

6. Sampling process for testing.

7. Analysis of xt on four datasets.

8. Visual Comparison with IR-SDE.

Dependenices Version

How to train?

datasets preparation and parameters configuration

train

test

About

Resources

Stars

Watchers

Forks