Unsupervised Image Dehazing Enhancement Method and System Based on Decoupling
By employing a decoupled unsupervised image dehazing method, and utilizing an improved atmospheric light scattering model and different loss functions to train the network directly on real data, the problem of image darkening and domain drift after dehazing is solved. This achieves efficient and clear restoration of foggy images, reducing the risk of traffic accidents.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SUN YAT SEN UNIV
- Filing Date
- 2023-03-22
- Publication Date
- 2026-05-26
AI Technical Summary
Existing methods using contrast-limited histogram equalization to generate pseudo-labels for fine-tuning the network can result in excessively high saturation in the dehazing results. Furthermore, the network needs to be trained on synthetic data and then fine-tuned on real data, which increases training costs.
A decoupled unsupervised image dehazing method is adopted. The feature map is obtained by inputting the foggy image into the encoder, and then input into the dehazing map, atmospheric light, residual coefficient and transmittance map networks respectively. The network is trained using an improved atmospheric light scattering model and different loss functions to avoid domain drift and train directly on real data.
It achieves efficient dehazing on real data, avoiding image darkening and domain drift after dehazing, improving image clarity, and is suitable for image restoration under foggy conditions, reducing the risk and loss of traffic accidents.
Smart Images
Figure CN116542867B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image restoration, and specifically to an unsupervised image dehazing enhancement method and system based on decoupling. Background Technology
[0002] Fog is a common atmospheric phenomenon. In foggy weather, images captured by lenses are blurry and have reduced contrast. Currently, most image-based applications rely on images obtained in good weather, which degrades the performance of advanced image-based applications such as object recognition and pedestrian tracking. Image dehazing has thus attracted considerable attention from researchers.
[0003] Furthermore, obtaining paired images of fog and their corresponding clear ones is nearly impossible, and researchers currently use atmospheric light scattering models to synthesize data. Deep learning-based dehazing methods have achieved significant performance. Due to the lack of paired datasets, the synthetic dataset RESIDE is often used for training. While this addresses the need for large amounts of data in deep learning, the distribution of synthetic datasets differs from that of real datasets. This often results in darker dehazed images using synthetic datasets because of domain drift, meaning the distribution of synthetic foggy images differs from that of real data. Although semi-supervised domain adaptation methods exist to address domain drift by applying constraints in the representation space to narrow the distribution distance between synthetic and real data, their generalization ability on real data is insufficient, and pseudo-labels need to be generated.
[0004] One current technique is the semi-supervised layer dehazing method proposed in the paper "Principled synthetic-to-real dehazing guided by physical priors". This method treats a foggy image as an entanglement of several simple layers, namely, a fog-free image layer, a transmission layer, and an atmospheric light layer. It first uses synthetic data for supervised training, and then outputs a fog-free image in the network output layer. Transmittance diagram and atmospheric light value Then, an atmospheric light scattering model is used to construct a loss function to train the network. After training on synthetic data, the same network is used on real data to generate pseudo-labels for fine-tuning. The disadvantage of this approach is that the use of pseudo-labels during fine-tuning, although it can enhance the brightness of the image, will cause the image to be re-foggy after defogging, resulting in overly saturated results; this approach requires training on synthetic data first and then fine-tuning on real data, which increases the training cost. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of existing methods and propose an unsupervised image dehazing enhancement method and system based on decoupling. The main problems addressed by this invention are: first, using contrast-limited histogram equalization to generate pseudo-labels for fine-tuning the network leads to re-hazing after dehazing, resulting in excessively high saturation; second, it requires training the model on synthetic data and then fine-tuning it using real data, increasing training costs.
[0006] To address the aforementioned problems, this invention proposes an unsupervised image dehazing enhancement method based on decoupling, the method comprising:
[0007] The foggy image is input into the encoder to obtain the foggy feature map z;
[0008] The foggy feature map is input into the J-net dehazing network layer to obtain the dehazing map.
[0009] The fog feature map is input into the atmospheric light network layer A-net, which outputs the decoupled atmospheric light value.
[0010] The foggy feature map is input into the residual coefficient network layer α-net to calculate the residual coefficients.
[0011] The hazy feature map is input into the transmittance map network layer t-net to obtain the transmittance map.
[0012] According to the defogging diagram The atmospheric light value The remaining coefficient and the transmittance map The foggy image is resynthesized, and the untangling of layers is guided by the fog reconstruction process. The network is then trained to obtain the defoggy image.
[0013] Preferably, the process involves inputting the foggy feature map into the J-net dehazing network layer to obtain the dehazing map. Specifically:
[0014] Using dark channel loss to constrain J-Net enables it to achieve dehazing capabilities. The dark channel prior reveals that in natural images, one of the three channels always has a pixel size of zero or close to zero, which is expressed as:
[0015]
[0016] Where D(·) is the dark channel operation, J c This is the output value of J-net, i.e., the dehazed image. c (y) is the pixel intensity of color channel c, and Ω(x) is the input image J. cA fixed-size local block in the image, dehazed using dark channel loss constraints, is defined as follows:
[0017]
[0018] N represents the amount of data. By reducing the dark channel value in the recovered image, an ideal haze-free image is obtained.
[0019] Preferably, the fog feature map is input into the atmospheric light network layer A-net, and the decoupled atmospheric light value is output. Specifically:
[0020] The pixel with the highest intensity in the input image I is selected as the true atmospheric light value A, and the loss is as follows:
[0021]
[0022] ‖·‖ F It uses the F-norm. For some white objects, pixel intensities are greater or less than atmospheric light values. Atmospheric light values are independent of the ground truth image values. A reparameterization technique is introduced, performing reparameterization on the variational lower bound to produce a lower bound estimate that can be directly optimized using the standard stochastic gradient method. Latent variables... It follows a standard normal distribution N(0,1), using KL divergence.
[0023]
[0024] Among them, z i It is the i-th dimension of z. During execution, the image is progressively input into A-net to obtain z, which is obtained by following a normal distribution. Gradually sample, by minimizing To enhance z, the constraints of the atmospheric light value network layer are:
[0025]
[0026] Preferably, the foggy feature map is input into the residual coefficient network layer α-net to calculate the residual coefficients. Specifically:
[0027] Residual coefficient max(d(x)) represents the maximum value of the depth map d(x), i.e., the farthest scene point in the image. Let the residual coefficients estimated by α-net be... The loss function of αβ-net is as follows:
[0028]
[0029] in‖·‖ p It is a p-norm.
[0030] Preferably, the process involves inputting the fog feature map into the transmittance map network layer t-net to obtain the transmittance map. Specifically:
[0031] To enhance the stability of the model, the transmittance map needs to be constrained. A transmittance map estimated using Laplacian smoothing is employed.
[0032]
[0033] Wherein, N(t) i ) is the second nearest neighbor of the i-th pixel for estimating transmittance, |N(t) i )| is the number of nearest neighbors, and m is the number of pixels.
[0034] Preferably, the step of determining the dehazing map The atmospheric light value The remaining coefficient and the transmittance map The fogged image is resynthesized, and the unwrapping of layers is guided by the fog reconstruction process. The network is then trained to obtain the dehazed image, specifically as follows:
[0035] The improved atmospheric scattering model (IASM) is expressed by the following formula:
[0036] I(x)=A·α(x)·R(x)·t(x)+A·(1-t(x))
[0037] Where I(x) represents the obtained foggy image, A refers to the atmospheric light value, R(x) is the scene reflectance, and t(x) = e -βd(x) This is a transmittance map, which represents the amount of light from the scene that passes through the fog and reaches the camera. β is the wavelength of the light, d(x) is the image depth, and α(x) is the residual coefficient, which expresses the amount of light remaining after the material absorbs the light. Light is absorbed by the texture, and the light absorption rate changes with the texture density. α(x)∈(0,1], so A·α(x)·R(x)≤A·R(x), which is the reason why it becomes darker after defogging.
[0038] According to the defogging diagram The atmospheric light value The remaining coefficient and the transmittance map Re-synthesized fog map The difference between the synthesized hazy image and the input image is calculated, then...
[0039]
[0040] ‖·‖2 is the 2-norm. The entire network, including sub-networks, is constrained. Layer unwrapping is guided by merging the fog reconstruction process. A supervised mechanism is used to train given J-net, A-net, t-net, and α-net, resulting in the total unsupervised defogging loss of the trained network.
[0041]
[0042] For an input image, iterate the network 100 times to minimize the total loss and obtain the dehazed image.
[0043] Accordingly, the present invention also provides an unsupervised image dehazing enhancement system based on decoupling, comprising:
[0044] The feature map acquisition unit is used to input the foggy image into the encoder to obtain the foggy feature map z;
[0045] The dehazing unit is used to input the hazy feature map into the dehazing network layer J-net to obtain the dehazing map.
[0046] The atmospheric light value unit is used to input the fog feature map into the atmospheric light network layer A-net and output the decoupled atmospheric light value.
[0047] The residual coefficient unit is used to input the foggy feature map into the residual coefficient network layer α-net to calculate the residual coefficients.
[0048] The transmittance mapping unit inputs the hazy feature map into the transmittance mapping network layer t-net to obtain the transmittance map.
[0049] The model training unit, based on the dehazed image The atmospheric light value The remaining coefficient and the transmittance map The foggy image is resynthesized, and the untangling of layers is guided by the fog reconstruction process. The network is then trained to obtain the defoggy image.
[0050] Implementing this invention has the following beneficial effects:
[0051] This method uses an improved atmospheric light scattering model, introducing residual coefficients to explain the darkening of the scene after image dehazing. Based on the improved atmospheric light scattering model, a new decoupled dehazing network is proposed, with different losses designed for different sub-networks to ensure that the sub-network outputs conform to physical laws. Training is performed on real data to avoid domain drift problems caused by synthetic data, resulting in clear dehazed images. In foggy public places such as stations and squares, the method effectively restores clear images, thereby effectively avoiding danger and reducing the occurrence of traffic accidents. Even in the event of a traffic accident, the dehazing results from the recorder can be used to clarify responsibility and reduce losses. This invention has high intelligence and employs a dimensionality reduction method for high computational efficiency, making it convenient and efficient. Attached Figure Description
[0052] Figure 1 This is a flowchart of the unsupervised image dehazing enhancement method based on decoupling according to an embodiment of the present invention;
[0053] Figure 2 This is a structural diagram of the unsupervised image dehazing enhancement system based on decoupling according to an embodiment of the present invention. Detailed Implementation
[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0055] Figure 1 This is a flowchart illustrating the overall process of the unsupervised image dehazing enhancement method based on decoupling according to an embodiment of the present invention. Figure 1 As shown, the method includes:
[0056] S1, Input the foggy image into the encoder to obtain the foggy feature map z;
[0057] S2, input the foggy feature map into the dehazing network layer J-net to obtain the dehazing map.
[0058] S3, input the fog feature map into the atmospheric light network layer A-net, and output the decoupled atmospheric light value.
[0059] S4, input the foggy feature map into the residual coefficient network layer α-net, and calculate the residual coefficients.
[0060] S5, input the fog feature map into the transmittance map network layer t-net to obtain the transmittance map.
[0061] S6, based on the dehazing map The atmospheric light value The remaining coefficient and the transmittance map The foggy image is resynthesized, and the untangling of layers is guided by the fog reconstruction process. The network is then trained to obtain the defoggy image.
[0062] Step S1 is as follows:
[0063] S1-1, Input the foggy image into the encoder to obtain the foggy feature map z.
[0064] Step S2 is as follows:
[0065] S2-1 uses dark channel loss to constrain J-net, enabling J-net to achieve dehazing capabilities. The dark channel prior reveals that in natural images, one of the three channels always has a pixel size of zero or close to zero, expressed as:
[0066]
[0067] Where D(·) is the dark channel operation, J c This is the output value of J-net, i.e., the dehazed image. c (y) is the pixel intensity of color channel c, and Ω(x) is the input image J. c A fixed-size local block in the image, dehazed using dark channel loss constraints, is defined as follows:
[0068]
[0069] N represents the amount of data. By reducing the dark channel value in the recovered image, an ideal haze-free image is obtained.
[0070] Step S3 is as follows:
[0071] S3-1, Select the pixel with the highest intensity in the input image I as the true atmospheric light value A, with the following loss:
[0072]
[0073] ‖·‖ F It uses the F-norm. For some white objects, pixel intensities are greater or less than atmospheric light values. Atmospheric light values are independent of the ground truth image values. A reparameterization technique is introduced, performing reparameterization on the variational lower bound to produce a lower bound estimate that can be directly optimized using the standard stochastic gradient method. Latent variables... It follows a standard normal distribution N(0,1), using KL divergence.
[0074]
[0075]
[0076] Among them, z i It is the i-th dimension of z. During execution, the image is progressively input into A-net to obtain z, which is obtained by following a normal distribution. Gradually sample, by minimizing To enhance z, the constraints of the atmospheric light value network layer are:
[0077]
[0078] Step S4 is as follows:
[0079] S4-1, Residual Coefficient max(d(x)) represents the maximum value of the depth map d(x), i.e., the farthest scene point in the image. Let the residual coefficients estimated by α-net be... The loss function of α-net is as follows:
[0080]
[0081] in‖·‖ p It is a p-norm.
[0082] Step S5 is as follows:
[0083] S5-1, To enhance the stability of the model, the transmittance map needs to be constrained. A transmittance map estimated using Laplacian smoothing is employed.
[0084]
[0085] Wherein, N(t) i ) is the second nearest neighbor of the i-th pixel for estimating transmittance, |N(t) i )| is the number of nearest neighbors, and m is the number of pixels.
[0086] Step S6 is as follows:
[0087] S6-1, the improved atmospheric scattering model (IASM), is expressed by the following formula:
[0088] I(x)=A·α(x)·R(x)·t(x)+A·(1-t(x))
[0089] Where I(x) represents the obtained foggy image, A refers to the atmospheric light value, R(x) is the scene reflectance, and t(x) = e -βd(x)This is a transmittance map, which represents the amount of light from the scene that passes through the fog and reaches the camera. β is the wavelength of the light, d(x) is the image depth, and α(x) is the residual coefficient, which expresses the amount of light remaining after the material absorbs the light. The light is absorbed by the texture, and the light absorption rate changes with the texture density. α(x)∈(0,1], so A·α(x)·R(x)≤A·R(x), which is the reason why it becomes darker after defogging.
[0090] S6-2, based on the dehazing diagram The atmospheric light value The remaining coefficient and the transmittance map Re-synthesized fog map The difference between the synthesized hazy image and the input image is calculated, then...
[0091]
[0092] ‖·‖2 is the 2-norm. The entire network, including sub-networks, is constrained. Layer unwrapping is guided by merging the fog reconstruction process. A supervised mechanism is used to train given J-net, A-net, t-net, and α-net, resulting in the total unsupervised defogging loss of the trained network.
[0093]
[0094] For an input image, iterate the network 100 times to minimize the total loss and obtain the dehazed image.
[0095] Accordingly, the present invention also provides an unsupervised image dehazing enhancement system based on decoupling, such as... Figure 2 As shown, it includes:
[0096] Feature map acquisition unit 1 is used to input the foggy image into the encoder to obtain the foggy feature map z;
[0097] Dehazing unit 2 is used to input the hazy feature map into the dehazing network layer J-net to obtain the dehazing map.
[0098] Specifically, dark channel loss is used to constrain J-net, enabling it to achieve dehazing capabilities. The dark channel prior reveals that in natural images, one of the three channels always has a pixel size of zero or close to zero, expressed as:
[0099]
[0100] Where D(·) is the dark channel operation, J c This is the output value of J-net, i.e., the dehazed image. c(y) is the pixel intensity of color channel c, and Ω(x) is the input image J. c A fixed-size local block in the image, dehazed using dark channel loss constraints, is defined as follows:
[0101]
[0102] N represents the amount of data. By reducing the dark channel value in the recovered image, an ideal haze-free image is obtained.
[0103] Atmospheric light value unit 3 is used to input the fog feature map into the atmospheric light network layer A-net and output the decoupled atmospheric light value.
[0104] Specifically, the pixel with the highest intensity in the input image I is selected as the true atmospheric light value A, and the loss is as follows:
[0105]
[0106] ‖·‖ F It uses the F-norm. For some white objects, pixel intensities are greater or less than atmospheric light values. Atmospheric light values are independent of the ground truth image values. A reparameterization technique is introduced, performing reparameterization on the variational lower bound to produce a lower bound estimate that can be directly optimized using the standard stochastic gradient method. Latent variables... It follows a standard normal distribution N(0,1), using KL divergence.
[0107]
[0108] Among them, z i It is the i-th dimension of z. During execution, the image is progressively input into A-net to obtain z, which is obtained by following a normal distribution. Gradually sample, by minimizing To enhance z, the constraints of the atmospheric light value network layer are:
[0109]
[0110] Residual coefficient unit 4 is used to input the foggy feature map into the residual coefficient network layer α-net to calculate the residual coefficients.
[0111] Specifically, the residual coefficient max(d(x)) represents the maximum value of the depth map d(x), i.e., the farthest scene point in the image. Let the residual coefficients estimated by α-net be... The loss function of α-net is as follows:
[0112]
[0113] in‖·‖p It is a p-norm.
[0114] Transmittance mapping unit 5 is used to input the hazy feature map into the transmittance mapping network layer t-net to obtain the transmittance map.
[0115] Specifically, to enhance the stability of the model, the transmittance map needs to be constrained using a transmittance map estimated with Laplacian smoothing.
[0116]
[0117] Wherein, N(t) i ) is the second nearest neighbor of the i-th pixel for estimating transmittance, |N(t) i )| is the number of nearest neighbors, and m is the number of pixels.
[0118] Model training unit 6 is used to train models based on the dehazed image. The atmospheric light value The remaining coefficient and the transmittance map The foggy image is resynthesized, and the untangling of layers is guided by the fog reconstruction process. The network is then trained to obtain the defoggy image.
[0119] Specifically, the improved atmospheric scattering model (IASM) is expressed as follows:
[0120] I(x)=A·α(x)·R(x)·t(x)+A·(1-t(x))
[0121] Where I(x) represents the obtained foggy image, A refers to the atmospheric light value, R(x) is the scene reflectance, and t(x) = e -βd(x) This is a transmittance map, which represents the amount of light from the scene that passes through the fog and reaches the camera. β is the wavelength of the light, d(x) is the image depth, and α(x) is the residual coefficient, which expresses the amount of light remaining after the material absorbs the light. Light is absorbed by the texture, and the light absorption rate changes with the texture density. α(x)∈(0,1], so A·α(x)·R(x)≤A·R(x), which is the reason why it becomes darker after defogging.
[0122] According to the defogging diagram The atmospheric light value The remaining coefficient and the transmittance map Re-synthesized fog map The difference between the synthesized hazy image and the input image is calculated, then...
[0123]
[0124] ‖·‖2 is the 2-norm. The entire network, including sub-networks, is constrained. Layer unwrapping is guided by merging the fog reconstruction process. A supervised mechanism is used to train given J-net, A-net, t-net, and α-net, resulting in the total unsupervised defogging loss of the trained network.
[0125]
[0126] For an input image, iterate the network 100 times to minimize the total loss and obtain the dehazed image.
[0127] Therefore, this method uses an improved atmospheric light scattering model, introducing residual coefficients to explain the darkening of the scene after image dehazing. Based on the improved atmospheric light scattering model, a new decoupled dehazing network is proposed, with different losses designed for different sub-networks to ensure that the sub-network outputs conform to physical laws. Training is performed on real data to avoid domain drift problems caused by synthetic data, resulting in clear dehazed images. In foggy public places such as stations and squares, the method effectively restores clear images, thereby effectively avoiding danger and reducing the occurrence of traffic accidents. Even in the event of a traffic accident, the dehazing results from the recorder can be used to clarify responsibility and reduce losses. This invention has high intelligence and employs a dimensionality reduction method, resulting in high computational efficiency and convenience.
[0128] The above provides a detailed description of the unsupervised image dehazing enhancement method and system based on decoupling provided by the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. An unsupervised image dehazing enhancement method based on decoupling, characterized in that, The method includes: The foggy image is input into the encoder to obtain the foggy feature map z; The foggy feature map is input into the J-net dehazing network layer to obtain the dehazing map. ; The fog feature map is input into the atmospheric light network layer A-net, which outputs the decoupled atmospheric light value. ; The foggy feature map is input into the residual coefficients network layer. In -net, calculate the residual coefficients. ; The hazy feature map is input into the transmittance map network layer t-net to obtain the transmittance map. ; According to the defogging diagram The atmospheric light value The remaining coefficient and the transmittance map The fogged image is resynthesized, and the unwrapping of layers is guided by the fog reconstruction process to train the network and obtain the dehazed image; specifically including: According to the defogging diagram The atmospheric light value The remaining coefficient and the transmittance map Re-synthesized fog map The synthesized hazy image is compared with the input image. Calculating the difference, we have: ； It is a 2-norm. The constraints include the entire network, including sub-networks, and the unwrapping of layers is guided by incorporating the haze reconstruction process. A supervised mechanism is used to train given J-net, A-net, t-net, and... -net, which gives the total loss for unsupervised dehazing of the trained network: ； in, This indicates atmospheric optical network layer constraints. This indicates the dark channel loss constraint. Represents the constraints of the residual coefficient network layer. This represents the Laplacian smoothing constraint; For an input image, iterate the network 100 times to minimize the total loss and obtain the dehazed image.
2. The unsupervised image dehazing enhancement method based on decoupling as described in claim 1, characterized in that, The foggy feature map is input into the J-net dehazing network layer to obtain the dehazing map. Specifically: Using dark channel loss to constrain J-Net enables it to achieve dehazing capabilities. The dark channel prior reveals that in natural images, one of the three channels always has a pixel size of zero or close to zero. Its expression is: ； in, It's a dark channel operation. This is the output value of J-net, i.e., the dehazed image. It is a color channel pixel intensity, It is the input image A fixed-size local block in the image, dehazed using dark channel loss constraints, is defined as follows: ； This indicates the amount of data. By reducing the dark channel value in the recovered image, an ideal haze-free image can be obtained.
3. The unsupervised image dehazing and enhancement method based on decoupling as described in claim 1, characterized in that, The fog feature map is input into the atmospheric light network layer A-net, and the decoupled atmospheric light value is output. Specifically: The pixel with the highest intensity in input image I is selected as the true value of atmospheric light. The losses are as follows: ； It uses the F-norm. For some white objects, pixel intensities are greater or less than atmospheric light values. Atmospheric light values are independent of the ground truth image values. A reparameterization technique is introduced, performing reparameterization on the variational lower bound to produce a lower bound estimate that can be directly optimized using the standard stochastic gradient method. Latent variables... Follows a standard normal distribution Using KL divergence, ； in, yes The i-th dimension, during execution, is obtained by progressively inputting the image into A-net. By using the normal distribution Gradually sample, by minimizing To enhance The constraints of the atmospheric light value network layer are: 。 4. The unsupervised image dehazing enhancement method based on decoupling as described in claim 1, characterized in that, The fog feature map is input into the residual coefficient network layer. In -net, calculate the residual coefficients. Specifically: Residual coefficient , Representing a depth map The maximum value, i.e., the farthest scene point in the image, is set as follows. The estimated residual coefficient is , The loss function for -net is as follows: ； in for Norm.
5. The unsupervised image dehazing enhancement method based on decoupling as described in claim 1, characterized in that, The process involves inputting the fog feature map into the transmittance map network layer t-net to obtain the transmittance map. Specifically: To enhance the stability of the model, the transmittance map needs to be constrained. A transmittance map estimated using Laplacian smoothing is employed. ； in, It is the second nearest neighbor of the i-th pixel when estimating transmittance. It is the number of neighbors. It refers to the number of pixels.
6. An unsupervised image dehazing and enhancement system based on decoupling, characterized in that, The system includes: The feature map acquisition unit is used to input the foggy image into the encoder to obtain a foggy feature map; The dehazing unit is used to input the hazy feature map into the dehazing network layer J-net to obtain the dehazing map. ; The atmospheric light value unit is used to input the fog feature map into the atmospheric light network layer A-net and output the decoupled atmospheric light value. ; The residual coefficient unit is used to input the foggy feature map into the residual coefficient network layer. In -net, calculate the residual coefficients. ; The transmittance mapping unit inputs the hazy feature map into the transmittance mapping network layer t-net to obtain the transmittance map. ; The model training unit, based on the dehazed image The atmospheric light value The remaining coefficient and the transmittance map The fogged image is resynthesized, and the unwrapping of layers is guided by the fog reconstruction process to train the network and obtain the dehazed image; specifically including: According to the defogging diagram The atmospheric light value The remaining coefficient and the transmittance map Re-synthesized fog map The synthesized hazy image is compared with the input image. Calculating the difference, we have: ； It is a 2-norm. The constraints include the entire network, including sub-networks, and the unwrapping of layers is guided by incorporating the haze reconstruction process. A supervised mechanism is used to train given J-net, A-net, t-net, and... -net, which gives the total loss for unsupervised dehazing of the trained network: ； in, This indicates atmospheric optical network layer constraints. This indicates the dark channel loss constraint. Represents the constraints of the residual coefficient network layer. This represents the Laplacian smoothing constraint; For an input image, iterate the network 100 times to minimize the total loss and obtain the dehazed image.
7. The unsupervised image dehazing and enhancement system based on decoupling as described in claim 6, characterized in that, The dehazing unit requires the use of dark channel loss to constrain J-net, enabling J-net to achieve dehazing capabilities. The dark channel prior reveals that in natural images, one of the three channels always has a pixel size of zero or close to zero, and its expression is: ； in, It's a dark channel operation. This is the output value of J-net, i.e., the dehazed image. It is a color channel pixel intensity, It is the input image A fixed-size local block in the image, dehazed using dark channel loss constraints, is defined as follows: ； This indicates the amount of data. By reducing the dark channel value in the recovered image, an ideal haze-free image can be obtained.
8. The unsupervised image dehazing and enhancement system based on decoupling as described in claim 6, characterized in that, The atmospheric light value unit needs to select the pixel with the highest intensity in the input image I as the true atmospheric light value. The losses are as follows: ； It uses the F-norm. For some white objects, pixel intensities are greater or less than atmospheric light values. Atmospheric light values are independent of the ground truth image values. A reparameterization technique is introduced, performing reparameterization on the variational lower bound to produce a lower bound estimate that can be directly optimized using the standard stochastic gradient method. Latent variables... Follows a standard normal distribution Using KL divergence, ； in, yes The i-th dimension, during execution, is obtained by progressively inputting the image into A-net. By using the normal distribution Gradually sample, by minimizing To enhance The constraints of the atmospheric light value network layer are: 。 9. The unsupervised image dehazing and enhancement system based on decoupling as described in claim 6, characterized in that, The remaining coefficient unit needs to input the foggy feature map into the remaining coefficient network layer. In -net, residual coefficients , Representing a depth map The maximum value, i.e., the farthest scene point in the image, is set as follows. The residual coefficients estimated by -net are , The loss function for -net is as follows: ； in for Norm.
10. The unsupervised image dehazing and enhancement system based on decoupling as described in claim 6, characterized in that, The transmittance map unit requires inputting the hazy feature map into the transmittance map network layer t-net. To enhance the stability of the model, the transmittance map needs to be constrained, using a transmittance map estimated by Laplacian smoothing. ； in, It is the second nearest neighbor of the i-th pixel when estimating transmittance. It is the number of neighbors. It refers to the number of pixels.