An image dehazing method combining course learning with teacher-student learning

Abstract

The present invention discloses an image defogging method that combines course learning with teacher-student learning, and belongs to the fields of computer vision and harsh natural environment processing technology. First, a teacher model is used to restore the image of a thin haze image that is easier to learn. After that, the teacher model is fixed, and then the student dense fog defogging model is trained. During the training process, not only the loss function of the student network itself is used, but also the supervised distillation loss of the teacher network on the student network is used. Finally, the trained student model can independently and efficiently perform image defogging. By organizing the learning process from thin fog to dense fog and from easy to difficult, better feature learning results can be obtained. This novel image defogging strategy can effectively overcome the problem that traditional defogging algorithms have poor defogging effects on real scenes with uneven haze distribution, and can obtain more thoroughly defogged images, better restore the detailed information of the image, and effectively improve the performance of the defogging model.

Description

Technical Field

[0001] The present invention belongs to the technical fields of computer vision and harsh natural environment processing, and specifically relates to an image defogging method that combines course learning with teacher-student learning. Background Art

[0002] Fog is caused by an aerosol system consisting of tiny water droplets or ice crystals suspended in the atmosphere near the ground, which interferes with light reflection. Haze is caused by solid particles such as dust in the atmosphere, which affect light reflection. With the development of industry, haze is becoming increasingly common in daily life, manifesting itself in poor air quality and reduced visibility. In terms of traffic safety, haze reduces vehicle visibility, making traffic accidents more likely. In terms of military security, haze can seriously affect the stability of military visual systems, reducing the accuracy of guidance and other weapon systems. In aerospace, haze can interfere with pilots' vision, impacting air travel safety. Therefore, the stable and effective removal of haze from images, restoring clear, fog-free images, is of great practical significance.

[0003] Image dehazing has seen significant research progress. Dehazing methods can be broadly categorized into two areas: learning-based and non-learning-based. Non-learning-based image dehazing algorithms generally include traditional enhancement-based and prior model-based algorithms. Traditional enhancement-based image dehazing algorithms, such as histogram equalization and Retinex dehazing, primarily enhance image contrast to achieve dehazing. However, these methods only improve the visual quality to a certain extent and fail to account for the underlying nature of haze formation. Prior model-based image dehazing algorithms, such as dark channel dehazing, primarily rely on atmospheric light scattering models. They estimate image transmittance and atmospheric light values ​​to achieve the transition between haze-free and hazy images. However, dehazing performance deteriorates when the prior conditions are not met. Deep learning-based image dehazing algorithms employ a designed network architecture and utilize fog image datasets for model training. This approach has become the mainstream dehazing method.

[0004] Curriculum learning advocates that the model should not directly learn more complex samples, but instead start with simple samples and gradually advance to complex samples. The model will spend more energy on learning simple samples in the early stages of training, and will only move on to learning complex samples after achieving good learning results on simple samples. This learning approach guides the model training towards a local optimum, achieving better generalization.

[0005] Teacher-student learning typically involves a teacher model and a student model. The core idea is to design a knowledge distillation method to transfer the knowledge of the superior teacher model to the student model, helping the student model achieve better performance than without the teacher model's assistance. This is often used for model compression. Specifically, distillation can be performed on the final outputs of the teacher's light fog dehazing network and the student's dense fog dehazing network. It can also be performed directly or indirectly on the feature maps of the intermediate layers between the teacher's and the student's dense fog dehazing networks. The mapping relationship between network layers can also be defined as a "flow," and a loss function can be designed for this mapping relationship for distillation. Summary of the Invention

[0006] Aiming at the problems of poor defogging effect and easy occurrence of fog residue in non-uniform haze scenes in current real scenes, the present invention provides an image defogging method that combines course learning with teacher-student learning.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] An image dehazing method combining course learning with teacher-student learning includes the following steps:

[0009] Step 1: Obtain publicly available synthetic and real-scene haze image datasets, and divide the collected haze image datasets into training sets, validation sets, and test sets;

[0010] Step 2: Divide the sample difficulty of the synthetic and real scene haze image datasets into easy and easy samples respectively. According to the curriculum learning strategy, the light haze images are divided into easy samples, and the heavy fog images are divided into difficult samples.

[0011] Step 3: Establish a teacher-student learning defogging framework, design the teacher's thin fog defogging network structure and the student's dense fog defogging network structure, and design the method of transferring the teacher's thin fog defogging network knowledge to the student's dense fog defogging network, as well as the model's optimizer and loss function;

[0012] Step 4: Input the haze image into the teacher haze dehazing network, use the clear haze-free image to supervise the dehazing result image output by the network, and train it in an end-to-end manner to obtain the trained teacher haze dehazing network model;

[0013] Step 5: Fix the trained teacher fog dehazing network model parameters, input the dense fog image into the student dense fog dehazing network, and use the loss function of the student dense fog dehazing network itself and the loss function supervised by the intermediate layer feature map of the teacher fog dehazing network model to train the student dense fog dehazing network model at the same time, and finally obtain the trained student dense fog dehazing network model;

[0014] Step 6: Input the haze image of the real scene into the finally trained student fog dehazing network model to obtain the dehazed result image.

[0015] Furthermore, step 1 obtains currently available synthetic and real-scene haze image datasets, and divides the collected haze image datasets into a training set, a validation set, and a test set. Specifically, the following steps are included:

[0016] Step 1.1: Obtain the currently available synthetic and real-scene haze image datasets;

[0017] In step 1.2, the fog image dataset that has been divided into training set, validation set, and test set can be used directly without any processing. For the undivided dataset, the training set, validation set, and test set are allocated in a ratio of 8:1:1.

[0018] Furthermore, step 2 divides the sample difficulty of the synthetic and real scene haze image datasets respectively. According to the curriculum learning strategy, the light haze images are divided into easy samples, and the dense fog images are divided into difficult samples. The specific steps include the following:

[0019] Step 2.1: For the synthetic haze image dataset, calculate the pixel mean of haze images of different concentrations corresponding to the same clear haze-free image. The image with the lowest pixel mean is regarded as the light haze image, and the remaining images are regarded as the heavy haze images.

[0020] In step 2.2, for the real-scene haze image dataset, set the appropriate illumination intensity A and atmospheric light scattering coefficient β, and then synthesize the light haze image according to the atmospheric light scattering model, while the real-scene haze image is used as the dense fog image.

[0021] Furthermore, the atmospheric light scattering model is expressed as:

[0022] I(x)=J(x)t(x)+A(1-t(x))

[0023] t(x)=e -βd(x)

[0024] Where I(x) represents the image to be defogged, J(x) represents the clear and fog-free image, t(x) represents the atmospheric light propagation rate map information, that is, the transmittance; A represents the global atmospheric light value, e is a natural constant, d(x) represents the image depth of field information, and β represents the atmospheric light scattering coefficient;

[0025] By controlling the global atmospheric light value A and the atmospheric light propagation rate map information t(x), and then based on the clear fog-free map J(x), the haze image in the real scene can be effectively synthesized.

[0026] Furthermore, in step 3, the teacher thin fog dehazing network and the student dense fog dehazing network adopt the same network structure, so that more consistent features can be obtained in the middle layer of the teacher thin fog dehazing network and the student dense fog dehazing network.

[0027] Furthermore, in step 3, the knowledge transfer from the teacher's thin fog dehazing network to the student's dense fog dehazing network utilizes a "knowledge review" strategy. Specifically, the features of the student's dense fog dehazing network are first fused layer by layer from deep to shallow layers. The resulting features of each layer of the student's dense fog dehazing network incorporate the features of each subsequent deeper layer. The features of each layer of the teacher's fog dehazing network then supervise the features of each subsequent layer of the student's dense fog dehazing network. By organizing the learning process from thin fog to dense fog and from easy to difficult, better feature learning results can be achieved.

[0028] Furthermore, the loss function in step 4 is expressed as:

[0029] L t =L c +λL p

[0030] Among them, L c Represents the image reconstruction loss, L p represents the perceptual loss of the image, and λ is a parameter that balances the ratio of the loss function; specifically, the reconstruction loss L c It is the minimum absolute value deviation between the result of dehazing the image and the clear image without haze:

[0031] L c =||D t (I)-J||1

[0032] Where I represents the haze image input to the teacher haze dehazing network, D t represents the teacher haze dehazing network, J represents the haze-free clear image;

[0033] The perceptual loss L p It is calculated based on the pre-trained VGG19 model and can be used to further reduce the domain difference between the dehazed image and the haze-free clear image, expressed as:

[0034] L p =∑W m *||Φ m (D t (I))-Φ m (J)||1,m∈{2,7,12,21,30}

[0035] Among them, Φ m () represents the feature map output by the mth layer in the pre-trained VGG19 model, Wm represents the weight of the mth layer.

[0036] Furthermore, in step 5, the parameters of the trained teacher fog dehazing network model are fixed, the dense fog image is input into the student dense fog dehazing network, and the student dense fog dehazing network model is trained simultaneously using the loss function of the student dense fog dehazing network itself and the loss function supervised by the intermediate layer feature map of the teacher fog dehazing network model, and finally a trained student dense fog dehazing network model is obtained. Specifically, the following steps are included:

[0037] Step 5.1: Get the trained teacher haze removal network and fix its parameters without gradient backpropagation or parameter update.

[0038] Step 5.2: Input the light fog image and heavy fog image corresponding to the same fog-free clear image into the teacher light fog dehazing network and the student heavy fog dehazing network respectively. The parameters of the student heavy fog dehazing network are adjusted together with the loss function of the student heavy fog dehazing network itself and the loss function of the intermediate layer feature map of the teacher fog dehazing network to supervise the student heavy fog dehazing network, so that the student heavy fog dehazing network achieves the optimal performance.

[0039] The loss function of the student fog dehazing network is expressed as follows:

[0040] L s =L c +λL p +γL distill

[0041] Among them, L c Represents the image reconstruction loss, L p represents the perceptual loss of the image, L distill represents the supervised distillation loss of the teacher's thin fog dehazing network on the student's dense fog dehazing network, and λ and γ are parameters that balance the ratio of the loss function;

[0042] The supervised distillation loss L distill The expression is as follows:

[0043]

[0044] Among them, N represents the number of intermediate output layers of the teacher's thin fog dehazing network and the student's dense fog dehazing network, Z represents the number of multi-scale pyramid pooling in each layer of the network, and P t i,j (F t i ) represents the output F of the i-th layer of the teacher mist dehazing network t i The features obtained in the j-th multi-scale pyramid pooling layer, It represents the fusion feature of the i-th layer after the student fog dehazing network is integrated layer by layer from deep to shallow Features obtained in the j-th multi-scale pyramid pooling layer.

[0045] Compared with the prior art, the present invention has the following advantages:

[0046] To address the problems of haze images, such as severe color cast, low contrast, and background blur, which can lead to residual fog in models used for dense fog dehazing and poor dehazing performance in real-world scenes, a dehazing method based on curriculum learning and a teacher-student learning framework was designed. A teacher model is first used to restore images of light haze, which is easier to learn. The teacher model is then fixed, and the student model for dense fog dehazing is trained. The training process not only uses the student network's own loss function but also uses the supervised distillation loss of the teacher network. Finally, the trained student model can independently and efficiently perform image dehazing. By organizing the learning process from light haze to dense haze and from easy to difficult, better feature learning results are achieved. This novel image dehazing strategy effectively overcomes the problem of poor dehazing performance of traditional dehazing algorithms in real-world scenes with uneven haze distribution, resulting in more thoroughly dehazed images, better restoration of image details, and significantly improving the performance of the dehazing model. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a diagram of the overall model structure of an image defogging method that combines course learning with teacher-student learning in the present invention.

[0048] Figure 2 This is a structural diagram of the enhancement module and the attention module in the teacher's thin fog defogging network and the student's thick fog defogging network of an image defogging method that combines course learning with teacher-student learning in the present invention. DETAILED DESCRIPTION

[0049] Example 1

[0050] An image dehazing method that combines course learning with teacher-student learning, such as Figure 1 As shown, the following steps are included:

[0051] Step 1: Obtain publicly available synthetic and real-scene haze image datasets, and divide the collected haze image datasets into training sets, validation sets, and test sets;

[0052] Step 1.1: Obtain the currently available synthetic and real-scene haze image datasets;

[0053] In step 1.2, the fog image dataset that has been divided into training set, validation set, and test set can be used directly without any processing. For the undivided dataset, the training set, validation set, and test set are allocated in a ratio of 8:1:1.

[0054] Step 2: A simple and efficient method is used to classify the sample difficulty of the synthetic and real scene haze image datasets. According to the curriculum learning strategy, the light haze image is divided into easy samples, and the dense fog image is divided into difficult samples.

[0055] In step 2.1, for the synthetic haze image dataset, it can be intuitively seen that the thicker the fog, the whiter the image will appear, i.e., the average pixel value of the image will increase. Conversely, the lighter the fog, the smaller the average pixel value of the image will be. Therefore, the present invention calculates the pixel mean of haze images of different concentrations corresponding to the same clear haze-free image, and uses the image with the lowest pixel mean as the light haze image, and the remaining images as the heavy haze images.

[0056] In step 2.2, obtaining a pair of clear, fog-free images and real, foggy images in real scenes is generally expensive, and obtaining a one-to-one correspondence is difficult. Currently, existing public datasets of real-world haze images typically only have one clear, fog-free image and one corresponding dense fog image, without multiple images of varying haze concentrations. Consequently, dataset diversity is limited. To learn dehazing from easy to difficult, the only way is to generate a light haze image using the atmospheric light scattering physical model. Specifically, we set appropriate illumination intensity A and atmospheric light scattering coefficient β, then synthesize a light haze image using the atmospheric light scattering model. The real-world haze images in the dataset serve as dense fog images.

[0057] The atmospheric light scattering model can be expressed as:

[0058] I(x)=J(x)t(x)+A(1-t(x))

[0059] t(x)=e -βd(x)

[0060] Where I(x) represents the image to be dehazed, J(x) represents the clear, haze-free image, A represents the global atmospheric light value, t(x) represents the atmospheric light propagation rate map (i.e., transmittance), d(x) represents the image depth, β represents the atmospheric light scattering coefficient, and e is a natural constant. By controlling the global atmospheric light value A and the propagation rate information t(x), a haze image in a real scene can be effectively synthesized based on the clear, haze-free image.

[0061] Step 3: Establish a teacher-student learning framework for defogging. Design the teacher's thin fog defogging network structure and the student's dense fog defogging network structure. Also, design a method for transferring knowledge from the teacher's thin fog defogging network to the student's dense fog defogging network, as well as the model's optimizer and loss function.

[0062] The teacher thin fog dehazing network and the student dense fog dehazing network adopt the same encoding-decoding network structure so that more consistent features can be obtained in the middle layer of the teacher thin fog dehazing network and the student dense fog dehazing network. First, the input haze image passes through an encoder consisting of three convolutional layers, the first of which is a 3*3 convolutional layer for preliminary feature extraction and increasing the number of channels. The other two are 4*4 convolutional layers for four-fold downsampling, so that the feature dimension changes from (256, 256, 3) to (64, 64, 64); the encoded features are then sent to the enhancement network, which consists of six identical enhancement modules; the obtained feature map is then input into the decoder, which consists of three convolutional layers. The enhanced features are first bilinearly upsampled by two times and then spliced ​​with the corresponding features in the encoder on the channel. After a 3*3 convolutional layer and ReLU activation function, the output is bilinearly upsampled by two times and then spliced ​​with the corresponding features in the encoder on the channel. The obtained features are also passed through a 3*3 convolutional layer and ReLU activation function. Finally, the features are passed through a layer of convolution for channel recovery to obtain the final output result.

[0063] like Figure 2 As shown in the figure, the enhancement module consists of an RDB, a convolution layer, and an attention module. The attention module decouples pixel attention from channel attention. First, the image is average pooled in both height and width directions to obtain features with dimensions (H, 1, C) and (1, W, C). The (1, W, C)-dimensional features are then transformed and concatenated with the (H, 1, C)-dimensional features in the first dimension to obtain (H+W, 1, C)-dimensional features. Two convolution layers are then applied to obtain (H+W, 1, 1)-dimensional features. The first dimension is then split in height and width directions, followed by a convolution layer and a Sigmoid activation function to obtain attention values ​​in the height and width directions. Finally, these are weighted onto the original input features, completing the pixel attention operation. The features obtained from the pixel attention operation are first average pooled, then passed through two fully connected layers and a Sigmoid activation function to obtain attention values ​​in the (1, 1, C)-dimensional dimensions. Finally, these are weighted onto the features obtained from the pixel attention operation to complete the channel attention operation.

[0064] The knowledge transfer from the teacher's thin fog dehazing network to the student's dense fog dehazing network utilizes a "knowledge review" strategy. This involves first fusing the features of the student's dense fog dehazing network layer by layer, from deep to shallow layers. This allows the features of each layer of the student's dense fog dehazing network to incorporate the features of each subsequent deeper layer. The features of each layer of the teacher's dense fog dehazing network then supervise the features of each subsequent layer of the student's dense fog dehazing network. This deep-to-shallow fusion strategy allows the shallow layers of the teacher's thin fog dehazing network to guide the learning of the deeper layers of the student's dense fog dehazing network during the initial stages of training. Conversely, the shallow features of the student's dense fog dehazing network can only achieve the same dehazing effect as the shallower layers of the teacher's dense fog dehazing network if they incorporate information from deeper layers.

[0065] Because the features of the student dense fog dehazing network after fusion from deep to shallow contain information from different stages, the present invention adopts a multi-scale pyramid pooling loss strategy in the knowledge transfer process. That is, multi-scale pyramid pooling is performed on the intermediate layer features of the teacher thin fog dehazing network and the features after fusion of the student dense fog dehazing network. Finally, L2 loss is applied to the pooled multi-scale feature map. This can fully utilize the rich knowledge of the intermediate layer features of the teacher thin fog dehazing network and the student dense fog dehazing network. By organizing the learning process from light fog to dense fog and from easy to difficult, better feature learning results can be obtained.

[0066] Step 4: Input the haze image into the teacher haze dehazing network, use the clear haze-free image to supervise the dehazing result image output by the network, and train it in an end-to-end manner to obtain the trained teacher haze dehazing network model;

[0067] The mathematical expression of the loss function of the teacher mist dehazing network end-to-end training is as follows:

[0068] L t =L c +λL p

[0069] Among them, L c Represents the image reconstruction loss, L p represents the perceptual loss of the image, and λ is a parameter that balances the ratio of the loss function; specifically, the reconstruction loss L c It is the minimum absolute value deviation between the result of dehazing the image and the clear image without haze:

[0070] L c =||D t (I)-J||1

[0071] Where I represents the haze image input to the teacher haze dehazing network, D t represents the teacher haze dehazing network, J represents the haze-free clear image;

[0072] The perceptual loss L p It can be used to further reduce the domain difference between the defogging image and the fog-free clear image. It is calculated based on the pre-trained VGG19 model. The specific expression is:

[0073] L p =∑W m *||Φ m (D t (I))-Φ m (J)||1,m∈{2,7,12,21,30}

[0074] Among them, Φ m () represents the feature map output by the mth layer in the pre-trained VGG19 model, W m represents the weight of the mth layer.

[0075] Step 5: Fix the trained teacher model parameters, input the dense fog image into the student dense fog dehazing network, and use the loss function of the student dense fog dehazing network itself and the loss function supervised by the intermediate layer feature map of the teacher model to train the student model simultaneously, and finally obtain the trained dense fog dehazing model;

[0076] Step 5.1: Get the trained teacher haze removal network and fix its parameters without gradient backpropagation or parameter update.

[0077] Step 5.2: Input the light fog image and heavy fog image corresponding to the same fog-free clear image into the teacher light fog dehazing network and the student heavy fog dehazing network respectively. The parameters of the student heavy fog dehazing network are adjusted together with the loss function of the student heavy fog dehazing network itself and the loss function of the intermediate layer feature map of the teacher fog dehazing network to supervise the student heavy fog dehazing network, so that the student heavy fog dehazing network achieves the optimal performance.

[0078] The loss function of the student fog dehazing network is expressed as follows:

[0079] L s =L c +λL p +γL distill

[0080] Among them, L c Represents the image reconstruction loss, L p represents the perceptual loss of the image, L distill represents the supervised distillation loss of the teacher's thin fog dehazing network on the student's dense fog dehazing network, and λ and γ are parameters that balance the ratio of the loss function;

[0081] Monitor distillation loss L distill The expression is as follows:

[0082]

[0083] Among them, N represents the number of intermediate output layers of the teacher's thin fog dehazing network and the student's dense fog dehazing network, Z represents the number of multi-scale pyramid pooling in each layer of the network, and P t i,j (F t i ) represents the output F of the i-th layer of the teacher mist dehazing network t i The features obtained in the j-th multi-scale pyramid pooling layer, It represents the fusion feature of the i-th layer after the student fog dehazing network is integrated layer by layer from deep to shallow Features obtained in the j-th multi-scale pyramid pooling layer.

[0084] In step 6, a haze image of a real scene is input into the finally trained dense fog dehazing model to obtain the dehazed result image.

[0085] Any matters not described in detail in this specification are prior art known to those skilled in the art. Although the above description of the present invention is based on specific embodiments to facilitate understanding of the present invention by those skilled in the art, it should be understood that the present invention is not limited to the scope of the specific embodiments. As long as various modifications are within the spirit and scope of the present invention as defined and determined by the appended claims, such modifications will be obvious to those skilled in the art, and all inventions and creations utilizing the concepts of the present invention are protected.

Claims

1. An image dehazing method that combines course learning with teacher-student learning, characterized by: The following steps are involved: Step 1: Obtain publicly available synthetic and real-scene haze image datasets, and divide the collected haze image datasets into training sets, validation sets, and test sets; Step 2: Divide the sample difficulty of the synthetic and real scene haze image datasets into easy and easy samples respectively. According to the curriculum learning strategy, the light haze images are divided into easy samples, and the heavy fog images are divided into difficult samples. Step 3: Establish a teacher-student learning defogging framework, design the teacher's thin fog defogging network structure and the student's dense fog defogging network structure, and design the method of transferring the teacher's thin fog defogging network knowledge to the student's dense fog defogging network, as well as the model's optimizer and loss function; Step 4: Input the haze image into the teacher haze dehazing network, use the clear haze-free image to supervise the dehazing result image output by the network, and train it in an end-to-end manner to obtain the trained teacher haze dehazing network model; Step 5: Fix the trained teacher fog dehazing network model parameters, input the dense fog image into the student dense fog dehazing network, and use the loss function of the student dense fog dehazing network itself and the loss function supervised by the intermediate layer feature map of the teacher fog dehazing network model to train the student dense fog dehazing network model at the same time, and finally obtain the trained student dense fog dehazing network model; Step 6: Input the haze image of the real scene into the finally trained student fog dehazing network model to obtain the dehazed result image.

2. The image dehazing method combining course learning with teacher-student learning according to claim 1 is characterized by: Step 1: Obtain the currently available synthetic and real-scene haze image datasets, and divide the collected haze image datasets into training sets, validation sets, and test sets. The specific steps include the following: Step 1.1: Obtain the currently available synthetic and real-scene haze image datasets; In step 1.2, the fog image dataset that has been divided into training set, validation set, and test set can be used directly without any processing. For the undivided dataset, the training set, validation set, and test set are allocated in a ratio of 8:1:

1.

3. The image dehazing method combining course learning with teacher-student learning according to claim 1 is characterized by: Step 2: Classify the samples of synthetic and real-scene haze image datasets into easy and difficult samples. According to the course learning strategy, the haze images are divided into easy samples and the dense fog images are divided into difficult samples. The specific steps include the following: Step 2.1: For the synthetic haze image dataset, calculate the pixel mean of haze images of different concentrations corresponding to the same clear haze-free image. The image with the lowest pixel mean is regarded as the light haze image, and the remaining images are regarded as the heavy haze images. Step 2.2: For the real scene haze image dataset, set the appropriate light intensity and atmospheric light scattering coefficient ,Then a light haze image is synthesized based on the atmospheric light scattering model, while the real scene haze image is used as the dense fog image.

4. The image dehazing method combining course learning with teacher-student learning according to claim 3 is characterized by: The atmospheric light scattering model is expressed as: ； ； in, represents the image to be dehazed, Indicates a clear and fog-free image. Represents the atmospheric light propagation rate map information, that is, transmittance; represents the global atmospheric light value, is a natural constant, Represents the image depth information, represents the atmospheric light scattering coefficient; By controlling the global atmospheric light value and atmospheric light propagation rate map information , then according to the clear fog-free map To effectively synthesize haze images in real scenes.

5. The image dehazing method combining course learning with teacher-student learning according to claim 1 is characterized by: In step 3, the teacher's thin fog dehazing network and the student's dense fog dehazing network adopt the same network structure.

6. The image dehazing method combining course learning with teacher-student learning according to claim 1, characterized in that: In step 3, the method of transferring the knowledge of the teacher's thin fog dehazing network to the student's dense fog dehazing network is to adopt a "knowledge review" strategy. Specifically, the features of the student's dense fog dehazing network are first fused layer by layer from deep to shallow. The features of each layer of the obtained student's dense fog dehazing network are fused with the features of each deeper layer. Then, the features of each layer of the teacher's mist dehazing network supervise the features of each layer of the corresponding student's dense fog dehazing network after fusion. By organizing the learning process from light fog to thick fog and from easy to difficult, better feature learning results can be obtained.

7. The image dehazing method combining course learning with teacher-student learning according to claim 1 is characterized by: The loss function of the teacher haze dehazing network trained in an end-to-end manner in step 4 is expressed as: ； in, represents the reconstruction loss of the image, represents the perceptual loss of the image, is a parameter that balances the ratio of the loss function; specifically, the reconstruction loss It is the minimum absolute value deviation between the result of dehazing the image and the clear image without haze: ； in, represents the haze map input to the teacher haze dehazing network, represents the teacher mist dehazing network, Indicates a fog-free and clear image; The perceptual loss It is calculated based on the pre-trained VGG19 model and is expressed as: ； in, Indicates the first The feature map output by the layer, Indicates the The weight of the layer.

8. The image dehazing method combining course learning with teacher-student learning according to claim 1 is characterized by: Step 5: Fix the trained teacher fog dehazing network model parameters, input the dense fog image into the student dense fog dehazing network, and use the loss function of the student dense fog dehazing network itself and the loss function supervised by the intermediate layer feature map of the teacher fog dehazing network model to train the student dense fog dehazing network model at the same time, and finally obtain the trained student dense fog dehazing network model. The specific steps include: Step 5.1: Get the trained teacher haze removal network and fix its parameters without gradient backpropagation or parameter update. Step 5.2: Input the light fog image and heavy fog image corresponding to the same fog-free clear image into the teacher light fog dehazing network and the student heavy fog dehazing network respectively. The parameters of the student heavy fog dehazing network are adjusted together with the loss function of the student heavy fog dehazing network itself and the loss function of the intermediate layer feature map of the teacher fog dehazing network to supervise the student heavy fog dehazing network, so that the student heavy fog dehazing network achieves the optimal performance. The loss function of the student fog dehazing network is expressed as follows: ； in, represents the reconstruction loss of the image, represents the perceptual loss of the image, represents the supervised distillation loss of the teacher thin fog dehazing network on the student dense fog dehazing network, and is the parameter that balances the ratio of the loss function; The supervised distillation loss The expression is as follows: ； in, represents the number of intermediate output layers of the teacher thin fog dehazing network and the student dense fog dehazing network, Indicates the number of multi-scale pyramid pooling in each layer in the middle of the network, Represents the output of the i-th layer of the teacher haze dehazing network The features obtained in the j-th multi-scale pyramid pooling layer, It represents the fusion feature of the i-th layer after the student fog dehazing network is integrated layer by layer from deep to shallow Features obtained in the j-th multi-scale pyramid pooling layer.

Images