Brightness-guided double-domain perception image defogging method and system

Abstract

The invention discloses a brightness-guided double-domain perception image defogging method, which comprises the following steps of: constructing a degradation model of a real haze image, and simulating a natural scene by using a haze synthesis pipeline so as to synthesize a haze data set required by training; the structure of a defogging network is designed, and the defogging network comprises an encoder, a dual-domain sensing module and a decoder; constructing a training strategy and a loss function required by the network, and forming a brightness-guided double-domain perception image defogging network; testing the defogging network by using a no-reference image quality evaluation index; and when a user inputs a to-be-processed haze image to the defogging network, an image defogging result is obtained. In addition, the invention also discloses a brightness-guided dual-domain perception image defogging system. The real haze scene can be effectively restored, compared with an existing image defogging method, the method has better defogging robustness, image structure, content details and brightness fidelity effects, and the restored image is more natural.

Description

Technical Field

[0001] The present invention relates to the fields of computer vision and image restoration, and in particular to a brightness-guided dual-domain perception image defogging method and system. Background Art

[0002] In severe weather conditions such as haze, smoke, and dust, images captured by outdoor vision systems often have obvious visibility and contrast degradation. For visual tasks that rely on imaging quality, such as traffic monitoring and object detection, image quality degradation caused by haze pollution will lead to significant performance degradation, and even cause serious safety issues for tasks such as autonomous driving assistance and near-ocean monitoring. In order to eliminate haze interference, although many representative defogging methods have been proposed, these methods are still relatively weak in the actual defogging problem.

[0003] The current defogging methods can be divided into defogging methods based on heuristic prior design and defogging methods based on deep learning in terms of implementation. For example, the prior ROP based on the two-dimensional matrix rank 1 of the transmittance map can achieve a computational complexity of O(N) in a uniform haze scene, where N is the image size. However, when the actual prior information obtained deviates from the perfect data assumption, the defogging effect is poor. In fact, affected by environmental factors such as atmospheric light intensity and wind speed changes, the haze distribution in real scenes often presents a non-uniform state.

[0004] In order to overcome the above problems, many deep learning-based methods have been proposed to obtain more robust dehazing methods. For example, RIDCP, a real dehazing method based on high-quality codebook prior, introduces a VQGAN discrete codebook of a network pre-trained on a large range of high-quality image datasets after feature encoding, and encapsulates it as a high-quality codebook prior to generate high-quality clear images. However, due to the lag of the prior network, this method can only process fog images of a specific concentration, and has poor dehazing effect on enhanced fog images or real non-uniform fog images. Another example is based on the depth before and after dehazing. Figure 1 The defogging method is consistent with the previous one, and this prior is combined with the transformer network to improve the defogging performance. However, this method is too dependent on the accuracy of depth map estimation. If the difference between the depth map before and after defogging is small, the defogging effect of the network will be significantly affected.

[0005] Due to the complexity of real haze scenes, the above methods still have great difficulties in dealing with haze of different depths and concentrations.

[0006] One of the current existing technologies is a transformer dehazing method based on transmission-aware 3D point embedding in Image Dehazing Transformer with Transmission-Aware 3D Position Embedding by C. Guo et al. This method focuses on two key issues of transformer-based dehazing methods: first, although the Transformer can provide long-range feature dependencies through cascaded self-attention, even in the early stages, it lacks the ability to retain local feature details, resulting in blurred and rough details of image reconstruction; second, previous position embedding methods ignore the differences between different haze density areas, affecting the image dehazing performance; and proposes two corresponding solutions: first, combining the global modeling ability of the Transformer and the local representation ability of the CNN for image dehazing. To achieve this goal, for an input haze image, this method extracts hierarchical global features through the Transformer module, and the corresponding hierarchical local features obtained by the CNN module. Moreover, this method does not simply connect or add Transformer features and CNN features, but integrates features by learning a feature modulation module of a modulation matrix, thereby solving the problem of feature inconsistency. Finally, the CNN decoder module is used to enlarge the image resolution and render the local details of the haze-free image. Secondly, this method designs a new transmission-aware 3D position embedding module, which introduces a prior related to haze density into Transformer. This module not only provides relative position, but also indicates haze density in different spatial regions. The disadvantages of this method are: it does not pay attention to the characteristics of the image in the frequency domain, and as a fully supervised training method, it does not use real fog images to prompt the network generalization ability.

[0007] The second existing technology is a lighting control dehazing method based on Retinex theory by J. Gui et al. in "Illumination Controllable Dehazing Network based on Unsupervised Retinex Embedding". This method mainly analyzes the clear image equal to the fog image multiplied by the lighting image based on the Retinex theory. According to this result, a transformer-based generation module CIDM is designed to learn the lighting image and control the brightness of the result through a hyperparameter alpha. The disadvantages of this method are: first, the Retinex theory used in this method is only applicable to the field of low-light enhancement. In fact, in the field of dehazing, the difference between the clear image and the fog image is not only the difference in illumination, but also the attenuation of illumination caused by haze, so the theoretical migration is far-fetched; second, the CIDM module is designed based on the transformer without further improvement. Its computational complexity is high, and its ability to restore local details of the image is poor, making it difficult to use for real dehazing. Summary of the invention

[0008] The purpose of the present invention is to overcome the shortcomings of the existing methods and propose a brightness-guided dual-domain perception image defogging method and system. The main problem solved by the present invention is how to overcome the problem that the existing technology is difficult to handle haze of different depths and concentrations, so as to restore a more natural and clear image in a real haze image.

[0009] In order to solve the above problems, the present invention proposes a brightness-guided dual-domain perception image defogging method, the method comprising:

[0010] Construct a degradation model of real haze images and use a haze synthesis pipeline to simulate haze in natural scenes, so as to design and synthesize a haze dataset containing multiple confounding factors;

[0011] Design a dehazing network, whose structure includes an encoder, a dual-domain perception module, and a decoder; the dual-domain perception module includes a frequency-domain attention module and a spatial attention module; the encoder is used to extract image features, the dual-domain perception module is used to calculate image features to obtain high-quality clear features, and the decoder is used to decode high-quality clear features to obtain a predicted clear image;

[0012] Designing the training strategy and loss function required for the defogging network, and using the haze dataset for training to form a brightness-guided dual-domain perception image defogging network;

[0013] The brightness-guided dual-domain perception image dehazing network is tested using a no-reference image quality evaluation index. When the performance is good on the above evaluation index, the network parameters are fixed to form the final brightness-guided dual-domain perception image dehazing network. The user inputs any haze image to be processed into the dehazing network to obtain the image dehazing result.

[0014] Preferably, the degradation model of the real haze image is constructed, and the haze synthesis pipeline is used to simulate the haze in the natural scene, so as to design and synthesize the haze data set containing multiple confounding factors, specifically:

[0015] The labels used to synthesize haze images in the haze dataset come from the images in the RESIDE dataset and the corresponding depth image information;

[0016] The haze dataset constructs a degradation model by adding factors such as noise N and low light γ, and then generates it through the degradation model. The construction process is described by the following formula:

[0017] I γ =γ(I 0 t+A(1-t)+N)

[0018] Among them, I γ Indicates the haze image to be synthesized, I 0 represents the input clear image, γ represents the constraint factor used to adjust the brightness of the generated haze data image, A represents the atmospheric light value, t represents the scene transmittance map, and N represents the noise term.

[0019] Preferably, the defogging network is designed, and its structure includes an encoder, a dual-domain perception module, and a decoder; wherein the dual-domain perception module includes a frequency domain attention module and a spatial attention module; the encoder is used to extract image features, the dual-domain perception module is used to calculate the image features to obtain high-quality clear features, and the decoder is used to decode the high-quality clear features to obtain a predicted clear image, specifically:

[0020] The encoder and decoder of the dehazing network adopt the Res2Net network architecture, which uses encoders pre-trained on a wide range of datasets to extract features;

[0021] In order to further obtain high-quality clear features, a dual-domain perception module is constructed, that is, the frequency domain attention module FAM is first designed, and then the designed frequency domain attention module FAM is applied to the spatial attention module; the spatial attention module is as follows:

[0022] In general, the features input to the encoder are first filtered in the frequency domain, and then the local self-attention mechanism is used to further focus on the decoupling of high-frequency image content and low-frequency haze in the geometric space, thereby achieving higher quality clear feature extraction;

[0023] Specifically, for the encoder features Through the layer normalization operation LN(·), the normalized feature z is obtained l = LN(z); Then, the normalized feature z l Q projected into the frequency domain f , K f and V f , whose projection is shown below:

[0024] Q f =FAM Q (z l ), K f =FAM K (z l ), V f =FAM V (z l )

[0025] Among them, FAM Q 、FAM K and FAM V Respectively represent three independent projection operations with learnable parameters. Through the above formula, the operation of applying the frequency domain attention module to the spatial attention module is realized;

[0026] In obtaining the characteristic Q considering the frequency domain information f , K f and V f Afterwards, feature learning is performed in the spatial domain from a local perspective, where the locally aware self-attention is calculated within an 8×8 non-overlapping window as follows:

[0027]

[0028] Where d and B represent the dimension and position deviation respectively, and the symbol Represents matrix multiplication;

[0029] Then, information is transmitted through residual connections to obtain the high-quality clear features, as follows:

[0030] z * =Attention(Q f , K f , V f )+z

[0031] Finally, the high-quality clear features are input into the decoder and decoded to obtain the predicted clear image.

[0032] Preferably, the design of the frequency domain attention module is specifically as follows:

[0033] For the frequency domain attention module, the encoder feature representation is in and Represent the height, width and number of channels respectively;

[0034] First, through Fourier transform F, each channel Projected into the frequency domain, specifically:

[0035]

[0036] Where (h, w) and (u, v) represent the coordinates in the spatial domain and frequency domain respectively. Indicates the channel index;

[0037] Accordingly, F -1 is the inverse Fourier transform, the real part of the Fourier transform and the imaginary part pass Get the amplitude spectrum of a single channel and phase spectrum It is obtained by applying the formula through the following channels:

[0038]

[0039] By applying the formula to each channel of the encoder feature, the amplitude spectrum of all channels is obtained and phase spectrum

[0040] Since fog and image content have different frequency characteristics in Fourier space, the channel attention mechanism is used for filtering to further remove unimportant fog information, which is recorded as: S(z)(u,v)∈{A(z)(u,v),P(z)(u,v)}. In order to focus on the spectral information on the channel, the channel weight is calculated using the following formula:

[0041] W(z)(u,v)=σ(MLP(AvgPool(S(z)(u,v)))+MLP(MaxPool(s(z)(u,v))))

[0042] Among them, AvgPool represents global average pooling, and MaxPool represents global maximum pooling. After the structures obtained by these two pooling methods are input into the shared fully connected layer MLP, two weight bars are obtained. These two weight bars are added and sigmoid normalized σ to obtain the final weight bar: The weight bar is multiplied pixel-wise with the feature map before pooling and filtered through a residual connection as follows:

[0043]

[0044] Based on the equation from S(z)(u,v) to The processing flow is as follows, and the filtered and The real and imaginary parts are obtained by the following formulas:

[0045]

[0046] After dynamic parameter learning in the frequency domain, the feature map is remapped to the spatial domain as follows:

[0047] z′=F -1 (R(z)(u,v),J(z)(u,v))

[0048] The characteristics Fourier transform and inverse Fourier transform are implemented using DFT and IDFT algorithms;

[0049] Preferably, the training strategy and loss function required for designing the defogging network are designed, and the haze data set is used for training to form a brightness-guided dual-domain perception image defogging network, specifically:

[0050] The brightness-guided dual-domain-aware image dehazing network is trained in two stages to sequentially obtain a teacher model and a student model, both of which share the same network architecture;

[0051] In the first stage, the network uses paired synthetic data to pre-train a teacher network, using the structural similarity loss L between the dehazed image and the clean label GT output by the network. SSIM To guide network updates, the formula is as follows:

[0052]

[0053] Among them, x and y represent the GT and the dehazed image output by the network, respectively, and μ x , μ y and σ x , σ y Represent the mean and standard deviation of the dehazed image of GT and network output, σ xy represents their covariance, C 1 and C 2 is a constant;

[0054] In order to compensate for the brightness distortion defect of the dehazed image, a brightness loss based on depth clustering is adopted, including using the blue channel of the dehazed image to approximate the depth map of the image, and then using the k-means clustering technique to divide the depth map into 10 regions. Finally, all pixel values ​​of the depth map in these 10 regions are averaged as the brightness value of the current region. The calculation formula is as follows:

[0055]

[0056] Where (h, w)∈Ω m represents the pixel coordinates in the mth region, c represents the number of channels of the image, M is the number of pixels in the mth region, and similarly, p corresponding to the dehazed image m The mth region Ω m Brightness Defined in the same way;

[0057] Based on the fact that the brightness of the dehazed image is lower than that of the haze image, the haze image is used to supervise the brightness of the dehazed image. In order to ensure the relative numerical relationship between the high-brightness and low-brightness areas before and after dehazing, a power function with a monotonically increasing property is used to preprocess the haze image. The formula reflecting the actual brightness loss of all areas is as follows:

[0058]

[0059] Where k represents the brightness intensity coefficient;

[0060] The loss L of the teacher network teacher Including structural similarity loss L SSIM And the actual brightness loss Its expression is as follows:

[0061]

[0062] In the second stage, the teacher network is first used to initialize the student network, and unpaired real fog images from the URHI dataset are introduced into the student network to improve the adaptability of the student network to actual haze scenes;

[0063] For learning on synthetic data, the student network uses the structural similarity loss L SSIM ; For learning on real data, we use the inter-domain alignment loss L DA , including using the high-quality and clear features to perform global average pooling on the synthetic data, and then calculating the KL divergence of the synthetic data and the real data respectively, and the inter-domain alignment loss L DA It can be obtained by the following formula:

[0064] LDA =KL(AvgPool(syn),AvgPool(real))

[0065] Wherein, AvgPool(syn) represents the global average pooling of synthetic data, and AvgPool(real) represents the global average pooling of real data;

[0066] The overall loss L of the student network student Include the structural similarity loss L for synthetic data SSIM , inter-domain alignment loss L for real data DA , and brightness loss L B , the student network is trained under the guidance of these three losses, specifically:

[0067] L student =λ 1 L SSIM +λ 2 L DA +λ 3 L B

[0068] Among them, λ 1 , 2 , 3 is a constant;

[0069] The total loss L of the brightness-guided dual-domain-aware image network total for:

[0070] L total =L teacher +L student .

[0071] Preferably, the brightness-guided dual-domain perception image defogging network is tested using the reference-free quality evaluation index. When the performance is good on the above evaluation index, the network parameters are fixed to form the final brightness-guided dual-domain perception image defogging network. The user inputs any haze image to be processed into the defogging network to obtain the image defogging result, which is specifically:

[0072] In the testing phase, the RTTS dataset from the RESIDE dataset is used. During the testing, haze images of any size are input into the brightness-guided dual-domain perception image network in batch order.

[0073] Loading the model weights trained on the haze dataset, inputting the haze images to be tested into the student network one by one, and outputting the predicted haze-free images;

[0074] The quality of the predicted fog-free output image is evaluated using non-reference image evaluation indicators, including PM2.5, FADE, and PIQE. If the evaluation result is good performance, the parameters of the student model are fixed to obtain the model training weights, thereby forming the brightness-guided dual-domain perception image defogging network;

[0075] The user inputs any haze image to be processed into the brightness-guided dual-domain perception image dehazing network to obtain the image dehazing result.

[0076] Accordingly, the present invention also provides a brightness-guided dual-domain perception image defogging system, comprising:

[0077] The dataset design unit constructs a degradation model of real haze images and uses a haze synthesis pipeline to simulate haze in natural scenes, thereby designing and synthesizing a haze dataset containing multiple confounding factors;

[0078] The network construction unit designs a defogging network, whose structure includes an encoder, a dual-domain perception module, and a decoder; the dual-domain perception module includes a frequency domain attention module and a spatial attention module; the encoder is used to extract image features, the dual-domain perception module is used to calculate image features to obtain high-quality clear features, and the decoder is used to decode high-quality clear features to obtain a predicted clear image;

[0079] A network training unit, which designs a training strategy and a loss function required for the defogging network, and uses the haze data set for training to form a brightness-guided dual-domain perception image defogging network;

[0080] The test application unit uses the reference-free image quality evaluation index to test the brightness-guided dual-domain perception image defogging network. When the performance is good on the above evaluation index, the network parameters are fixed to form the final brightness-guided dual-domain perception image defogging network. The user inputs any haze image to be processed into the defogging network to obtain the image defogging result.

[0081] The implementation of the present invention has the following beneficial effects:

[0082] The present invention can effectively restore real haze scenes. Compared with existing image defogging methods, this method has better effects in defogging robustness, image structure, content details, and brightness fidelity, and the restored haze-free image is more natural;

[0083] The dual-domain perception module of the present invention integrates the information of haze images in frequency domain and geometric space. The channel attention mechanism in frequency domain helps to retain high-frequency image information and filter out low-frequency haze information. The frequency domain attention module is used to map the feature map of the self-attention mechanism, and the extracted frequency domain information is cleverly combined with the self-attention mechanism.

[0084] The true brightness loss formula constructed in the present invention can make the brightness of the dehazed image closer to the original true brightness of the image, the inter-domain alignment loss formula can make the distribution of synthetic data and real data closer, and the structural similarity loss formula can make the image content structure information before and after dehazing more completely preserved. These three losses enable the overall network training to be updated in the natural direction of dehazing. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 This is a flow chart of a brightness-guided dual-domain perception image defogging method according to an embodiment of the present invention;

[0086] Figure 2 This is a general structure diagram of a brightness-guided dual-domain perception image defogging network according to an embodiment of the present invention;

[0087] Figure 3 is a flow chart of a frequency domain attention module according to an embodiment of the present invention;

[0088] Figure 4 is a flow chart of a dual-domain perception module according to an embodiment of the present invention;

[0089] Figure 5 This is a structural diagram of a brightness-guided dual-domain perception image defogging system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0090] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0091] Figure 1 is a flow chart of a brightness-guided dual-domain perception image defogging method according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0092] S1, constructs a degradation model of real haze images and uses a haze synthesis pipeline to simulate haze in natural scenes, thereby designing and synthesizing a haze dataset containing multiple confounding factors;

[0093] S2, design a defogging network, whose structure includes an encoder, a dual-domain perception module, and a decoder; the dual-domain perception module includes a frequency domain attention module and a spatial attention module; the encoder is used to extract image features, the dual-domain perception module is used to calculate image features to obtain high-quality clear features, and the decoder is used to decode high-quality clear features to obtain a predicted clear image;

[0094] S3, designing a training strategy and loss function required for the defogging network, and using the haze dataset for training to form a brightness-guided dual-domain perception image defogging network;

[0095] S4, using the reference-free image quality evaluation index to test the brightness-guided dual-domain perception image defogging network. When the performance is good on the above evaluation index, the network parameters are fixed to form the final brightness-guided dual-domain perception image defogging network. The user inputs any haze image to be processed into the defogging network to obtain the image defogging result.

[0096] Step S1 is as follows:

[0097] S1-1: Considering that the degradation of haze images in real scenes will be affected by multiple confounding factors, the existing data synthesized based on the atmospheric light scattering model (ASM) only considers a single factor of haze image degradation, making it difficult for the models trained on these haze data to adapt to complex real haze scenes. Therefore, in the present invention, we reconstruct the degradation model of real haze images and propose a new haze synthesis pipeline to simulate haze in natural scenes. In this step, the clear labels used to synthesize haze images come from 2061 clear images and corresponding depth image information in the RESIDE dataset;

[0098] The degradation model is constructed by adding noise N, low light γ and other factors to the RESIDE haze dataset, and then the generation is performed using the degradation model. The construction process is described by the following formula:

[0099] I γ =γ(I 0 t+A(1-t)+N)

[0100] Among them, I γ Indicates the haze image to be synthesized, I 0 represents the input clear image, γ represents the constraint factor used to adjust the brightness of the generated haze data image, and is uniformly sampled in the interval [0.2, 0.8], A represents the atmospheric light value, and is randomly selected in the interval [0.75, 0.9], t represents the scene transmittance map, N represents the noise term, which obeys the mean of 0, and the formula is a normal distribution in the interval [0.0001, 5]. It can be found that the data synthesized by this formula can simulate multiple confounding factors in the real haze image degradation model, such as haze, low light, noise, etc.

[0101] Step S2 is as follows:

[0102] S2-1: The overall structure of the brightness-guided dual-domain perception image dehazing network is shown in Figure 2As shown in the figure, in this network, the encoder and decoder can be any existing efficient lightweight codec network. Here, the encoder and decoder of the dehazing network adopt the Res2Net network architecture, and use the encoder pre-trained in a wide range of data sets to extract features;

[0103] In order to further obtain high-quality clear features, a dual-domain perception module is constructed, that is, the frequency domain attention module FAM is first designed, and then the designed frequency domain attention module FAM is applied to the spatial attention module;

[0104] S2-2: The flowchart of the frequency domain attention module is as follows Figure 3 As shown, the characteristic representation of the encoder is in and Represent the height, width and number of channels respectively;

[0105] First, through Fourier transform F, each channel Projected into the frequency domain, specifically:

[0106]

[0107] Where (h, w) and (u, v) represent the coordinates in the spatial domain and frequency domain respectively. Indicates the channel index;

[0108] Accordingly, F -1 is the inverse Fourier transform, the real part of the Fourier transform and the imaginary part pass Get the amplitude spectrum of a single channel and phase spectrum It is obtained by applying the formula through the following channels:

[0109]

[0110] By applying the formula to each channel of the encoder feature, the amplitude spectrum of all channels is obtained and phase spectrum

[0111] Since fog and image content have different frequency characteristics in Fourier space, the channel attention mechanism is used for filtering to further remove unimportant fog information, which is recorded as: S(z)(u,v)∈{A(z)(u,v),P(z)(u,v)}. In order to focus on the spectral information on the channel, the channel weight is calculated using the following formula:

[0112] W(z)(u,v)=σ(MLP(AvgPool(S(z)(u,v)))+MLP(MaxPool(S(z)(u,v))))

[0113] Among them, AvgPool represents global average pooling, and MaxPool represents global maximum pooling. After the structures obtained by these two pooling methods are input into the shared fully connected layer MLP, two weight bars are obtained. These two weight bars are added and sigmoid normalized σ to obtain the final weight bar: The weight bar is multiplied pixel-wise with the feature map before pooling and filtered through a residual connection as follows:

[0114]

[0115] Based on the equation from S(z)(u,v) to The processing flow is as follows, and the filtered and The real and imaginary parts are obtained by the following formulas:

[0116]

[0117] After dynamic parameter learning in the frequency domain, the feature map is remapped to the spatial domain as follows:

[0118] z′=F -1 (R(z)(u,v),J(z)(u,v))

[0119] Features Fourier transform and inverse Fourier transform can be implemented using DFT and IDFT algorithms;

[0120] S2-3: Apply the frequency domain attention module to the spatial attention module, and the two together constitute the dual-domain perception module. The flowchart of the dual-domain perception module is as follows: Figure 4 As shown, in general, the features input to the encoder are first filtered in the frequency domain, and then the local self-attention mechanism is used to further focus on the decoupling of high-frequency image content and low-frequency haze in the geometric space, thereby achieving higher quality clear feature extraction;

[0121] Specifically, for the characteristics of the encoder Through the layer normalization operation LN(·), the normalized feature z is obtained l = LN(z); Then, the normalized feature z l Q projected into the frequency domain f (Query), K f (Key) and V f (value), whose projection is as follows:

[0122] Q f =FAM Q (z l ), K f =FAM K (z l ), V f =FAM V (z l )

[0123] Among them, FAM Q 、FAM K and FAM V Respectively represent three independent projection operations with learnable parameters. Through the above formula, the operation of applying the frequency domain attention module to the spatial attention module is realized;

[0124] In obtaining the characteristic Q considering the frequency domain information f , K f and V f Afterwards, feature learning is performed in the spatial domain from a local perspective, where the locally aware self-attention is calculated within an 8×8 non-overlapping window as follows:

[0125]

[0126] Where d and B represent the dimension and position deviation respectively, and the symbol Represents matrix multiplication;

[0127] Then, information is transmitted through residual connections to obtain the high-quality clear features, as follows:

[0128] z * =Attention(Q f , K f , V f )+z.

[0129] Finally, the high-quality clear features are input into the decoder and decoded to obtain the predicted clear image.

[0130] Step S3 is as follows:

[0131] S3-1: The brightness-guided dual-domain-aware image dehazing network is trained in two stages to sequentially obtain a teacher model and a student model, both of which share the same network architecture;

[0132] In the first stage, the network uses paired synthetic data to pre-train a teacher network, using the structural similarity loss L between the dehazed image and the clean label GT output by the network. SSIM To guide network updates, the formula is as follows:

[0133]

[0134] Among them, x and y represent the GT and the dehazed image output by the network, respectively, and μ x , μ y and σ x , σ y Represent the mean and standard deviation of the dehazed image of GT and network output, σ xy represents their covariance, C 1 and C 2 is a constant;

[0135] In order to compensate for the brightness distortion defect of the network for the dehazed image, a brightness loss based on depth clustering is adopted, including using the blue channel of the dehazed image to approximate the depth map of the image, and then using k- means The clustering technique divides the depth map into 10 regions, and finally averages all pixel values ​​of the depth map in these 10 regions as the brightness value of the current region. The calculation formula is as follows:

[0136]

[0137] Where (h, w)∈Ω m represents the pixel coordinates in the mth region, c represents the number of channels of the image, M is the number of pixels in the mth region, and similarly, p corresponding to the dehazed image m The mth region Ω m Brightness Defined in the same way;

[0138] Based on the fact that the brightness of the dehazed image is lower than that of the haze image, the haze image is used to supervise the brightness of the dehazed image. In order to ensure the relative numerical relationship between the high-brightness and low-brightness areas before and after dehazing, a power function with a monotonically increasing property is used to preprocess the haze image. The formula reflecting the actual brightness loss of all areas is as follows:

[0139]

[0140] Where, (k≥1) represents the brightness intensity coefficient;

[0141] The loss L of the teacher network teacher Including structural similarity loss L SSIM And the actual brightness loss After training it on our proposed NLHaze dataset for 520 cycles, its expression is as follows:

[0142]

[0143] S3-2: In the second stage, the teacher network is first used to initialize the student network. The student network not only uses NLHaz e The 3,800 pairs of synthetic data in the dataset also introduce 4,000 unpaired real fog images from the URHI dataset to improve the adaptability of the student network to actual haze scenes;

[0144] For learning on synthetic data, the student network uses the structural similarity loss L SSIM ; For learning on real data, we use the inter-domain alignment loss L DA , including using the high-quality and clear features to perform global average pooling on the synthetic data, and then calculating the KL divergence of the synthetic data and the real data respectively, and the inter-domain alignment loss L DA It can be obtained by the following formula:

[0145] L DA =KL(AvgPool(syn),AvgPool(real))

[0146] Wherein, AvgPool(syn) represents the global average pooling of synthetic data, and AvgPool(real) represents the global average pooling of real data;

[0147] The overall loss L of the student network student Include the structural similarity loss L for synthetic data SSIM , inter-domain alignment loss L for real data DA , and brightness loss L B , the student network is trained for 20 cycles under the guidance of these three losses, specifically:

[0148] L student =λ 1 L SSIM +λ 2 L DA +λ 3 L B

[0149] Among them, λ 1 , 2 , 3 are constants, taking the values ​​of 20, 0.001 and 1 respectively;

[0150] The total loss L of the brightness-guided dual-domain-aware image network total for:

[0151] L total =L teacher +L student .

[0152] Step S4 is as follows:

[0153] In the test phase, the RTTS dataset from the RESIDE dataset is used, which contains 4322 haze images in real scenes, including but not limited to road traffic haze scenes, natural landscape haze, night haze, etc.; in the test, haze images of any size are input into the brightness-guided dual-domain perception image network in batch order;

[0154] Loading the model weights trained on the haze dataset, inputting the haze images to be tested into the student network one by one, and outputting the predicted haze-free images;

[0155] The quality of the predicted fog-free output image is evaluated using a variety of authoritative non-reference image quality evaluation indicators such as PM2.5, FADE, PIQE, etc. If the evaluation result is good performance, the parameters of the student model are fixed to obtain the model training weights to form the brightness-guided dual-domain perception image defogging network;

[0156] The user inputs any haze image to be processed into the brightness-guided dual-domain perception image dehazing network to obtain the image dehazing result;

[0157] In addition, if satisfactory requirements are not met on the multiple no-reference graph evaluation indicators used above, the model training weight parameters are replaced and retested until the network model performs well on multiple authoritative no-reference graph indicators.

[0158] Accordingly, the present invention also provides a brightness-guided dual-domain perception image defogging system, such as Figure 5 As shown, including:

[0159] Dataset design unit 1 is used to construct a degradation model of real haze images and use the haze synthesis pipeline to simulate haze in natural scenes, so as to design and synthesize a haze dataset containing multiple confounding factors;

[0160] The network construction unit 2 is used to design a defogging network, and its structure includes an encoder, a dual-domain perception module, and a decoder; wherein the dual-domain perception module includes a frequency domain attention module and a spatial attention module; the encoder is used to extract image features, the dual-domain perception module is used to calculate the image features to obtain high-quality clear features, and the decoder is used to decode the high-quality clear features to obtain a predicted clear image;

[0161] A network training unit 3 is used to design a training strategy and loss function required for the defogging network, and use the haze data set for training to form a brightness-guided dual-domain perception image defogging network;

[0162] The test application unit 4 is used to test the brightness-guided dual-domain perception image defogging network using a reference-free quality evaluation index. When the performance is good on the above evaluation index, the network parameters are fixed to form the final brightness-guided dual-domain perception image defogging network. The user inputs any haze image to be processed into the defogging network to obtain the image defogging result.

[0163] Therefore, the present invention can effectively restore the real haze scene. Compared with the existing image defogging methods, this method has better effects in defogging robustness, image structure, content details, and brightness fidelity, and the restored haze-free image is more natural; the present invention integrates the information of haze images in the frequency domain and geometric space in the dual-domain perception module, and the channel attention mechanism in the frequency domain helps to retain high-frequency image information and filter out low-frequency haze information; the frequency domain attention module is used to map the feature map of the self-attention mechanism, and the extracted frequency domain information is cleverly combined with the self-attention mechanism; the real brightness loss formula constructed in the present invention can make the brightness of the defogged image closer to the original real brightness of the image, the inter-domain alignment loss formula can make the distribution of synthetic data and real data closer, and the structural similarity loss formula can make the image content structure information before and after defogging more complete. These three losses update the overall network training in the direction of natural defogging.

[0164] The above is a detailed introduction to a brightness-guided dual-domain perception image defogging method and system provided by an embodiment of the present invention. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for general technical personnel in this field, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A brightness-guided dual-domain perception image defogging method, characterized in that: The method comprises: Construct a degradation model of real haze images and use a haze synthesis pipeline to simulate haze in natural scenes, so as to design and synthesize a haze dataset containing multiple confounding factors; Design a dehazing network, whose structure includes an encoder, a dual-domain perception module, and a decoder; the dual-domain perception module includes a frequency-domain attention module and a spatial attention module; the encoder is used to extract image features, the dual-domain perception module is used to calculate image features to obtain high-quality clear features, and the decoder is used to decode high-quality clear features to obtain a predicted clear image; Designing the training strategy and loss function required for the defogging network, and using the haze dataset for training to form a brightness-guided dual-domain perception image defogging network; The brightness-guided dual-domain perception image dehazing network is tested using a no-reference image quality evaluation index. When the performance is good on the above evaluation index, the network parameters are fixed to form the final brightness-guided dual-domain perception image dehazing network. The user inputs any haze image to be processed into the dehazing network to obtain the image dehazing result.

2. The brightness-guided dual-domain perception image defogging method according to claim 1, characterized in that: The degradation model of the real haze image is constructed, and the haze synthesis pipeline is used to simulate the haze in the natural scene, so as to design and synthesize the haze data set containing multiple confounding factors, specifically: The labels used to synthesize haze images in the haze dataset come from the images in the RESIDE dataset and the corresponding depth image information; The haze dataset constructs a degradation model by adding factors such as noise N and low light γ, and then generates it through the degradation model. The construction process is described by the following formula: I γ =γ(I0t+A(1-t)+N) Among them, I γ represents the haze image to be synthesized, I0 represents the input clear image, γ represents the constraint factor used to adjust the brightness of the generated haze data image, A represents the atmospheric light value, t represents the scene transmittance map, and N represents the noise term.

3. The brightness-guided dual-domain perception image defogging method according to claim 1, characterized in that: The defogging network is designed, and its structure includes an encoder, a dual-domain perception module, and a decoder; wherein the dual-domain perception module includes a frequency domain attention module and a spatial attention module; the encoder is used to extract image features, the dual-domain perception module is used to calculate the image features to obtain high-quality clear features, and the decoder is used to decode the high-quality clear features to obtain a predicted clear image, specifically: The encoder and decoder of the dehazing network adopt the Res2Net network architecture, which uses encoders pre-trained on a wide range of datasets to extract features; In order to further obtain high-quality clear features, a dual-domain perception module is constructed, that is, the frequency domain attention module FAM is first designed, and then the designed frequency domain attention module FAM is applied to the spatial attention module; the spatial attention module is as follows: In general, the features input to the encoder are first filtered in the frequency domain, and then the local self-attention mechanism is used to further focus on the decoupling of high-frequency image content and low-frequency haze in the geometric space, thereby achieving higher quality clear feature extraction; Specifically, for the encoder features Through the layer normalization operation LN(·), the normalized feature z is obtained l = LN(z); Then, the normalized feature z l Q projected into the frequency domain f , K f and V f , whose projection is shown below: Q f =FAM Q (With l ),K f =FAM K (With l ),V f =FAM V (With l ) Among them, FAM Q 、FAM K and FAM V Respectively represent three independent projection operations with learnable parameters. Through the above formula, the operation of applying the frequency domain attention module to the spatial attention module is realized; In obtaining the characteristic Q considering the frequency domain information f , K f and V f Afterwards, feature learning is performed in the spatial domain from a local perspective, where the locally aware self-attention is calculated within an 8×8 non-overlapping window as follows: Where d and B represent the dimension and position deviation respectively, and the symbol Represents matrix multiplication; Then, information is transmitted through residual connections to obtain the high-quality clear features, as follows: z * =Attention(Q f ,K f ,V f )+z Finally, the high-quality clear features are input into the decoder and decoded to obtain the predicted clear image.

4. The brightness-guided dual-domain perception image defogging method according to claim 3, characterized in that: The designed frequency domain attention module is specifically: For the frequency domain attention module, the features of the encoder are used in and Represent the height, width and number of channels respectively; First, through Fourier transform F, each channel Projected into the frequency domain, specifically: Among them, (h,w) and (u,v) represent the coordinates in the spatial domain and frequency domain respectively. Indicates the channel index; Accordingly, F -1 is the inverse Fourier transform, the real part of the Fourier transform and the imaginary part pass Get the amplitude spectrum of a single channel and phase spectrum It is obtained by applying the formula through the following channels: By applying the formula to each channel of the encoder feature, the amplitude spectrum of all channels is obtained and phase spectrum Since fog and image content have different frequency characteristics in Fourier space, the channel attention mechanism is used for filtering to further remove unimportant fog information, which is recorded as: S(z)(u,v)∈{A(z)(u,v),P(z)(u,v)}. In order to focus on the spectral information on the channel, the channel weight is calculated using the following formula: W(z)(u,v)=σ(MLP(AvgPool(S(z)(u,v)))+MLP(MaxPool(S(Z)(u,v))))W ... The weight bar is multiplied pixel-wise with the feature map before pooling and filtered through a residual connection as follows: Based on the equation from S(z)(u,v) to The processing flow is as follows, and the filtered and The real and imaginary parts are obtained by the following formulas: After dynamic parameter learning in the frequency domain, the feature map is remapped to the spatial domain as follows: z′=F -1 (R(z)(u,v),J(z)(u,v)) The characteristics Fourier transform and inverse Fourier transform are implemented using DFT and IDFT algorithms.

5. The brightness-guided dual-domain perception image defogging method according to claim 1, characterized in that: The training strategy and loss function required for designing the defogging network are used, and the haze data set is used for training to form a brightness-guided dual-domain perception image defogging network, specifically: The brightness-guided dual-domain-aware image dehazing network is trained in two stages to sequentially obtain a teacher model and a student model, both of which share the same network architecture; In the first stage, the network uses paired synthetic data to pre-train a teacher network, using the structural similarity loss L between the dehazed image and the clean label GT output by the network. SSIM To guide network updates, the formula is as follows: Among them, x and y represent the GT and the dehazed image output by the network, respectively, and μ x , μ y and σ x , σ y Represent the mean and standard deviation of the dehazed image of GT and network output, σ xy represents their covariance, C1 and C2 are constants; In order to compensate for the brightness distortion defect of the dehazed image, a brightness loss based on depth clustering is adopted, including using the blue channel of the dehazed image to approximate the depth map of the image, and then using the k-means clustering technique to divide the depth map into 10 regions. Finally, all pixel values ​​of the depth map in these 10 regions are averaged as the brightness value of the current region. The calculation formula is as follows: Where (h,w)∈Ω m represents the pixel coordinates in the mth region, c represents the number of channels of the image, M is the number of pixels in the mth region, and similarly, p corresponding to the dehazed image m The mth region Ω m Brightness Defined in the same way; Based on the fact that the brightness of the dehazed image is lower than that of the haze image, the haze image is used to supervise the brightness of the dehazed image. In order to ensure the relative numerical relationship between the high-brightness and low-brightness areas before and after dehazing, a power function with a monotonically increasing property is used to preprocess the haze image. The formula reflecting the actual brightness loss of all areas is as follows: Where k represents the brightness intensity coefficient; The loss L of the teacher network teacher Including structural similarity loss L SSIM And the actual brightness loss Its expression is as follows: In the second stage, the teacher network is first used to initialize the student network, and unpaired real fog images from the URHI dataset are introduced into the student network to improve the adaptability of the student network to actual haze scenes; For learning on synthetic data, the student network uses the structural similarity loss L SSIM ; For learning on real data, we use the inter-domain alignment loss L DA , including using the high-quality and clear features to perform global average pooling on the synthetic data, and then calculating the KL divergence of the synthetic data and the real data respectively, and the inter-domain alignment loss L DA It can be obtained by the following formula: L DA =KL(AvgPool(syn),AvgPool(real)) Wherein, AvgPool(syn) represents the global average pooling of synthetic data, and AvgPool(real) represents the global average pooling of real data; The overall loss L of the student network student Include the structural similarity loss L for synthetic data SSIM , inter-domain alignment loss L for real data DA , and brightness loss L B , the student network is trained under the guidance of these three losses, specifically: L student =λ1L SSIM +λ2L DA +λ3L B Among them, λ1, λ2, λ3 are constants; The total loss L of the brightness-guided dual-domain-aware image network total for: L total =L teacher +L student .

6. The brightness-guided dual-domain perception image defogging method according to claim 5, characterized in that: The brightness-guided dual-domain perception image defogging network is tested using the reference-free image quality evaluation index. When the performance is good on the above evaluation index, the network parameters are fixed to form the final brightness-guided dual-domain perception image defogging network. The user inputs any haze image to be processed into the defogging network to obtain the image defogging result, which is specifically: In the testing phase, the RTTS dataset from the RESIDE dataset is used. During the testing, haze images of any size are input into the brightness-guided dual-domain perception image network in batch order. Loading the model weights trained on the haze dataset, inputting the haze images to be tested into the student network one by one, and outputting the predicted haze-free images; The quality of the predicted fog-free output image is evaluated using non-reference image evaluation indicators, including PM2.5, FADE, and PIQE. If the evaluation result is good performance, the parameters of the student model are fixed to obtain the model training weights, thereby forming the brightness-guided dual-domain perception image defogging network; The user inputs any haze image to be processed into the brightness-guided dual-domain perception image dehazing network to obtain the image dehazing result.

7. A brightness-guided dual-domain perception image defogging system, characterized in that: The system comprises: The dataset design unit constructs a degradation model of real haze images and uses a haze synthesis pipeline to simulate haze in natural scenes, thereby designing and synthesizing a haze dataset containing multiple confounding factors; The network construction unit designs a defogging network, whose structure includes an encoder, a dual-domain perception module, and a decoder; the dual-domain perception module includes a frequency domain attention module and a spatial attention module; the encoder is used to extract image features, the dual-domain perception module is used to calculate image features to obtain high-quality clear features, and the decoder is used to decode high-quality clear features to obtain a predicted clear image; A network training unit, which designs a training strategy and a loss function required for the defogging network, and uses the haze data set for training to form a brightness-guided dual-domain perception image defogging network; The test application unit uses the reference-free image quality evaluation index to test the brightness-guided dual-domain perception image defogging network. When the performance is good on the above evaluation index, the network parameters are fixed to form the final brightness-guided dual-domain perception image defogging network. The user inputs any haze image to be processed into the defogging network to obtain the image defogging result.

8. The brightness-guided dual-domain perception image defogging system according to claim 7, characterized in that: The degradation model of the real haze image is constructed, and the haze synthesis pipeline is used to simulate the haze in the natural scene, so as to design and synthesize the haze data set containing multiple confounding factors, specifically: The labels used to synthesize haze images in the haze dataset come from the images in the RESIDE dataset and the corresponding depth image information; The haze dataset constructs a degradation model by adding factors such as noise N and low light γ, and then generates it through the degradation model. The construction process is described by the following formula: I γ =γ(I0t+A(1-t)+N) Among them, I γ represents the haze image to be synthesized, I0 represents the input clear image, γ represents the constraint factor used to adjust the brightness of the generated haze data image, A represents the atmospheric light value, t represents the scene transmittance map, and N represents the noise term.

9. The brightness-guided dual-domain perception image defogging system according to claim 7, characterized in that: The defogging network is designed, and its structure includes an encoder, a dual-domain perception module, and a decoder; wherein the dual-domain perception module includes a frequency domain attention module and a spatial attention module; the encoder is used to extract image features, the dual-domain perception module is used to calculate the image features to obtain high-quality clear features, and the decoder is used to decode the high-quality clear features to obtain a predicted clear image, specifically: The encoder and decoder of the dehazing network adopt the Res2Net network architecture, which uses encoders pre-trained on a wide range of datasets to extract features; In order to further obtain high-quality clear features, a dual-domain perception module is constructed, that is, the frequency domain attention module FAM is first designed, and then the designed frequency domain attention module FAM is applied to the spatial attention module; the spatial attention module is as follows: In general, the features input to the encoder are first filtered in the frequency domain, and then the local self-attention mechanism is used to further focus on the decoupling of high-frequency image content and low-frequency haze in the geometric space, thereby achieving higher quality clear feature extraction; Specifically, for the encoder features Through the layer normalization operation LN(·), the normalized feature z is obtained l = LN(z); Then, the normalized feature z l Q projected into the frequency domain f , K f and V f , whose projection is shown below: Q f =FAM Q (With l ),K f =FAM K (With l ),V f =FAM V (With l ) Among them, FAM Q 、FAM K and FAM V Respectively represent three independent projection operations with learnable parameters. Through the above formula, the operation of applying the frequency domain attention module to the spatial attention module is realized; In obtaining the characteristic Q considering the frequency domain information f , K f and V f Afterwards, feature learning is performed in the spatial domain from a local perspective, where the locally aware self-attention is calculated within an 8×8 non-overlapping window as follows: Where d and B represent the dimension and position deviation respectively, and the symbol Represents matrix multiplication; Then, information is transmitted through residual connections to obtain the high-quality clear features, as follows: z * =Attention(Q f ,K f ,V f )+z Finally, the high-quality clear features are input into the decoder and decoded to obtain the predicted clear image.

10. The brightness-guided dual-domain perception image defogging system according to claim 9, characterized in that: The designed frequency domain attention module is specifically: For the frequency domain attention module, the encoder feature representation is in and Represent the height, width and number of channels respectively; First, through Fourier transform F, each channel Projected into the frequency domain, specifically: Among them, (h,w) and (u,v) represent the coordinates in the spatial domain and frequency domain respectively. Indicates the channel index; Accordingly, F -1 is the inverse Fourier transform, the real part of the Fourier transform and the imaginary part pass Get the amplitude spectrum of a single channel and phase spectrum It is obtained by applying the formula through the following channels: By applying the formula to each channel of the encoder feature, the amplitude spectrum of all channels is obtained and phase spectrum Since fog and image content have different frequency characteristics in Fourier space, the channel attention mechanism is used for filtering to further remove unimportant fog information, which is recorded as: S(z)(u,v)∈{A(z)(u,v),P(z)(u,v)}. In order to focus on the spectral information on the channel, the channel weight is calculated using the following formula: W(z)(u,v)=σ(MLP(AvgPool(S(z)(u,v)))+MLP(MaxPool(S(Z)(u,v)))) Among them, AvgPool represents global average pooling, and MaxPool represents global maximum pooling. After the structures obtained by these two pooling methods are input into the shared fully connected layer MLP, two weight bars are obtained. These two weight bars are added and sigmoid normalized σ to obtain the final weight bar: The weight bar is multiplied pixel-wise with the feature map before pooling and filtered through a residual connection as follows: Based on the equation from S(z)(u,v) to The processing flow is as follows, and the filtered and The real and imaginary parts are obtained by the following formulas: After dynamic parameter learning in the frequency domain, the feature map is remapped to the spatial domain as follows: z′=F -1 (R(z)(u,v),J(z)(u,v)) The characteristics Fourier transform and inverse Fourier transform are implemented using DFT and IDFT algorithms.

11. The brightness-guided dual-domain perception image defogging system according to claim 7, characterized in that: The training strategy and loss function required for designing the defogging network are used, and the haze data set is used for training to form a brightness-guided dual-domain perception image defogging network, specifically: The brightness-guided dual-domain-aware image dehazing network is trained in two stages to sequentially obtain a teacher model and a student model, both of which share the same network architecture; In the first stage, the network uses paired synthetic data to pre-train a teacher network, using the structural similarity loss L between the dehazed image and the clean label GT output by the network. SSIM To guide network updates, the formula is as follows: Among them, x and y represent the GT and the dehazed image output by the network, respectively, and μ x , μ y and σ x , σ y Represent the mean and standard deviation of the dehazed image of GT and network output, σ xy represents their covariance, C1 and C2 are constants; In order to compensate for the brightness distortion defect of the dehazed image, a brightness loss based on depth clustering is adopted, including using the blue channel of the dehazed image to approximate the depth map of the image, and then using the k-means clustering technique to divide the depth map into 10 regions. Finally, all pixel values ​​of the depth map in these 10 regions are averaged as the brightness value of the current region. The calculation formula is as follows: Where (h,w)∈Ω m represents the pixel coordinates in the mth region, c represents the number of channels of the image, M is the number of pixels in the mth region, and similarly, p corresponding to the dehazed image m The mth region Ω m Brightness Defined in the same way; Based on the fact that the brightness of the dehazed image is lower than that of the haze image, the haze image is used to supervise the brightness of the dehazed image. In order to ensure the relative numerical relationship between the high-brightness and low-brightness areas before and after dehazing, a power function with a monotonically increasing property is used to preprocess the haze image. The formula reflecting the actual brightness loss of all areas is as follows: Where k represents the brightness intensity coefficient; The loss L of the teacher network teacher Including structural similarity loss L SSIM And the actual brightness loss Its expression is as follows: In the second stage, the teacher network is first used to initialize the student network, and unpaired real fog images from the URHI dataset are introduced into the student network to improve the adaptability of the student network to actual haze scenes; For learning on synthetic data, the student network uses the structural similarity loss L SSIM ; For learning on real data, we use the inter-domain alignment loss L DA , including using the high-quality and clear features to perform global average pooling on the synthetic data, and then calculating the KL divergence of the synthetic data and the real data respectively, and the inter-domain alignment loss L DA It can be obtained by the following formula: L DA =KL(AvgPool(syn),AvgPool(real)) Wherein, AvgPool(syn) represents the global average pooling of synthetic data, and AvgPool(real) represents the global average pooling of real data; The overall loss L of the student network student Include the structural similarity loss L for synthetic data SSIM , inter-domain alignment loss L for real data DA , and brightness loss L B , the student network is trained under the guidance of these three losses, specifically: L student =λ1L SSIM +λ2L DA +λ3L B Among them, λ1, λ2, λ3 are constants; The total loss L of the brightness-guided dual-domain-aware image network total for: L total =L teacher +L student .

12. The brightness-guided dual-domain perception image defogging system according to claim 11, characterized in that: The brightness-guided dual-domain perception image defogging network is tested using the reference-free image quality evaluation index. When the performance is good on the above evaluation index, the network parameters are fixed to form the final brightness-guided dual-domain perception image defogging network. The user inputs any haze image to be processed into the defogging network to obtain the image defogging result, which is specifically: In the testing phase, the RTTS dataset from the RESIDE dataset is used. During the testing, haze images of any size are input into the brightness-guided dual-domain perception image network in batch order. Loading the model weights trained on the haze dataset, inputting the haze images to be tested into the student network one by one, and outputting the predicted haze-free images; The quality of the predicted fog-free output image is evaluated using non-reference image evaluation indicators, including PM2.5, FADE, and PIQE. If the evaluation result is good performance, the parameters of the student model are fixed to obtain the model training weights, thereby forming the brightness-guided dual-domain perception image defogging network; The user inputs any haze image to be processed into the brightness-guided dual-domain perception image dehazing network to obtain the image dehazing result.

Images