An image defogging method, device, equipment and storage medium

Abstract

The application discloses an image defogging method and device, equipment and storage medium, and relates to the technical field of artificial intelligence, and comprises the following steps: in the process of defogging the target fog image collected in the current foggy environment by using a target image defogging model based on a neural network, a target fog light coverage map corresponding to the target fog image is generated, and a target fusion feature is generated based on the target fog light coverage map; a target feature manifold representation corresponding to the target fusion feature is generated based on Lie group operation and manifold learning method; the target feature manifold representation is mapped to a plurality of low-dimensional sub-manifolds, so as to generate a target candidate projection set based on each low-dimensional sub-manifold, and a target projection is selected from the target candidate projection set by using a preset loss function; a target defogging image feature corresponding to the target projection is generated, and the target defogging image feature is decoded to obtain a corresponding target defogging image. The application can improve the accuracy of image defogging in complex and changeable scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to an image defogging method, device and equipment and a storage medium. BACKGROUND

[0002] In the past few years, defogging imaging technology has made significant progress in many fields, providing important technical support for solving perception problems in various complex environments. In foggy weather, atmospheric suspended particles such as water droplets and dust have scattering and absorption effects on light, causing image blurring, reduced contrast, color distortion and other problems, which seriously affect the reliability and safety of practical applications such as autonomous driving, security monitoring, remote sensing and mapping. Traditional defogging methods are mainly based on atmospheric scattering physical models, such as dark channel prior and color attenuation prior, which estimate atmospheric light and transmittance to restore the image. However, these methods rely on artificially designed prior assumptions and are prone to failure in complex scenes, making it difficult to recover detailed information such as non-uniform fog concentration and dynamic lighting scenes. Changes in lighting conditions, background noise interference and other factors can affect the quality and accuracy of imaging, making it difficult to meet the adaptive needs of variable and complex scenes.

[0003] In summary, how to improve the accuracy of image defogging in complex and variable scenes is a technical problem that needs to be solved at present. SUMMARY

[0004] Therefore, the purpose of the present application is to provide an image defogging method, device, equipment and storage medium that can improve the accuracy of image defogging in complex and variable scenes. The specific scheme is as follows:

[0005] In the first aspect, the present application provides an image defogging method, comprising:

[0006] In the process of defogging the target fog image collected in the current foggy environment using a neural network-based target image defogging model, a target fog lighting coverage map corresponding to the target fog image is generated, and a target fusion feature is generated based on the target fog lighting coverage map;

[0007] A target feature manifold representation corresponding to the target fusion feature is generated based on Lie group operation and manifold learning method;

[0008] The target feature manifold representation is mapped to a plurality of low-dimensional submanifolds to generate a target candidate projection set based on each low-dimensional submanifold, and a target projection is selected from the target candidate projection set using a preset loss function;

[0009] A target defogging image feature corresponding to the target projection is generated, and the target defogging image feature is decoded to obtain a corresponding target defogging image.

[0010] Optionally, the generating the target foggy image corresponding target foggy light coverage map comprises:

[0011] The target low-level features of the target foggy image are extracted through the encoder of the backbone network in the target image defogging model, and a preset size of a receptive field convolution kernel, a shared parameter network layer and a residual connection are used.

[0012] The target foggy image corresponding target foggy light coverage map is generated based on the target low-level features, and the target foggy image is labeled based on the target foggy light coverage map and a preset binarization method to generate a target foggy mask feature tensor.

[0013] Optionally, the generating the target fusion feature based on the target foggy light coverage map comprises:

[0014] The target foggy light coverage map and the target foggy mask feature tensor are subjected to stack convolution operation, batch normalization operation and nonlinear activation operation to generate corresponding target high semantic features;

[0015] The channel weights corresponding to the feature channels of the target high semantic features are generated through a channel attention mechanism to obtain corresponding channel attention features;

[0016] The spatial weights of the channel attention features are determined through a spatial attention mechanism to obtain corresponding spatial attention features; the spatial attention features are features of different scales;

[0017] The spatial attention features are subjected to feature fusion based on a preset cross-scale attention mechanism to obtain the target fusion feature corresponding to the target foggy image.

[0018] Optionally, the generating the target feature manifold representation corresponding to the target fusion feature based on the Lie group operation and the manifold learning method comprises:

[0019] The target fusion feature and the target foggy light coverage map are spliced to generate a target composite feature and obtain foggy weather scene characteristic information in the current environment through a feature manifold estimation and embedding module in the target image defogging model;

[0020] The target composite feature is subjected to distribution estimation to obtain target distribution estimation information corresponding to the target composite feature;

[0021] The Lie group operation is used, and a target Lie group space is constructed based on the foggy weather scene characteristic information and the target distribution estimation information; and a target manifold structure corresponding to the target distribution estimation information is established in the target Lie group space based on the manifold learning method.

[0022] The target feature manifold representation is mapped to a plurality of low-dimensional sub-manifolds, a target candidate projection set is generated based on each low-dimensional sub-manifold, and a target projection is screened from the target candidate projection set by using a preset loss function.

[0023] Optionally, the target feature manifold representation is mapped to a plurality of low-dimensional sub-manifolds, a target candidate projection set is generated based on each low-dimensional sub-manifold, and a target projection is screened from the target candidate projection set by using a preset loss function.

[0024] The target feature manifold representation is mapped to a plurality of low-dimensional sub-manifolds based on a preset manifold geometric structure condition and a preset feature constraint condition by a projection selection module in the target image defogging model, and the target candidate projection set is generated based on each low-dimensional sub-manifold; each candidate projection in the target candidate projection set is a low-dimensional feature distribution result of the target feature manifold representation.

[0025] A first loss corresponding to each candidate projection is determined by using a first preset loss function, and a second loss corresponding to each candidate projection is determined by using a second preset loss function.

[0026] The first loss and the second loss are respectively configured with corresponding weights, and the first loss and the second loss are weighted and summed based on the weight corresponding to the first loss and the weight corresponding to the second loss, so as to obtain a target loss corresponding to each candidate projection.

[0027] The candidate projection corresponding to the smallest target loss is determined as the target projection.

[0028] The first preset loss function is used to ensure that adjacent feature points in the target feature manifold representation still maintain a neighboring relationship in the low-dimensional sub-manifold, and the second preset loss function is used to compress redundant projections in the low-dimensional sub-manifold.

[0029] Optionally, the target defogging image feature is decoded to obtain a corresponding target defogging image.

[0030] The target defogging image feature is processed by using a current upsampling layer of a decoder of the backbone network in the target image defogging model, so as to obtain a current upsampling feature.

[0031] The target low-level feature corresponding to the target fog image generated by an encoder of the backbone network in the target image defogging model is obtained based on a skip connection, and the current upsampling feature and the target low-level feature are spliced to obtain a current splicing feature.

[0032] determine the next up-sampling layer as the current up-sampling layer, and jump to the step of processing the target defogging image feature by using the current up-sampling layer until the spatial size corresponding to the current stitching feature is the same as the preset spatial resolution corresponding to the target foggy image, to obtain a target stitching feature;

[0033] compress feature channels corresponding to the target stitching feature based on a preset channel number condition, and perform a nonlinear activation operation on the compressed target stitching feature, to obtain the target defogging image.

[0034] Optionally, the image defogging method further comprises:

[0035] obtaining a fog-free image and a first foggy image, and inputting the fog-free image and the first foggy image into the target image defogging model respectively;

[0036] generating a first feature manifold representation corresponding to the fog-free image by using the target image defogging model, and generating a second feature manifold representation corresponding to the first foggy image by using the target image defogging model;

[0037] determining a loss value between the first feature manifold representation and the second feature manifold representation according to a third preset loss function, and determining a target gradient based on the loss value;

[0038] optimizing network parameters in the target image defogging model based on a preset high-dimensional tendency score method and the target gradient, to defog a second foggy image by using the optimized target image defogging model.

[0039] In a second aspect, the present application provides an image defogging device, comprising:

[0040] a target fusion feature generation module configured to generate a target foggy light coverage atlas corresponding to a target foggy image in a process of defogging the target foggy image collected in a current foggy environment by using a target image defogging model based on a neural network, and generate a target fusion feature based on the target foggy light coverage atlas;

[0041] a target feature manifold representation generation module configured to generate a target feature manifold representation corresponding to the target fusion feature based on Lie group operation and manifold learning method;

[0042] a target projection determination module configured to map the target feature manifold representation to a plurality of low-dimensional sub-manifolds, generate a target candidate projection set based on each of the low-dimensional sub-manifolds, and screen a target projection from the target candidate projection set by using a preset loss function.

[0043] a target dehazed image determination module, configured to generate a target dehazed image feature corresponding to the target projection, and decode the target dehazed image feature to obtain a corresponding target dehazed image.

[0044] In a third aspect, the present application provides an electronic device, comprising:

[0045] a memory configured to store a computer program;

[0046] a processor configured to execute the computer program to implement the image dehazing method.

[0047] In a fourth aspect, the present application provides a computer readable storage medium configured to store a computer program; wherein the computer program is executed by a processor to implement the image dehazing method.

[0048] In the present application, in the process of dehazing the target foggy image collected in the current foggy environment by using the target image dehazing model based on neural network, first, a target foggy light coverage map corresponding to the target foggy image is generated, and a target fusion feature is generated based on the target foggy light coverage map; then, a target feature manifold representation corresponding to the target fusion feature is generated based on Lie group operation and manifold learning method; subsequently, the target feature manifold representation is mapped to a plurality of low-dimensional sub-manifolds to generate a target candidate projection set based on each low-dimensional sub-manifold, and a target projection is selected from the target candidate projection set by using a preset loss function; finally, a target dehazed image feature corresponding to the target projection is generated, and a corresponding target dehazed image is obtained by decoding the target dehazed image feature. As can be seen from the above, in the present application, the target image dehazing model is constructed based on neural network, the target foggy image collected in the current foggy environment is input into the target image dehazing model, and the target image dehazing model performs a series of operations on the target foggy image. First, a target foggy light coverage map corresponding to the target foggy image is generated, and a target fusion feature is generated based on the target foggy light coverage map. Then, the target fusion feature is processed by using Lie group operation and manifold learning method to generate a corresponding target feature manifold representation. Then, the low-dimensional sub-manifold corresponding to the target feature manifold representation is determined, and a corresponding target candidate projection set is generated. The target projection in the target candidate projection set is determined by using a preset loss function. Finally, a target dehazed image feature is generated based on the target projection, and a target dehazed image is obtained by decoding the target dehazed image feature. In this way, the image is dehazed by using the target image dehazing model, which can realize the adaptive reconstruction task of dehazed image in multiple scenes and variable scenes, can solve the problem that the traditional dehazing imaging method is only suitable for single light condition and has insufficient adaptive ability, and can improve the accuracy of image dehazing. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description only represent some embodiments of the present application, and all other drawings obtained by those of ordinary skill in the art without creative effort based on the drawings are within the scope of the present application.

[0050] Figure 1 A flow chart of an image defogging method provided by the present application;

[0051] Figure 2 A working flow diagram of a specific target image defogging model provided by the present application;

[0052] Figure 3 A working flow diagram of a specific feature manifold estimation and embedding module provided by the present application;

[0053] Figure 4 A specific image acquisition diagram provided by the present application;

[0054] Figure 5 A specific target image defogging model optimization flow chart provided by the present application;

[0055] Figure 6 A specific optimization constraint condition diagram provided by the present application;

[0056] Figure 7 A structural diagram of an image defogging device provided by the present application;

[0057] Figure 8 A structural diagram of an electronic device provided by the present application. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments only represent some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of the present application.

[0059] In the past few years, dehazing imaging technology has made significant progress in many fields, providing important technical support for solving perception problems in various complex environments. In hazy weather, atmospheric suspended particles have scattering and absorption effects on light, resulting in image blur, reduced contrast, color distortion and other problems, seriously affecting the reliability and safety of practical applications such as autonomous driving, security monitoring, and remote sensing mapping. Traditional dehazing methods are mainly based on the physical model of atmospheric scattering, and image restoration is performed by estimating atmospheric light and transmittance. However, these methods rely on artificially designed prior assumptions, are prone to failure in complex scenes, and are difficult to restore detailed information. In addition, factors such as changes in lighting conditions and interference from background noise will affect the quality and accuracy of imaging, and it is impossible to meet the adaptive requirements of changing and complex scenes. To this end, the present application provides an image dehazing solution that can improve the accuracy of image dehazing in complex and changing scenes.

[0060] See also Figure 1 As shown, an embodiment of the present invention discloses an image defogging method, which may include:

[0061] Step S11: In the process of defogging a target fog image collected in the current foggy environment using a target image defogging model based on a neural network, a target fog illumination coverage map corresponding to the target fog image is generated, and a target fusion feature is generated based on the target fog illumination coverage map.

[0062] In this embodiment, a target image dehazing model is first constructed based on a neural network, including a backbone network, a feature manifold estimation and embedding module (FEM), and a projection selection module. The workflow of the target image dehazing model is shown in Figure 2 As shown in the figure, the backbone network of the target image dehazing model utilizes a deep neural network architecture with an encoder-decoder structure. Typical examples include U-Net (Convolutional Networks for Biomedical Image Segmentation) and its improved variants that incorporate attention mechanisms. This architecture is particularly effective in low-visibility foggy image reconstruction tasks because it can simultaneously capture local details and global contextual information, and achieve efficient feature representation and fusion at multi-scale semantic levels. The backbone network consists of a symmetrical encoder and decoder, with skip connections between them enabling the fusion and transfer of features at different levels.

[0063] It should be noted that the encoder part of the backbone network is divided into a shallow encoder and a deep encoder in structure, which respectively model the information modeling requirements of different levels. The above generating the target fog light coverage map corresponding to the target fog image can include: first, through the encoder of the backbone network of the target image defogging model, and using a preset size of a receptive field convolution kernel, a shared parameter network layer and a residual connection to extract target low-level features of the target fog image; then, based on the target low-level features, the target fog light coverage map corresponding to the target fog image is generated, and based on the target fog light coverage map and a preset binarization method, the target fog image is labeled to generate a target fog mask feature tensor. It should be noted that the shallow encoder extracts global optical features from the input target fog image, mainly including physical properties such as atmospheric scattering components, direct light and ambient light. Specifically, the shallow encoder can realize the perception of low-level features through the network structure with small convolution kernel receptive field and high parameter sharing degree, and meanwhile, the texture continuity in the original target fog image is maintained by means of residual connection. The core output of the shallow encoder is the target fog light coverage map, which represents the distribution density of the fog in each region of the image and the change of the optical thickness in the feature space, and establishes a basis for subsequent modeling of fog removal and reflection compensation. At the same time, the shallow encoder can label the target fog image based on the target fog light coverage map and the binarization method, thereby generating the target fog mask feature tensor.

[0064] In the embodiment, the target fusion feature based on the target fog light coverage atlas can be generated by: firstly, performing stack convolution operation, batch normalization operation and nonlinear activation operation on the target fog light coverage atlas and the target fog mask feature tensor to generate corresponding target high semantic features; then, generating channel weights corresponding to feature channels of the target high semantic features through a channel attention mechanism to obtain corresponding channel attention features; then, determining spatial weights of the channel attention features through a spatial attention mechanism to obtain corresponding spatial attention features; the spatial attention features are features of different scales; finally, performing feature fusion on the spatial attention features based on a preset cross-scale attention mechanism to obtain the target fusion feature corresponding to the target fog image. Specifically, the deep encoder further captures and outputs high semantic features in the target fog image, and the high semantic features cover information such as local structure texture, edge details and key target regions. The deep encoder generates high semantic features through stack convolution operation, normalization operation and nonlinear activation operation. The feature expression capability is enhanced while the dimension is reduced, which can effectively model the nonlinear degradation relationship introduced by non-uniform scattering, occlusion and complex background in the image. The deep encoder has high sensitivity to the microstructure of the image, such as contour, texture and reflective edge, and has the ability to capture the relationship between image details and real scene semantics. Then, on the basis of the encoder, the backbone network introduces a multi-scale attention mechanism, including a channel attention mechanism, a spatial attention mechanism and a cross-scale fusion attention mechanism, to further improve the robustness of feature extraction and fusion. The channel attention mechanism can reshape the channel dimension of the target high semantic feature, the spatial attention mechanism can extract spatial position features, and the cross-scale attention mechanism can fuse the performance features of the target high semantic feature in different channels and spaces to obtain a reshaped unified feature that contains channel features such as target surface scattering characteristics, reflection characteristics and texture characteristics, and spatial information, i.e., the target fusion feature.

[0065] In step S12, the target feature manifold representation corresponding to the target fusion feature is generated based on Lie group operation and manifold learning method.

[0066] In the embodiment, referring to Figure 3As shown, the feature manifold estimation and embedding module of the target image defogging model performs further high-dimensional space mapping and information fusion processing on the target fusion feature obtained by the above encoding, and finally establishes the corresponding feature manifold representation. The feature manifold estimation and embedding module includes an IGAB module (Illumination-Guided Attention Block), which is composed of an LN (Layer Normalization), an MLA (Multi-head Latent Attention), an LN, and an FFN (Feed-Forward Network). The target feature manifold representation corresponding to the target fusion feature generated based on the Lie group operation and manifold learning method can include the following steps: first, the feature manifold estimation and embedding module in the target image defogging model is used to splice the target fusion feature and the target fog and light coverage map to generate a target composite feature, and obtain the foggy scene characteristic information under the current environment; then, the distribution of the target composite feature is estimated to obtain the target distribution estimation information corresponding to the target composite feature; then, the Lie group operation is used, and based on the foggy scene characteristic information and the target distribution estimation information, a target Lie group space is constructed, and based on the manifold learning method, a target manifold structure corresponding to the target distribution estimation information is established in the target Lie group space; finally, the target composite feature and the target distribution estimation information are interacted to obtain a target interaction feature using the preset information interaction term in the feature manifold estimation and embedding module and the target manifold structure, and the target feature manifold representation is generated based on the target interaction feature. Specifically, the foggy scene characteristic information can be directly obtained from network training. The feature manifold estimation and embedding module estimates the distribution of the target composite feature to obtain the target distribution estimation information corresponding to the target composite feature, and then introduces the foggy scene characteristic information to realize information interaction and fusion of different feature sources in the Lie group space and the manifold structure. The information interaction term in the feature manifold estimation and embedding module is used to map and fuse the target composite feature and the target distribution estimation information item by item to obtain the target interaction feature. Finally, an image feature representation with Lie group structure constraint and manifold continuity, i.e., the target feature manifold representation, can be constructed.

[0067] In step S13, the target feature manifold representation is mapped to a plurality of low-dimensional submanifolds to generate a target candidate projection set based on each low-dimensional submanifold, and a target projection is selected from the target candidate projection set using a preset loss function.

[0068] In the embodiment, the projection selection module of the target image defogging model performs a low-dimensional projection operation on the constructed target feature manifold representation, and selects a target projection closest to the target fog image from a plurality of possible low-dimensional submanifolds.

[0069] It should be noted that the above mapping of the target feature manifold representation to a plurality of low-dimensional submanifolds, generating a target candidate projection set based on each low-dimensional submanifold, and screening a target projection from the target candidate projection set using a preset loss function can include: first, mapping the target feature manifold representation to a plurality of low-dimensional submanifolds based on a preset manifold geometric structure condition and a preset feature constraint condition through the projection selection module in the target image defogging model, and generating the target candidate projection set based on each low-dimensional submanifold; each candidate projection in the target candidate projection set is a low-dimensional feature distribution result of the target feature manifold representation; then, determining a first loss corresponding to each candidate projection using a first preset loss function, and determining a second loss corresponding to each candidate projection using a second preset loss function; subsequently, configuring corresponding weights for the first loss and the second loss, and performing weighted summation on the first loss and the second loss based on the weight corresponding to the first loss and the weight corresponding to the second loss to obtain a target loss corresponding to each candidate projection; finally, determining the candidate projection corresponding to the smallest target loss as the target projection; wherein the first preset loss function is used to ensure that adjacent feature points in the target feature manifold representation still maintain a proximity relationship in the low-dimensional submanifold, and the second preset loss function is used to compress redundant projections in the low-dimensional submanifold. Specifically, the projection selection module maps the high-dimensional target feature manifold representation to a low-dimensional space according to the manifold geometric structure rule and the specific constraint condition, obtains a plurality of low-dimensional submanifolds, and generates a target candidate projection set based on each low-dimensional submanifold. The projection selection module is internally provided with a manifold expansion partial loss function and a manifold contraction partial loss function. The manifold expansion partial loss function is used to measure the expansion and adaptability of the projection in the local space, i.e., it can be used to ensure that adjacent feature points in the target feature manifold representation still maintain a proximity relationship in the low-dimensional submanifold; the manifold contraction partial loss function is used to compress invalid projection structures and maintain local consistency of the manifold, i.e., it is used to compress redundant projections in the low-dimensional submanifold. Through the synergistic effect of the two, each candidate projection is pruned and supplemented in terms of local expansion and global consistency to obtain a target loss corresponding to each candidate projection. Further, each candidate projection distribution is evaluated and screened to finally obtain an optimal projection representation closest to the distribution feature of the target fog image, i.e., the target projection.

[0070] Step S14: generating a target defogging image feature corresponding to the target projection, and decoding the target defogging image feature to obtain a corresponding target defogging image.

[0071] In the embodiment, firstly, the probability estimation and the manifold reconstruction are performed for the target projection to obtain the target defogging image features. Then, the corresponding sampling and decoding operations are performed on the target defogging image features by using the decoder module in the backbone network, and the shallow detail features transmitted by the skip connection are combined to restore the spatial resolution, the detail structure and the real color of the multi-scene foggy image, to generate a high-quality restored image, to realize the image clarification and the visible information enhancement in the multi-scene foggy environment. That is, the decoder module can gradually restore the spatial resolution of the image through a plurality of up-sampling layers. Specifically, firstly, the decoder of the backbone network in the target image defogging model is used to process the target defogging image features by using the current up-sampling layer to obtain the current up-sampled features; then, the target low-level features corresponding to the target fog image generated by the encoder of the backbone network in the target image defogging model are obtained based on the skip connection, and the current up-sampled features and the target low-level features are spliced to obtain the current spliced features; then, the next up-sampling layer is determined as the current up-sampling layer, and the step of processing the target defogging image features by using the current up-sampling layer is jumped to until the spatial size corresponding to the current spliced features is the same as the preset spatial resolution corresponding to the target fog image, to obtain the target spliced features; finally, the feature channels corresponding to the target spliced features are compressed based on the preset channel number condition, and the compressed target spliced features are subjected to a nonlinear activation operation to obtain the target defogging image.

[0072] As can be seen from the above, in the process of using the target image defogging model based on the neural network to defog the target fog image collected in the current foggy environment, first, the target fog light coverage map corresponding to the target fog image is generated, and the target fusion feature is generated based on the target fog light coverage map; then, the target feature manifold representation corresponding to the target fusion feature is generated based on the Lie group operation and the manifold learning method; then, the target feature manifold representation is mapped to a plurality of low-dimensional sub-manifolds, so as to generate a target candidate projection set based on each low-dimensional sub-manifold, and a target projection is screened from the target candidate projection set by using a preset loss function; finally, the target defogging image feature corresponding to the target projection is generated, and the corresponding target defogging image is obtained by decoding the target defogging image feature. As can be seen from the above, in the embodiment, the target image defogging model is constructed based on the neural network, the target fog image collected in the current foggy environment is input into the target image defogging model, and the target image defogging model performs a series of operations on the target fog image. First, the target fog light coverage map corresponding to the target fog image is generated, and the target fusion feature is generated based on the target fog light coverage map. Then, the target fusion feature is processed by using the Lie group operation and the manifold learning method to generate the corresponding target feature manifold representation. Then, the low-dimensional sub-manifold corresponding to the target feature manifold representation is determined, and the corresponding target candidate projection set is generated. The target projection in the target candidate projection set is determined by using the preset loss function. Finally, the target defogging image feature is generated based on the target projection, and the target defogging image is obtained by decoding the target defogging image feature. In this way, the target image defogging model is used to defog the image, which can realize the defogging image adaptive reconstruction task in multiple scenes and variable scenes, can solve the problem that the traditional defogging imaging method is only suitable for single light condition and has insufficient adaptive ability, and can improve the accuracy of image defogging.

[0073] Based on the above embodiment, the target image defogging model can be used to defog the fog image. Next, in order to improve the accuracy and stability of image defogging, the optimization process of the target image defogging model will be described in detail.

[0074] In the embodiment, the optimization process of the target image defogging model can specifically include: first, obtaining a haze-free image and a first haze image, and inputting the haze-free image and the first haze image into the target image defogging model respectively; then, generating a first feature manifold representation corresponding to the haze-free image by using the target image defogging model, and generating a second feature manifold representation corresponding to the first haze image by using the target image defogging model; subsequently, determining a loss value between the first feature manifold representation and the second feature manifold representation according to a third preset loss function, and determining a target gradient based on the loss value; finally, optimizing network parameters in the target image defogging model based on a preset high-dimensional tendency score method and the target gradient, so as to defog a second haze image by using the target image defogging model after optimization. Specifically, referring to FIG. 2, Figure 4 As shown in FIG. 2, the haze-free image and the first haze image are obtained by: arranging three different light environments in a laboratory environment, namely “bright light”, “dim light” and “local light source”, and placing a small fogging machine at the side wall of each scene module for generating haze in real time under the light condition; installing a camera on the opposite wall surface, which is configured in pairs with the fogging machine, to ensure that the camera has the same acquisition frequency for taking images after adding haze and for taking images after removing the fogging machine. The camera collects the scene to obtain the haze-free image and the first haze image. It should be noted that the optimization process of the target image defogging model mainly depends on two key loss functions of the feature manifold estimation and embedding module: one is a reconstruction part loss function, which is used to redefine the structural features and semantic attributes of the target haze image in the Lie group and manifold space; the other is a preliminary cognition part loss function, which mainly acts on the feature fusion process between the haze-free image and the haze image to enhance the adaptability to the scene diversity and the blur distribution characteristics. The network parameters are optimized through two paths: one is to input the haze-free image into the target image defogging model to obtain the first feature manifold representation, and the other is to input the first haze image into the target image defogging model to obtain the second feature manifold representation. The first feature manifold representation and the second feature manifold representation are calculated by the loss function to calculate the loss and gradient backpropagation, and the network parameters in the target image defogging model are optimized to redefine the structural features and semantic attributes of the haze image and enhance the feature fusion process.

[0075] In a specific embodiment, referring to FIG. 3, Figure 5 As shown in FIG. 3, the first feature manifold representation corresponding to the haze-free image, i.e., x manifold, is generated, and the second feature manifold representation corresponding to the first haze image, i.e., z manifold, is generated. Then, in order to ensure the stability and consistency of the feature distribution in the projection transformation process, the training and constraint phase of the operator f is needed, which is further divided into four parts.

[0076] The first part mainly focuses on the optimization strategy in the HDPS method (High-Dimensional Propensity Score Approach). The core goal is to train a mapping operator that can generate a data distribution as close as possible to the expected output given the input. Set an ideal scenario: when the input data itself is not distorted, i.e. , the mapping problem becomes simple and clear. In this case, the optimal mapping operator should satisfy . Further derivation shows that when the input is , there should also be , thus achieving the ideal effect of information reconstruction. The Lie group and manifold space established in this process actually correspond to the internal relationship between the original distribution and the distorted distribution . To quantitatively characterize the difference between the input processed by the operator and the ideal output, introduce the distance metric to evaluate the degree of deviation between the two. Specifically, the drift error can be expressed as:

[0077] ;

[0078] is the parameter to be trained in the operator . In order to obtain the most perfect solution in the special solution case, the optimization method needs to continuously optimize and minimize this drift, and the mapping operator will gradually converge to a state that can accurately restore the ideal distribution.

[0079] The second part discusses the information modeling challenges in diverse foggy environments. Due to the influence of light attenuation, scattering noise, and significant distortion in complex weather conditions, the images collected by the camera system are often affected, resulting in a large deviation between the original distribution and the disturbed distribution . In the context of limited data acquisition or limited computing resources, if we try to directly establish a mapping between and through a single constraint mechanism, it is often difficult to achieve ideal results. Therefore, the goal of this stage is to train a parameterized operator to achieve the output as close as possible to after mapping the distorted input . This process actually completes the key information integration, i.e., establishing a stable data correlation relationship within the feature manifold estimation and embedding module. In the ideal state, this process requires minimizing and the distribution difference between them while strictly following the preset optimization constraints. Accordingly, the second optimization constraint condition is derived in this embodiment:

[0080] ;

[0081] The third and fourth parts mainly focus on further optimizing the output characteristics to achieve a more accurate approximation of the target. In the third part, to ensure that the processed after the operator can fully reflect the similar properties of the original target , based on the aforementioned characteristic analysis of , the following strategy is proposed: when the input is not degraded, i.e., the original image state, the operator should apply the least change to the data. Therefore, when has met the first two optimization constraints, it can be reasonably inferred that is relatively close to . Based on this premise, further requires after being constructed by Lie group and manifold learning, as a low-dimensional projection output, to have the ability to select the optimal projection result, thus deriving the third optimization constraint:

[0082] ;

[0083] where is the frozen parameter, and no gradient backpropagation is performed; is the function identifier.

[0084] However, the introduction of the third part also brings new problems: two paths involving the parameter in the objective function appear, corresponding to the upper and lower processing paths respectively, each forming a different gradient update direction. Therefore, these two paths must be distinguished. To this end, in the preliminary optimization stage, only the upper path is optimized, and the lower path parameters are frozen, aiming to gradually approach the output to the target manifold characteristics through training, which defines the optimization objective of the third part:

[0085] ;

[0086] In the fourth part, to further converge the output to a smaller target neighborhood, an additional contraction mechanism needs to be introduced. That is, only when very close to the target manifold, the is accepted. To achieve this purpose, under the condition of freezing the upper path, the is regarded as the current fixed output of the operator , and the is maximized distance between the current and the target to push the operator in the subsequent update to generate a result closer to the target . Thus, referring to Figure 6 , the optimization objective of the fourth part can be expressed as:

[0087] ;

[0088] wherein, is a model parameter, is a function identifier.

[0089] In this way, the embodiment can realize high-quality image reconstruction under the support of a small-scale data set, significantly reducing the dependence on a large amount of prior data; at the same time, it supports direct application in miniaturization deployment without additional research and development costs, ensuring the accuracy and stability of image defogging in miniaturization application. The embodiment is suitable for adaptive image restoration in complex and variable foggy environment conditions, has good scene migration and environmental adaptability, and can maintain stable image defogging effect under conditions such as illumination change, background noise interference, and different types of fog droplets.

[0090] Correspondingly, referring to Figure 7 , the embodiment of the application also provides an image defogging device, which can include:

[0091] The target fusion feature generation module 11 is configured to generate a target fog and light coverage map corresponding to the target fog image in the process of defogging the target fog image collected in the current foggy environment by using the target image defogging model based on the neural network, and generate a target fusion feature based on the target fog and light coverage map.

[0092] The target feature manifold representation generation module 12 is configured to generate a target feature manifold representation corresponding to the target fusion feature based on Lie group operation and manifold learning method.

[0093] The target projection determination module 13 is configured to map the target feature manifold representation to a plurality of low-dimensional submanifolds, generate a target candidate projection set based on each low-dimensional submanifold, and screen a target projection from the target candidate projection set by using a preset loss function.

[0094] The target defogging image determination module 14 is configured to generate a target defogging image feature corresponding to the target projection, and decode the target defogging image feature to obtain a corresponding target defogging image.

[0095] As can be seen from the above, in the process of defogging the target fog image collected in the current foggy environment using the neural network-based target image defogging model in this application, the target fog illumination coverage map corresponding to the target fog image is first generated, and the target fusion feature is generated based on the target fog illumination coverage map; then, the target feature manifold representation corresponding to the target fusion feature is generated based on Lie group operation and manifold learning method; then, the target feature manifold representation is mapped to several low-dimensional sub-manifolds to generate a target candidate projection set based on each of the low-dimensional sub-manifolds, and the target projection is screened out from the target candidate projection set using a preset loss function; finally, the target defogging image feature corresponding to the target projection is generated, and the target defogging image feature is decoded to obtain the corresponding target defogging image. As can be seen from the above, in this application, a target image defogging model is constructed based on a neural network, and the target fog image collected in the current foggy environment is input into the target image defogging model. The target image defogging model performs a series of operations on the target fog image. First, a target fog illumination coverage map corresponding to the target fog image is generated, and a target fusion feature is generated based on the target fog illumination coverage map. Then, the target fusion feature is processed using Lie group operations and manifold learning methods to generate a corresponding target feature manifold representation. Then, the low-dimensional submanifold corresponding to the target feature manifold representation is determined, and a corresponding target candidate projection set is generated. The target projection in the target candidate projection set is determined using a preset loss function. Finally, a target defogging image feature is generated based on the target projection, and the target defogging image feature is decoded to obtain a target defogging image. In this way, the present application uses the target image defogging model to defog the image, which can realize the adaptive reconstruction task of defogging images in multiple scenes and variable scenes, and can solve the problem that the traditional defogging imaging method is only adapted to a single lighting condition and has insufficient adaptive ability, and improves the accuracy of image defogging.

[0096] In some specific implementations, the target fusion feature generation module 11 may include:

[0097] A target low-level feature extraction unit is used to extract target low-level features of the target fog image through the encoder of the backbone network in the target image defogging model and using a receptive field convolution kernel of a preset size, a shared parameter network layer and a residual connection;

[0098] A target fog illumination coverage map generation unit is used to generate the target fog illumination coverage map corresponding to the target fog image based on the target low-level features, and mark the target fog image based on the target fog illumination coverage map and a preset binarization method to generate a target fog mask feature tensor.

[0099] In some specific implementations, the target fusion feature generation module 11 may include:

[0100] a target high semantic feature generation unit configured to perform a stack convolution operation, a batch normalization operation and a nonlinear activation operation on the target fog light coverage map and the target fog mask feature tensor to generate a corresponding target high semantic feature;

[0101] a channel attention feature determination unit configured to generate channel weights corresponding to feature channels of the target high semantic feature through a channel attention mechanism to obtain a corresponding channel attention feature;

[0102] a spatial attention feature determination unit configured to determine spatial weights of the channel attention feature through a spatial attention mechanism to obtain a corresponding spatial attention feature; the spatial attention feature is a feature of several different scales;

[0103] a target fusion feature generation unit configured to perform feature fusion on the spatial attention feature based on a preset cross-scale attention mechanism to obtain the target fusion feature corresponding to the target fog image.

[0104] In some embodiments, the target feature manifold representation generation module 12 can include:

[0105] a target composite feature generation unit configured to perform splicing on the target fusion feature and the target fog light coverage map through a feature manifold estimation and embedding module in the target image defogging model to generate a target composite feature and obtain foggy scene characteristic information in a current environment;

[0106] a target distribution estimation information determination unit configured to perform distribution estimation on the target composite feature to obtain target distribution estimation information corresponding to the target composite feature;

[0107] a target manifold structure establishment unit configured to utilize the Lie group operation, construct a target Lie group space based on the foggy scene characteristic information and the target distribution estimation information, and establish a target manifold structure corresponding to the target distribution estimation information in the target Lie group space based on the manifold learning method;

[0108] a target feature manifold representation generation unit configured to interact the target composite feature and the target distribution estimation information through a preset information interaction term in the feature manifold estimation and embedding module and the target manifold structure to obtain a target interaction feature, and generate the target feature manifold representation based on the target interaction feature.

[0109] In some embodiments, the target projection determination module 13 can include:

[0110] The target candidate projection set generation unit is configured to map the target feature manifold representation to a plurality of low-dimensional sub-manifolds based on preset manifold geometric structure conditions and preset feature constraint conditions through a projection selection module in the target image defogging model, and generate the target candidate projection set based on each low-dimensional sub-manifold; each candidate projection in the target candidate projection set is a low-dimensional feature distribution result of the target feature manifold representation;

[0111] The second loss determination unit is configured to determine a first loss corresponding to each candidate projection by using a first preset loss function, and determine a second loss corresponding to each candidate projection by using a second preset loss function;

[0112] The target loss determination unit is configured to configure respective weights for the first loss and the second loss respectively, and perform weighted summation on the first loss and the second loss based on the weight corresponding to the first loss and the weight corresponding to the second loss, to obtain a target loss corresponding to each candidate projection;

[0113] The target projection determination unit is configured to determine a candidate projection corresponding to the smallest target loss as the target projection; the first preset loss function is configured to ensure that adjacent feature points in the target feature manifold representation still maintain a proximity relationship in the low-dimensional sub-manifold, and the second preset loss function is configured to compress redundant projections in the low-dimensional sub-manifold.

[0114] In some embodiments, the target defogging image determination module 14 can include:

[0115] The up-sampling feature determination unit is configured to process the target defogging image feature by using a current up-sampling layer through a decoder of the backbone network in the target image defogging model, to obtain a current up-sampling feature;

[0116] The feature concatenation unit is configured to obtain a target low-level feature corresponding to the target fog image generated by an encoder of the backbone network in the target image defogging model based on a skip connection, and concatenate the current up-sampling feature and the target low-level feature to obtain a current concatenation feature;

[0117] The target concatenation feature determination unit is configured to determine a next up-sampling layer as the current up-sampling layer, and jump to the step of processing the target defogging image feature by using the current up-sampling layer, until a spatial size corresponding to the current concatenation feature is the same as a preset spatial resolution corresponding to the target fog image, to obtain a target concatenation feature;

[0118] The target defogging image determination unit is configured to compress feature channels corresponding to the target splicing feature based on a preset channel number condition, and perform a nonlinear activation operation on the compressed target splicing feature to obtain the target defogging image.

[0119] In some specific embodiments, the image defogging device can further include:

[0120] The image acquisition module is configured to acquire a haze-free image and a first haze image, and input the haze-free image and the first haze image into the target image defogging model, respectively.

[0121] The second feature manifold representation generation module is configured to generate a first feature manifold representation corresponding to the haze-free image by using the target image defogging model, and generate a second feature manifold representation corresponding to the first haze image by using the target image defogging model.

[0122] The target gradient determination module is configured to determine a loss value between the first feature manifold representation and the second feature manifold representation according to a third preset loss function, and determine a target gradient based on the loss value.

[0123] The target image defogging model optimization module is configured to optimize network parameters in the target image defogging model based on a preset high-dimensional tendency score method and the target gradient, so as to defog a second haze image by using the optimized target image defogging model.

[0124] Further, the embodiment of the present application further discloses an electronic device, Figure 8 is an electronic device 20 structure diagram shown according to an exemplary embodiment, the contents in the figure cannot be considered as any limitation on the use range of the present application. The electronic device 20, specifically can include: at least one processor 21, at least one memory 22, power supply 23, communication interface 24, input output interface 25 and communication bus 26. Wherein, the memory 22 is used to store computer program, the computer program is loaded and executed by the processor 21, to realize the related steps in the image defogging method disclosed in any preceding embodiment. In addition, the electronic device 20 in the embodiment specifically can be electronic computer.

[0125] In the embodiment, the power supply 23 is used to provide working voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and the external device, and the communication protocol followed by the communication interface 24 can be any communication protocol applicable to the technical solution of the present application, which is not limited specifically herein; the input output interface 25 is used to acquire external input data or output data to the outside world, and the specific interface type can be selected according to the specific application needs, which is not limited specifically herein.

[0126] In addition, the memory 22 can be a read-only memory, a random access memory, a magnetic disk or an optical disk, etc., as a carrier for storing resources, and the resources stored thereon can include an operating system 221, a computer program 222, etc., and the storage mode can be temporary storage or permanent storage.

[0127] The operating system 221 is used to manage and control each hardware device on the electronic device 20 and the computer program 222, and can be Windows Server, Netware, Unix, Linux, etc. In addition to the computer program capable of completing the image defogging method executed by the electronic device 20 disclosed in any of the foregoing embodiments, the computer program 222 can further include a computer program capable of completing other specific work.

[0128] Further, the present application also discloses a computer readable storage medium for storing a computer program; wherein the computer program is executed by a processor to implement the image defogging method disclosed above. For the specific steps of the method, refer to the corresponding content disclosed in the foregoing embodiments, which will not be described here.

[0129] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. For the same or similar parts between the embodiments, refer to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and refer to the method part for the relevant part.

[0130] The skilled person can further realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware, computer software or a combination of the two. In order to clearly show the interchangeability of hardware and software, the components and steps of each example have been described in the above description. Whether the functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0131] The steps of the method or algorithm described in combination with the embodiments disclosed herein can be directly implemented by hardware, a software module executed by a processor, or a combination of the two. The software module can be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0132] Finally, it needs to be pointed out that in this document, relational terms such as first and second and the like can only be intended to distinguish one entity or operation from another entity or operation without necessarily requiring or implying any actual such relationship or order between such entities or operations. Moreover, the terms "comprising", "including", or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without more limitations, an element defined by the statement "comprising a" does not exclude the existence of additional identical elements in the process, method, article, or apparatus including the stated element.

[0133] The above detailed description of the technical solutions provided by the present application has been provided, and the principles and implementation manners of the present application have been described by applying specific examples. The above description of the examples is only for the purpose of helping to understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation manner and application range will be changed, and the above description of the content of the specification should not be understood as a limitation of the present application.

Claims

1. An image defogging method, characterized in that: include: In the process of defogging a target fog image collected in the current foggy environment using a target image defogging model based on a neural network, a target fog illumination coverage map corresponding to the target fog image is generated, and a target fusion feature is generated based on the target fog illumination coverage map; Generate a target feature manifold representation corresponding to the target fusion feature based on Lie group operations and manifold learning methods; Mapping the target feature manifold representation to a plurality of low-dimensional submanifolds to generate a target candidate projection set based on each of the low-dimensional submanifolds, and screening the target projection from the target candidate projection set using a preset loss function; Generating target defogging image features corresponding to the target projection, and decoding the target defogging image features to obtain a corresponding target defogging image; The step of generating a target feature manifold representation corresponding to the target fusion feature based on Lie group operations and a manifold learning method includes: The target fusion feature and the target fog illumination coverage map are spliced ​​together to generate target composite features through the feature manifold estimation and embedding module in the target image defogging model, and foggy scene characteristic information in the current environment is obtained; Performing distribution estimation on the target composite feature to obtain target distribution estimation information corresponding to the target composite feature; Using the Lie group operation, constructing a target Lie group space based on the foggy scene characteristic information and the target distribution estimation information, and establishing a target manifold structure corresponding to the target distribution estimation information in the target Lie group space based on the manifold learning method; Utilizing the preset information interaction items in the feature manifold estimation and embedding module and the target manifold structure, the target composite features and the target distribution estimation information are interacted to obtain target interaction features, and the target feature manifold representation is generated based on the target interaction features.

2. The image defogging method according to claim 1, characterized in that: Generating a target fog illumination coverage map corresponding to the target fog image includes: Extracting target low-level features of the target fog image through the encoder of the backbone network in the target image defogging model, and using a preset receptive field convolution kernel, a shared parameter network layer, and a residual connection; The target fog illumination coverage map corresponding to the target fog image is generated based on the target low-level features, and the target fog image is marked based on the target fog illumination coverage map and a preset binarization method to generate a target fog mask feature tensor.

3. The image defogging method according to claim 2, characterized in that: Generating target fusion features based on the target fog illumination coverage map includes: Performing stacked convolution operations, batch normalization operations, and nonlinear activation operations on the target fog illumination coverage map and the target fog mask feature tensor to generate corresponding target high-semantic features; Generate the channel weight corresponding to the feature channel of the target high semantic feature through the channel attention mechanism to obtain the corresponding channel attention feature; Determine the spatial weight of the channel attention feature through a spatial attention mechanism to obtain a corresponding spatial attention feature; the spatial attention feature is a feature of several different scales; The spatial attention features are fused based on a preset cross-scale attention mechanism to obtain the target fusion features corresponding to the target fog image.

4. The image defogging method according to claim 1, wherein: Mapping the target feature manifold representation to a plurality of low-dimensional submanifolds to generate a target candidate projection set based on each of the low-dimensional submanifolds, and screening the target projection from the target candidate projection set using a preset loss function, includes: The target feature manifold representation is mapped to a plurality of low-dimensional submanifolds based on a preset manifold geometry condition and a preset feature constraint condition through a projection selection module in the target image dehazing model, and a target candidate projection set is generated based on each of the low-dimensional submanifolds; each candidate projection in the target candidate projection set is a low-dimensional feature distribution result of the target feature manifold representation; Determine a first loss corresponding to each candidate projection using a first preset loss function, and determine a second loss corresponding to each candidate projection using a second preset loss function; Configuring corresponding weights for the first loss and the second loss, respectively, and performing a weighted summation of the first loss and the second loss based on the weight corresponding to the first loss and the weight corresponding to the second loss, so as to obtain a target loss corresponding to each candidate projection; Determine the candidate projection corresponding to the minimum target loss as the target projection; Among them, the first preset loss function is used to ensure that adjacent feature points in the target feature manifold representation still maintain a proximity relationship in the low-dimensional sub-manifold, and the second preset loss function is used to compress redundant projections in the low-dimensional sub-manifold.

5. The image defogging method according to claim 1, characterized in that: The decoding of the target defogging image features to obtain a corresponding target defogging image includes: Processing the target dehazed image features through the decoder of the backbone network in the target image dehazing model and using the current upsampling layer to obtain the current upsampling features; Obtaining target low-level features corresponding to the target fog image generated by the encoder of the backbone network in the target image defogging model based on a skip connection, and splicing the current up-sampled features and the target low-level features to obtain current spliced ​​features; Determining the next upsampling layer as the current upsampling layer, and jumping to the step of processing the target defogging image feature using the current upsampling layer until the spatial size corresponding to the current stitching feature is the same as the preset spatial resolution corresponding to the target fog image, thereby obtaining the target stitching feature; The feature channels corresponding to the target splicing features are compressed based on a preset channel quantity condition, and a nonlinear activation operation is performed on the compressed target splicing features to obtain the target defogging image.

6. The image defogging method according to any one of claims 1 to 5, characterized in that: Also includes: Acquire a fog-free image and a first fog image, and input the fog-free image and the first fog image into the target image defogging model respectively; Generating a first characteristic manifold representation corresponding to the fog-free image using the target image defogging model, and generating a second characteristic manifold representation corresponding to the first foggy image using the target image defogging model; Determining a loss value between the first feature manifold representation and the second feature manifold representation according to a third preset loss function, and determining a target gradient based on the loss value; The network parameters in the target image defogging model are optimized based on a preset high-dimensional propensity score method and the target gradient, so as to defog the second foggy image using the optimized target image defogging model.

7. An image defogging device, characterized in that: include: a target fusion feature generation module for generating a target fog illumination coverage map corresponding to the target fog image during defogging of the target fog image collected in the current foggy environment using the neural network-based target image defogging model, and generating target fusion features based on the target fog illumination coverage map; A target feature manifold representation generation module is used to generate a target feature manifold representation corresponding to the target fusion feature based on Lie group operations and manifold learning methods; a target projection determination module, configured to map the target feature manifold representation to a plurality of low-dimensional submanifolds, generate a set of target candidate projections based on each of the low-dimensional submanifolds, and select a target projection from the set of target candidate projections using a preset loss function; a target defogging image determination module, configured to generate target defogging image features corresponding to the target projection, and decode the target defogging image features to obtain a corresponding target defogging image; The target feature manifold representation generation module includes: a target composite feature generation unit, configured to generate target composite features by splicing the target fusion features and the target fog illumination coverage map through the feature manifold estimation and embedding module in the target image defogging model, and obtain characteristic information of the foggy scene in the current environment; a target distribution estimation information determining unit, configured to perform distribution estimation on the target composite feature to obtain target distribution estimation information corresponding to the target composite feature; a target manifold structure establishing unit, configured to construct a target Lie group space based on the foggy scene characteristic information and the target distribution estimation information using the Lie group operation, and to establish a target manifold structure corresponding to the target distribution estimation information in the target Lie group space based on the manifold learning method; A target feature manifold representation generation unit is used to utilize the preset information interaction items in the feature manifold estimation and embedding module and the target manifold structure to interact the target composite features and the target distribution estimation information to obtain target interaction features, and generate the target feature manifold representation based on the target interaction features.

8. An electronic device, characterized in that: The electronic device includes a processor and a memory; wherein the memory is used to store a computer program, and the computer program is loaded and executed by the processor to implement the image defogging method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that Used to store a computer program, which, when executed by a processor, implements the image defogging method according to any one of claims 1 to 6.

Images