Haze image enhancement method and device

Abstract

The invention discloses a haze image enhancement method and device. The method comprises the following steps: constructing a haze image enhancement network; wherein the haze image enhancement network comprises a defogging sub-network and a chroma enhancement sub-network; training the haze image enhancement network according to a preset joint loss function to obtain a haze image reconstruction model; and performing defogging enhancement processing on a to-be-enhanced image through the haze image reconstruction model to obtain a reconstructed image. According to the invention, the visual perception effect of the enhanced haze image can be improved.

Description

Technical Field

[0001] The present invention relates to the field of image enhancement, and in particular to a haze image enhancement method and device. Background Art

[0002] For humans, vision is one of the main channels for obtaining information and an important basis for people to perceive the world. The core carrier of visual information is the image, which is an indispensable medium in people's daily life. Digital image enhancement technology aims to improve image quality and image clarity to meet people's needs for obtaining information through visual perception. From a physical perspective, haze weather is static fog-like weather. In such scenarios, the transmission medium in the air usually has a small scale, so that a single medium particle cannot be captured and recorded by the imaging system. Unlike dynamic environments such as rain and snow, static fog-like weather usually has low variability, and suspended particles in the atmosphere generally exhibit isotropic characteristics when imaging.

[0003] Images acquired by imaging equipment in haze weather usually show low contrast, dull colors, and blurred details. Due to the scattering of light by particles suspended in the air, the overall image appears grayish or bluish, and distant objects appear hazy. In addition, uneven brightness leads to a weakened sense of image hierarchy and blurred edges, making target recognition and information extraction difficult, resulting in a significant reduction in the accuracy of visual information acquisition and processing. This problem is particularly prominent in the fields of traffic monitoring, autonomous driving, and remote sensing imaging. Therefore, how to effectively eliminate the visual degradation caused by haze to improve the stability and practicality of outdoor computer vision systems has become an urgent problem to be solved in the field of image visual enhancement. Summary of the invention

[0004] In order to overcome the defects of the prior art, the present invention provides a haze image enhancement method and device, which can improve the visual perception effect of the haze image after image enhancement.

[0005] An embodiment of the present invention provides a haze image enhancement method, comprising the following steps:

[0006] Constructing a haze image enhancement network; wherein the haze image enhancement network includes a dehazing sub-network and a chroma enhancement sub-network;

[0007] The haze image enhancement network is trained according to a preset joint loss function to obtain a haze image reconstruction model;

[0008] The image to be enhanced is input into the haze image reconstruction model, and the image to be enhanced is dehazed and enhanced by the dehazing sub-network and the chroma enhancement sub-network to obtain a reconstructed image.

[0009] Furthermore, the haze image enhancement network includes a disassembly module, a dehazing sub-network, a chroma enhancement sub-network and a splicing module;

[0010] The decomposition module is connected to the input ends of the dehazing sub-network and the chroma enhancement sub-network respectively, and is used to decompose the input image into a luminance component and a chroma component;

[0011] The defogging sub-network includes a sampling module, a convolution correction module, the feature extraction module, a pixel attention module and a fusion module, which are used to perform defogging on the brightness component to obtain a fog-free brightness component; wherein the feature extraction module is constructed by a third-order Runge-Kutta method;

[0012] The chroma enhancement subnetwork includes a convolution correction module, a feature extraction module and a channel attention module, which are used to perform equalization processing on the chroma component to obtain a chroma equalization feature;

[0013] The splicing module is connected to the output ends of the defogging sub-network and the chroma enhancement sub-network respectively, and is used to perform splicing processing on the haze-free brightness component and the chroma balance feature, and map the splicing result into an RGB image.

[0014] Preferably, the function expression of the feature extraction module is specifically:

[0015]

[0016] Among them, y k Indicates the input state, y k+1 represents the output state, G1, G2, and G3 are all trainable adaptive nonlinear representation units, h represents the step size between adjacent independent variables, and f(x k ,y k ) represents the point (x k ,y k )’s instantaneous slope.

[0017] Furthermore, the haze image enhancement network is trained according to a preset joint loss function to obtain a haze image reconstruction model, which specifically includes:

[0018] According to a preset training set, the haze image enhancement network is iteratively trained by a preset Adam optimizer; wherein the preset training set includes a plurality of haze images and clear images corresponding to the plurality of haze images, and the loss function in the preset Adam optimizer is specifically the preset joint loss function;

[0019] When the number of iterations reaches a preset iteration threshold, the haze image reconstruction model is outputted.

[0020] Preferably, the preset joint loss function is specifically:

[0021] L T =λ1L1+λ2L c

[0022]

[0023] Among them, L T is the total loss value, λ1 and λ2 are L1 and L c The corresponding preset weight coefficient, N is the total number of training samples in the training set, I is the haze image, J is the clear image corresponding to the haze image, J′ is the output image corresponding to the haze image, and J i is the clear image corresponding to the i-th haze image in the training set, J i ′ is the output image corresponding to the i-th haze image in the training set, Φ is the preset pre-training model, M is the number of hidden layers in the preset pre-training model, α i To preset the weight coefficient, D[x,y] represents the calculation of the Euclidean distance between x and y in the high-dimensional space.

[0024] Furthermore, the image to be enhanced is input into the haze image reconstruction model, and the image to be enhanced is dehazed and enhanced by the dehazing sub-network and the chroma enhancement sub-network to obtain a reconstructed image, specifically including:

[0025] Mapping the image to be enhanced from an RGB image to a YUV image through the decomposition module, and decomposing the mapped YUV image to obtain corresponding brightness components and chrominance components;

[0026] Inputting the brightness component into the defogging sub-network for defogging processing to obtain the fog-free brightness component;

[0027] Inputting the chroma component into the chroma enhancement subnetwork for enhancement processing to obtain the chroma balance feature;

[0028] The fog-free brightness component and the chromaticity balance feature are spliced ​​by the splicing module, and the splicing result is mapped from the YUV space to the RGB space to obtain the reconstructed image.

[0029] Furthermore, the step of inputting the brightness component into the defogging sub-network for defogging to obtain the fog-free brightness component specifically includes:

[0030] The luminance component is sequentially upsampled and downsampled by the sampling module to obtain an upsampled component and a downsampled component respectively;

[0031] The brightness component, the up-sampling component and the down-sampling component are processed simultaneously and in sequence by the convolution correction module, the feature extraction module and the pixel attention module, and a first brightness processing result, an up-sampling processing result and a down-sampling processing result are obtained correspondingly;

[0032] Performing down-sampling processing on the up-sampling processing result to obtain a second brightness processing result, and performing up-sampling processing on the down-sampling processing result to obtain a third brightness processing result;

[0033] The first brightness processing result, the second brightness processing result and the third brightness processing result are linearly fused by the fusion module, and the fusion result is corrected by the convolution correction module to obtain a fog-free brightness component.

[0034] Furthermore, the step of inputting the chroma component into the chroma enhancement subnetwork for enhancement processing to obtain the chroma equalization feature specifically includes:

[0035] Processing the chrominance component sequentially through the convolution correction module, the feature extraction module and the channel attention module to obtain an attention processing result;

[0036] The attention processing result is corrected by the convolution correction module to obtain the chromaticity balance feature.

[0037] Another embodiment of the present invention provides a haze image enhancement device, including: a construction module, a training module and a reconstruction module;

[0038] The construction module is used to construct a haze image enhancement network; wherein the haze image enhancement network includes a dehazing sub-network and a chroma enhancement sub-network;

[0039] The training module is used to train the haze image enhancement network according to a preset joint loss function to obtain a haze image reconstruction model;

[0040] The reconstruction module is used to input the image to be enhanced into the haze image reconstruction model, and perform dehazing and enhancement processing on the image to be enhanced through the dehazing sub-network and the chroma enhancement sub-network to obtain a reconstructed image.

[0041] Furthermore, the training module is used to train the haze image enhancement network according to a preset joint loss function to obtain a haze image reconstruction model, specifically including:

[0042] According to a preset training set, the haze image enhancement network is iteratively trained by a preset Adam optimizer; wherein the preset training set includes a plurality of haze images and clear images corresponding to the plurality of haze images, and the loss function in the preset Adam optimizer is specifically the preset joint loss function;

[0043] When the number of iterations reaches a preset iteration threshold, the haze image reconstruction model is outputted.

[0044] Compared with the prior art, the beneficial effects of the present invention are:

[0045] By processing the haze image simultaneously through the dehazing sub-network and the chroma enhancement sub-network, it is possible to eliminate the blurring interference caused by the haze while performing adaptive equalization on the chroma information in the image, thereby improving the visual perception effect of the haze image after image enhancement. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A schematic flow chart of a haze image enhancement method provided in one embodiment of the present invention.

[0047] Figure 2 A schematic diagram of a haze image enhancement network provided by an embodiment of the present invention

[0048] Figure 3 A schematic diagram of the structure of a feature extraction module constructed by a third-order Runge-Kutta method provided in one embodiment of the present invention.

[0049] Figure 4 A schematic diagram of the principle of a contrastive learning loss function provided by an embodiment of the present invention.

[0050] Figure 5 A schematic diagram showing comparison of processing results of different image enhancement methods on the RESIDE SOTS test set provided by an embodiment of the present invention.

[0051] Figure 6 A schematic structural diagram of a haze image enhancement device provided in another embodiment of the present invention. DETAILED DESCRIPTION

[0052] The drawings are for illustrative purposes only and should not be construed as limiting the present patent;

[0053] It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0054] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0055] Reference Figure 1, is a flow chart of a haze image enhancement method provided by an embodiment of the present invention, comprising the following steps:

[0056] S1: constructing a haze image enhancement network; wherein the haze image enhancement network includes a dehazing sub-network and a chroma enhancement sub-network;

[0057] S2: training the haze image enhancement network according to a preset joint loss function to obtain a haze image reconstruction model;

[0058] S3: Input the image to be enhanced into the haze image reconstruction model, and perform dehazing and enhancement processing on the image to be enhanced through the dehazing sub-network and the chroma enhancement sub-network to obtain a reconstructed image.

[0059] For step S1, specifically, the haze image enhancement network includes a disassembly module, a dehazing sub-network, a chroma enhancement sub-network and a splicing module;

[0060] The decomposition module is connected to the input ends of the dehazing sub-network and the chroma enhancement sub-network respectively, and is used to decompose the input image into a luminance component and a chroma component;

[0061] The defogging sub-network includes a sampling module, a convolution correction module, the feature extraction module, a pixel attention module and a fusion module, which are used to perform defogging on the brightness component to obtain a fog-free brightness component; wherein the feature extraction module is constructed by a third-order Runge-Kutta method;

[0062] The chroma enhancement subnetwork includes a convolution correction module, a feature extraction module and a channel attention module, which are used to perform equalization processing on the chroma component to obtain a chroma equalization feature;

[0063] The splicing module is connected to the output ends of the defogging sub-network and the chroma enhancement sub-network respectively, and is used to perform splicing processing on the haze-free brightness component and the chroma balance feature, and map the splicing result into an RGB image.

[0064] In a preferred embodiment, referring to Figure 2 , which is a schematic diagram of the structure of a haze image enhancement network provided by an embodiment of the present invention. Among them, "Conv2D" represents a two-dimensional convolution operation, "PReLU" represents a parameterized rectified linear unit, k is the convolution kernel size, and s is the convolution step size; "TRK" is the feature extraction module; "PA Block" represents a pixel attention PA module, "CA Block" represents a channel attention CA module, and "CE Block" represents a chrominance enhancement (CE) module.

[0065] Depend on Figure 2It can be seen that the dehazing subnetwork adopts a pyramid structure design to enhance the feature extraction capability of the model under different receptive fields. In addition, for the chromaticity features that are linearly attenuated, if they are directly dimensionally spliced ​​with the fog-free brightness component, the brightness and contrast of the output result will be greatly suppressed, resulting in the image visually presenting a grayscale trend. To this end, a chromaticity enhancement subnetwork is designed, which dynamically assigns weight coefficients to the characterization information according to the local degradation intensity by cascading TRK modules and CA modules, and outputs chromaticity equalization features. As a result, the haze image enhancement network can perform adaptive equalization processing on the chromaticity information in the image while eliminating the blurring interference caused by haze, thereby improving the visual perception quality of the output image.

[0066] For step S1, further, the function expression of the feature extraction module is specifically:

[0067]

[0068] Among them, y k Indicates the input state, y k+1 represents the output state, G1, G2, and G3 are all trainable adaptive nonlinear representation units, h represents the step size between adjacent independent variables, and f(J k ,y k ) indicates point (J k ,y k )’s instantaneous slope.

[0069] In a preferred embodiment, from the perspective of the dynamic system, the information flow transmission process between network layers is modeled as the state evolution of features at different time nodes, and a feature extraction module TRK based on the third-order Runge-Kutta method is constructed to enhance the learning and understanding ability of the network model for abstract representations in high-dimensional space. The specific process is as follows:

[0070] (1) According to the mean value theorem of differential calculus, the average slope of a function on the interval is equal to the instantaneous slope at a point in the interval. We can get the following formula:

[0071] y(x k+1 )=y(x k )+hf(x k +εh,y(x k +εh)),0<ε<1

[0072] Among them, h is the step size between adjacent independent variables, ε is a variable parameter, and f(x k ,y k ) represents the point (x k ,y k )’s instantaneous slope;

[0073] For the solution process of any S-order Runge-Kutta method, the corresponding mathematical expression is as follows:

[0074]

[0075] Among them, G i is the slope between the independent variables, which is defined here as a trainable adaptive nonlinear representation unit; a i , b i , c i To correspond to the weight coefficient, the specific value of the weight coefficient can be determined by Taylor Series Expansion (TSE) according to the different solution orders.

[0076] (2) In order to balance the algorithm performance and model processing efficiency, the present invention adopts the third-order Runge-Kutta solution method and fixes the corresponding order of the local truncation error of the hidden variable in the state transition process to o(h 4 ). On the one hand, we can transform y(x k+1 ) at point (x k ,y k ) is expanded into a power series using TSE, as shown below:

[0077]

[0078] On the other hand, let S = 3, and replace G1 and G2 in the general expression of Runge-Kutta method at the point (x k ,y k ) is expanded into a power series using TSE to obtain y k+1 The specific calculation process is:

[0079]

[0080] (3) In y(x k+1 )-y k =o(h 4 ) under the condition of solving the corresponding weight coefficient, the specific solution process of the third-order Runge-Kutta method can be obtained:

[0081]

[0082] (4) Based on the above formula, a third-order Runge-Kutta module is constructed, and its internal structure is as follows: Figure 3 As shown. k Represents the input state of high-dimensional space features; y k+1It represents the output state of the input features after a series of numerical calculations; G1, G2 and G3 represent trainable adaptive nonlinear representation units, which need to undergo multiple iterations of training to obtain the optimal weight parameters and bias values. As a method of approximating numerical integration based on multiple intermediate steps, the Runge-Kutta method has high stability and fast convergence. During the training phase, the iterative optimization process of neural networks for parameters usually involves a large number of numerical solution calculations, such as matrix operations, activation functions, etc. Embedding the TRK module into the backbone model can speed up network convergence, improve the stability of the algorithm during training and inference, and avoid numerical instability problems caused by calculation fluctuations.

[0083] For step S2, specifically, training the haze image enhancement network according to a preset joint loss function to obtain a haze image reconstruction model specifically includes:

[0084] According to a preset training set, the haze image enhancement network is iteratively trained by a preset Adam optimizer; wherein the preset training set includes a plurality of haze images and clear images corresponding to the plurality of haze images, and the loss function in the preset Adam optimizer is specifically the preset joint loss function;

[0085] When the number of iterations reaches a preset iteration threshold, the haze image reconstruction model is outputted.

[0086] In a preferred embodiment, a haze image dataset of different scenes and different concentrations is obtained and divided into a training set and a test set to train the haze image enhancement network. In the training stage, 5000 pairs of haze images and corresponding clear images are randomly selected from the outdoor haze dataset of RESIDE for model training, and the input images are cropped to a resolution of 256×256 and pixel normalized to finally obtain the preset training set.

[0087] In this preferred embodiment, the haze image enhancement network is based on the Pytorch deep learning development framework and runs on the NVIDIA GeForce RTX2070 GPU. The Adam optimizer (β1=0.9, β2=0.999) is used for training, the initial learning rate is set to 0.001, and the network model is optimized for 100 iterations. The weight coefficients λ1 and λ2 corresponding to the total loss function are set to 0.7 and 0.3 respectively.

[0088] In order to reduce the inherent loss in the experimental process, the conversion process from RGB space to YUV space involved in the algorithm is completely reversible. The training process of the network proposed in the present invention is shown in Table 1.

[0089] Table 1. Haze image enhancement network training process

[0090]

[0091] For step S2, further, the preset joint loss function is specifically:

[0092] L T =λ1L1+λ2L c

[0093]

[0094] Among them, L T is the total loss value, λ1 and λ2 are L1 and L c The corresponding preset weight coefficient, N is the total number of training samples in the training set, I is the haze image, J is the clear image corresponding to the haze image, J′ is the output image corresponding to the haze image, and J i is the clear image corresponding to the i-th haze image in the training set, J i ′ is the output image corresponding to the i-th haze image in the training set, Φ is the preset pre-training model, M is the number of hidden layers in the preset pre-training model, α i To preset the weight coefficient, D[x,y] represents the calculation of the Euclidean distance between x and y in the high-dimensional space.

[0095] In a preferred embodiment, the loss function in the preset Adam optimizer is set to the joint loss function, including L1 loss and contrast loss L c , optimizing the built-in parameters in the network model at the pixel level and the perception level respectively.

[0096] L1 loss is used to reduce the pixel error between the reconstructed image and the target image. For a given N pairs of training samples It can be described using the following formula:

[0097]

[0098] In addition, the contrastive learning loss L c is a constraint function used to train similarity learning models. Its basic principle is as follows Figure 4 This function aims to minimize the Euclidean distance between the reconstructed image and the clear image at the perceptual level, while maintaining the maximum difference between the reconstructed features and the degraded features. The corresponding mathematical expression is as follows:

[0099]

[0100] Where Φ represents the pre-trained model VGG-19 for perceptual feature extraction; M is the number of hidden layers in the pre-trained model; α i is the trade-off coefficient; Represents the Euclidean distance between x and y in high-dimensional space.

[0101] In summary, the total loss function (i.e., the joint loss function) can be expressed by the following formula, where λ1 and λ2 are weight coefficients for balancing pixel loss and contrast loss, respectively:

[0102] L T =λ1L1+λ2L c

[0103] For step S3, specifically, the step of inputting the image to be enhanced into the haze image reconstruction model, performing dehazing and enhancement processing on the image to be enhanced through the dehazing sub-network and the chroma enhancement sub-network to obtain a reconstructed image specifically includes:

[0104] Mapping the image to be enhanced from an RGB image to a YUV image through the decomposition module, and decomposing the mapped YUV image to obtain corresponding brightness components and chrominance components;

[0105] Inputting the brightness component into the defogging sub-network for defogging processing to obtain the fog-free brightness component;

[0106] Inputting the chroma component into the chroma enhancement subnetwork for enhancement processing to obtain the chroma balance feature;

[0107] The fog-free brightness component and the chromaticity balance feature are spliced ​​by the splicing module, and the splicing result is mapped from the YUV space to the RGB space to obtain the reconstructed image.

[0108] Furthermore, the step of inputting the brightness component into the defogging sub-network for defogging to obtain the fog-free brightness component specifically includes:

[0109] The luminance component is sequentially upsampled and downsampled by the sampling module to obtain an upsampled component and a downsampled component respectively;

[0110] The brightness component, the up-sampling component and the down-sampling component are processed simultaneously and in sequence by the convolution correction module, the feature extraction module and the pixel attention module, and a first brightness processing result, an up-sampling processing result and a down-sampling processing result are obtained correspondingly;

[0111] Performing down-sampling processing on the up-sampling processing result to obtain a second brightness processing result, and performing up-sampling processing on the down-sampling processing result to obtain a third brightness processing result;

[0112] The first brightness processing result, the second brightness processing result and the third brightness processing result are linearly fused by the fusion module, and the fusion result is corrected by the convolution correction module to obtain a fog-free brightness component.

[0113] Furthermore, the step of inputting the chroma component into the chroma enhancement subnetwork for enhancement processing to obtain the chroma equalization feature specifically includes:

[0114] Processing the chrominance component sequentially through the convolution correction module, the feature extraction module and the channel attention module to obtain an attention processing result;

[0115] The attention processing result is corrected by the convolution correction module to obtain the chromaticity balance feature.

[0116] In a preferred embodiment, referring to Figure 5 , which is a schematic diagram of the comparison of processing results of different image enhancement methods provided by an embodiment of the present invention on the RESIDE SOTS test set. Figure 5 It can be seen that although the SLA algorithm can eliminate the visual haziness caused by haze, it does not consider the changes in brightness characteristics during the enhancement process, which makes the local detail texture of the output image unclear, and there is a certain color distortion problem in the sky area; the results of the TOENet algorithm have obvious halo artifact problems, and there is also a certain degree of "stratification" phenomenon in the sky area; the output results of the DehazeFormer algorithm face more haze residual problems. PMD-Net, T-Net and MSBDN have relatively good performance in balancing image color perception and maintaining visual fidelity, but the three algorithms cannot completely eliminate the visual impact caused by haze, especially in scene areas with a large depth of field, and the results are still accompanied by varying degrees of haziness. Compared with the aforementioned method, the method proposed in the present invention can effectively eliminate the visual interference caused by haze while improving the overall clarity of the image, and the reconstruction result is more realistic and natural visually.

[0117] In order to more comprehensively verify the effectiveness of the present invention, a quantitative comparison is made with the aforementioned dehazing model on the RESIDE SOTS test set. The software platforms involved in the experiment are Matlab2016a and PyCharm, and the hardware platforms are NVIDIA Geforce RTX2070 GPU, 8GB video memory, Intel(R) Core(TM) i5-9400 CPU, 2.90GHz main frequency, and 16G RAM.

[0118] Table 2 Quantitative comparison of the contrast methods on the RESIDE SOTS dataset

[0119]

[0120] As shown in Table 2, the evaluation results are the average values ​​of all samples included in the RESIDE SOTS test set for each comparison algorithm. Through comparative analysis, it can be seen that the output results of the proposed method have the smallest error in pixel reconstruction and color balance, and the optimal index value is achieved on the test set. Compared with the suboptimal algorithm PMD-Net in the comparison algorithm, the method proposed in the present invention has performance advantages of 0.36dB, 0.31% and 1.91 in PSNR, SSIM and CIEDE2000 indicators respectively.

[0121] Reference Figure 6 , is a schematic structural diagram of a haze image enhancement device provided by another embodiment of the present invention, comprising: a construction module 101, a training module 102 and a reconstruction module 103;

[0122] The construction module 101 is used to construct a haze image enhancement network; wherein the haze image enhancement network includes a dehazing sub-network and a chroma enhancement sub-network;

[0123] The training module 102 is used to train the haze image enhancement network according to a preset joint loss function to obtain a haze image reconstruction model;

[0124] The reconstruction module 103 is used to input the image to be enhanced into the haze image reconstruction model, and perform dehazing and enhancement processing on the image to be enhanced through the dehazing sub-network and the chroma enhancement sub-network to obtain a reconstructed image.

[0125] Furthermore, the training module 102 is used to train the haze image enhancement network according to a preset joint loss function to obtain a haze image reconstruction model, which specifically includes:

[0126] According to a preset training set, the haze image enhancement network is iteratively trained by a preset Adam optimizer; wherein the preset training set includes a plurality of haze images and clear images corresponding to the plurality of haze images, and the loss function in the preset Adam optimizer is specifically the preset joint loss function;

[0127] When the number of iterations reaches a preset iteration threshold, the haze image reconstruction model is outputted.

[0128] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A haze image enhancement method, characterized in that: The steps include: Constructing a haze image enhancement network; wherein the haze image enhancement network includes a dehazing sub-network and a chroma enhancement sub-network; The haze image enhancement network is trained according to a preset joint loss function to obtain a haze image reconstruction model; The image to be enhanced is input into the haze image reconstruction model, and the image to be enhanced is dehazed and enhanced by the dehazing sub-network and the chroma enhancement sub-network to obtain a reconstructed image.

2. The haze image enhancement method according to claim 1, characterized in that: The haze image enhancement network specifically includes a disassembly module, a dehazing sub-network, a chroma enhancement sub-network and a splicing module; The decomposition module is connected to the input ends of the dehazing sub-network and the chroma enhancement sub-network respectively, and is used to decompose the input image into a luminance component and a chroma component; The defogging sub-network includes a sampling module, a convolution correction module, a feature extraction module, a pixel attention module and a fusion module, which are used to perform defogging on the brightness component to obtain a fog-free brightness component; wherein the feature extraction module is constructed by a third-order Runge-Kutta method; The chroma enhancement subnetwork includes a convolution correction module, a feature extraction module and a channel attention module, which are used to perform equalization processing on the chroma component to obtain a chroma equalization feature; The splicing module is connected to the output ends of the defogging sub-network and the chroma enhancement sub-network respectively, and is used to perform splicing processing on the haze-free brightness component and the chroma balance feature, and map the splicing result into an RGB image.

3. The haze image enhancement method according to claim 2, characterized in that: The function expression of the feature extraction module is specifically: Among them, y k Indicates the input state, y k+1 represents the output state, G1, G2, and G3 are all trainable adaptive nonlinear representation units, h represents the step size between adjacent independent variables, and f(x k ,y k ) represents the point (x k ,y k )’s instantaneous slope.

4. The haze image enhancement method according to claim 1, characterized in that: The training of the haze image enhancement network according to the preset joint loss function to obtain the haze image reconstruction model specifically includes: According to a preset training set, the haze image enhancement network is iteratively trained by a preset Adam optimizer; wherein the preset training set includes a plurality of haze images and clear images corresponding to the plurality of haze images, and the loss function in the preset Adam optimizer is specifically the preset joint loss function; When the number of iterations reaches a preset iteration threshold, the haze image reconstruction model is outputted.

5. The haze image enhancement method according to claim 4, characterized in that: The preset joint loss function is specifically: L T =λ1L1+λ2L c Among them, L T is the total loss value, λ1 and λ2 are L1 and L c The corresponding preset weight coefficient, N is the total number of training samples in the training set, I is the haze image, J is the clear image corresponding to the haze image, J′ is the output image corresponding to the haze image, and J i is the clear image corresponding to the i-th haze image in the training set, J i ′ is the output image corresponding to the i-th haze image in the training set, Φ is the preset pre-training model, M is the number of hidden layers in the preset pre-training model, α i To preset the weight coefficient, D[x,y] represents the calculation of the Euclidean distance between x and y in the high-dimensional space.

6. The haze image enhancement method according to claim 2, characterized in that: The step of inputting the image to be enhanced into the haze image reconstruction model, performing dehazing and enhancement processing on the image to be enhanced through the dehazing sub-network and the chroma enhancement sub-network to obtain a reconstructed image specifically includes: Mapping the image to be enhanced from an RGB image to a YUV image through the decomposition module, and decomposing the mapped YUV image to obtain corresponding brightness components and chrominance components; Inputting the brightness component into the defogging sub-network for defogging processing to obtain the fog-free brightness component; Inputting the chroma component into the chroma enhancement subnetwork for enhancement processing to obtain the chroma balance feature; The fog-free brightness component and the chromaticity balance feature are spliced ​​by the splicing module, and the splicing result is mapped from the YUV space to the RGB space to obtain the reconstructed image.

7. The haze image enhancement method according to claim 6, characterized in that: The step of inputting the brightness component into the defogging sub-network for defogging to obtain the fog-free brightness component specifically includes: The luminance component is sequentially upsampled and downsampled by the sampling module to obtain an upsampled component and a downsampled component respectively; The brightness component, the up-sampling component and the down-sampling component are processed simultaneously and in sequence by the convolution correction module, the feature extraction module and the pixel attention module, and a first brightness processing result, an up-sampling processing result and a down-sampling processing result are obtained correspondingly; Performing down-sampling processing on the up-sampling processing result to obtain a second brightness processing result, and performing up-sampling processing on the down-sampling processing result to obtain a third brightness processing result; The first brightness processing result, the second brightness processing result and the third brightness processing result are linearly fused by the fusion module, and the fusion result is corrected by the convolution correction module to obtain a fog-free brightness component.

8. The haze image enhancement method according to claim 6, characterized in that: The step of inputting the chroma component into the chroma enhancement subnetwork for enhancement processing to obtain the chroma equalization feature specifically includes: Processing the chrominance component sequentially through the convolution correction module, the feature extraction module and the channel attention module to obtain an attention processing result; The attention processing result is corrected by the convolution correction module to obtain the chromaticity balance feature.

9. A haze image enhancement device, characterized in that: include: Building modules, training modules, and refactoring modules; The construction module is used to construct a haze image enhancement network; wherein the haze image enhancement network includes a dehazing sub-network and a chroma enhancement sub-network; The training module is used to train the haze image enhancement network according to a preset joint loss function to obtain a haze image reconstruction model; The reconstruction module is used to input the image to be enhanced into the haze image reconstruction model, and perform dehazing and enhancement processing on the image to be enhanced through the dehazing sub-network and the chroma enhancement sub-network to obtain a reconstructed image.

10. The haze image enhancement device according to claim 9, characterized in that: The training module is used to train the haze image enhancement network according to a preset joint loss function to obtain a haze image reconstruction model, and specifically includes: According to a preset training set, the haze image enhancement network is iteratively trained by a preset Adam optimizer; wherein the preset training set includes a plurality of haze images and clear images corresponding to the plurality of haze images, and the loss function in the preset Adam optimizer is specifically the preset joint loss function; When the number of iterations reaches a preset iteration threshold, the haze image reconstruction model is outputted.

Images