Underwater image turbidity removing method based on intermediate domain model

Abstract

The invention provides an underwater image turbidity removing method based on an intermediate domain model, and belongs to the technical field of underwater image processing. Constructing a real underwater turbid image data set; using the real underwater turbid image data set and the paired non-turbid image data sets to generate paired underwater turbid image data sets through knowledge conversion skills; constructing an underwater image turbidity removal model; the method for performing model pre-training by using the paired turbid image data sets comprises the following steps of: obtaining an intermediate domain image through a turbid image estimation network; obtaining a compensation graph through a compensation graph estimation network; adding the intermediate domain image and the compensation image; a high-quality underwater image after turbidity removal is obtained through a convolution block; constraining the underwater image turbidity removal model through a joint loss function; and inputting a real underwater turbid image, and obtaining a high-quality underwater image after turbid removal by using the trained model. According to the method, high-quality underwater images can be obtained, and the method has better universality.

Description

Technical Field

[0001] The invention relates to the technical field of underwater image processing, and in particular to an underwater image de-turbidity method based on an intermediate domain model. Background Art

[0002] Underwater images contain important underwater information, which provides a basis for underwater robots to perceive complex underwater environments and perform underwater operations. Therefore, in order to provide accurate information to robots, clear and high-quality water images are urgently needed. However, there are a large number of suspended particles and floating algae in the underwater turbid environment, which produce complex scattering and reflection effects on the propagation of light, resulting in low visibility and severe blurring of the underwater images taken. Processing such images faces more complex and diverse problems.

[0003] Common methods for underwater image processing in the prior art include image enhancement, image restoration, etc. Among them, the image enhancement method does not consider the modeling of the underwater imaging process during the image processing process, and usually reassigns the pixel values ​​of a given image to increase the contrast or achieve color correction. Its advantage is that the processing speed is relatively fast. Its disadvantage is that since the image value of the image is directly operated without considering the underwater imaging model, there are often problems such as over-enhancement, color distortion and offset. The image restoration method takes into account the physical mechanisms such as light propagation and imaging process during the image processing process. First, an underwater imaging model must be established, and then the key parameters in the imaging model must be estimated based on some prior knowledge or observations. Finally, the original clear image is restored by inverting the imaging model. Its advantage is that considering the underwater imaging model, the restored image is more theoretically reasonable. Its disadvantage is that since it involves a lot of prior knowledge, the underwater image restoration algorithm is highly targeted and has poor universality. At the same time, it involves a large number of complex calculations and takes a long time to reason.

[0004] Therefore, an effective and robust underwater image dehazing method is urgently needed for the special underwater turbid environment. Summary of the invention

[0005] In view of this, the present invention provides an underwater image dehazing method based on an intermediate domain model to solve the problems of low visibility and severe blur of underwater turbid images; at the same time, it alleviates the problems of poor universality, poor contrast and color correction effects of underwater turbid image processing methods.

[0006] To this end, the present invention provides the following technical solutions:

[0007] An underwater image de-turbidity method based on an intermediate domain model, comprising:

[0008] Construct a real underwater turbidity image dataset;

[0009] Construct and train an underwater image de-turbidity model; the model input is a real underwater turbidity image, and the output is a high-quality underwater image after de-turbidity; the model includes:

[0010] Intermediate Domain Image Estimation Network:

[0011] I Int (x) = IT γ

[0012] Where I represents the underwater turbidity image, T γ represents a turbidity map to estimate the turbidity level of turbid underwater images;

[0013] Compensation map estimation network:

[0014] I C (x) = A*(I max -I Int (x))*(1-I Int (x))*I max

[0015] Among them, I max =max(I Int (x)); A represents the compensation parameter used to control the compensation degree of the compensation image to the intermediate domain image;

[0016] The intermediate domain image is added to the compensation map and passed through a convolution block to obtain a high-quality underwater image after de-turbidity:

[0017] I DT (x) = I Int (x)+I C (x)

[0018] Among them, I Int (x) represents the intermediate domain image; I C (x) indicates compensation map;

[0019] The convolution block includes a convolution operation and a Tanh activation function.

[0020] Furthermore, the model training includes:

[0021] Using a real underwater turbid image dataset and a paired non-turbid image dataset, a paired underwater turbid image dataset is generated through the knowledge conversion technique;

[0022] The paired non-turbid image data set includes: a non-turbid original underwater image and a corresponding reference image; the paired underwater turbid image data set includes: an underwater image having turbidity characteristics of a real underwater turbid image data set and a corresponding reference image;

[0023] Using paired turbid image datasets, an intermediate domain image is obtained through an intermediate domain image estimation network;

[0024] Using paired turbid image datasets, a compensation map is obtained through a compensation map estimation network;

[0025] The underwater image dehazing model is constrained by a joint loss function.

[0026] Further, the intermediate domain image estimation network includes a turbidity map estimation network;

[0027] The turbidity map estimation network comprises:

[0028] Through 6 convolution blocks, the input paired underwater turbidity image features are extracted; each convolution block includes a convolution operation and a ReLU activation function;

[0029] Generate the parameter graph to be solved through a convolution block, where the convolution block includes a convolution operation and a Tanh activation function;

[0030] Use skip connections between symmetrical convolutional blocks to prevent the loss of feature information;

[0031] Output 6-channel parameter map, namely T and γ.

[0032] Furthermore, the compensation map estimation network comprises:

[0033] Through 6 convolution blocks, the input paired underwater turbidity image features are extracted; each convolution block includes a convolution operation and a ReLU activation function;

[0034] Generate the parameter graph to be solved through a convolution block, where the convolution block includes a convolution operation and a Tanh activation function;

[0035] Use skip connections between symmetrical convolutional blocks to prevent the loss of feature information;

[0036] The output of the compensation map is a 3-channel parameter map, namely A.

[0037] Furthermore, the underwater image de-hazing model is constrained by a joint loss function:

[0038]

[0039] Among them, λ1, λ2, λ3 and λ4 represent empirical weight parameters, which are used to balance the weight of each loss function; {X, Y} are the input pairs of turbid underwater images, X represents the underwater image with turbidity characteristics, and Y represents the corresponding reference image.

[0040] Furthermore, the joint loss function includes:

[0041] Color consistency loss function It is used to constrain the intermediate domain image and eliminate the color attenuation of the intermediate domain image, which is expressed as:

[0042]

[0043] in, and They represent the pixel means of color channels C1 and C2 respectively, and ε∈{(R,G),(R,B),(G,B)}, where R, G, B represent the red, green, and blue color channels respectively.

[0044] Furthermore, the joint loss function also includes:

[0045] Adversarial Loss Function Promote the network to learn the mapping relationship between input image and reference image in an adversarial learning manner:

[0046]

[0047] Where G and D represent the generator and discriminator respectively; the generator G generates the input underwater turbid image into a clear high-quality image, and the discriminator D is used to distinguish the image generated by the generator G from the high-quality reference image; G(X) represents the de-turbidified underwater image;

[0048] L1 Loss Function Used to improve the visual quality of underwater images at the pixel level:

[0049]

[0050] Where X(i,j) represents the pixel value of the turbid underwater image at position (i,j); Y(i,j) represents the pixel value of the reference image at position (i,j);

[0051] Perceptual Loss Function Used to promote feature consistency between generated and reference images:

[0052]

[0053] Among them, Φ k Represents the feature map output by the kth convolutional layer of the VGG19 network pre-trained on the ImageNet dataset; the output features of the relu5-2 layer of VGG19 are used to calculate the perceptual loss.

[0054] Advantages and positive effects of the present invention:

[0055] Through the joint action of four loss functions, the model of the present invention can correctly learn the turbidity characteristics of underwater turbid images, effectively generate intermediate domain images, eliminate the differences between underwater images in different domains, robustly remove turbidity in underwater images, improve the clarity of underwater images, and correct the color cast in underwater images. The compensation map is further used to act on the intermediate domain image to improve the brightness of the generated underwater image and enhance its contrast. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0057] Figure 1 A real underwater turbidity dataset proposed for this invention;

[0058] Figure 2 This is a structural diagram of an underwater image de-hazing method based on an intermediate domain model proposed by the present invention;

[0059] Figure 3 It is a structural diagram of the turbidity map estimation network and the compensation map estimation network proposed by the present invention;

[0060] Figure 4 This is a subjective evaluation of the method of the present invention and the comparative method on the UTID dataset. DETAILED DESCRIPTION

[0061] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0062] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0063] The existing common models for underwater image processing use synthetic turbidity datasets instead of real underwater turbidity datasets for learning and training, which is not enough to simulate complex underwater environments, let alone support the learning of underwater image processing methods. In addition, compared with color attenuation and contrast reduction, underwater turbid images have more complex and diverse image problems. The existing restoration and enhancement models have poor universality and are not suitable for processing turbidity problems.

[0064] The present invention provides an underwater image de-hazing method based on an intermediate domain model. By using a real underwater turbidity dataset and introducing intermediate domain image processing, turbidity in underwater images can be effectively removed, thereby improving the clarity of underwater images. Specifically, a real underwater turbidity dataset is constructed, which contains 1,000 underwater images with different turbidity levels. In view of the problem that existing underwater image processing methods are not good at processing underwater turbidity, the underwater image de-hazing method based on the intermediate domain model can effectively improve the quality of underwater turbid images. The specific steps include:

[0065] S1. Build a real underwater turbidity image dataset;

[0066] Among them, the underwater robot used to capture underwater videos and some images in the UTID dataset are as follows: Figure 1 As shown. Figure 1 It can be seen that the underwater turbidity images in UTID not only have different turbidity levels, but are also severely affected by absorption and scattering. Specifically, these images show severe color attenuation and low contrast. The real large-scale underwater turbidity dataset provides a basis and contributes to the study of underwater image processing methods.

[0067] S2, build a network model;

[0068] It includes two stages: training image generation and underwater image de-turbidity.

[0069] S21. Because the acquired real turbidity data set only contains the original underwater turbidity images and there are no high-quality reference images, which are not sufficient for the subsequent deturbidity method training. Therefore, the role of the training image generation stage is to generate paired (original images and corresponding reference images) turbidity images for the subsequent deturbidity method training through knowledge conversion techniques.

[0070] Combination Figure 2 Further explanation of the training image generation stage:

[0071] In this embodiment, the published paired but non-turbid Underwater Image Enhancement Benchmark (UIEB) underwater dataset and the constructed unpaired but turbid UTID dataset are selected as input, and CycleGAN is used as a knowledge conversion model to add the natural attribute of turbidity to the paired but non-turbid underwater images. Through knowledge conversion, paired underwater turbid images can be generated in this embodiment.

[0072] The generated paired turbid images not only have problems such as color cast and low contrast, but also have different degrees of turbidity, which meets the training requirements of subsequent underwater image dehazing methods.

[0073] S22, the underwater image de-hazing stage is used to establish an underwater image de-hazing model, solve the parameters in the model, generate an intermediate domain image, and further enhance the intermediate domain image to obtain a clear underwater image after de-hazing.

[0074] In this embodiment, the deturbidity model established is as follows:

[0075] I DT (x) = I Int (x)+I C (x)

[0076] Among them I DT (x) represents the high-quality underwater image after de-turbidity; Int (x) represents the intermediate domain image, which has effectively removed the turbidity in the original image and corrected the color deviation of the image. C (x) represents a compensation map, which is used to compensate the middle domain image to improve the contrast of the middle domain image and enhance the details of the middle domain image.

[0077] Intermediate domain image I Int (x) is:

[0078] I Int (x) = IT γ

[0079] Where I represents the original underwater turbidity image, T γIndicates turbidity map;

[0080] Compensation Image I C (x) is:

[0081] I C (x) = A*(I max -I Int (x))*(1-I Int (x))*I max

[0082] I max =max(I Int (x)

[0083] Wherein, A represents the compensation parameter used to control the compensation degree of the compensation map to the intermediate domain image.

[0084] S23. According to the above underwater image de-turbidity model, if the turbidity in the underwater image is to be effectively removed, the unknown parameters T, γ, and A need to be estimated.

[0085] To this end, the parameters are estimated through an underwater image dehazing generative adversarial network; the network consists of two branch networks: a turbidity map estimation network and a compensation map estimation network.

[0086] The turbidity map estimation network is used to solve T and γ, and the compensation map estimation network is used to solve A.

[0087] Construct a turbidity map estimation network and a compensation map estimation network with the same design structure, and combine Figure 3 The structures of the turbidity map estimation network and the compensation map estimation network are further explained:

[0088] Both the turbidity map estimation network and the compensation map estimation network consist of 7 convolution blocks. The first 6 convolution blocks consist of convolution operations and ReLU activation functions to extract the features of the input image; the last convolution block consists of convolution operations and Tanh activation functions to generate estimated parameters. In addition, jump connections are applied between symmetrical convolution blocks to prevent the loss of feature information. The inputs of the turbidity map estimation network and the compensation map estimation network are both original underwater images. The output of the turbidity map estimation network is a 6-channel parameter map, namely T and γ, while the output of the compensation map is a 3-channel parameter map, namely A.

[0089] By solving the parameters T, γ, and A, we can further obtain the intermediate domain image I Int (x) and compensation diagram I C (x).

[0090] S3. By adding the intermediate domain image to the compensation map and passing it through a convolution block consisting of a convolution operation and a Tanh activation function, a high-quality underwater image after de-turbidity can be obtained.

[0091] S4. Design loss function;

[0092] Four loss functions are used to constrain the above network so that it can correctly learn the mapping between turbid underwater images and high-quality reference images. The four loss functions are: adversarial loss, L1 loss, perceptual loss, and color consistency loss.

[0093] {X, Y} is a pair of input turbid underwater images, X represents the turbid underwater image, and Y represents the corresponding reference image.

[0094] Adversarial Loss Function The mapping relationship between the input image and the reference image is promoted by the network learning in an adversarial learning manner, which can be expressed as:

[0095]

[0096] Among them, G and D represent the generator and discriminator respectively. The generator G generates the input underwater turbid image into a clear high-quality image. The discriminator D is used to distinguish the image generated by the generator G from the high-quality reference image. The generator G and the discriminator D promote the improvement of each other's performance by confronting each other. G(X) represents the de-turbid underwater image.

[0097] L1 Loss Function It is used to improve the visual quality of underwater images at the pixel level, which can be expressed as:

[0098]

[0099] Among them, X(i,j) and Y(i,j) represent the pixel values ​​of the corresponding image at position (i,j).

[0100] Perceptual Loss Function It is used to promote the feature consistency between the generated image and the reference image, which can be expressed as:

[0101]

[0102] Among them, Φ k Represents the feature map output by the kth convolutional layer of the VGG19 network pre-trained on the ImageNet dataset. Specifically, the proposed network uses the output features of the relu5-2 layer of VGG19 to calculate the perceptual loss.

[0103] Color consistency loss function Applied to constrain the intermediate domain image, it can eliminate the color attenuation of the intermediate domain image and make it better follow the color constancy, that is, the mean values ​​of the three color channels are as close as possible, reducing the difference between images in different waters and improving the generalization ability of the model. It can be expressed as:

[0104]

[0105] in, and They represent the pixel means of color channels C1 and C2 respectively, and ε∈{(R,G),(R,B),(G,B)}, where R, G, B represent the red, green, and blue color channels respectively.

[0106] In summary, the loss function used to train the proposed underwater image de-hazing network can be expressed as:

[0107]

[0108] Among them, λ1,λ2,λ3 and λ4 represent empirical weight parameters, which are used to balance the weight of each loss function. Under the joint action of the four loss functions, the proposed underwater image dehazing method can effectively generate an intermediate domain image, remove the turbidity in the underwater image, correct the color cast in the underwater image, and further improve the brightness of the underwater image and enhance the contrast through the compensation map.

[0109] The effect of the method of the present invention is verified by the following specific application examples:

[0110] Step 1: In this embodiment, a large real turbidity dataset is collected, named (Underwater TurbidImage Dataset, UTID), specifically:

[0111] Prepare 26 underwater videos, totaling 110 hours. These underwater videos were shot when multiple underwater robots were performing underwater cruising, underwater target grabbing and other tasks. The videos contain a variety of underwater scenes and underwater targets, including: starfish, scallops, sea urchins, fish, corals, cages, pipes, etc.

[0112] 1,200 underwater images with different turbidity levels were captured from non-repetitive underwater videos.

[0113] The 1200 underwater turbid images were selected and 200 images with extremely similar scenes or without objects were removed to obtain 1000 real underwater turbid images as the real turbidity dataset.

[0114] It is divided into a training set and a test set. The number of images in the training set is 800, named: train_UTID; the number of images in the test set is 200, named: test_UTID.

[0115] Step 2: Select 800 pairs of underwater images from the paired but non-turbid UIEB dataset and train CycleGan together with train_UTID in step 1 to generate 800 paired underwater turbid images. These 800 paired underwater turbid images are used for subsequent de-turbidity network training.

[0116] Step 3-1: Use Pytorch deep learning tool to build the proposed underwater image deturbidity network; train the model with GTX3080 graphics processing unit on Linux workstation; Adam algorithm is used to optimize the training process; learning efficiency is fixed to 0.0001; the input image size is 256×256×3, and the pixel value is normalized to [-1,1]; BatchSize size is 8.

[0117] Step 3-2: Use the 800 paired underwater turbidity images in step 2 to train the constructed deturbidity network to obtain the trained model weight parameters.

[0118] Step 4: Selection of comparison method and evaluation index. The present invention tests the trained model from both subjective and objective perspectives, and selects the most recent underwater image processing methods as comparison methods, including: ColorBalance, ICSP, FUnIE, UWCNN, CWR, Ucolor, and U-Shape. Four objective underwater image evaluation indicators are selected to evaluate the results of various methods, including UIQM, UCIQE, natural image quality (NIQE), and information entropy (Entropy). Among them, the smaller the value of the NIQE indicator, the more consistent the color distribution of the processed underwater image is with the natural image. The larger the values ​​of the three evaluation indicators of UIQM, UCIQE, and Entropy, the more uniform the color distribution of the processed underwater image, the higher the contrast, the less noise, and the more consistent with the visual aesthetic standards.

[0119] Step 5: Conduct subjective evaluation on the trained underwater image dehazing network. Figure 4 The subjective results of the method of the present invention and various comparison methods on the test_UTID test set are shown in Figure . Figure 4 It can be seen that most underwater image processing methods are not good at dealing with underwater turbidity, which is directly related to the lack of real underwater turbidity datasets and the fact that the proposed underwater model is not suitable for underwater turbid environments. In contrast, the method of the present invention can not only effectively remove turbidity from underwater images and improve the clarity of underwater images, but also correctly correct the color cast of underwater images and compensate for the brightness and contrast of underwater images.

[0120] Step 6: Objectively evaluate the trained underwater image de-turbidity network. The results are shown in Table 1. The bolded ones represent the best results on each evaluation index. Therefore, the method of the present invention achieves the best results on all four objective evaluation indexes, which is consistent with the subjective results.

[0121] Table 1

[0122] method UIQM↑ UCIQE↑ Entropy↑ NIQE↓ Colorbalance 0.943 0.470 7.200 5.223 ICSP 1.158 0.428 4.821 10.027 FUnIE 0.978 0.470 7.514 5.778 UWCNN 0.926 0.432 6.839 5.938 CWR 1.213 0.500 7.167 5.611 Ucolor 1.031 0.494 7.466 6.116 U-shape 1.053 0.499 7.454 5.416 Method of the present invention 1.329 0.600 7.658 5.145

[0123] In summary, the excellent results in subjective and objective evaluations show that the proposed method can effectively deal with turbidity, color cast, low contrast and other problems in various underwater environments and improve the quality of underwater images. This is due to the proposed correct underwater image de-turbidity model and a diverse set of real underwater turbidity datasets.

[0124] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for dehazing underwater images based on an intermediate domain model, characterized in that: include: Construct a real underwater turbidity image dataset; Build and train a model for underwater image de-turbidity; The model input is a real underwater turbid image, and the output is a high-quality underwater image after de-turbidity; the model includes: Intermediate Domain Image Estimation Network: I Int (x)=I-T γ Where I represents the underwater turbidity image, T γ represents a turbidity map to estimate the turbidity level of turbid underwater images; Compensation map estimation network: I C (x)=A*(I max -I Int (x))*(1-I Int (x))*I max Among them, I max =max(I Int (x)); A represents the compensation parameter used to control the compensation degree of the compensation image to the intermediate domain image; The intermediate domain image is added to the compensation map and passed through a convolution block to obtain a high-quality underwater image after de-turbidity: I DT (x)=I Int (x)+I C (x) Among them, I Int (x) represents the intermediate domain image; I C (x) indicates compensation map; The convolution block includes a convolution operation and a Tanh activation function.

2. The method for dehazing underwater images based on an intermediate domain model according to claim 1, characterized in that: The model training includes: Using a real underwater turbid image dataset and a paired non-turbid image dataset, a paired underwater turbid image dataset is generated through the knowledge conversion technique; The paired non-turbid image data set includes: a non-turbid original underwater image and a corresponding reference image; the paired underwater turbid image data set includes: an underwater image having turbidity characteristics of a real underwater turbid image data set and a corresponding reference image; Using paired turbid image datasets, an intermediate domain image is obtained through an intermediate domain image estimation network; Using paired turbid image datasets, a compensation map is obtained through a compensation map estimation network; The underwater image dehazing model is constrained by a joint loss function.

3. The method for dehazing underwater images based on an intermediate domain model according to claim 1, characterized in that: The intermediate domain image estimation network includes a turbidity map estimation network; The turbidity map estimation network comprises: Through 6 convolution blocks, the input paired underwater turbidity image features are extracted; each convolution block includes a convolution operation and a ReLU activation function; Generate the parameter graph to be solved through a convolution block, where the convolution block includes a convolution operation and a Tanh activation function; Use skip connections between symmetrical convolutional blocks to prevent the loss of feature information; Output 6-channel parameter map, namely T and γ.

4. The method for dehazing underwater images based on an intermediate domain model according to claim 1, characterized in that: The compensation map estimation network comprises: Through 6 convolution blocks, the input paired underwater turbidity image features are extracted; each convolution block includes a convolution operation and a ReLU activation function; Generate the parameter graph to be solved through a convolution block, where the convolution block includes a convolution operation and a Tanh activation function; Use skip connections between symmetrical convolutional blocks to prevent the loss of feature information; The output of the compensation map is a 3-channel parameter map, namely A.

5. The method for dehazing underwater images based on an intermediate domain model according to claim 2, characterized in that: The underwater image de-hazing model is constrained by a joint loss function: Among them, λ1, λ2, λ3 and λ4 represent empirical weight parameters, which are used to balance the weight of each loss function; {X, Y} are the input pairs of turbid underwater images, X represents the underwater image with turbidity characteristics, and Y represents the corresponding reference image.

6. The method for dehazing underwater images based on an intermediate domain model according to claim 5, characterized in that: The joint loss function includes: Color consistency loss function It is used to constrain the intermediate domain image and eliminate the color attenuation of the intermediate domain image, which is expressed as: in, and They represent the pixel means of color channels C1 and C2 respectively, and ε∈{(R,G),(R,B),(G,B)}, where R, G, B represent the red, green, and blue color channels respectively.

7. The method for dehazing underwater images based on an intermediate domain model according to claim 5, characterized in that: The joint loss function also includes: Adversarial Loss Function Promote the network to learn the mapping relationship between input image and reference image in an adversarial learning manner: Where G and D represent the generator and discriminator respectively; the generator G generates the input underwater turbid image into a clear high-quality image, and the discriminator D is used to distinguish the image generated by the generator G from the high-quality reference image; G(X) represents the de-turbidified underwater image; L1 loss function Used to improve the visual quality of underwater images at the pixel level: Where X(i,j) represents the pixel value of the turbid underwater image at position (i,j); Y(i,j) represents the pixel value of the reference image at position (j,j); Perceptual Loss Function Used to promote feature consistency between generated and reference images: Among them, Φ k Represents the feature map output by the kth convolutional layer of the VGG19 network pre-trained on the ImageNet dataset; the output features of the relu5-2 layer of VGG19 are used to calculate the perceptual loss.

Images