Unsupervised lightweight underwater image enhancement method based on frequency domain

Abstract

The invention provides an unsupervised lightweight underwater image enhancement method based on a frequency domain, and belongs to the technical field of underwater image enhancement and restoration. The method comprises the following steps: receiving a degraded underwater image, and obtaining an enhanced image through a lightweight unsupervised network: converting the degraded underwater image into a frequency domain image, and splitting the frequency domain image into amplitude information and phase information; extracting amplitude characteristics based on the amplitude information through a color correction network, and fusing the amplitude characteristics with the phase information to obtain a color correction image; extracting phase features based on the phase information through a detail sharpening network, and fusing the phase features and the amplitude information to obtain a detail sharpening image; and extracting a weight parameter corresponding to the color correction image and a weight parameter corresponding to the detail sharpening image based on the degraded underwater image through the spatial feature extraction network, and performing weighted fusion based on the weight parameter corresponding to the color correction image and the weight parameter corresponding to the detail sharpening image to obtain an enhanced image. And accurate and efficient underwater image enhancement is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater image enhancement and restoration, and in particular to an unsupervised lightweight underwater image enhancement method based on frequency domain. Background Art

[0002] The vast oceans and abundant marine resources have necessitated the emergence of a wide range of intelligent marine equipment, including underwater robots. These intelligent marine devices rely on real-time underwater video or image capture for underwater exploration. However, water selectively absorbs and scatters light, which affects underwater imaging quality and causes color decay and blurred details in underwater videos or images. These issues severely reduce the operational efficiency of intelligent marine devices. Therefore, the invention of effective and efficient underwater image enhancement technology to post-process degraded underwater images and improve their color decay and blurred details is of great significance for improving the operational efficiency of intelligent marine devices and promoting the development of the marine industry.

[0003] Existing underwater image enhancement methods can be categorized as statistically based, physical imaging model-based, and deep learning-based. Statistical methods utilize traditional image processing methods (such as histogram stretching and gamma correction) to process underwater images, directly altering the pixel distribution. These methods do not rely on specific underwater imaging models, have low dataset requirements, and can rapidly process underwater images. However, these methods lack specific standards for processing underwater images and often over-focus on enhancing a single characteristic while neglecting overall visual quality, resulting in under- and over-enhancement in the processed results. Physical imaging model-based methods focus on simulating the underwater imaging process, constructing specific mathematical formulas to model the process, and then solving these models to restore the pre-degraded underwater image. To account for the diverse parameters in the imaging model, various physical priors and assumptions (such as dark channel priors and red channel priors) have been proposed. Underwater image enhancement methods based on physical imaging models can effectively restore underwater image quality when the proposed physical priors and assumptions satisfy the intended underwater environment. However, when faced with unknown and variable underwater environments, the assumptions and priors of such methods often make it difficult to accurately estimate the parameters of the imaging model, resulting in poor processing results. Consequently, such methods have low generalization capabilities and their performance fluctuates significantly due to environmental influences.

[0004] Compared to the two aforementioned methods, deep learning-based underwater image enhancement methods demonstrate superior performance. By learning from a large number of underwater datasets, these methods can establish a mapping between degraded underwater images and high-quality images, demonstrating high generalization capabilities. However, these methods are demanding in terms of dataset requirements. Most supervised methods rely on paired underwater images for training, requiring both degraded underwater images and their corresponding high-quality underwater images. Furthermore, most deep learning-based underwater image enhancement methods suffer from poor applicability, resulting from high model complexity, heavy computational burdens, and significant hardware requirements, making them difficult to deploy for real-time application in resource-constrained marine intelligent equipment, such as underwater robots.

[0005] Therefore, an unsupervised and efficient underwater image enhancement method is needed. Summary of the Invention

[0006] In light of this, the present invention provides an unsupervised, lightweight underwater image enhancement method based on the frequency domain. This method, based on deep learning technology, extracts global features of underwater images from the frequency domain and processes the amplitude and phase information contained in the frequency domain to correct the image's color and sharpen its details, achieving underwater image enhancement.

[0007] To this end, the present invention provides the following technical solutions:

[0008] An unsupervised lightweight underwater image enhancement method based on frequency domain, comprising:

[0009] Receive degraded underwater images and perform image enhancement through a lightweight unsupervised network;

[0010] The lightweight unsupervised network includes:

[0011] Convert the degraded underwater image into a frequency domain image and split it into amplitude information and phase information;

[0012] Amplitude features are extracted based on amplitude information through a color correction network, and the amplitude features are fused with phase information to obtain a color-corrected image.

[0013] The detail sharpening network extracts phase features based on phase information and fuses the phase features with amplitude information to obtain a detail sharpened image.

[0014] The weight parameters corresponding to the color correction image and the weight parameters corresponding to the detail sharpening image are extracted based on the degraded underwater image through the spatial feature extraction network;

[0015] The color-corrected image and the detail-sharpened image are weightedly fused based on the corresponding weight parameters of the color-corrected image and the corresponding weight parameters of the detail-sharpened image to obtain an enhanced image.

[0016] Furthermore, it also includes:

[0017] Unsupervised training of lightweight unsupervised networks via no-reference loss functions;

[0018] The no-reference loss function:

[0019]

[0020] in, represents the color loss function, represents the perceptual loss function, represents the brightness loss function; λ1, λ2, λ3 represent empirical constants used to constrain the contributions of three different loss functions; I R represents the degraded underwater image, I c represents the color-corrected image, I s represents the detail-sharpened image, I E Represents an enhanced image.

[0021] Furthermore, the step of converting the degraded underwater image into a frequency domain image and splitting the image into amplitude information and phase information includes:

[0022]

[0023] Where k represents a complex unit; (x, y) represents the position coordinates of the spatial domain image pixel; (i, j) represents the signal coordinates of the frequency domain image; H and W represent the height and width of the degraded underwater image, respectively;

[0024] The amplitude information reflects the global color characteristics of underwater images, and the formula is expressed as:

[0025]

[0026] Phase information reflects the global detail characteristics of underwater images, and the formula is expressed as:

[0027]

[0028] Among them, F amp Indicates amplitude information; F pha Represents phase information, R(I f (i,j)) and S(I f (i,j)) represent the frequency domain image I f The real and imaginary parts of (i,j).

[0029] Furthermore, the color correction network includes:

[0030] (F amp ,F pha )=FFT(I R ),

[0031] I c =IFFT(Comp(Pw_Lr_3(F amp ),F pha )),

[0032] Among them, I R Represents a degraded underwater image; I c Represents the color-corrected image; Pw_Lr_3(F amp ) represents feature extraction of amplitude information, performing three point-by-point convolution operations and three LeakyReLU operations; Comp(·) represents a composite operation; FFT represents fast Fourier transform.

[0033] Furthermore, the detail sharpening network includes:

[0034] (F amp ,F pha )=FFT(I R ),

[0035] I s =IFFT(Comp(F amp ,Pw_Lr_3(F pha ))),

[0036] Among them, I s It represents detail-sharpened image; IFFT represents inverse fast Fourier transform.

[0037] Furthermore, the spatial feature extraction network includes:

[0038] (W c ,W s )=Dw_Lr_4(I R ),

[0039] Among them, Dw_Lr_4(I R ) represents feature extraction of the degraded underwater image in the spatial domain, performing four depthwise convolution operations and four LeakyReLU operations; (W c ,W s ) represents the parameter map of the weight parameters corresponding to the color correction image and the detail sharpening image.

[0040] Furthermore, performing weighted fusion of the color-corrected image and the detail-sharpened image based on the weight parameters corresponding to the color-corrected image and the weight parameters corresponding to the detail-sharpened image to obtain the enhanced image includes:

[0041]

[0042] Among them, I Erepresents the enhanced image, Conv_Tanh(·) represents the convolution operation and Tanh activation function; represents point-by-point multiplication, Represents point-by-point addition.

[0043] Furthermore, the color distribution of the enhanced image and the color-corrected image is constrained by the color loss function;

[0044] The color loss function:

[0045]

[0046] in, Represents the enhanced image I E The mean value of pixels in color channel p; Represents the enhanced image I E The mean value of pixels on color channel q; Represents the color-corrected image I c The mean of pixels on channel p; Represents the color-corrected image I c The mean value of pixels on channel q; the value of (p,q) comes from ε∈{(R,G),(R,B),(G,B)}, where R, G, B represent the red, green, and blue color channels respectively.

[0047] Furthermore, the perceptual loss function is used to constrain the enhanced image and the detail sharpened image from the feature level to maintain the consistency of the perceptual content between the images;

[0048] The perceptual loss function:

[0049]

[0050] Among them, Φ j (·) represents the feature extraction function, which is calculated by the relu5-2 layer in the VGG-19 network.

[0051] Furthermore, the global brightness of the image is enhanced by constraining the brightness loss function;

[0052] The brightness loss function:

[0053]

[0054] Where A represents the enhanced image I E The number of preset areas in represents the mean grayscale value of all pixels in region k, and B is the brightness empirical parameter.

[0055] Advantages and positive effects of the present invention:

[0056] This method achieves efficient and effective underwater image enhancement by constructing a lightweight unsupervised network. Specifically, the color correction network is used to extract the amplitude features in the frequency domain to improve the color attenuation of the underwater image. The detail information of the underwater image is enhanced by extracting the phase features in the frequency domain through the detail sharpening network. The image fusion weight is obtained through the spatial feature extraction network, and the color-corrected image and the detail-sharpened image are adaptively fused to obtain the enhanced underwater image. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. 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 any creative labor.

[0058] Figure 1 This is a flowchart of the unsupervised lightweight underwater image enhancement method based on frequency domain;

[0059] Figure 2 This is the lightweight unsupervised network structure diagram in this method;

[0060] Figure 3 Schematic diagram of fast Fourier transform and inverse fast Fourier transform in this method;

[0061] Figure 4 This is the subjective evaluation of this method and the comparison method on the test set. DETAILED DESCRIPTION

[0062] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions 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 embodiments described 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 making creative efforts should fall within the scope of protection of the present invention.

[0063] It should be noted that the terms "first", "second", etc. in the description 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 numbers 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 clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0064] This paper provides a frequency-domain, unsupervised, lightweight underwater image enhancement method. This method extracts global features of underwater images from the frequency domain and processes the amplitude and phase information in the frequency domain to correct the color and sharpen the details of the underwater images. Furthermore, this method constructs a lightweight network structure and designs a reference-free loss function to enable unsupervised training of the network model, alleviating the network model's dependence on paired datasets. The effectiveness and efficiency of this method have been verified on multiple underwater image datasets.

[0065] S1. Dataset selection and construction.

[0066] Three public underwater datasets and one unpublished underwater turbidity dataset are used to train and test the method.

[0067] The three public underwater datasets are UIEB, EUVP, and RUIE. The unpublished underwater turbidity dataset is UTIEB.

[0068] This method randomly selects 900 underwater images from each of the four underwater image datasets as training sets, which are denoted as trainU, trainE, trainR, and trainT. At the same time, 100 underwater images are selected from each of the four datasets as test sets, which are denoted as testU, testE, testR, and testT.

[0069] S2. Build a lightweight unsupervised network to perform underwater image enhancement based on the frequency domain.

[0070] Combine Figure 1 and Figure 2 As shown, the lightweight unsupervised network in this method is further explained:

[0071] Frequency domain information can effectively capture the global features of underwater images. In the frequency domain, the global color information of underwater images can be reflected by the amplitude component, while the global detail information of underwater images can be reflected by the phase component.

[0072] Through a lightweight unsupervised water network, the amplitude and phase information of the frequency domain underwater image are enhanced simultaneously, so as to achieve the purpose of correcting the underwater image color and sharpening the underwater image details, thereby realizing the enhancement of low-quality underwater images.

[0073] The lightweight unsupervised network structure in this method is as follows Figure 2 As shown, it includes: color correction network, detail sharpening network and spatial feature extraction network.

[0074] 1) The color correction network extracts the amplitude features of underwater images to improve the color of degraded underwater images and generate color-corrected images.

[0075] The color correction network includes: fast Fourier transform, three point-by-point convolutions, three LeakyReLU activation functions and inverse fast Fourier transform; the formula for the color correction process is expressed as:

[0076] (F amp ,F pha )=FFT(I R ),

[0077] I c =IFFT(Comp(Pw_Lr_3(F amp ),F pha )),

[0078] Among them, I R Represents a degraded underwater image; I c Represents the color-corrected image; Pw_Lr_3(F amp ) represents feature extraction of amplitude information, performing three point-by-point convolution operations and three LeakyReLU operations; Comp(·) represents a composite operation; FFT represents fast Fourier transform.

[0079] like Figure 3 As shown in the figure, the degraded underwater image is transformed into a frequency domain representation by fast Fourier transform and decomposed to obtain amplitude information and phase information;

[0080] Applying Fast Fourier Transform to transform the spatial domain degraded underwater image I R Converted to the frequency domain and represented as I f .

[0081] The same inverse fast Fourier transform converts the frequency domain image to the spatial domain represented as .

[0082] The fast Fourier transform process, that is, the FFT formula is expressed as:

[0083]

[0084] Among them, the spatial domain image of the degraded underwater image is represented as I R； k represents a complex unit, (x, y) represents the position coordinates of the pixels in the spatial domain image, (i, j) represents the signal coordinates of the frequency domain image, H and W represent the spatial domain image I R The height and width of kθ =cosθ+ksinθ, The formula is:

[0085]

[0086] Representation of Degraded Underwater Images in Frequency Domain I f (i, j) can be further decomposed into phase and amplitude information F amp With phase information F pha Two parts.

[0087] The amplitude information reflects the global color characteristics of underwater images, and its formula is expressed as:

[0088]

[0089] Phase information reflects the global detail characteristics of underwater images, and its formula is expressed as:

[0090]

[0091] Among them, F amp Indicates amplitude information; F pha Represents phase information, R(I f (i,j)) and S(I f (i,j)) represent the frequency domain image I f The real and imaginary parts of (i,j).

[0092] 2) The detail sharpening network extracts the phase features of the underwater image to enhance the details of the underwater image and generates a detail-sharpened image.

[0093] The structure of the detail sharpening network is the same as that of the color correction network;

[0094] The detail sharpening process formula is expressed as:

[0095] (F amp ,F pha )=FFT(I R ),

[0096] I s=IFFT(Comp(F amp ,Pw_Lr_3(F pha ))),

[0097] Among them, I s It represents detail-sharpened image; IFFT represents inverse fast Fourier transform.

[0098] 3) The spatial feature extraction network extracts the spatial features of the underwater image and generates weight parameters to adaptively fuse the color-corrected image with the detail-sharpened image.

[0099] The spatial feature extraction network consists of four convolution blocks. Each convolution block consists of a depthwise convolution and a LeakyReLU activation function. The formula for generating weight parameters is expressed as follows:

[0100] (W c ,W s )=Dw_Lr_4(I R ),

[0101] Among them, (W c ,W s ) represents two parameter maps for adaptive fusion of color-corrected image and detail-sharpened image, Dw_Lr_4(I R ) means extracting features of the input image in the spatial domain, performing four depthwise convolution operations and four LeakyReLU operations.

[0102] Finally, the color-corrected image is weightedly fused with the detail-sharpened image, and the enhanced image is obtained through a convolution operation and Tanh activation function. The formula of the process is expressed as:

[0103]

[0104] Among them, I E represents the enhanced image, and Conv_Tanh(·) represents the convolution operation and Tanh activation function. represents point-by-point multiplication, Represents point-by-point addition.

[0105] S3. Design a no-reference loss function.

[0106] In order to achieve unsupervised training of lightweight unsupervised underwater image enhancement network, this method designs a no-reference loss function The formula is:

[0107]

[0108] in, represents the color loss function, represents the perceptual loss function, represents the brightness loss function; λ1, λ2, λ3 represent empirical constants used to constrain the contributions of three different loss functions; I R represents the degraded underwater image, I c represents the color-corrected image, I s represents the detail-sharpened image, I E Represents an enhanced image.

[0109] Color loss function For constrained enhanced image I E With color corrected image I c The color distribution of is designed based on the grayscale world assumption, which assumes that the pixel means of the red, green, and blue color channels in non-degraded natural images are similar. Based on this, The formula is expressed as:

[0110]

[0111] in, Represents the enhanced image I E The mean value of pixels in color channel p; Represents the enhanced image I E The mean value of pixels on color channel q; Represents the color-corrected image I c The mean of pixels on channel p; Represents the color-corrected image I c The mean value of pixels on channel q; the value of (p,q) comes from ε∈{(R,G),(R,B),(G,B)}, where R, G, B represent the red, green, and blue color channels respectively.

[0112] Perceptual loss function Used to constrain the enhanced image I from the feature level E With detail sharpening image I c , used to maintain consistency in perceptual content between images. The formula is:

[0113]

[0114] Among them, Φ j (·) represents the feature extraction function, which is calculated by the relu5-2 layer in the VGG-19 network.

[0115] Brightness loss function For constrained enhanced image I E The global brightness and darkness of the image are adjusted to maintain the uniformity of the overall brightness of the enhanced image. The formula is:

[0116]

[0117] Where A represents the enhanced image I E The number of regions with a size of 16×16, represents the mean grayscale value of all pixels in region k, and B is the brightness empirical parameter, which is set to 0.5 and is used to control the overall brightness of the enhanced image.

[0118] Under the combination of three loss functions, the proposed underwater image enhancement network can achieve unsupervised training and generate underwater images with vivid colors, clear details and uniform brightness.

[0119] The effects of this method are described with specific examples:

[0120] Step 1: Construct underwater datasets for training and testing.

[0121] The training dataset includes: trainU, trainE, trainR, trainT. The test dataset includes: testU, testE, testR, testT. The test sets testU and testE contain paired underwater images, while the test sets testR and testT contain only degraded underwater images and lack corresponding high-quality reference images.

[0122] Step 2: Parameter setting.

[0123] Step 2-1: Specific training parameter settings are as follows: the proposed network model was built using the Pytorch deep learning framework; the model was trained on a GTX 3080 graphics processing unit on a Linux workstation; the Adam algorithm was used to optimize the training process; the learning efficiency was fixed at 0.0001; the input image size was 256 × 256 × 3, and the pixel values ​​were normalized to [0, 1]; and the batch size was 8. The empirical constants λ1, λ2, and λ3 were set to 1, 10, and 10, respectively.

[0124] Step 2-2: Build an unsupervised lightweight underwater image enhancement network based on the frequency domain and perform unsupervised training on it.

[0125] Step 3: Conduct subjective and objective evaluations on the test sets testU, testE, testR, and testT, and evaluate the efficiency and computational complexity of the method. Six recent representative underwater image enhancement methods are selected to conduct comparative experiments with this method to verify the effectiveness of this method. The selected comparison methods are:

[0126] FUnIE (Fast underwaterimage enhancement for improved visualperception), UWCNN (Underwater scene prior inspired deep underwater image and videoenhancement), Ucolor (Underwater image enhancement via medium transmission-guided multi-color space embedding), UIEWD (A wavelet-based dual-stream network for underwater image enhancement), U-shape (U-shape transformer for underwater image enhancement) and TUDA (Domain adaptation for underwater image enhancement).

[0127] Step 3-1: Subjective evaluation. The subjective evaluation of this method and the comparison method on the test set is as follows Figure 4 As shown. Figure 4 It can be seen that the underwater images enhanced by the six different contrast methods all suffer from some degree of under-enhancement or over-enhancement. For example, the image enhanced by UWCNN is insufficiently bright, and the image enhanced by FUnIE exhibits non-uniform color cast. In contrast, the frequency-domain-based unsupervised lightweight underwater image enhancement method proposed in this paper can effectively remove color casts in degraded images, enhance blurred details, and improve image brightness.

[0128] Step 3-2: Objective evaluation.

[0129] All compared methods and our method were objectively evaluated using four reference evaluation metrics: mean squared error (MSE), structural similarity (SSIM), peak signal-to-noise ratio (PSNR), and patch-based contrast quality evaluation (PCQI); and two non-reference evaluation metrics: natural image quality (NIQE) and information entropy (ENTROPY). Lower MSE and NIQE values ​​indicate higher enhanced image quality and a more natural visual perception; higher PSNR, SSIM, PCQI, and ENTROPY values ​​indicate a closer similarity between the enhanced image and the reference image and richer image detail. The objective evaluation results of each underwater image enhancement method are summarized in Table 1. As shown in Table 1, our method achieved the best results on all four reference evaluation metrics and the two non-reference evaluation metrics, demonstrating its effectiveness in enhancing the visual perception and improving the quality of degraded underwater images.

[0130] Step 3-3: Efficiency and computational complexity evaluation.

[0131] The time required to enhance a single underwater image using this method and the comparison methods under the same experimental conditions was calculated using the Time metric. The computational effort of this method and the comparison methods was calculated using the Params and FLOPs metrics. Table 2 shows the results for each method on these three metrics. As Table 2 demonstrates, this method not only effectively enhances underwater image quality but also possesses a lightweight model framework, meeting the requirements for real-time underwater image processing and making it suitable for resource-constrained marine intelligent equipment such as underwater robots.

[0132] Table 1

[0133]

[0134]

[0135] Table 2

[0136] Methods Params(M)↓ FLOPs(G)↓ Time(s)↓ FUnIE 7.020 10.240 0.003 UWCNN 0.400 5.230 0.078 Ucolor 148.771 2805.340 1.570 UIEWD 13.700 60.170 0.007 U-shape 22.817 2.983 0.087 TUDA 5.940 169.200 0.376 This method 0.469 7.901 0.036

[0137] 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 above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above 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. An unsupervised lightweight underwater image enhancement method based on frequency domain, characterized in that: include: Receive degraded underwater images and perform image enhancement through a lightweight unsupervised network; The lightweight unsupervised network includes: Convert the degraded underwater image into a frequency domain image and split it into amplitude information and phase information; Amplitude features are extracted based on amplitude information through a color correction network, and the amplitude features are fused with phase information to obtain a color-corrected image. The detail sharpening network extracts phase features based on phase information and fuses the phase features with amplitude information to obtain a detail sharpened image. The weight parameters corresponding to the color correction image and the weight parameters corresponding to the detail sharpening image are extracted based on the degraded underwater image through the spatial feature extraction network; The color-corrected image and the detail-sharpened image are weightedly fused based on the corresponding weight parameters of the color-corrected image and the corresponding weight parameters of the detail-sharpened image to obtain an enhanced image.

2. The method according to claim 1, characterized in that Also includes: Unsupervised training of lightweight unsupervised networks via no-reference loss functions; The no-reference loss function: in, represents the color loss function, represents the perceptual loss function, represents the brightness loss function; λ1, λ2, λ3 represent empirical constants used to constrain the contributions of three different loss functions; I R represents the degraded underwater image, I c represents the color-corrected image, I s represents the detail-sharpened image, I E Represents an enhanced image.

3. The method according to claim 1, characterized in that The step of converting the degraded underwater image into a frequency domain image and splitting the image into amplitude information and phase information includes: Where k represents a complex unit; (x, y) represents the position coordinates of the spatial domain image pixel; (i, j) represents the signal coordinates of the frequency domain image; H and W represent the height and width of the degraded underwater image, respectively; The amplitude information reflects the global color characteristics of underwater images, and the formula is expressed as: Phase information reflects the global detail characteristics of underwater images, and the formula is expressed as: Among them, F amp Indicates amplitude information; F pha Represents phase information, R(I f (i,j)) and S(I f (i,j)) represent the frequency domain image I f The real and imaginary parts of (i,j).

4. The method according to claim 1, wherein The color correction network comprises: (F amp ,F pha )=FFT(I R ), I c =IFFT(Comp(Pw_Lr_3(F amp ),F pha )), Among them, I R Represents a degraded underwater image; I c Represents the color-corrected image; Pw_Lr_3(F amp ) represents feature extraction of amplitude information, performing three point-by-point convolution operations and three LeakyReLU operations; Comp(·) represents a composite operation; FFT represents fast Fourier transform.

5. The method according to claim 1, wherein The detail sharpening network includes: (F amp ,F pha )=FFT(I R ), I s =IFFT(Comp(F amp ,Pw_Lr_3(F pha ))), Among them, I s It represents detail-sharpened image; IFFT represents inverse fast Fourier transform.

6. The method according to claim 1, characterized in that The spatial feature extraction network includes: (W c ,W s )=Dw_Lr_4(I R ), Among them, Dw_Lr_4(I R ) represents feature extraction of the degraded underwater image in the spatial domain, performing four depthwise convolution operations and four LeakyReLU operations; (W c ,W s ) represents the parameter map of the weight parameters corresponding to the color correction image and the detail sharpening image.

7. The method according to claim 1, characterized in that The step of performing weighted fusion of the color-corrected image and the detail-sharpened image based on the weight parameters corresponding to the color-corrected image and the weight parameters corresponding to the detail-sharpened image to obtain an enhanced image includes: Among them, I E represents the enhanced image, Conv_Tanh(·) represents the convolution operation and Tanh activation function; represents point-by-point multiplication, Represents point-by-point addition.

8. The method according to claim 2, characterized in that By means of the color loss function, the color distribution of the enhanced image and the color-corrected image is constrained; The color loss function: in, Represents the enhanced image I E The mean value of pixels in color channel p; Represents the enhanced image I E The mean value of pixels on color channel q; Represents the color-corrected image I c The mean of pixels on channel p; Represents the color-corrected image I c The mean value of pixels on channel q; the value of (p,q) comes from ε∈{(R,G),(R,B),(G,B)}, where R, G, B represent the red, green, and blue color channels respectively.

9. The method according to claim 2, characterized in that By using the perceptual loss function, the enhanced image and the detail sharpened image are constrained from the feature level to maintain the consistency of the perceptual content between images; The perceptual loss function: Among them, Φ j (·) represents the feature extraction function, which is calculated by the relu5-2 layer in the VGG-19 network.

10. The method according to claim 2, characterized in that The global brightness and darkness of the image is enhanced by constraining the brightness loss function; The brightness loss function: Where A represents the enhanced image I E The number of preset areas in represents the mean grayscale value of all pixels in region k, and B is the brightness empirical parameter.

Images