Underwater image enhancement method and system

Abstract

The invention discloses an underwater image enhancement method, and the method comprises the steps: inputting a degraded underwater image and a clear reference image into an encoder, and mapping the degraded underwater image and the clear reference image to a potential space, thereby obtaining corresponding potential features; performing frequency domain decomposition on the potential features by using an amplitude-phase decomposition network, performing enhancement processing on amplitude and phase components, and recovering the amplitude and phase components to a spatial domain to obtain preliminarily enhanced features; taking the preliminary enhanced features as input, taking the potential representation of the degraded underwater image as a guide condition, and generating enhanced features through reverse denoising of a conditional diffusion model; the enhanced features are input into a decoder, and a final enhanced underwater image is obtained through reconstruction; according to the method, frequency domain information and diffusion model generation capability are effectively combined, and efficient conversion from a degraded underwater image to a clear image is realized through two-stage processing.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a technology for clarifying underwater images. Background Art

[0002] Previous underwater image enhancement methods are still mainly focused on the spatial domain, and the exploration of frequency domain information is relatively limited. This limitation has also, to a certain extent, restricted the further improvement of the UIE model performance. Due to the influence of environmental factors, underwater images often exhibit problems such as color distortion and low contrast. Frequency domain information can provide global context and rich multi-scale features, which plays an important role in enhancing image contrast, detail restoration, and color restoration. Therefore, the integration of frequency domain information can provide more feature dimensions for existing models, helping the model to better capture and restore complex image details. However, the combination of frequency domain features and generative models is still rare. How to effectively integrate frequency domain features and spatial domain information has become a key challenge that needs to be urgently addressed in the current field of underwater image enhancement. Summary of the Invention

[0003] The purpose of the present invention is to design an underwater image enhancement method to improve the problems of color cast, blur and the like in underwater images.

[0004] The purpose of the present invention is achieved as follows: the first stage includes a pre-trained encoder ε(·) and an amplitude-phase decomposition network (APDNet). The second stage includes a conditional diffusion denoising model (APD-LDM) and a decoder D(·) for clear image reconstruction. First, a pair of degraded underwater images and their corresponding reference images are taken as input and mapped to the latent space through the encoder ε(·). The amplitude-phase decomposition network (APDNet) then uses FFT to decompose the latent features into phase components ρ deg and ρ clear , and the amplitude component A deg and A clear , these components are enhanced separately. Then, the phase component ρ of the degraded image is transformed into deg The amplitude component A of the reference image clear Combined with the inverse fast Fourier transform (IFFT), the features in the frequency domain are restored to the spatial domain. On this basis, the restored feature representation x0 is used as input to degrade the potential representation of the underwater image. As a guide, enter the conditional diffusion denoising model to generate enhanced features Finally, the generated enhanced features It is input into the decoder D(·) for reconstruction to obtain the final restored image The architecture achieves efficient conversion from degraded images to clear images through staged processing.

[0005] As a further preferred embodiment of the present invention, in order to ensure that underwater images can be converted from high-dimensional pixel space to latent space, a network architecture of detail enhancement attention network (DEA-Net) is introduced. Its core is the feature conversion module, which uses stacked detail enhancement attention modules (DEAB) to capture and enhance the features of underwater degraded images. In the implementation, the encoder uses two DEB stacked convolution blocks and downsampling to extract features from the input. The decoder reconstructs features through two DEB stacked convolution blocks and upsampling, while integrating downsampled features to mitigate feature loss. The middle layer adopts the DEAB module, in which CGA plays a key role by combining spatial and channel attention mechanisms.

[0006] CGA solves the problem of uneven color distribution in underwater images by assigning two weights and to each channel and mixing channel and spatial attention mechanisms through broadcast rules. This mechanism highlights important areas, enhances feature encoding and improves enhancement performance. During decoding, each layer integrates features from the downsampling stage through a fusion module to minimize feature loss and preserve image details. Let denote the input features, and the goal of CGA is to generate channel-specific spatial importance maps (SIMs) with the same dimension as X. According to the calculation of the corresponding w c and w s .

[0007]

[0008] Among them, C k×k represents a convolutional layer with a k×k convolution kernel, [·] represents a channel-level connection operation, They represent the global average pooling operation across spatial dimensions, the global average pooling operation across channel dimensions, and the global maximum pooling operation across channel dimensions, respectively.

[0009] As a further preferred embodiment of the present invention, the amplitude-phase decoupling network performs a Fourier transform in the latent space, decomposes the input features into amplitude and phase components, and independently models and enhances them. This design effectively alleviates the coupling and phase-amplitude mismatch problems that occur when directly operating spatial domain features. Specifically, APDNet aims to enhance structural information (encoded by phase) and energy distribution (encoded by amplitude) through decoupling and refinement operations. First, use formula (1-2) to apply a fast Fourier transform (FFT) to the input image features to obtain their corresponding amplitude and phase representations. For each input image X H×W×C , the calculated amplitude A x and phase P x The processing is done through two independent branches. Specifically, a series of convolutional layers are first used to extract the embedding features, denoted as A'x =Convs(A x ) and P' x =Convs(P x ). Subsequently, the self-attention module is introduced to further refine the amplitude feature A' x , highlighting key information and suppressing redundant features, thereby solving the coupling problem in spatial domain operations. Finally, through additional convolutional layers, the refined amplitude feature is represented as A" x =Convs(A' x ).

[0010] Fast Fourier Transform (FFT) is often used in image processing to achieve more efficient calculations. FFT is represented by F, and it transforms a given image X into H×W×C Convert to frequency domain, where each color channel is represented as a complex value. Its mathematical expression is as follows:

[0011]

[0012] Where u and v represent the coordinates in the frequency domain, the amplitude component in the frequency domain is represented as A(x)(u,v), and the phase component is represented as P(x)(u,v), which are derived from the real part and imaginary part, respectively.

[0013]

[0014] As a further preferred embodiment of the present invention, the conditional diffusion denoising model (APD-LDM) utilizes the generative power of the diffusion model to compensate for content loss and mitigate the situation where amplitude cannot be fully refined. In other words, it further mitigates the unnatural artifacts that may be produced in the reconstructed image when the high-frequency amplitude of the red channel does not match the phase of the degraded image. The method proposed in this invention follows the standard diffusion model, performing forward diffusion and backward denoising processes to generate the final reconstruction result. The process of the second stage is as follows:

[0015] (1) Forward process

[0016] Given the decomposed components of the paired images, the phase component of the degraded underwater image and the amplitude component of the clear underwater image are taken as input, expressed as x0 = F -1 (A clear ,P deg ). To perform the forward diffusion process, a predefined variance scheduling parameter set {β1,β2,β3,...,β T}, gradually transform x0 into Gaussian noise X through T steps T , which can be expressed as:

[0017]

[0018] Among them, x trepresents the noise image at time step t(t∈[0,T]), β t ∈(0,1) is a variable that controls the noise variance of each iteration. By combining and optimizing multiple Gaussian distribution terms, we can directly input the image x0 and then obtain the noise image x at any time step. t , using the reparameterization technique, formula (1-6) can be simplified to: where α t =1-β t ,

[0019] And ε t ~N(0,1).

[0020] (2) Reverse diffusion

[0021] The inverse denoising process is defined as a reverse Markov process, which is the inverse of the forward process. t To begin with, the whole process can be expressed as:

[0022]

[0023] here represents the denoised image at time t, represents the potential representation of the degraded underwater image And θ represents the parameters of the model.

[0024] The goal of the inverse process is to gradually denoise the random Gaussian noise X T , and convert it into a latent representation of a clearer and more realistic underwater image This process is affected by the latent representation of degraded underwater images. This inference process consists of T iterative refinement steps (Formula 1-7), each of which uses the conditional distribution p θ (x t-1 |x t ,F deg )(shown in Formula 1-8), the distribution is determined by the neural network ε θ The variance is shown in formula (1-9), and the mean is shown in formula (1-10).

[0025]

[0026] During the training phase, the goal of the diffusion model is to optimize the parameters of the network θ to ensure that the estimated noise vector Close to the true Gaussian distribution, this goal can be expressed as:

[0027]

[0028] In the inference phase, the distribution learned by the diffusion model is used to obtain the potential representation of the restored underwater image using the implicit sampling strategy in the reverse denoising process. This representation is then fed into the decoder to produce the final result In addition, as mentioned before, although the first stage can effectively restore most underwater scenes, some stubborn samples may be affected by artifacts caused by phase and amplitude mismatch, which are difficult to remove. To solve this problem, this paper proposes a self-constrained consistency L scc , making the restored image closer to the real underwater image. scc The purpose of is to constrain feature similarity and guide the reconstruction of the diffusion model So that it satisfies the following form:

[0029]

[0030] In general, the training strategy of APD-LDM is summarized in Algorithm 1-1, and the optimized redefined objective function is:

[0031] L=L diff +λ1L scc (1-13)

[0032] Algorithm 1-1APD-LDM training pseudo code

[0033] Algorithm1-1APD-LDMTrainingPseudo-code

[0034]

[0035] As a further preferred embodiment of the present invention, a two-stage network training strategy is adopted. In the first stage, paired underwater images from the UIEBD and LSUI datasets are used, denoted as and To optimize the encoder ε(·), APDNet and decoder D(·), while freezing the parameters of the diffusion model. The encoder and decoder are optimized by the content loss L con Optimize the form:

[0036]

[0037] APDNet decomposes L dec Optimize, L dec By reconstruction loss L rec , Fourier frequency domain consistency loss L fft and spatial consistency loss L spatial Composition. recThe purpose is to ensure that the decomposed components can reconstruct the input features, which is specifically expressed as:

[0038] L rec =∑||F j -F -1 (P j ,A j )||1 (1-15)

[0039] Here, F j represents the potential representation of the image, and P j and A j Represents the phase and amplitude of the image respectively. -1 (·) is the inverse Fourier transform, which converts the image from the frequency domain to the image space domain. The phase spectrum of the restored image is enhanced by using loss functions in the spatial domain and frequency domain, which can be expressed as:

[0040] L fft =||A(F deg )-A(F clear ))||1 (1-16)

[0041] L spatial =||F deg -F clear ||1 (1-17)

[0042] The total loss formula for optimizing APDNet is L dec =λ2L rec +λ3L fft +λ4L spatial .

[0043] BatchSize is set to 16, and the dataset size is processed to 256×256. The network converges after two stages of training with 5000 and 1000 iterations respectively. This paper uses the Adam optimizer for optimization, and the initial learning rate is set to 1×10 -4 , and decayed by a factor of 0.8 after each stage, while the learning rate was reinitialized to a fixed value of 2×10 -5 The hyperparameters λ1, λ2, λ3, and λ4 are empirically set to 1, 1, 0.1, and 0.01, respectively. Furthermore, the first-stage and second-stage networks are trained separately, rather than jointly. During the second-stage training, the parameters of the first-stage APDNet network must remain frozen.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] 1. The present invention effectively combines frequency domain information and diffusion model generation capabilities to achieve efficient conversion from degraded underwater images to clear images through two-stage processing.

[0046] 2. The designed amplitude-phase decomposition network decomposes the processing features in the latent space, alleviates the coupling and phase-amplitude mismatch problems of spatial domain operations, and enhances the structural information and energy distribution.

[0047] 3. The conditional diffusion model compensates for content loss, alleviates the lack of amplitude refinement, and reduces reconstructed image artifacts.

[0048] 4. Verified by multiple comparative experiments and ablation experiments, the method of the present invention has fewer parameters, faster inference time, and balanced model performance among underwater image enhancement methods based on diffusion models. It performs well in cross-domain datasets and has significant effects in color restoration, contrast, and clarity improvement. It can provide high-quality images for subsequent underwater image tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the technical description of the embodiments. Obviously, the drawings described below are only 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 work.

[0050] Figure 1 This is the overall flow chart of the APD-LDM model framework.

[0051] Figure 2 It is a codec structure diagram.

[0052] Figure 3 This is the APDNet network architecture diagram.

[0053] Figure 4 This is the APDNet effect diagram (paired image test).

[0054] Figure 5 This is the APDNet effect diagram (unpaired image test).

[0055] Figure 6 It is the full reference result comparison (LSUI dataset)

[0056] Figure 7 It is a full reference result comparison (UIEBD dataset).

[0057] Figure 8 This is a graph of subjective perception comparison results on the U45 dataset.

[0058] Figure 9This is a comparison chart of subjective perception of the Challenge60 test set.

[0059] Figure 10 This is a comparison chart of the subjective perception of EUVP test samples.

[0060] Figure 11 It is the residual image between the reconstruction results of different ablation losses and the true reference image.

[0061] Figure 12 This is a comparison chart of depth estimation of the enhancement results of different underwater image enhancement methods.

[0062] Figure 13 It is key point detection. DETAILED DESCRIPTION

[0063] The following is a clear and complete description of the technical solutions in the embodiments of the present invention, in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0064] like Figure 1-13 As shown, a latent diffusion model for underwater enhancement based on amplitude-phase decoupling is proposed. The method first corrects the color cast by exchanging the amplitudes of the degraded underwater image and the clear underwater image in the frequency domain, and then further refines the corrected underwater image using the conditional diffusion model to eliminate the generated artifacts.

[0065] Here’s how:

[0066] S1: A pair of degraded underwater images and their corresponding reference images are taken as input and mapped to the latent space through the encoder ε(·).

[0067] S2: The latent representations of the degraded underwater image and the corresponding reference image in the obtained latent space are input into the amplitude-phase decomposition network (APDNet), and the latent features are decomposed into phase components ρ using FFT deg and ρ clear , and the amplitude component A deg and A clear , these components are enhanced separately.

[0068] S3: By transforming the phase component ρ of the degraded image deg The amplitude component A of the reference image clear In combination, the Inverse Fast Fourier Transform (IFFT) is applied to recover the latent features from the frequency domain to the spatial domain, and an enhanced representation of the preliminary recovery is obtained.

[0069] S4: The recovered feature representation x0 is used as input to the latent representation of the degraded underwater image As a guide, enter the conditional diffusion denoising model to generate enhanced features

[0070] S5: The restored enhanced features are input into the decoder for reconstruction to obtain the final restored image.

[0071] The structure diagram of the encoder ε(·) is shown in the attached figure. Figure 2 It consists of an encoder, a feature conversion module, and a decoder. The core of the module is the feature conversion module, which uses a stacked detail enhancement attention module (DEAB) to capture and enhance the features of underwater degraded images. In the implementation, the encoder uses two DEB stacked convolution blocks and downsampling from the input X H×W×C The decoder reconstructs features by stacking two DEB convolution blocks and upsampling, while integrating downsampled features to mitigate feature loss. The middle layer adopts the DEAB module, in which CGA plays a key role by combining spatial and channel attention mechanisms. CGA assigns two weights w to each channel. c and w s The algorithm addresses the uneven color distribution in underwater images by blending channels with a broadcast rule and a spatial attention mechanism. This mechanism highlights important areas, enhances feature encoding, and improves enhancement performance. During decoding, each layer integrates features from the downsampling stage through a fusion module, minimizing feature loss and preserving image details.

[0072] The Amplitude-Phase Decomposition Network (APDNet) network architecture is as follows:

[0073] APDNet aims to achieve this by decoupling and refining operations such as Figure 3 As shown, the structural information (encoded by phase) and energy distribution (encoded by amplitude) are enhanced. First, Fast Fourier Transform (FFT) is applied to the input image features to obtain their corresponding amplitude and phase representations. For each input image X H×W×C , the calculated amplitude A x and phase P x The processing is done through two independent branches. Specifically, a series of convolutional layers are first used to extract the embedding features, denoted as A' x =Convs(A x ) and P' x =Convs(P x ). Subsequently, the self-attention module is introduced to further refine the amplitude feature A' x , highlighting key information and suppressing redundant features, thereby solving the coupling problem in spatial domain operations. Finally, through additional convolutional layers, the refined amplitude feature is represented as A" x=Convs(A' x ). Figure 4 and attached Figure 5 This is the effect of using APDNet to solve the artifact problem after exchanging amplitudes.

[0074] The model training batch size is set to 16, and the dataset size is processed into 256×256. The network converges after two stages of training with 5000 and 1000 iterations respectively. This paper uses the Adam optimizer for optimization, and the initial learning rate is set to 1×10 -4 , and decayed by a factor of 0.8 after each stage, while the learning rate was reinitialized to a fixed value of 2×10 -5 The hyperparameters λ1, λ2, λ3, and λ4 are empirically set to 1, 1, 0.1, and 0.01, respectively. Furthermore, the first-stage and second-stage networks are trained separately, rather than jointly. During the second-stage training, the parameters of the first-stage APDNet network must remain frozen.

[0075] The second-stage process of the underwater enhanced potential diffusion model based on amplitude-phase decoupling is as follows:

[0076] (1) Forward process

[0077] Given the decomposed components of the paired images, the phase component of the degraded underwater image and the amplitude component of the clear underwater image are taken as input, expressed as x0 = F -1 (A clear ,P deg ). To perform the forward diffusion process, a predefined variance scheduling parameter set {β1,β2,β3,...,β T}, gradually transform x0 into Gaussian noise X through T steps T .

[0078] (2) Reverse diffusion

[0079] The goal of the inverse process is to gradually denoise the random Gaussian noise X T , and convert it into a latent representation of a clearer and more realistic underwater image This process is affected by the latent representation of degraded underwater images. This inference process consists of T iterative refinement steps (Formula 1-7), each of which uses the conditional distribution p θ (x t-1 |x t ,F deg )(shown in Formula 1-8), the distribution is determined by the neural network ε θDuring the training phase, the goal of the diffusion model is to optimize the parameters of the network θ to ensure that the estimated noise vector Close to a true Gaussian distribution.

[0080] In the inference phase, the distribution learned by the diffusion model is used to obtain the potential representation of the restored underwater image using the implicit sampling strategy in the reverse denoising process. This representation is then fed into the decoder to produce the final result Data set selection and evaluation index method:

[0081] This paper uses three benchmark datasets to evaluate the performance of the proposed method and other UIE methods. (a) LUSI, a paired underwater image dataset, contains 4279 pairs of real-world underwater images and reference images, of which 3879 pairs are randomly selected as training sets and the remaining 400 pairs are used as test sets (LSUI-400). (b) UIEBD, contains 890 pairs of real underwater images and reference images and 60 challenging unpaired real underwater images; 800 pairs of images are used for training (UIEBD-800), 90 pairs of images are used for testing (UIEBD-90), and the remaining 60 unpaired images are used as unpaired test sets (UIEBD-60). (c) U45, a public underwater image test dataset, contains 45 underwater images with different degradation types and different scenes, involving color distortion, low contrast, haze effect and blue color cast. Through these datasets, the robustness and applicability of the proposed method can be comprehensively evaluated. (d) EUVP, which contains a large number of paired and unpaired underwater images with poor and good perceptual quality. These underwater images were collected during ocean exploration at different locations and under different visibility conditions. This paper uses the test sample set provided by EUVP, which contains 515 pairs of paired underwater image test samples.

[0082] To evaluate the performance of different UIE methods, this chapter uses four full-reference metrics: MSE, PSNR, SSIM, and LPIPS, as well as two no-reference metrics: UCIQE and UIQM.

[0083] Comparison of experimental results:

[0084] The method of the present invention respectively carried out full-reference underwater image comparison, no-reference underwater image comparison and cross-domain underwater image comparison. The full-reference underwater image comparison used the LSUI dataset and the UIEBD dataset. The no-reference underwater image comparison used the U45 and challenge60 datasets. In order to fairly and more comprehensively evaluate the effectiveness of the proposed method, this paper also carried out a cross-domain underwater image comparison. The cross-domain dataset used the EUVP dataset, while following the experimental settings in the original literature of other underwater image enhancement models. This paper compares other UIE methods, such as WWPE, FUnIE-GAN, PA-Diff, U-shape, DM-Water, and WF-Diff. The subjective perception results of the full-reference underwater image comparison are shown in Figure 2. Figure 6 and Figure 7 ,The comparison of objective evaluation indicators is shown in Table 1. The subjective perception results of the comparison without reference underwater images are shown in Figure 8 and Figure 9 The objective evaluation indicators are shown in Table 2 and Table 3. At the same time, in order to verify the domain generalization performance of the model, the model was tested on the EUVP test dataset. The subjective perception results are shown in Figure 10 The objective evaluation results are shown in Table 4.

[0085] Table 1 Quantitative comparison of different underwater enhancement methods on UIEBD and LSUI datasets

[0086]

[0087] The proposed method performs exceptionally well in color restoration, effectively restoring the original colors of underwater objects and significantly improving the overall image contrast and clarity. Table 1 shows the objective evaluation metrics, with the proposed method achieving the best MSE score.

[0088] Table 2 Quantitative comparison results on the U45 test set

[0089]

[0090] The model obtained by the method of the present invention on UIEBD training has the highest UCIQE score when tested on the U45 and Challenge60 datasets, which illustrates the effectiveness of the method.

[0091] Table 3 Quantitative comparison results on the Challenge60 test set

[0092]

[0093]

[0094] In order to fully verify the cross-domain generalization performance of each model, the EUVP dataset was tested. The models trained on UIEBD and LSUI were tested on the EUVP test set. MSE, PSNR, LPIPS, SSIM, UIQM, and UCIQE indicators were used for comprehensive evaluation. The objective indicator results are shown in Table 4, and the subjective perception results are shown in Table 4. Figure 10 The results show that the invented method is undoubtedly the best, and the image details are well restored.

[0095] Table 4 Quantitative comparison of cross-domain generalization capabilities on the EUVP dataset

[0096]

[0097] In the field of underwater image enhancement, the goal of improving image quality is usually to make subsequent tasks (such as target detection, depth estimation, etc.) more accurate. To this end, the present invention designs an experiment to compare the effect of the proposed method with the existing methods on underwater image enhancement. Figure 12 As described above, the depth estimation image obtained by the method of the present invention shows smoother gradient changes at object boundaries and long-distance areas compared to other methods.

[0098] The SIFT method is used to detect the key points in all underwater enhancement results. Figure 13 It can be seen that the number of key points detected by the proposed method is higher than that of DM-water, PA-Diff and WF-Diff of the same type (based on diffusion model).

[0099] The proposed framework consists of two main stages: an amplitude-phase decomposition network (APDNet) and a diffusion generation module based on conditional latent variable diffusion. In the first stage, color distortion in underwater images is addressed by leveraging amplitude exchange in the frequency domain. In the second stage, the latent representation of the enhanced image obtained in the first stage is further refined using a conditional diffusion model. Here, the latent representation of the degraded underwater image serves as a condition to guide the denoising process, ensuring that the final enhanced image more closely resembles the actual underwater scene. Figures 6 to 10 Tables 1 to 4 demonstrate that the model proposed in this study can work effectively in most scenarios and has significant effects on color restoration and clarity improvement of underwater images.

[0100] The above description is only a specific embodiment of the present invention, but the scope of protection of the invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be covered by the scope of protection of the invention.

Claims

1. A method for underwater image enhancement, characterized in that: The method comprises: S1: Input the degraded underwater image and its clear reference image into the encoder, map them into the latent space, and obtain the corresponding latent features; S2: The Amplitude-Phase Decomposition Network (APDNet) is used to decompose the latent features of the degraded underwater image and its clear reference image in the frequency domain. After enhancing the amplitude of the latent representation of the clear reference image and the phase components of the latent representation of the degraded underwater image, the inverse Fourier transform is used to restore them to the spatial domain to obtain preliminary enhanced features. S3: Using the preliminary enhanced features as input and the potential representation of the degraded underwater image as a guide, the enhanced features are generated through inverse denoising using the conditional diffusion model; S4: The enhanced features are input into the decoder to reconstruct the final enhanced underwater image.

2. The method according to claim 1, characterized in that The amplitude-phase decomposition network includes: A fast Fourier transform module to decompose the latent representation into amplitude and phase components; Amplitude enhancement branch, which includes convolutional layers and self-attention modules to refine the amplitude component and suppress redundant features; Phase preservation branch, which directly preserves the phase component to maintain image structure information; The inverse Fourier transform module combines the enhanced amplitude component with the original phase component to reconstruct the spatial domain features.

3. The method according to claim 1, characterized in that The enhancement processing of the amplitude and phase components includes: performing convolutional layer feature extraction and self-attention module refinement on the amplitude component; and performing convolutional layer feature extraction on the phase component.

4. The method according to claim 1, wherein The encoder and decoder adopt the DEA-Net network architecture, including: the encoder extracts features through two detail enhancement modules and downsampling; the decoder reconstructs features through two detail enhancement modules and upsampling, and fuses the features of the downsampling stage to retain image details; the middle layer adopts a detail enhancement attention module, which balances color distribution by combining spatial and channel attention mechanisms.

5. The method according to claim 1, wherein The conditional diffusion model includes: Forward diffusion process: The phase component of the degraded underwater image and the amplitude component of the clear underwater image are taken as input, which is expressed as: x0=F -1 (P deg ,A clear ); Based on the predefined variance scheduling parameter set {β1,β2,β3,...,β T }, gradually transform x0 into Gaussian noise X through T steps T , the announcement is: where α t =1-β t , And ε t ~N(0,1). Inverse denoising process: Based on the potential representation of the degraded image, the conditional distribution is learned through the neural network θ, and the Gaussian noise is gradually denoised into the potential representation of the clear image; the optimization goal is to minimize the estimated noise With the real noise ε t The mean square error is expressed as:

6. The method according to claim 1, characterized in that Also includes self-constraint consistency L scc , which is used to suppress artifacts caused by phase-amplitude mismatch, is expressed as: The overall optimization goal is L = L diff +λ1L scc .

7. The method according to claim 1, characterized in that The frequency domain decomposition of the phase component P(x)(u,v) and the amplitude component A(x)(u,v) is expressed as:

8. An underwater image enhancement system, characterized in that: include: The encoder module maps the input image to the latent space; Amplitude-phase decomposition network module for frequency domain feature decomposition, enhancement and reconstruction; The decoder module is used to reconstruct the enhanced underwater image.

Images