Underwater image enhancement method based on probability uncertainty modeling

Abstract

The invention discloses an underwater image enhancement method based on probability uncertainty modeling, and aims to solve the problems that a traditional image enhancement method is poor in robustness in a real and complex underwater environment and an existing deep learning method neglects an uncertainty enhancement result in an underwater image enhancement task. According to the method, the underwater image enhancement process is decoupled into two stages of'distribution estimation 'and'consensus decision'. The first stage is distribution estimation, potential spatial modeling is carried out on an input image by introducing hidden variables and adopting a conditional variation auto-encoder (CVAE), and diversity enhancement distribution of a degraded image is learned; and then, in combination with a probabilistic adaptive instance normalization module (PADIN), mapping the latent variables into style adjustment parameters, and further generating a plurality of image enhancement candidate results conforming to potential distribution. And the second stage is a consensus decision, and an optimal enhancement result is selected from the candidate images through a maximum likelihood estimation algorithm so as to improve the authenticity and visual quality of final output. Experimental results show that the method has good generalization ability and robustness in different underwater scenes, and can adapt to complex and changeable underwater illumination and degradation conditions. According to the method, the brightness, the color rendition degree, the texture detail definition and the structural information fidelity of the underwater image can be remarkably improved, and the generated enhanced image is natural, real and excellent in visual effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of underwater image processing, and particularly relates to an underwater image enhancement method based on probabilistic uncertainty modeling. BACKGROUND

[0002] Underwater images have wide applications in marine exploration, underwater robots, underwater photography, etc. However, due to the light scattering and absorption effects in water, underwater images usually have serious degradation phenomena, including insufficient brightness, color shift, reduced contrast, blurred details, etc., which seriously affect the performance of subsequent image processing and recognition tasks. Therefore, underwater image enhancement technology has become a hot research direction in recent years. Traditional underwater image enhancement methods mainly include image processing-based enhancement algorithms and physical model-based image restoration algorithms. The former, such as histogram equalization, white balance, Retinex, etc., enhances the visual effect by enhancing certain aspects of the image (such as contrast or color), but lacks modeling of the degradation mechanism, which can easily lead to color distortion or over-enhancement; the latter is based on the light transmission model, trying to estimate the water scattering parameters or the transmission map to restore the original image, but this kind of method relies on strong assumptions (such as water homogeneity), and has poor robustness in real complex underwater environments.

[0003] In recent years, deep learning-based methods have been widely applied in underwater image enhancement due to their strong end-to-end fitting ability and visual perception characteristics. Most of these methods rely on convolutional neural networks (CNN) or generative adversarial networks (GAN) to learn the mapping relationship between degraded images and reference images, improving the visual quality and perceptual consistency of the enhanced images. However, there is a key problem with existing deep learning methods: the reference images used in the training process are uncertain, but a deterministic mapping method is used for modeling. The reference images in almost all current underwater image enhancement datasets cannot directly obtain the real "ground truth images", but are constructed through artificial synthesis. For example, the original underwater images are processed using multiple image enhancement algorithms, and then the best version of the enhanced effect is selected by artificial subjective selection, or the version captured by a certain device is used as the reference image. This construction method is influenced by multiple factors, including: (1) image enhancement algorithm parameter settings; (2) subjective preferences of different judges; (3) parameter characteristics and lighting conditions of the synthesis device. Therefore, for the same underwater degraded image, there may be multiple enhanced versions that are reasonable in appearance but different in style, i.e., inherent uncertainty. However, existing deep learning enhancement methods almost all establish a one-to-one deterministic mapping relationship between degraded images and reference images, ignoring the ambiguity and multi-solution nature of the image enhancement process. This modeling method not only limits the network's ability to adapt to the diversity of real scenes, but also may lead to overfitting, distortion, poor generalization performance, and other problems in non-training samples. Therefore, a method is needed that can explicitly model uncertainty, support diversified output, and have structural stability for underwater image enhancement, to adapt to the enhancement needs in actual complex environments and improve the generation quality and robustness of the model. SUMMARY

[0004] The present application provides an underwater image enhancement method based on probabilistic uncertainty modeling, aiming to solve the problems of poor robustness of traditional image enhancement methods in real complex underwater environments and the neglect of uncertain enhancement results in existing deep learning methods. In view of the actual situation that the reference images in current enhancement datasets are mainly selected by artificial algorithms or synthesized by devices, and there is significant subjectivity and uncertainty, the present application discards the traditional one-to-one mapping method of "degraded image to unique reference image", and instead introduces an implicit variable modeling mechanism and a diversity sampling strategy to explicitly model the uncertainty in the image enhancement process, thereby improving the authenticity, diversity, and robustness of the enhanced results.

[0005] The method decouples the underwater image enhancement process into two stages: "distribution estimation" and "consensus decision". In the distribution estimation stage, a conditional variational autoencoder (CVAE) is used to model the latent variable space, multiple sets of latent style control parameters are generated by sampling, and a frequency domain dynamic filtering module (DynamicFilter) is introduced to enhance the expression ability of style features. Combined with the probabilistic adaptive normalization module (PAdaIN), style modulation and structure alignment are realized, and finally multiple candidate enhanced images are output. In the consensus decision stage, according to the maximum likelihood estimation strategy, the optimal enhanced result is selected from all candidate images as the final output image. This method can effectively generate multiple structure-consistent but style-diverse image enhancement versions, and is suitable for the multi-solution modeling needs of real visual recovery in complex underwater environments.

[0006] The method specifically includes the following technical principles and implementation steps: S1 Input image: input the underwater degraded image to be enhanced For a three-channel underwater visible light image, it is defined as: , where , , respectively, represent the height and width of the image. The image usually has color shift, low contrast, structure blur and other degradation problems, and is the input basis for subsequent enhancement modeling.

[0007] S2 Multi-scale feature extraction and attention modeling: input the image into the encoder to extract multi-scale semantic features , where each level of feature is embedded in the Efficient Multi-scale Attention (EMA) module. The EMA module calculates the direction attention enhancement result based on the direction pooling and channel group normalization strategy on the input feature . . , where represents the Sigmoid function, represents the convolution after average pooling along the height direction, represents the convolution after average pooling along the width direction, and is the feature map channel number, represents the feature map size. The EMA enhanced feature has stronger spatial direction structure perception ability, which is beneficial to retain the image texture and contour.

[0008] S3 Constructing prior distribution and posterior distribution: use CVAE structure to establish the prior distribution and posterior distribution of the latent variable, both of which are modeled as Gaussian distribution, and their mathematical forms are respectively: . . where, and are the mean and standard deviation output by the prior encoding network, and are the mean and standard deviation output by the posterior encoding network, dimensional latent variable for controlling the enhanced style; this modeling approach can effectively characterize the diversity and uncertainty of underwater image enhancement.

[0009] S4 Latent variable modeling: In the training phase, the latent variable is sampled from the posterior using the reparameterization method, and its calculation method is: . where, and are parameters from the posterior or prior network, denotes the standard normal distribution noise.

[0010] S5 Style control feature generation: The sampled latent variable is mapped to the mean map and the standard deviation map by 1x1 convolution, and the size is: .

[0011] S6 Frequency domain filtering enhancement: input and into the dynamic filter (DynamicFilter) respectively, and perform fast Fourier transform (FFT), weighted fusion and inverse transform (iFFT). First, perform fast Fourier transform (FFT): . . where , are the style mean and standard deviation maps mapped from the latent variable , respectively, denote the complex Fourier representation in the frequency domain, FFT represents fast Fourier transform, which is used to convert the image from the spatial domain to the frequency domain. Then use multiple complex convolution kernels and dynamic weights ∈[0,1] weighted filtering: . . where, denotes element-wise complex multiplication, denotes the number of complex convolution kernel groups, and​ represents the result of the weighted filtering operation on the frequency domain feature map. The weighted filtered feature map is inverse Fourier transformed (iFFT) to recover to the spatial domain: . . wherein, represents the style mean and standard deviation map after frequency modulation, and iFFT represents a two-dimensional inverse fast Fourier transform. The DynamicFilter module enhances the semantic expression ability and high-frequency structure modeling ability of the style control feature through frequency domain feature modeling and weighted filtering operation.

[0012] S7 Probabilistic normalization modulation: inject the style parameter after frequency domain modulation into the feature map, and perform PAdaIN normalization operation, and the specific calculation formula is: . wherein is the feature map of the current decoder stage, is the mean value of each channel in the channel dimension, is the channel standard deviation, represents the style mean and standard deviation map derived from the latent variable and frequency modulated (from S6), is the normalized output feature map after modulation, that is, the underwater image enhancement result. This normalization method preserves the basic features of the image structure, and controls the style difference through the latent variable, thereby realizing the generation of style diversified, structure consistent enhancement results.

[0013] S8 Diversified enhancement image generation: the generation process of the plurality of candidate enhancement images is based on repeated sampling from the latent variable prior distribution, and the generation method includes: S8.1: Sample a plurality of latent variables from the prior distribution: . wherein, is the input underwater degraded image, represents the th latent variable sample, and is the mean and standard deviation output by the prior encoding network, represents the sampling number (i.e. the number of candidate enhancement images); S8.2: Each latent variable sample is mapped, normalized modulated and decoded to obtain a candidate enhancement image: . wherein, denotes the complete enhancement network, including style mapping (by ), frequency domain filtering (DynamicFilter), PAdaIN normalization and decoder module; denotes the th candidate enhanced image, which has the same resolution as the input image. This step generates multiple structurally reasonable and stylistically diverse underwater image enhancement results through multiple latent variable sampling and reconstruction, forming a candidate image set: .

[0014] S9 Optimal enhancement result selection: The consensus decision-making strategy based on maximum likelihood estimation method selects the final enhancement result from the candidate image set, and the calculation formula is: . wherein, denotes the th candidate enhanced image, is the input underwater degraded image, denotes the likelihood estimation of the candidate image under the condition of the input image, denotes the final output enhanced image, denotes the log-likelihood. This maximum likelihood consensus mechanism selects the optimal one from multiple potential enhancement results, ensuring that the output image has the best visual quality and enhancement stability while maintaining structural authenticity.

[0015] The loss function of the underwater image enhancement method based on probability uncertainty modeling consists of enhancement loss and KL divergence loss . Enhancement loss is used to measure the reconstruction error and perceptual difference between the network output image and the reference image, which consists of mean square error loss (MSE) and perceptual loss (Perceptual Loss), and its expression is: . . . wherein, denotes the pixel-wise mean square error between the network output image and the reference image , denotes the perceptual loss calculated based on the VGG16 network feature extractor, is the intermediate feature extracted by the pre-trained VGG16 network, is the balance weight of perceptual loss. Enhancement loss The network output is encouraged to be consistent with the reference image in low-level pixels and high-level perceptual space, improving the visual quality and structural consistency of the enhanced image. KL divergence loss for constraining the latent variable distribution in CVAE, making the posterior distribution align with the prior distribution , improving the generation consistency of the enhancement model and the stability during inference. The KL divergence loss is defined on the mean branch and the variance branch respectively: . . wherein, represents the KL divergence, used to measure the difference between two normal distributions , ; and is a parameter of the posterior distribution, is a parameter of the prior distribution. The weighted combination of the above loss terms constitutes the total training objective function: . wherein is the weighted coefficient of the KL divergence.

[0016] The present application proposes an underwater image enhancement method based on probabilistic uncertainty modeling. By introducing a conditional variational autoencoder (CVAE) structure, the image enhancement process is expanded from traditional single deterministic mapping to a generative enhancement framework with diversity latent variable modeling and consensus optimal selection. The method innovatively integrates multi-scale attention mechanism (EMA), frequency domain dynamic filtering module (DynamicFilter) and probabilistic adaptive normalization (PAdaIN) structure, effectively improving the model's perception of image details, colors and structures. Compared with existing methods, the present application has the following significant advantages: 1. Modeling uncertainty enhancement space: multiple candidate enhancement results are generated by latent variable sampling, truly reflecting the multiple reasonable possibilities of underwater image in visual recovery; 2. Improve enhancement quality and stability: combined with frequency domain filtering and structure perception attention mechanism, effectively preserve image edge texture and structural details, enhance image naturalness; 3. Support result controllable and consensus optimization: the optimal enhanced image is automatically selected from the candidate set through the maximum likelihood strategy, ensuring the optimal output quality; 4. Good universality and expansibility: this method does not rely on real ground images and can be directly applied to real complex underwater environments and adapt to various image enhancement tasks. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to make the technical solutions of the embodiments of the present application clearer, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0018] Figure 1 The flow chart of the underwater image enhancement method based on probability uncertainty modeling provided by an embodiment of the present application.

[0019] Figure 2 The overall network structure diagram of the underwater image enhancement method based on probability uncertainty modeling provided by an embodiment of the present application.

[0020] Figure 3 The underwater image enhancement result display diagram of the underwater image enhancement method based on probability uncertainty modeling provided by an embodiment of the present application. DETAILED DESCRIPTION

[0021] The following will make a detailed description of the embodiments of the present application. It should be emphasized that the following description is only exemplary and is not intended to limit the scope of the present application and its applications.

[0022] Referring to Figure 1 , the present application provides a probability uncertainty modeling underwater image enhancement method, which comprises the following steps: S1, inputting a to-be-enhanced underwater degraded image ; S2, extracting multi-scale feature information of the image through an encoder, and introducing a multi-scale attention mechanism at each scale stage to enhance the image spatial structure perception ability; S3, inputting the degraded image into a prior distribution network to construct a prior distribution of a latent variable ; in the training stage, the degraded image and a reference enhanced image are spliced and input into a posterior distribution network to construct a posterior distribution of the latent variable ; S4, sampling the latent variable from the posterior or prior distribution by using a reparameterization method , and generating style control features: a mean map and a standard deviation map ; S5, inputting the and into a frequency domain dynamic filtering module respectively, performing fast Fourier transform (FFT), filtering weighting, inverse fast Fourier transform (iFFT) to obtain a frequency-enhanced style feature map; S6, input the style feature map into the decoder feature map, and use a probabilistic adaptive instance normalization (PAdaIN) method to modulate the image features in style: S7, recover the image spatial resolution through the decoder, and output a plurality of candidate enhanced images; S8, select the optimal enhanced image as the final output result from the candidate images based on a maximum likelihood estimation criterion; Referring to Figure 2 , the network structure diagram of the underwater image enhancement method based on probabilistic uncertainty modeling is provided for the embodiment of the application, which includes the implementation technical details of steps S1-S8.

[0023] Step S1 input image: input the underwater degraded image to be enhanced It is a three-channel underwater visible light image, defined as: , wherein , respectively represent the height and width of the image. The image usually has degradation problems such as color offset, low contrast, and structure blur, which is the input basis for subsequent enhancement modeling.

[0024] Step S2 multi-scale feature extraction and attention modeling: input the input image into the encoder to extract multi-scale semantic features , wherein each level of feature is embedded in an Efficient Multi-scale Attention (EMA) module, and the EMA module calculates the direction attention enhancement result based on the direction pooling and channel group normalization strategy on the input feature . . , wherein represents a Sigmoid function, represents a convolution after average pooling along the height direction, represents a convolution after average pooling along the width direction, and is the number of feature map channels, represents the size of the feature map. The EMA enhanced feature has stronger spatial direction structure perception ability, which is beneficial to retain the image texture and contour.

[0025] Step S3 construct prior distribution and posterior distribution: use the CVAE structure to establish the prior distribution and posterior distribution of the latent variable, and the prior distribution and posterior distribution of the latent variable are modeled as Gaussian distribution, and the mathematical forms are respectively: . . , wherein and is the mean and standard deviation output by the prior encoding network, and is the mean and standard deviation output by the posterior encoding network, dimensional latent variable for controlling the enhanced style; this modeling approach can effectively characterize the diversity and uncertainty of underwater image enhancement.

[0026] Step S4: Latent variable modeling: In the training phase, the latent variable is sampled from the posterior using the reparameterization method, which is calculated as: . where, and are parameters from the posterior or prior network, denotes the standard normal distribution noise.

[0027] Step S5: Style control feature generation: The sampled latent variable is mapped to the mean map and the standard deviation map by 1x1 convolution, with dimensions: .

[0028] Step S6: Frequency domain filtering enhancement: input and into the dynamic filter (DynamicFilter) for fast Fourier transform (FFT), weighted fusion, and inverse transform (iFFT). First, perform fast Fourier transform (FFT): . . where , are the style mean and standard deviation maps mapped from the latent variable , respectively, denote the complex Fourier representation in the frequency domain, FFT denotes fast Fourier transform, which is used to convert the image from the spatial domain to the frequency domain. Then use multiple complex convolution kernels and dynamic weights ∈[0,1] for weighted filtering: . . where, denotes element-wise complex multiplication, denotes the number of complex convolution kernel groups, and ​denotes the result of the weighted filtering operation on the frequency domain feature map. The weighted filtered feature map is inverse Fourier transformed (iFFT) to recover to the spatial domain: . . wherein, denotes the style mean and standard deviation map after frequency modulation, and iFFT denotes a two-dimensional inverse fast Fourier transform. The DynamicFilter module enhances the semantic expression ability and high-frequency structure modeling ability of the style control feature through frequency domain feature modeling and weighted filtering operation.

[0029] Step S7: Probabilistic normalization modulation: inject the style parameter after frequency domain modulation into the feature map, and perform PAdaIN normalization operation, the specific calculation formula is: . wherein, is the feature map of the current decoder stage, is the mean value of each channel in the channel dimension, is the channel standard deviation, denotes the style mean and standard deviation map derived from the latent variable and after frequency domain modulation (from S6, is the normalized output feature map after modulation, that is, the underwater image enhancement result. This normalization method preserves the basic features of the image structure, and controls the style difference through the latent variable, thereby realizing the generation of style diversified, structure consistent enhancement results.

[0030] Step S8: Diversified enhancement image generation: the generation process of the plurality of candidate enhancement images is based on repeated sampling from the latent variable prior distribution, and the generation method includes: S8.1: sampling a plurality of latent variables from the prior distribution: . wherein, is the input underwater degraded image, denotes the latent variable sample, and are the mean and standard deviation output by the prior encoding network, denotes the sampling number (i.e. the number of candidate enhancement images); S8.2: each latent variable sample is mapped, normalized and modulated, and processed by the decoder to obtain a candidate enhancement image: . wherein, denotes the complete enhancement network, including style mapping (through ), frequency domain filtering (DynamicFilter), PAdaIN normalization and decoder module; Indicates the This step generates multiple underwater image enhancement results with reasonable structures and diverse styles through multiple latent variable sampling and reconstruction to form a candidate image set: .

[0031] Step S9: Optimal enhancement result selection: The consensus decision strategy is based on the maximum likelihood estimation method to select the final enhancement result from the candidate image set. The calculation formula is: . in, Indicates the candidate enhanced images, is the input underwater degraded image, represents the likelihood estimate of the candidate image under the input image condition, represents the final output enhanced image, The maximum likelihood consensus mechanism selects the best among multiple potential enhancement results, ensuring that the output image has the best visual quality and enhancement stability while maintaining structural authenticity.

[0032] The loss function of the underwater image enhancement method based on probabilistic uncertainty modeling is composed of the enhancement loss , KL divergence loss Composition. Enhancement loss It is used to measure the reconstruction error and perceptual difference between the network output image and the reference image. It consists of mean square error loss (MSE) and perceptual loss. Its expression is: . . . in, Represents the network output image With reference image The pixel-wise mean square error, represents the perceptual loss calculated based on the VGG16 network feature extractor, is the intermediate feature extracted from the pre-trained VGG16 network, is the balance weight of the perceptual loss. Enhancement loss Encourage the network output to be consistent with the reference image in low-level pixel and high-level perceptual space, improving the visual quality and structural consistency of the enhanced image. KL divergence loss to constrain the latent variable distribution in CVAE, make the posterior distribution to the prior distribution Align, enhance the generation consistency of the model and the stability at inference. The KL divergence loss is defined on the mean branch and the variance branch respectively: . . where, represents the KL divergence, used to measure the difference between two normal distributions , , is the posterior distribution parameter, is the prior distribution parameter. The weighted combination of the above multiple loss terms constitutes the total training target function: . where is the weighted coefficient of the KL divergence.

[0033] The present application proposes an underwater image enhancement method based on probability uncertainty modeling. By introducing a conditional variational autoencoder (CVAE) structure, the image enhancement process is expanded from traditional single deterministic mapping to a generative enhancement framework with diversity latent variable modeling and consensus optimal selection. The method innovatively integrates multi-scale attention mechanism (EMA), frequency domain dynamic filtering module (DynamicFilter) and probabilistic adaptive normalization (PAdaIN) structure, effectively improving the model's perception ability of image details, colors and structures. Compared with existing methods, the present application has the following significant advantages: 1. Modeling uncertainty enhancement space: multiple candidate enhancement results are generated by latent variable sampling, which truly reflects the multiple reasonable possibilities of underwater image in visual recovery; 2. Improve the quality and stability of enhancement: combined with frequency domain filtering and structure perception attention mechanism, the image edge texture and structural details are effectively preserved, and the natural feeling of the image is enhanced; 3. Support result controllable and consensus optimization: the optimal enhanced image is automatically selected from the candidate set through the maximum likelihood strategy, ensuring the optimal output quality; 4. Good universality and expansibility: this method does not depend on real ground images and can be directly applied to real complex underwater environments and adapt to various image enhancement tasks.

[0034] In summary, the present application method realizes the organic unity of underwater image enhancement quality, diversity and stability, and has wide application prospect and popularization value in the fields of image visual recovery, ocean exploration, intelligent underwater perception, etc.

Claims

1. An underwater image enhancement method based on probabilistic uncertainty modeling, characterized in that: The method comprises the following steps: S1, receiving an underwater degraded image to be enhanced as input; S2, extracting multi-scale feature information of the image by an encoder, and introducing a multi-scale attention mechanism at each scale stage to enhance the spatial structure perception ability of the image; S3, inputting the degraded image into a prior distribution network to construct a prior distribution of the latent variable ; in a training phase, inputting the degraded image and the reference enhanced image into the posterior distribution network to construct a posterior distribution of the latent variable ; S4, sampling latent variables from the posterior or prior distribution using a reparameterization method and generate style control features by convolutional transformation: mean map and standard deviation map ; S5, the and The results are input into the frequency domain dynamic filtering module for fast Fourier transform (FFT), filter weighting, and inverse Fourier transform (iFFT) to obtain the style feature map after frequency enhancement. S6, inputting the style feature map into a decoder feature map, and using a probabilistic adaptive instance normalization (PAdaIN) method to modulate the image features in style: S7, restoring the image spatial resolution through the decoder, and outputting a plurality of candidate enhanced images; S8, selecting an optimal enhanced image as a final output result from the candidate images based on a maximum likelihood estimation criterion.

2. The underwater image enhancement method of claim 1, wherein, The input image in step S1 For three-channel underwater visible light images, it is defined as: wherein , respectively represent the height and width of the image. The image usually has degradation problems such as color offset, low contrast, structure blur, etc., which is the input basis for subsequent enhancement modeling.

3. The underwater image enhancement method of claim 1, wherein, The encoder in step S2 includes a plurality of convolution modules, and outputs multi-scale features Wherein each level of features is embedded in an Efficient Multi-scale Attention (EMA) module, and the EMA module calculates the direction attention enhancement result for the input features The calculation direction attention enhancement result is: . wherein, denotes a Sigmoid function, denotes a convolution after average pooling along the height direction, denotes a convolution after average pooling along the width direction, and is the number of feature map channels, denotes the size of the feature map. The EMA enhanced feature has a stronger spatial direction structure perception ability, which is conducive to retaining the image texture and contour.

4. The underwater image enhancement method of claim 1, wherein, The prior distribution and the posterior distribution of the latent variable in step S3 are both modeled as Gaussian distributions, and their mathematical forms are respectively: . . in, and is the mean and standard deviation output by the prior encoding network, and is the mean and standard deviation output by the posterior encoding network, The dimension is The latent variables are used to control the enhancement style; this modeling method can effectively characterize the diversity and uncertainty of underwater image enhancement.

5. The underwater image enhancement method of claim 1, wherein, the latent variables in step S4 By reparameterization trick sampling, the computation is: . where, and are parameters from the posterior or prior network, denotes a standard normal distributed noise; after sampling is mapped to a mean map by a 1 x 1 convolution with a standard deviation map with dimensions: .

6. The underwater image enhancement method of claim 1, wherein, The frequency domain dynamic filtering module (DynamicFilter) in step S5 first performs two-dimensional Fourier transform (FFT) on the style feature map and ​ . . wherein , represent the style mean and standard deviation maps, respectively, obtained by mapping the latent variables represent the complex Fourier representation in the frequency domain, FFT denotes the Fast Fourier Transform, used to convert an image from the spatial domain to the frequency domain. The image is then filtered using a plurality of complex convolution kernels and dynamic weights ∈ [0, 1]​ . . wherein, represents element-wise complex multiplication, represents the number of complex convolution kernel groups, and represents the result after the weighted filtering operation on the frequency domain feature map. The feature map after the weighted filtering is inverse Fourier transformed (iFFT) to recover to the spatial domain: . . wherein, The DynamicFilter module enhances the semantic expression ability and high-frequency structure modeling ability of the style control features through frequency domain feature modeling and weighted filtering operations.

7. The underwater image enhancement method of claim 1, wherein, The style injection normalization operation in step S6 is based on probabilistic adaptive instance normalization (PAdaIN), and its specific calculation formula is: . wherein, is the feature map of the current decoder stage, is the mean value for each channel in the channel dimension, is the channel standard deviation, denotes the style mean and standard deviation maps derived from the latent variable and modulated in the frequency domain (from claim 6, is the modulated normalized output feature map, i.e. the underwater image enhancement result. This normalization way preserves the basic features of the image structure, while controlling the style difference through the latent variable, thereby achieving the generation of enhancement results with style diversity and consistent structure.

8. The underwater image enhancement method of claim 1, wherein, The generation process of the plurality of candidate enhanced images in step S7 is based on repeated sampling from the prior distribution of the latent variable, and the generation method includes: S7.1: sampling a plurality of latent variables from the prior distribution: . wherein, is an input underwater degraded image, represents the th latent variable sample, and is the mean and standard deviation output by the prior encoding network, denotes the number of samplings (i.e., the number of candidate enhanced images); S7.2: Each latent variable sample After mapping, normalization modulation, and decoder processing, the candidate augmented image is obtained: . wherein, represents the complete enhancement network, including style mapping (through ), frequency domain filtering (DynamicFilter), PAdaIN normalization, and the decoder module; represents the first candidate enhanced image, which has the same resolution as the input image. This step generates multiple structurally reasonable and stylistically diverse underwater image enhancement results through multiple latent variable sampling and reconstruction, forming a candidate image set: 。 9. The underwater image enhancement method of claim 1, wherein, The consensus decision strategy in step S8 is based on the maximum likelihood estimation method, and the final enhanced result is selected from the candidate image set, and its calculation formula is: . wherein, represents the i-th candidate enhanced image, is the input underwater degraded image, represents the likelihood estimate of the candidate image under the input image condition, represents the final output enhanced image, represents the log-likelihood. This maximum likelihood consensus mechanism selects the best one among multiple potential enhancement results, ensuring that the output image has the best visual quality and enhancement stability while maintaining structural fidelity.

10. A computer readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to realize the underwater image enhancement method according to any one of claims 1 to 9.

Images