Underwater image enhancement method based on adaptive composite degradation model

Abstract

The present application relates to the field of image enhancement in image data processing, and particularly relates to an underwater image enhancement method based on an adaptive composite degradation model, which comprises the following steps: constructing an adaptive composite degradation model; constructing residual subband sets of each clear image and corresponding each underwater image based on wavelet packet decomposition; constructing a generator based on the residual subband sets, inputting the underwater image into the generator to obtain a restored image; calculating a total loss function; optimizing the total loss function using a large amount of underwater image data sets to obtain a trained generator; inputting a newly collected underwater image sequence into the trained generator to obtain an enhanced newly collected underwater image sequence. The existing underwater image enhancement method has the problems of low underwater scene universality and low detail precision of the enhanced underwater image, and the underwater image enhancement method based on the adaptive composite degradation model has high underwater scene universality and high detail precision of the enhanced underwater image.

Description

Technical Field

[0001] This invention relates to the field of image enhancement in image data processing, and specifically to an underwater image enhancement method based on an adaptive composite degradation model. Background Technology

[0002] Underwater images often suffer from problems such as high noise, low contrast, and color distortion, which seriously affect underwater vision tasks. Therefore, underwater image enhancement is used to clarify degraded, low-quality images into high-quality images.

[0003] Underwater image enhancement methods can be mainly divided into two categories based on the technology used: underwater image enhancement based on traditional image processing and underwater image enhancement based on deep learning.

[0004] Underwater image enhancement based on traditional image processing mainly includes histogram equalization (and its variants), white balance algorithms, contrast stretching, and sharpening filtering. These techniques restore the true color and structural information of an image through color correction, contrast enhancement, and noise suppression, but they have poor adaptability to different underwater scenes.

[0005] Deep learning-driven underwater image enhancement mainly utilizes underwater image datasets to train deep neural networks to automatically learn the end-to-end mapping from underwater images to clear images. This avoids explicitly modeling physical processes and instead learns restoration rules implicitly from the data. However, the performance of this method heavily depends on the quantity and quality of the training data. Obtaining real-world paired underwater image datasets is quite difficult, which makes the images enhanced by the trained model perform poorly in detail restoration.

[0006] Therefore, existing underwater image enhancement methods suffer from low versatility for underwater scenes and low detail in the enhanced underwater images. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide an underwater image enhancement method based on an adaptive composite degradation model that has high versatility in underwater scenes and enhances the detail of underwater images.

[0008] To address the aforementioned technical problems, the underwater image enhancement method based on an adaptive composite degradation model provided by this invention includes:

[0009] S1. Construct an adaptive composite degradation model to obtain the underwater images corresponding to each clear image;

[0010] S2. Based on wavelet packet decomposition, construct the residual subband set of each clear image and its corresponding underwater image;

[0011] S3. Construct a generator based on the residual subband set, input the underwater image into the generator to obtain the restored image, i.e. the enhanced image;

[0012] S4. Calculate the total loss function based on reconstruction loss, frequency domain loss, perceptual loss, adversarial loss, and physical consistency loss;

[0013] S5. Optimize the total loss function using a large underwater image dataset to train the generator and discriminator. After training, save the generator's parameters to obtain the trained generator.

[0014] S6. Input the newly acquired underwater image sequence into the trained generator to obtain an enhanced sequence of newly acquired underwater images.

[0015] As a further improvement of the present invention: step S1 includes:

[0016] S101. Calculate parameters related to water quality and depth, including attenuation coefficient, signal-dependent noise intensity parameters, and Gaussian noise basic intensity parameters.

[0017] S102. Calculate the parameters related to the motion state, including the motion direction angle and the motion fuzz length;

[0018] S103. Calculate the parameters related to the depth map, including the defocus blur radius and the transmission map;

[0019] S104. An adaptive composite degradation model for generating corresponding underwater images for clear images is constructed based on parameters related to water quality and depth, parameters related to motion state, and parameters related to depth map.

[0020] Preferably, step S101 includes:

[0021] S1011. Create three two-dimensional lookup tables: an attenuation coefficient lookup table, a signal-dependent noise intensity parameter lookup table, and a Gaussian noise basic intensity parameter lookup table.

[0022] S1012. Calculate the attenuation coefficient based on the attenuation coefficient lookup table;

[0023] S1013. Calculate the intensity parameters of signal-dependent noise based on the intensity parameter lookup table of signal-dependent noise;

[0024] S1014. Calculate the basic intensity parameters of Gaussian noise based on the lookup table of basic intensity parameters of Gaussian noise;

[0025] S1015. Calculate the Gaussian noise at each pixel location based on the basic intensity parameters of Gaussian noise.

[0026] As a further improvement of the present invention: step S102 includes:

[0027] S1021. Calculate motion vector components based on optical flow calculation;

[0028] S1022. Calculate motion fuzziness length based on motion vector components;

[0029] S1023. Calculate the motion direction angle based on the motion vector components.

[0030] As a further improvement of the present invention: step S103 includes:

[0031] S1031. Calculate the transmission diagram based on the attenuation coefficient;

[0032] S1032. Calculate the defocus blur radius based on water quality type.

[0033] As a further improvement of the present invention: step S2 includes:

[0034] S201. Calculate the set of wavelet subbands of a clear image;

[0035] S202. Calculate the wavelet packet set of each underwater image and the set of the two adjacent underwater images before and after it;

[0036] S203. Calculate the frequency domain residual based on the wavelet packet set of the clear image and the wavelet packet set of the underwater image;

[0037] S204. Construct a set of residual subbands based on frequency domain residuals.

[0038] As a further improvement of the present invention: step S4 includes:

[0039] S401. Calculate the reconstruction loss based on the sharp image and the corresponding restored image;

[0040] S402. Calculate frequency domain loss based on wavelet packet transform;

[0041] S403. Calculate the sensing loss based on the VGG network;

[0042] S404. Introduce a discriminator to calculate adversarial loss;

[0043] S405. Calculate physical consistency loss based on regenerated image;

[0044] S406. Calculate the total loss function based on reconstruction loss, frequency domain loss, perceptual loss, adversarial loss, and physical consistency loss.

[0045] Preferably, the formula for calculating the total loss function in S406 is:

[0046] ,

[0047] In the formula, To reconstruct the loss weights; For reconstruction loss; For frequency domain loss weights; For frequency domain loss; Weights for perceived loss; To perceive loss; To counteract the loss of weight; To combat the losses; Weights for physical consistency loss; This represents a loss of physical consistency.

[0048] The beneficial effects of the present invention are as follows: The underwater image enhancement method based on the adaptive composite degradation model provided by the present invention has high versatility in underwater scenes and the enhanced underwater image has high detail.

[0049] First, the method constructs an adaptive composite degradation model to degrade clear images. The adaptive composite degradation model integrates motion blur, spatial variation defocus blur, light transmission attenuation, and signal-dependent noise, and has high physical accuracy. The model achieves adaptation to different underwater scenarios through three two-dimensional lookup tables. This dynamic parameter adjustment mechanism can avoid the limitations of traditional fixed models and improve the underwater scenario versatility of the method.

[0050] Secondly, this method constructs a set of residual subbands for each sharp image and its corresponding underwater image based on wavelet packet decomposition, and decomposes each sharp image and underwater image into a set using wavelet packet transformation. The layer extracts multi-scale sub-bands and calculates residuals to complete frequency domain feature extraction and dual residual learning. It combines spatial frequency domain residuals with frequency domain inverse transform to protect the path of high-frequency detail recovery and improve the detail fineness of the enhanced underwater image.

[0051] Finally, the overall loss function of this method is a hybrid loss function that integrates reconstruction loss, frequency domain loss, perceptual loss, adversarial loss, and physical consistency loss. It is suitable for training on unpaired data. The self-supervised mechanism and multi-level optimization strategy also ensure that the trained generator has strong interpretability and robustness, enabling this method to stably enhance underwater images. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the principle of the present invention. Detailed Implementation

[0053] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0054] like Figure 1 As shown, the underwater image enhancement method based on an adaptive composite degradation model provided by this invention includes:

[0055] S1. Construct an adaptive composite degradation model to obtain the underwater images corresponding to each clear image;

[0056] S101. Calculate parameters related to water quality and depth. These parameters are mainly affected by water quality and depth, and include attenuation coefficients. Signal-dependent noise intensity parameters and the basic intensity parameters of Gaussian noise ;

[0057] S1011. Create three readable two-dimensional lookup tables. The three two-dimensional lookup tables are: a lookup table for attenuation coefficients providing corresponding attenuation coefficients for different water quality types and depths; and an attenuation coefficient lookup table. A lookup table for signal-dependent noise intensity parameters, providing corresponding parameters for different water quality types and depths. And a lookup table of basic Gaussian noise intensity parameters that provides corresponding basic intensity parameters for different water quality types and average depths of different underwater images. ;

[0058] S10111. Use water quality type express, , This indicates that the water quality type is open water, relatively clear, and characterized by low attenuation, low scattering, and low noise. This indicates that the water quality type is coastal water, characterized by moderate attenuation and moderate scattering. This refers to port waters, which are relatively turbid and characterized by high attenuation, high scattering, and high noise.

[0059] The average depth of the underwater image is represented as... ;

[0060] S10112. Through experiments or numerical simulations, the attenuation coefficient, signal-dependent noise intensity parameters, and basic Gaussian noise intensity parameters for different water quality types and depths are measured. An attenuation coefficient lookup table is then compiled based on the different water quality types, depths, and attenuation coefficients. A lookup table of signal-dependent noise intensity parameters was compiled by organizing the intensity parameters of different water quality types, different depths, and signal-dependent noise. A lookup table of basic intensity parameters for Gaussian noise was compiled by organizing the basic intensity parameters of Gaussian noise for different water quality types, different depths, and Gaussian noise levels. This allows for the calculation of attenuation coefficients, signal-dependent noise intensity parameters, or Gaussian noise fundamental intensity parameters for a given water area. The solution involves simply searching the relevant two-dimensional lookup tables based on the water quality type and depth of the water area, enabling adaptive changes to these parameters without requiring complex derivations each time.

[0061] S1012. Calculate the attenuation coefficient based on the attenuation coefficient lookup table;

[0062] Attenuation coefficient The calculation formula is:

[0063] ;

[0064] S1013. Calculate the intensity parameters of signal-dependent noise based on the intensity parameter lookup table of signal-dependent noise. Signal-dependent noise includes Poisson noise and shot noise.

[0065] Signal-dependent noise intensity parameters The calculation formula is:

[0066] ;

[0067] S1014. Calculate the basic intensity parameters of Gaussian noise based on the lookup table of basic intensity parameters of Gaussian noise;

[0068] Basic intensity parameters of Gaussian noise The calculation formula is:

[0069] ;

[0070] S1015. Calculate the Gaussian noise at each pixel location based on the basic intensity parameters of Gaussian noise;

[0071] The Gaussian noise at each pixel location is the spatial variation of the noise standard deviation. The calculation formula is:

[0072] ,

[0073] In the formula, This is a spatial variability intensity parameter used to control the degree to which noise varies with the transmission rate. Typically, ; This is a transmission map, representing pixel positions. The transmission rate value at that location, ; These are the spatial coordinates of the pixel, i.e., the pixel position; This represents the horizontal position of a pixel. The vertical position of the pixel;

[0074] S102. The motion-related parameters are adaptively calculated using the optical flow between underwater images from the input video. These motion-related parameters are determined by the camera's motion state and include motion direction angles. Motion blur length ;

[0075] S1021. Calculate motion vector components based on optical flow calculation;

[0076] The horizontal component of the motion vector components between each underwater image obtained through optical flow analysis. With vertical component ;

[0077] The formula for calculating the motion vector components is:

[0078] ,

[0079] In the formula, It is a median function that calculates the median of all motion vectors in the optical flow field and is robust to outliers. For optical flow calculation function, Item calculation from the first Zhang underwater images to the first The dense optical flow field of an underwater image, where the optical flow displacement needs to be multiplied by a scaling factor; For the first Zhang underwater images; For the first Zhang underwater images, and The relationship is that they are two consecutive underwater images in time;

[0080] S1022. Calculate motion fuzziness length based on motion vector components;

[0081] Motion blur length The calculation formula is:

[0082] ,

[0083] In the formula, This is the scaling factor; The Euclidean distance is the distance between the horizontal component and the vertical component of the motion vector between underwater images.

[0084] S1023. Calculate the motion direction angle based on motion vector components;

[0085] Direction and angle of motion The calculation formula is:

[0086] ,

[0087] In the formula, It is the arctangent function in the four quadrants;

[0088] S103. Calculate the parameters related to the depth map, which depend on the depth of each pixel. These parameters include the defocus blur radius. With transmission diagram ;

[0089] S1031. Calculate the transmission diagram based on the attenuation coefficient;

[0090] The transmission graph represents the light transmission rate at a pixel location, with values ​​ranging from [0,1]. The calculation formula is:

[0091] ,

[0092] In the formula, It is an exponential function; The attenuation coefficient, also known as the beam attenuation coefficient, characterizes the degree of light attenuation in water. This is a depth map, representing the scene depth at pixel locations. Scene depth value (actual scene depth);

[0093] S1032. Calculate the defocusing blur radius based on water quality type;

[0094] The defocus blur radius is the radius of the defocus blur circle at the pixel location. The calculation formula is:

[0095] ,

[0096] In the formula, The maximum fuzzy radius; This refers to the focusing distance, specifically the focusing distance of the camera lens. The water quality scaling factor is a fuzzy intensity scaling coefficient based on the water quality type. At that time, open water can be considered as pure geometric optics, therefore, the following is set: ,when At 2 o'clock, scattering in coastal waters intensifies blurring, therefore, the setting is... ,when At 3 o'clock, the forward scattering effect in the port waters exacerbates optical blur, making the defocus more pronounced even with accurate focus, resulting in severe blur. Therefore, the setting is... After introducing the water quality scaling factor, the defocus blur radius is directly related to the water quality type. Applying the defocus blur radius to the adaptive composite degradation model enables the adaptive composite degradation model to adapt to different underwater environments, from clear open water to turbid port waters, significantly improving the model's practicality and accuracy.

[0097] S104. An adaptive composite degradation model for generating corresponding underwater images for clear images is constructed based on parameters related to water quality and depth, parameters related to motion state, and parameters related to depth map.

[0098] Underwater images are degraded images, the first Zhang underwater images The calculation formula (i.e., the expression for the adaptive composite degradation model) is:

[0099] ,

[0100] In the formula, For the first Zhang Qingquan image; This is the convolution operator; The diffusion function for motion blur points; The spread function for defocused and blurred points; This is an element-wise multiplication (Hadamard product). For transmission graph functions; This is a signal-dependent noise term; The image after blurring and attenuation; This is a Gaussian noise term; The standard deviation of spatially varying Gaussian noise;

[0101] The adaptive composite degradation model simulates the image blurring effects caused by motion blur and optical defocus through convolution operations, fully characterizing the point spread characteristics of the underwater imaging system. It employs a transfer graph function to perform element-wise attenuation processing on the blurred image, accurately reflecting the exponential attenuation law of light waves in water. In the noise simulation stage, the model innovatively integrates signal-dependent noise related to signal strength and Gaussian noise related to depth and water quality, comprehensively covering the noise sources of underwater imaging. This is a highly accurate and adaptive physical process of underwater image degradation. All key parameters used in the model possess environmental adaptability, giving the model excellent self-adaptability.

[0102] S2. Based on wavelet packet decomposition, construct the residual subband set of each clear image and its corresponding underwater image;

[0103] S201. Calculate the set of wavelet subbands of a clear image;

[0104] Set of wavelet packet bands of clear images The calculation formula is:

[0105] ,

[0106] In the formula, For the first clear image A small wave of buns; Wavelet packet transform;

[0107] S202. Calculate the wavelet packet set for each underwater image and its two adjacent preceding and following underwater images, that is, for three consecutive underwater images (the first one...), the wavelet packet set is calculated. Zhang underwater images , No. Zhang underwater images , No. Zhang underwater images Perform wavelet packet transform to decompose to the nth layer, To determine the number of decomposition layers, N is usually set to 3 or N=4;

[0108] No. Wavelet packet set of underwater images The calculation formula is:

[0109] ,

[0110] In the formula, For the first The first underwater image A small wave of buns; This is the index for the number of wavelet packet bands. , The total number of sub-bands, ;

[0111] From the above formula, we can obtain:

[0112] No. Wavelet packet set of underwater images The calculation formula is:

[0113] ;

[0114] No. Wavelet packet set of underwater images The calculation formula is:

[0115] ;

[0116] S203. Calculate the frequency domain residual based on the wavelet bag set of the clear image and the wavelet bag set of the underwater image to achieve feature extraction;

[0117] The calculation of frequency domain residuals is based on the difference in wavelet subband sets between the sharp image and the underwater image. The calculation formula is:

[0118] ;

[0119] S204. Constructing a set of residual subbands based on frequency domain residuals;

[0120] Residual subband set Represented as:

[0121] ;

[0122] S3. Construct a generator based on the residual subband set, input the underwater image into the generator to obtain the restored image, i.e. the enhanced image;

[0123] The generator takes as input a set of residual subbands. With three consecutive underwater images;

[0124] The generator outputs a restored image. ;

[0125] The generator employs an encoder-bottleneck-decoder backbone architecture, which is a U-Net-like architecture. U-Net is a CNN (Convolutional Neural Network) model proposed in 2015, primarily used to solve pixel-level image prediction problems, such as image segmentation and edge extraction. The generator includes a multimodal input module, a physically guided residual block group, and a dual-path output head, reconstructing the enhanced image through identity mapping and inverse wavelet transform.

[0126] The multimodal input module includes the fusion of the original image, wavelet subbands, physical parameters, and optical flow features from adjacent underwater images;

[0127] The dual-path output head is used to generate the image spatial domain residual and wavelet subbands, respectively;

[0128] S301. The first The underwater image and its corresponding wavelet packet (after stitching) are input into a multi-scale encoder, and multi-scale features are extracted through the convolution module and downsampling module in the multi-scale encoder.

[0129] S302. The first Zhang underwater images and the first The multi-scale features of underwater images are introduced into the inter-frame attention module. The inter-frame attention module uses optical flow alignment and attention mechanism to fuse complementary information from adjacent images to compensate for motion blur and information loss.

[0130] S3021. Collect the results calculated in S202. The wavelet packet set of underwater images, the first The wavelet packet set of underwater images, the first The set of wavelet packet bands in an underwater image;

[0131] S3022. Use an upsampling algorithm to unify the size of all wavelet packet subbands in each underwater image;

[0132] The upsampling algorithm is bilinear interpolation;

[0133] With the first Zhang underwater images (the first) Zhang underwater image, the first Taking the underwater image as an example (the same operation is performed on the same image), we will explain in detail: using the upsampling algorithm to... The first underwater image The size of the small wavelet packet is reset to the first... The height of Zhang's underwater images With width , obtained the The first underwater image The size of the small wavelet has been standardized. ;

[0134] S3023. Stitch together all wavelet packets of uniform size from each underwater image along the channel dimension to form a single multichannel tensor;

[0135] With the first Taking an underwater image as an example, the wavelet packet bands with uniform size obtained by upsampling all of its wavelet packet bands are: the first... The first underwater image The size of the small wavelet has been standardized. , No. The first underwater image The size of the small wavelet has been standardized. ,..., No. The first underwater image The size of the small wavelet has been standardized. By splicing along the channel dimension, the first part is formed. Multichannel tensors of underwater images , No. The dimension of the multichannel tensor of the underwater image is: .

[0136] S303. The estimated transmission graph is obtained by predicting using the intermediate layer of the generator. Motion fuzzy kernel Depth map The estimated transmission map Motion fuzzy kernel Depth map In the residual blocks guided by the input physical model, the weights of the convolutional layers are dynamically modulated, so that the network learning process is constrained by the physical model; through the residual learning framework, the learning objective of the network is limited to reversing the physical degradation process.

[0137] S304. Use a wavelet packet decoder to decode the upsampled features and predict the image spatial domain residual. Frequency domain residual ∆Wl;

[0138] S305. Obtain the restored image based on the spatial domain residual and frequency domain residual, and complete the image reconstruction;

[0139] No. The formula for calculating the restored image is:

[0140] ,

[0141] In the formula, For generator functions; For image spatial domain residuals; This is the inverse wavelet packet transform;

[0142] S4. Calculate the total loss function based on reconstruction loss, frequency domain loss, perceptual loss, adversarial loss, and physical consistency loss;

[0143] S401. Calculate the reconstruction loss based on the clear image and the corresponding restored image to ensure pixel-level accuracy;

[0144] Reconstruction loss The calculation formula is:

[0145] ,

[0146] In the formula, To restore the image; It is an L2 norm; This is a smoothing parameter with a value of 10. -6 ~10 -3 ;

[0147] S402. Calculate frequency domain loss based on wavelet packet transform. Frequency domain loss is used to measure the difference in frequency domain features to force the restored image to be consistent with the clear image in the frequency domain. Special attention is paid to high-frequency sub-bands, which directly constrains the network to recover the correct high-frequency information.

[0148] Frequency domain loss The calculation formula is:

[0149] ,

[0150] In the formula, It is an L1 norm;

[0151] S403. Calculate perceptual loss based on VGG network, and use pre-trained VGG (Visual Geometry Group) network to extract features to ensure semantic and structural similarity.

[0152] Perceived loss The calculation formula is:

[0153] ,

[0154] In the formula, This is the extraction function for the VGG network, where VGG represents deep features; For network layer indexing;

[0155] S404. Introduce a discriminator to calculate adversarial loss;

[0156] Adversarial loss The calculation formula is:

[0157] ,

[0158] In the formula, To maximize the output value of the discriminator; This is the expectation operator, representing the average of all samples. This is the discriminator function, used to output the probability that the input image is a real image;

[0159] By introducing a discriminator, advanced semantic constraints on the realism of the restored image are provided, serving as a powerful regularization method.

[0160] S405. Calculate physical consistency loss based on regenerated image;

[0161] Physical consistency loss The calculation formula is:

[0162] ,

[0163] In the formula, To regenerate the image, that is, to re-degrade the image using the restored image;

[0164] Substituting the restored image into the adaptive composite degradation model in S1, the result should again approximate the first... Zhang underwater images;

[0165] Regenerated Image The calculation formula is:

[0166] ,

[0167] In the formula, To estimate the motion fuzzy point spread function estimated by the motion fuzzy kernel network; To estimate the defocused blur point spread function estimated by the defocused blur kernel network; To estimate the transmission graph, i.e., the optical transmission graph estimated by the network; To estimate the noise, i.e. the noise term estimated by the network, it includes signal-dependent noise and Gaussian noise;

[0168] Estimated transmission graph The calculation formula is:

[0169] ,

[0170] In the formula, To estimate the water body attenuation coefficient; To estimate the scene depth map;

[0171] S406. Calculate the total loss function;

[0172] Total loss function The calculation formula is:

[0173] ,

[0174] In the formula, The reconstructed loss weights are used to control the importance of pixel-level fidelity. These are frequency domain loss weights used to control the importance of frequency domain details; The perceived loss weights are used to control the importance of semantic consistency; To counteract loss weights, used to control the importance of generating realism; The physical consistency loss weight is used to control the importance of physical consistency;

[0175] S5. Optimize the total loss function using a large underwater image dataset to train the generator and discriminator. After training, save the generator's parameters to obtain the trained generator.

[0176] The underwater image dataset includes paired and unpaired data;

[0177] The paired data includes a large number of corresponding underwater image pairs with sharp images;

[0178] The data consists of a large number of unpaired underwater images and clear images that do not correspond to each other;

[0179] Training is complete when the total loss function converges and the discriminator cannot effectively distinguish between different loss functions.

[0180] The generator's parameters include all weights, all biases, and the state of all parameterized modules.

[0181] S6. Input the newly acquired underwater image sequence into the trained generator to obtain the enhanced newly acquired underwater image sequence in real time.

Claims

1. An underwater image enhancement method based on an adaptive composite degradation model, characterized in that, include: S1. Construct an adaptive composite degradation model to obtain the underwater images corresponding to each clear image; The steps in S1 include: S101. Calculate parameters related to water quality and depth, including attenuation coefficient, signal-dependent noise intensity parameters, and Gaussian noise basic intensity parameters. S102. Calculate the parameters related to the motion state, including the motion direction angle and the motion fuzz length; S103. Calculate the parameters related to the depth map, including the defocus blur radius and the transmission map; S104. An adaptive composite degradation model for generating corresponding underwater images for clear images is constructed based on parameters related to water quality and depth, parameters related to motion state, and parameters related to depth map. S2. Based on wavelet packet decomposition, construct the residual subband set of each clear image and its corresponding underwater image; S3. Construct a generator based on the residual subband set, input the underwater image into the generator to obtain the restored image, i.e. the enhanced image; S4. Calculate the total loss function based on reconstruction loss, frequency domain loss, perceptual loss, adversarial loss, and physical consistency loss; S5. Optimize the total loss function using a large underwater image dataset to train the generator and discriminator. After training, save the generator's parameters to obtain the trained generator. S6. Input the newly acquired underwater image sequence into the trained generator to obtain an enhanced sequence of newly acquired underwater images.

2. The underwater image enhancement method based on an adaptive composite degradation model according to claim 1, characterized in that, The steps in S101 include: S1011. Create three two-dimensional lookup tables: an attenuation coefficient lookup table, a signal-dependent noise intensity parameter lookup table, and a Gaussian noise basic intensity parameter lookup table. S1012. Calculate the attenuation coefficient based on the attenuation coefficient lookup table; S1013. Calculate the intensity parameters of signal-dependent noise based on the intensity parameter lookup table of signal-dependent noise; S1014. Calculate the basic intensity parameters of Gaussian noise based on the lookup table of basic intensity parameters of Gaussian noise; S1015. Calculate the Gaussian noise at each pixel location based on the basic intensity parameters of Gaussian noise.

3. The underwater image enhancement method based on an adaptive composite degradation model according to claim 1, characterized in that, The steps in S102 include: S1021. Calculate motion vector components based on optical flow calculation; S1022. Calculate motion fuzziness length based on motion vector components; S1023. Calculate the motion direction angle based on the motion vector components.

4. The underwater image enhancement method based on an adaptive composite degradation model according to claim 1, characterized in that, The steps in S103 include: S1031. Calculate the transmission diagram based on the attenuation coefficient; S1032. Calculate the defocus blur radius based on water quality type.

5. The underwater image enhancement method based on an adaptive composite degradation model according to claim 1, characterized in that, Step S2 includes: S201. Calculate the set of wavelet subbands of a clear image; S202. Calculate the wavelet packet set of each underwater image and the set of the two adjacent underwater images before and after it; S203. Calculate the frequency domain residual based on the wavelet packet set of the clear image and the wavelet packet set of the underwater image; S204. Construct a set of residual subbands based on frequency domain residuals.

6. The underwater image enhancement method based on an adaptive composite degradation model according to claim 1, characterized in that, The steps in S4 include: S401. Calculate the reconstruction loss based on the sharp image and the corresponding restored image; S402. Calculate frequency domain loss based on wavelet packet transform; S403. Calculate the sensing loss based on the VGG network; S404. Introduce a discriminator to calculate adversarial loss; S405. Calculate physical consistency loss based on regenerated image; S406. Calculate the total loss function based on reconstruction loss, frequency domain loss, perceptual loss, adversarial loss, and physical consistency loss.

7. The underwater image enhancement method based on an adaptive composite degradation model according to claim 6, characterized in that, The formula for calculating the total loss function in S406 is as follows: ， In the formula, To reconstruct the loss weights; For reconstruction loss; For frequency domain loss weights; For frequency domain loss; Weights for perceived loss; To perceive loss; To counteract the loss of weight; To combat the losses; Weights for physical consistency loss; This represents a loss of physical consistency.

Images