Polarization physics constraint and pseudo-label guided unsupervised underwater image enhancement method

Abstract

The application discloses a polarization physical constraint and pseudo-label guided unsupervised underwater image enhancement method, belongs to the field of computer vision underwater image enhancement, and is used for underwater image enhancement and comprises the following steps: acquiring an underwater RGB image as a data set, constructing a double-branch collaborative network, calculating cross-reconstruction consistency loss, clear content consistency loss and polarization degree transmission map space consistency loss in a training stage, combining the three kinds of consistency losses to construct an overall loss function, and optimizing network parameters of the double-branch collaborative network based on the overall loss function; in an inference stage, inputting the underwater RGB image into only a polarization physical prior constraint main branch of the double-branch collaborative network, and outputting a final scene radiance map. Through the double-branch collaborative network, the application can effectively improve the underwater image de-scattering capability, the detail recovery capability and the visual clarity under the unsupervised condition, and is suitable for underwater image enhancement in a high turbidity medium environment.

Description

Technical Field

[0001] This invention discloses an unsupervised underwater image enhancement method guided by polarization physics constraints and pseudo-tags, belonging to the field of underwater image enhancement in computer vision. Background Technology

[0002] Underwater image enhancement is a key fundamental problem in underwater visual perception, and it is widely used in scenarios such as marine resource exploration, underwater robot navigation, and seabed ecological monitoring. Due to the selective absorption of light by water and the scattering effect of suspended particles, underwater images often suffer from reduced contrast, color distortion, and blurred details, which affect subsequent downstream tasks such as recognition, detection, and localization.

[0003] Most current mainstream underwater image enhancement methods rely primarily on RGB information for image restoration. While these methods can improve image brightness, contrast, and color performance to some extent, they struggle to effectively separate reflected light from backscattered light when relying solely on RGB intensity information. In contrast, polarization information provides physical clues related to scattering, which helps improve the accuracy of underwater degradation modeling.

[0004] Existing underwater image enhancement methods incorporating polarization information mostly rely on supervised training with paired, clear reference images, but such data is difficult to obtain in real turbid waters. While existing unsupervised methods alleviate label dependence to some extent, they still suffer from inaccurate estimation of background light and transmittance under high turbidity conditions, and insufficient stability of the enhancement results. Furthermore, some methods utilize existing enhancement algorithms to generate pseudo-labels to assist network training; however, due to limitations of the enhancement methods, pseudo-labels often contain noise, artifacts, or color distortion, which can easily mislead the network learning if directly used for supervision. Therefore, it is necessary to propose an unsupervised underwater image enhancement method based on polarization physical constraints and indirect guidance by pseudo-labels. Summary of the Invention

[0005] The purpose of this invention is to provide an unsupervised underwater image enhancement method guided by polarization physical constraints and pseudo-tags, so as to solve the problems in the prior art where the lack of real reference images and the insufficient adaptability of single physical priors lead to poor enhancement effects.

[0006] Polarization physics constraints and pseudo-label-guided unsupervised underwater image enhancement methods include: Underwater RGB images were acquired as a dataset, and a two-branch collaborative network was constructed. The dual-branch collaborative network consists of a main branch constrained by polarization physics priors, an auxiliary branch indirectly guided by pseudo-labels, and a loss layer indirectly guided by pseudo-labels. During the training phase, the pseudo-label indirectly guided loss layer calculates the cross-reconstruction consistency loss, clear content consistency loss, and polarization degree transmission map spatial consistency loss based on the processing results of the main branch and the auxiliary branch indirectly guided by the pseudo-label, respectively. The three consistency losses are combined to construct an overall loss function, and the network parameters of the dual-branch collaborative network are optimized based on the overall loss function. During the inference phase, the underwater RGB image is input only into the main branch of the polarization physics prior constraint of the dual-branch collaborative network, and the final scene radiation map is output. The main branch of the polarization physics prior constraint includes a polarization prior-guided transmission map estimation network and a background light estimation module; The polarization prior-guided transmission map estimation network includes a polarization feature extraction module, an RGB encoder, a polarization-guided sensing module, a decoder, and a transmission map estimation head; The auxiliary branches indirectly guided by pseudo-tags include a pseudo-tag transmissive estimation network and a background light estimation module; the pseudo-tag transmissive estimation network includes an encoder, a decoder, and a transmissive estimation head.

[0007] Acquiring underwater RGB images And calculate the corresponding polarization degree diagram. Several underwater image enhancement methods were used to respectively... Enhancement is performed to obtain corresponding candidate pseudo-label images. The final pseudo-label image is then determined using preset no-reference image quality evaluation metrics and photometric safety discrimination. ; Will and The main branch of the input polarization physics prior constraints will The auxiliary branch is indirectly guided by the input pseudo-label.

[0008] Will Input the RGB encoder for feature extraction and output the RGB backbone features; The input polarization feature extraction module extracts features and outputs polarization features. ; RGB backbone features and Input polarization-guided sensing module, The polarization-guided fusion feature is injected into the RGB backbone feature as a conditional guide term to obtain the polarization-guided fusion feature. The polarization-guided fusion feature is then input into the decoder and the transmission map estimation head in sequence to output the estimated transmission map. ; Will Input the background light estimation module, first construct a brightness map : ; In the formula, for In coordinates The value at that location, for In the The pixel values ​​of each color channel. For color channel index, It is a collection of three channels: red, green, and blue. An adaptive two-dimensional Gaussian kernel is used for smoothing filtering, and a reflection filling strategy is employed during the filtering process to obtain the Gaussian low-frequency components. : ; In the formula, for In coordinates The value at that location, It is a two-dimensional Gaussian kernel. This is a convolution operation; extract Global median of pixel values ,by As a robust brightness benchmark for background light estimation, combined with Calculate the background light grayscale estimation image : ; In the formula, for In coordinates The value at that location, These are weighting coefficients used to balance low-frequency spatial variations with global brightness stability; calculate In the Global mean across each color channel : ; In the formula, for height, for The width; Construct color scaling coefficients based on the global mean of each color channel. : ; In the formula, for In the Global mean across each color channel For temporary color channel indexes, It is a numerically stable term; Calculate the first Background light estimation results for each color channel : ; Background light estimation results from the red channel Green channel background light estimation results And blue channel background light estimation results The main branch background light estimation results, which together constitute the polarization physics prior constraints, are as follows: : ; Introducing an underwater degradation imaging model: ; In the formula, for In coordinates Observations at that location This is a scene radiation map. for In coordinates The value at that location, for In coordinates The estimated value at that location; use and right Perform reverse reconstruction to obtain the scene radiation map. : ; based on Constructing a transmission map reference item : ; In the formula, As a preset constant, for In coordinates The value at; right and Normalization is performed to obtain the normalized transmission image. and normalized transmission reference terms ,use and Constructing a polarization degree-transmission map spatial consistency loss function : ; In the formula, This is the spatial consistency constraint function.

[0009] The encoder, decoder, and transmission map estimation head of the pseudo-label transmission map estimation network are input sequentially, and the output is the pseudo-label transmission map. ;Will Input background light estimation module, output pseudo-label background light estimation result ; Based on the underwater degradation imaging model, and right Perform reverse recovery to obtain the pseudo-label scene radiation map. : ; In the formula, for In coordinates Observations at that location for In coordinates The value at that location, for In coordinates The estimated value at that location.

[0010] Cross-consistency reconstruction loss includes, based on , and Construct the first cross-reconstruction result : ; In the formula, For pixel-by-pixel multiplication, This is a gradient truncation operation; based on , and Construct the second cross-reconstruction result : ; Constructing cross-reconstruction consistency loss : ; In the formula, for Norm; based on and Constructing a clear content consistency loss : ; Constructing pseudo-labels to indirectly guide loss : .

[0011] Construct the overall loss function : ; In the formula, for Weighting coefficients; based on Backpropagation and network parameter updates are performed on the two-branch cooperative network.

[0012] The polarization prior-guided transmission map estimation network includes four polarization-guided sensing modules, and the polarization feature extraction module receives... Output and to Perform downsampling once, twice, and three times respectively. Input the first polarization-guided sensing module, input the first downsampling result into the second polarization-guided sensing module, input the second downsampling result into the third polarization-guided sensing module, and input the third downsampling result into the fourth polarization-guided sensing module.

[0013] The RGB encoder consists of four stages in sequence. The first stage of the RGB encoder is the Conv layer and the ReLU activation function. The second stage of the RGB encoder is the MaxPool layer, the Conv layer and the ReLU activation function. The third stage of the RGB encoder is the MaxPool layer, the Conv layer and the ReLU activation function. The fourth stage of the RGB encoder is the MaxPool layer, the Conv layer and the ReLU activation function. The RGB encoder receives underwater RGB images and processes them in four stages. The result of the first stage is input to the first polarization-guided sensing module, the result of the second stage is input to the second polarization-guided sensing module, the result of the third stage is input to the third polarization-guided sensing module, and the result of the fourth stage is input to the fourth polarization-guided sensing module.

[0014] The decoder consists of four stages in sequence: the first stage is the Resblock layer, the second stage is the Bilinear layer, the Conv layer and the ReLU activation function, the third stage is the Bilinear layer, the Conv layer and the ReLU activation function, and the fourth stage is the Bilinear layer, the Conv layer and the ReLU activation function. The decoder in the first stage receives and processes the output of the fourth polarization-guided sensing module; the decoder in the second stage receives and processes the output of the third polarization-guided sensing module and the output of the first stage; the decoder in the third stage receives and processes the output of the second polarization-guided sensing module and the output of the second stage; and the decoder in the fourth stage receives and processes the output of the first polarization-guided sensing module and the output of the third stage.

[0015] The transmission estimation head consists of a Resblock layer, a 1×1 Conv layer, and a Sigmoid activation function. The transmission estimation head receives the decoder's processing results and finally outputs a transmission map.

[0016] Compared with existing technologies, this invention has the following advantages: By jointly modeling polarization information and RGB image information, this invention introduces polarization physical constraints, improving the accuracy and physical rationality of transmission map estimation; by designing auxiliary branches indirectly guided by pseudo-labels and a cross-consistency reconstruction mechanism, it realizes the indirect guidance of the main branch by pseudo-labels, reducing the error propagation caused by direct supervision by imperfect pseudo-labels; furthermore, by constructing an overall loss function composed of cross-reconstruction consistency loss, clear content consistency loss, and polarization degree-transmission map spatial consistency loss, the network can achieve joint optimization under the condition of no real clear reference image; at the same time, only the output enhancement results of the main branch are retained in the inference stage, thereby reducing model complexity and improving inference efficiency while ensuring the enhancement effect. Attached Figure Description

[0017] Figure 1 This is a diagram of the dual-branch collaborative network architecture of the present invention; Figure 2 This is a diagram of the polarization prior-guided transmission map estimation network architecture of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] Polarization physics constraints and pseudo-label-guided unsupervised underwater image enhancement methods include: Underwater RGB images were acquired as a dataset, and a two-branch collaborative network was constructed. The dual-branch collaborative network consists of a main branch constrained by polarization physics priors, an auxiliary branch indirectly guided by pseudo-labels, and a loss layer indirectly guided by pseudo-labels. During the training phase, the pseudo-label indirectly guided loss layer calculates the cross-reconstruction consistency loss, clear content consistency loss, and polarization degree transmission map spatial consistency loss based on the processing results of the main branch and the auxiliary branch indirectly guided by the pseudo-label, respectively. The three consistency losses are combined to construct an overall loss function, and the network parameters of the dual-branch collaborative network are optimized based on the overall loss function. During the inference phase, the underwater RGB image is input only into the main branch of the polarization physics prior constraint of the dual-branch collaborative network, and the final scene radiation map is output. The main branch of the polarization physics prior constraint includes a polarization prior-guided transmission map estimation network and a background light estimation module; The polarization prior-guided transmission map estimation network includes a polarization feature extraction module, an RGB encoder, a polarization-guided sensing module, a decoder, and a transmission map estimation head; The auxiliary branches indirectly guided by pseudo-tags include a pseudo-tag transmissive estimation network and a background light estimation module; the pseudo-tag transmissive estimation network includes an encoder, a decoder, and a transmissive estimation head.

[0020] Acquiring underwater RGB images And calculate the corresponding polarization degree diagram. Several underwater image enhancement methods were used to respectively... Enhancement is performed to obtain corresponding candidate pseudo-label images. The final pseudo-label image is then determined using preset no-reference image quality evaluation metrics and photometric safety discrimination. ; Will and The main branch of the input polarization physics prior constraints will The auxiliary branch is indirectly guided by the input pseudo-label.

[0021] Will Input the RGB encoder for feature extraction and output the RGB backbone features; The input polarization feature extraction module extracts features and outputs polarization features. ; RGB backbone features and Input polarization-guided sensing module, The polarization-guided fusion feature is injected into the RGB backbone feature as a conditional guide term to obtain the polarization-guided fusion feature. The polarization-guided fusion feature is then input into the decoder and the transmission map estimation head in sequence to output the estimated transmission map. ; Will Input the background light estimation module, first construct a brightness map : ; In the formula, for In coordinates The value at that location, for In the The pixel values ​​of each color channel. For color channel index, It is a collection of three channels: red, green, and blue. An adaptive two-dimensional Gaussian kernel is used for smoothing filtering, and a reflection filling strategy is employed during the filtering process to obtain the Gaussian low-frequency components. : ; In the formula, for In coordinates The value at that location, It is a two-dimensional Gaussian kernel. This is a convolution operation; extract Global median of pixel values ,by As a robust brightness benchmark for background light estimation, combined with Calculate the background light grayscale estimation image : ; In the formula, for In coordinates The value at that location, These are weighting coefficients used to balance low-frequency spatial variations with global brightness stability; calculate In the Global mean across each color channel : ; In the formula, for height, for The width; Construct color scaling coefficients based on the global mean of each color channel. : ; In the formula, for In the Global mean across each color channel For temporary color channel indexes, It is a numerically stable term; Calculate the first Background light estimation results for each color channel : ; Background light estimation results from the red channel Green channel background light estimation results And blue channel background light estimation results The main branch background light estimation results, which together constitute the polarization physics prior constraints, are as follows: : ; Introducing an underwater degradation imaging model: ; In the formula, for In coordinates Observations at that location This is a scene radiation map. for In coordinates The value at that location, for In coordinates The estimated value at that location; use and right Perform reverse reconstruction to obtain the scene radiation map. : ; based on Constructing a transmission map reference item : ; In the formula, As a preset constant, for In coordinates The value at; right and Normalization is performed to obtain the normalized transmission image. and normalized transmission reference terms ,use and Constructing a polarization degree-transmission map spatial consistency loss function : ; In the formula, This is the spatial consistency constraint function.

[0022] The encoder, decoder, and transmission map estimation head of the pseudo-label transmission map estimation network are input sequentially, and the output is the pseudo-label transmission map. ;Will Input background light estimation module, output pseudo-label background light estimation result ; Based on the underwater degradation imaging model, and right Perform reverse recovery to obtain the pseudo-label scene radiation map. : ; In the formula, for In coordinates Observations at that location for In coordinates The value at that location, for In coordinates The estimated value at that location.

[0023] Cross-consistency reconstruction loss includes, based on , and Construct the first cross-reconstruction result : ; In the formula, For pixel-by-pixel multiplication, This is a gradient truncation operation; based on , and Construct the second cross-reconstruction result : ; Constructing cross-reconstruction consistency loss : ; In the formula, for Norm; based on and Constructing a clear content consistency loss : ; Constructing pseudo-labels to indirectly guide loss : .

[0024] Construct the overall loss function : ; In the formula, for Weighting coefficients; based on Backpropagation and network parameter updates are performed on the two-branch cooperative network.

[0025] The polarization prior-guided transmission map estimation network includes four polarization-guided sensing modules, and the polarization feature extraction module receives... Output and to Perform downsampling once, twice, and three times respectively. Input the first polarization-guided sensing module, input the first downsampling result into the second polarization-guided sensing module, input the second downsampling result into the third polarization-guided sensing module, and input the third downsampling result into the fourth polarization-guided sensing module.

[0026] The RGB encoder consists of four stages in sequence. The first stage of the RGB encoder is the Conv layer and the ReLU activation function. The second stage of the RGB encoder is the MaxPool layer, the Conv layer and the ReLU activation function. The third stage of the RGB encoder is the MaxPool layer, the Conv layer and the ReLU activation function. The fourth stage of the RGB encoder is the MaxPool layer, the Conv layer and the ReLU activation function. The RGB encoder receives underwater RGB images and processes them in four stages. The result of the first stage is input to the first polarization-guided sensing module, the result of the second stage is input to the second polarization-guided sensing module, the result of the third stage is input to the third polarization-guided sensing module, and the result of the fourth stage is input to the fourth polarization-guided sensing module.

[0027] The decoder consists of four stages in sequence: the first stage is the Resblock layer, the second stage is the Bilinear layer, the Conv layer and the ReLU activation function, the third stage is the Bilinear layer, the Conv layer and the ReLU activation function, and the fourth stage is the Bilinear layer, the Conv layer and the ReLU activation function. The decoder in the first stage receives and processes the output of the fourth polarization-guided sensing module; the decoder in the second stage receives and processes the output of the third polarization-guided sensing module and the output of the first stage; the decoder in the third stage receives and processes the output of the second polarization-guided sensing module and the output of the second stage; and the decoder in the fourth stage receives and processes the output of the first polarization-guided sensing module and the output of the third stage.

[0028] The transmission estimation head consists of a Resblock layer, a 1×1 Conv layer, and a Sigmoid activation function. The transmission estimation head receives the decoder's processing results and finally outputs a transmission map.

[0029] The following description, in conjunction with the accompanying drawings, further illustrates the present invention's dual-branch cooperative network as follows: Figure 1As shown, the underwater unsupervised dual-branch collaborative enhancement network includes a main branch with polarization physics prior constraints, an auxiliary branch indirectly guided by pseudo-labels, and an indirect guidance loss constraint part. The dual-branch collaborative network includes a main branch with polarization physics prior constraints, an auxiliary branch indirectly guided by pseudo-labels, and an auxiliary constraint method. The main branch with polarization physics prior constraints includes a polarization prior-guided transmission map estimation network and a background light estimation module. The polarization prior-guided transmission map estimation network includes a polarization feature extraction module, an RGB encoder, a polarization-guided sensing module, a decoder, and a transmission map estimation head. First, polarization degree maps are calculated for the polarization images of the original water at four angles, and the polarization degree maps are input into the polarization feature extraction module. The original underwater RGB images are input into the encoder and the background light estimation module respectively, and the background light estimation module outputs the background light... The polarization-guided sensing module receives and processes the results from the polarization feature extraction module and the encoder. The processed results are then sequentially input into the decoder and the transmission image estimation head, outputting a transmission image. ; The auxiliary branch indirectly guided by the pseudo-tags includes a pseudo-tag transmissive map estimation network and a background light estimation module; the pseudo-tag transmissive map estimation network includes an encoder, a decoder, and a transmissive map estimation head; the pseudo-tag RGB images are input into the pseudo-tag transmissive map estimation network and the background light estimation module respectively; the pseudo-tag transmissive map estimation network, including an encoder, a decoder, and a transmissive map estimation head, outputs the transmissive map. The background light estimation module outputs the background light. ; Auxiliary constraint method based on , , and Calculate scene radiation graph J1 and scene radiation graph J2, perform content consistency constraints and cross-reconstruction consistency constraints, and then return to update the network parameters.

[0030] In this embodiment of the invention, degraded images of the same underwater scene at multiple polarization angles and the corresponding turbid underwater degraded images are first acquired. Specifically, original degraded images at four polarization angles can be acquired, and then... Figure 1 The polarization calculation module in the image calculates the corresponding polarization degree map φ; simultaneously, it acquires the corresponding underwater original degraded RGB image of the scene, denoted as φ. Among them, underwater degradation images The polarization map φ is used to provide color, brightness, and texture information, while the polarization map φ provides prior polarization physics information related to backscattering and medium transport. Meanwhile, to construct auxiliary branch inputs, several existing underwater image enhancement methods are used to enhance the RGB images of the degraded underwater image. Enhancement is performed to generate multiple candidate enhancement results; then, image quality evaluation metrics are used to evaluate and filter these candidate enhancement results, and the result with the best quality is selected as the pseudo-label image. The pseudo-label image During the training phase, auxiliary branches are input to provide clear reference information to the main branch.

[0031] The polarization prior-guided transmission map estimation network of this invention is as follows: Figure 2 As shown, the polarization prior-guided transmission map estimation network includes four polarization-guided sensing modules, and the polarization feature extraction module receives... Output and to Perform downsampling once, twice, and three times respectively. The first polarization-guided sensing module receives the image, the second receives the result of a single downsampling, the third receives the result of a second downsampling, and the fourth receives the result of a third downsampling. The RGB encoder receives the underwater RGB image and processes it through four stages: the first stage result is input to the first polarization-guided sensing module, the second stage result to the second, the third stage result to the third, and the fourth stage result to the fourth. The decoder processes the output of the fourth polarization-guided sensing module in its first stage; it processes the outputs of the third and first stages in its second stage; it processes the outputs of the second and second stages in its third stage; and it processes the outputs of the first and third stages in its fourth stage. The transmission estimation head consists of a Resblock layer, a 1×1 Conv layer, and a Sigmoid activation function. It receives the decoder's processing results and finally outputs a transmission map.

[0032] In this embodiment of the invention, the polarization prompting module performs convolutional mapping on RGB features and polarization degree features respectively. Specifically, the RGB features, after convolution, generate a guiding branch; the polarization degree features, after processing by convolution, average pooling, and a fully connected activation module, generate polarization guiding weights; subsequently, through operations such as channel multiplication, element-wise addition, and element-wise multiplication, the polarization features are injected step-by-step into the RGB features to obtain the polarization-guided fused features. Further, Figure 2The polarization feature extraction module shown is used for global modeling of the polarization map. This module first maps the input polarization map to a feature embedding space using an image patch embedding module. Then, it constructs query features, key features, and value features respectively. Using operations such as the Softmax function and element-wise multiplication, it extracts polarization features with global discriminative capabilities. These polarization features are then fed into various levels of polarization-aware modulation modules for multi-scale injection. In the decoding stage, the fused multi-scale features are progressively restored to spatial resolution through a bilinear upsampling module and a convolutional activation module, and then fused with skip connection features at the corresponding scale. Finally, a transmission map is output through a pointwise convolutional module and an activation function. Meanwhile, the background light estimation module in the main branch performs background light estimation on the murky RGB image. Perform background light estimation to obtain the background light Specifically, the background light estimation module takes the input image brightness channel as input. The global median is used as the benchmark, and a local smoothing filter is applied using a two-dimensional Gaussian kernel. A reflection-filling method is then used to suppress dark edge artifacts at the boundary, resulting in the background light estimation result. .

[0033] The following is a further explanation with reference to an embodiment. In the training phase of Embodiment 1 of the present invention, the Adam optimizer is used, the input image patch size is set to 384×384, the batch size is 1, and the initial learning rate is... A total of 30 training rounds were conducted, and the following settings were implemented. During training, the main branch and auxiliary branches are jointly optimized end-to-end, while only the main branch is retained for inference during the testing phase. The dataset contains 1759 original samples, of which 1423 are used for training and 336 for testing. The image turbidity ranges from 1 to 10 NTU and includes underwater objects of different angles and materials. Considering the difficulty in obtaining clear ground truth images with strict registration for this dataset, UIQM, UCIQE, TM, NIQE, and URanker are used as no-reference evaluation metrics.

[0034] The method of this invention was compared with 13 other underwater image enhancement methods, including traditional methods Color_balance, HLRP, MMLE, PCFB, ROP, UNTV, and PDS, as well as deep learning methods HCLR-Net, ERD, Semi-UIR, UDNet, USUIR, and UPGD. For traditional methods, their publicly available implementations were directly used for testing; for trainable deep learning methods, they were all retrained under the same data partitioning before testing. Since this dataset does not provide real paired labels, the comparative methods involving supervised training all used the constructed pseudo-labels as training supervision. The comparison results are shown in Table 1: Table 1. Comparison results on the TJUP-USOD dataset (original resolution) ; As shown in Table 1, the method of this invention achieves the best overall performance under the original settings of the TJUP-USOD dataset, with UIQM, UCIQE, TM, NIQE, and URanker scores of 3.049, 0.612, 4718.178, 5.461, and 0.825, respectively. Compared with various comparative methods, the method of this invention maintains its leading position in multiple dimensions such as image visual quality, detail restoration, structure preservation, and perceptual evaluation, indicating that the method of this invention can more effectively suppress scattering, improve color shift, and restore scene details in complex and murky scenes.

[0035] Furthermore, this invention also conducted visualization comparisons under typical degradation scenarios such as low turbidity, medium turbidity, high turbidity, and blue-green bias. The visualization results show that existing methods are prone to color distortion, local overexposure, residual haze, or uneven brightness distribution in some scenarios, while the method of this invention can maintain a more natural brightness distribution, clearer structural details, and more stable descattering effect under different degradation scenarios. All methods were trained and tested at a uniform 256×256 resolution to ensure the fairness of the comparison. The comparison methods include HCLR-Net, UPGD, and the method of this invention, and the comparison results are shown in Table 2. Table 2 Comparison results on the TJUP-USOD dataset (256*256 resolution) ; As shown in Table 2, under the constraint of a uniform 256×256 resolution, the method of this invention achieved the best results in 3 out of 5 evaluation metrics and the second best in 2 out of 5 metrics. The UIQM, UCIQE, TM, NIQE, and URanker scores were 3.047, 0.600, 5998.545, 6.039, and 1.077, respectively. Notably, the method of this invention achieved the best results in UCIQE, TM, and URanker, indicating its strong advantages in improving color contrast, restoring texture details, and evaluating overall perceptual quality. Supplementary visualization results further demonstrate that, compared to methods that only support fixed input sizes, the method of this invention still exhibits better visual balance and enhanced stability under uniform resolution conditions.

[0036] In Embodiment 2 of this invention, the dataset consists of monochrome images. During testing, single-channel images are copied to three channels as input and then converted to grayscale images after output, to maintain consistency with the comparison methods designed for color images. All comparison methods are tested directly using publicly available code and default parameter settings, without training or fine-tuning on this dataset. The experiment uses AG, EI, and NIQE as evaluation metrics, and the comparison results are shown in Table 3. Table 3 Comparison results on real underwater polarization datasets (original resolution) ; The test results at the original resolution show that the method of this invention achieves AG of 7.948, EI of 56.933, and NIQE of 6.187 on a real underwater polarization dataset. The AG and EI scores are superior to those of HLRP, PDS, PCFB, UNTV, Semi-UIR, UDNet, and USUIR, indicating that the method of this invention has stronger edge enhancement, detail recovery, and target structure representation capabilities in real underwater polarization scenarios.

[0037] Furthermore, supplementary comparison results after uniformly scaling to 256×256 are shown in Table 4: Table 4 Comparison results on real underwater polarization datasets (256*256 resolution) ; The method of this invention achieves AG, EI, and NIQE scores of 17.054, 118.756, and 5.912, respectively. AG and EI are significantly superior to HCLR-Net and UPGD, indicating that even under fixed low-resolution input conditions, the method of this invention can still maintain strong detail recovery and edge enhancement capabilities. Corresponding visualization results show that the method of this invention can more effectively suppress scattering interference in real polarization scenes and better recover coral outlines, texture levels, and structural details of the target area, further verifying the good generalization ability and practical application value of the method of this invention.

[0038] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A polarization-physical constraint and pseudo-tag-guided unsupervised underwater image enhancement method, characterized in that, include: Underwater RGB images were acquired as a dataset, and a two-branch collaborative network was constructed. The dual-branch collaborative network consists of a main branch constrained by polarization physics priors, an auxiliary branch indirectly guided by pseudo-labels, and a loss layer indirectly guided by pseudo-labels. During the training phase, the pseudo-label indirectly guided loss layer calculates the cross-reconstruction consistency loss, clear content consistency loss, and polarization degree transmission map spatial consistency loss based on the processing results of the main branch and the auxiliary branch indirectly guided by the pseudo-label, respectively. The three consistency losses are combined to construct an overall loss function, and the network parameters of the dual-branch collaborative network are optimized based on the overall loss function. During the inference phase, the underwater RGB image is input only into the main branch of the polarization physics prior constraint of the dual-branch collaborative network, and the final scene radiation map is output. The main branch of the polarization physics prior constraint includes a polarization prior-guided transmission map estimation network and a background light estimation module; The polarization prior-guided transmission map estimation network includes a polarization feature extraction module, an RGB encoder, a polarization-guided sensing module, a decoder, and a transmission map estimation head; The auxiliary branches indirectly guided by pseudo-tags include a pseudo-tag transmittance estimation network and a background light estimation module; The pseudo-label transparency estimation network consists of an encoder, a decoder, and a transparency estimation head; Acquiring underwater RGB images And calculate the corresponding polarization degree diagram. Several underwater image enhancement methods were used to respectively... Enhancement is performed to obtain corresponding candidate pseudo-label images. The final pseudo-label image is then determined using preset no-reference image quality evaluation metrics and photometric safety discrimination. ; Will and The main branch of the input polarization physics prior constraints will The auxiliary branch is indirectly guided by the input pseudo-label.

2. The polarization physics constraint and pseudo-tag guided unsupervised underwater image enhancement method according to claim 1, characterized in that, Will Input the RGB encoder for feature extraction and output the RGB backbone features; The input polarization feature extraction module extracts features and outputs polarization features. ; RGB backbone features and Input polarization-guided sensing module, Injected as a conditional guiding term into the RGB backbone features, polarization-guided fusion features are obtained; The polarization-guided fusion features are sequentially input into the decoder and the transmission map estimation head, and the estimated transmission map is output. ; Will Input the background light estimation module, first construct a brightness map. : ； In the formula, for In coordinates The value at that location, for In the The pixel values ​​of each color channel. For color channel index, It is a set of three channels: red, green, and blue. An adaptive two-dimensional Gaussian kernel is used for smoothing filtering, and a reflection filling strategy is employed during the filtering process to obtain the Gaussian low-frequency components. : ； In the formula, for In coordinates The value at that location, It is a two-dimensional Gaussian kernel. This is a convolution operation; extract Global median of pixel values ,by As a robust brightness benchmark for background light estimation, combined with Calculate the background light grayscale estimation image : ； In the formula, for In coordinates The value at that location, These are weighting coefficients used to balance low-frequency spatial variations with global brightness stability; calculate In the Global mean across each color channel : ； In the formula, for height, for The width; Construct color scaling coefficients based on the global mean of each color channel. : ； In the formula, for In the Global mean across each color channel For temporary color channel indexes, It is a numerically stable term; Calculate the first Background light estimation results for each color channel : ； Background light estimation results from the red channel Green channel background light estimation results And blue channel background light estimation results The main branch background light estimation results, which together constitute the polarization physics prior constraints, : ； Introducing an underwater degradation imaging model: ； In the formula, for In coordinates Observations at that location This is a scene radiation map. for In coordinates The value at that location, for In coordinates The estimated value at that location; use and right Perform reverse reconstruction to obtain the scene radiation map. : ； based on Constructing a transmission map reference item : ； In the formula, As a preset constant, for In coordinates The value at; right and Normalization is performed to obtain the normalized transmission image. and normalized transmission reference terms ,use and Constructing a polarization degree-transmission map spatial consistency loss function : ； In the formula, This is the spatial consistency constraint function.

3. The polarization physics constraint and pseudo-tag guided unsupervised underwater image enhancement method according to claim 2, characterized in that, The encoder, decoder, and transmission map estimation head of the pseudo-label transmission map estimation network are input sequentially, and the pseudo-label transmission map is output. ;Will Input background light estimation module, output pseudo-label background light estimation result ; Based on the underwater degradation imaging model, and right Perform reverse recovery to obtain the pseudo-label scene radiation map. : ； In the formula, for In coordinates Observations at that location for In coordinates The value at that location, for In coordinates The estimated value at that location.

4. The polarization physics constraint and pseudo-tag guided unsupervised underwater image enhancement method according to claim 3, characterized in that, Cross-consistency reconstruction loss includes, based on , and Construct the first cross-reconstruction result : ； In the formula, For pixel-by-pixel multiplication, This is a gradient truncation operation; based on , and Construct the second cross reconstruction result : ； Constructing cross-reconstruction consistency loss : ； In the formula, for Norm; based on and Constructing a clear content consistency loss : ； Constructing pseudo-labels to indirectly guide loss : 。 5. The polarization physics constraint and pseudo-tag guided unsupervised underwater image enhancement method according to claim 4, characterized in that, Construct the overall loss function : ； In the formula, for Weighting coefficients; based on Backpropagation and network parameter updates are performed on the two-branch cooperative network.

6. The polarization physics constraint and pseudo-tag guided unsupervised underwater image enhancement method according to claim 5, characterized in that, The polarization prior-guided transmission map estimation network includes four polarization-guided sensing modules, and the polarization feature extraction module receives... Output and to Perform downsampling once, twice, and three times respectively. Input the first polarization-guided sensing module, input the first downsampling result into the second polarization-guided sensing module, input the second downsampling result into the third polarization-guided sensing module, and input the third downsampling result into the fourth polarization-guided sensing module.

7. The polarization physics constraint and pseudo-tag guided unsupervised underwater image enhancement method according to claim 6, characterized in that, The RGB encoder consists of four stages in sequence. The first stage of the RGB encoder is the Conv layer and the ReLU activation function. The second stage of the RGB encoder is the MaxPool layer, the Conv layer and the ReLU activation function. The third stage of the RGB encoder is the MaxPool layer, the Conv layer and the ReLU activation function. The fourth stage of the RGB encoder is the MaxPool layer, the Conv layer and the ReLU activation function. The RGB encoder receives underwater RGB images and processes them in four stages. The result of the first stage is input to the first polarization-guided sensing module, the result of the second stage is input to the second polarization-guided sensing module, the result of the third stage is input to the third polarization-guided sensing module, and the result of the fourth stage is input to the fourth polarization-guided sensing module.

8. The polarization physics constraint and pseudo-tag guided unsupervised underwater image enhancement method according to claim 7, characterized in that, The decoder consists of four stages in sequence: the first stage is the Resblock layer, the second stage is the Bilinear layer, the Conv layer and the ReLU activation function, the third stage is the Bilinear layer, the Conv layer and the ReLU activation function, and the fourth stage is the Bilinear layer, the Conv layer and the ReLU activation function. The decoder receives and processes the output of the fourth polarization-guided sensing module in the first stage. The second stage of the decoder receives and processes the outputs of the third polarization-guided sensing module and the first stage of the decoder. The third stage of the decoder receives and processes the outputs of the second polarization-guided sensing module and the second stage of the decoder. The fourth stage of the decoder receives and processes the outputs of the first polarization-guided sensing module and the third stage of the decoder.

9. The polarization physics constraint and pseudo-tag guided unsupervised underwater image enhancement method according to claim 8, characterized in that, The transmission estimation head consists of a Resblock layer, a 1×1 Conv layer, and a Sigmoid activation function in sequence; the transmission estimation head receives the decoder processing results and finally outputs a transmission map.

Images