An underwater polarized image despeckling method based on two-stage semi-supervised learning

Abstract

The application discloses a kind of underwater polarized image depolarization methods based on two-stage semi-supervised learning, comprising: constructing the underwater polarized image depolarization network including Fourier modulation refinement module, the underwater polarized image depolarization network is enhanced polarized feature expression by space domain and frequency domain joint modeling, improve global structure perception ability and inhibit scattering interference, while combining detail compensation and edge enhancement strategy, improve the texture detail recovery ability and visual clarity of depolarization image;The underwater polarized image depolarization network uses two-stage semi-supervised learning framework, further optimizes model;The present application can make full use of potential information in unlabelled sample, effectively inhibit underwater scattering interference, improve global structure keeping ability and edge texture detail recovery quality, so as to obtain higher quality underwater depolarization image.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to an underwater polarization image descattering method based on two-stage semi-supervised learning. Background Technology

[0002] When light travels through water, it is subject to scattering and absorption, resulting in reduced contrast, color distortion, and blurred details in underwater images, which affects subsequent target identification and image analysis. Polarization imaging technology utilizes the difference in polarization characteristics between reflected and backscattered light from a target to separate target information from scattering interference to a certain extent, thus having significant application value in the field of underwater polarization image restoration.

[0003] Existing underwater polarization image descattering methods mainly include those based on polarization difference imaging and those based on Stokes vector analysis. These methods are typically built upon specific polarization imaging physical models, requiring estimation of polarization parameters, scene conditions, or medium properties. While these methods can achieve certain restoration results under ideal conditions, in complex underwater environments, due to the complex composition of water, variable lighting conditions, strong noise interference, and uncertainties in the imaging process, relevant physical parameters are often difficult to obtain accurately. This leads to problems such as insufficient robustness, limited adaptability, and unstable restoration results in practical applications.

[0004] In recent years, deep learning methods have been gradually introduced into underwater polarization image descattering tasks. By learning the mapping relationship between degraded and target images through data-driven approaches, these methods reduce the reliance on precise physical models and parameter estimations to some extent, achieving good results in image contrast enhancement, detail restoration, and visual quality improvement. However, most existing deep learning-based underwater polarization image descattering methods rely on a large amount of paired labeled data for supervised training, requiring the simultaneous acquisition of both the degraded polarization image and its corresponding high-quality reference image. Due to the difficulty in acquiring clear reference images in real underwater environments, the high cost of acquisition, and the complexity of data annotation, the number of paired labeled samples available for training is limited, thus restricting the training effectiveness, generalization ability, and engineering application scope of existing methods.

[0005] Furthermore, image degradation in complex underwater environments involves both global scattering interference and local detail loss. Existing methods fail to fully utilize the potential information in unlabeled polarized images, resulting in shortcomings in global structure preservation, edge restoration, and texture reconstruction. Therefore, a semi-supervised underwater polarized image restoration method is needed that can simultaneously utilize both limited labeled and unlabeled data to improve image restoration performance in complex underwater environments. Summary of the Invention

[0006] In view of the shortcomings of existing technologies, this invention provides an underwater polarization image descattering method based on two-stage semi-supervised learning. This invention fully utilizes the advantages of underwater polarization imaging in distinguishing between reflected and backscattered light from a target, and combines this with a semi-supervised learning strategy to effectively mine potential information in unlabeled polarization images under limited labeled data conditions, thereby improving the model's adaptability to complex underwater degradation scenarios. This invention not only enhances the global structural representation of images but also improves the restoration of edge and texture details, and effectively suppresses underwater background scattering interference, thus generating clear underwater restored images.

[0007] The technical means employed in this invention are as follows:

[0008] Structure of the Polarization Image Restoration Model: The underwater polarization image descattering model of this invention adopts a two-stage semi-supervised learning framework, with the encoder-decoder backbone network as the main body of the model. During the training phase, images acquired at four different polarization directions (0°, 45°, 90°, and 135°) in the same scene are used as input, organized into multi-polarization inputs, and then fed into the encoder-decoder backbone network. The input polarization images are first convolved and mapped to a high-dimensional feature space to complete shallow feature extraction; subsequently, multi-scale feature extraction is performed through progressive downsampling to expand the receptive field and extract deep semantic information at different scales. During feature extraction, the extracted features are fed into a Fourier modulation thinning module for modeling and enhancement to improve the network's ability to represent complex degradation patterns and key structural information. In the decoding phase, the spatial resolution of the feature map is restored through progressive upsampling, and the corresponding layer features of the encoder and decoder are fused through skip connections and feature concatenation operations to reduce the loss of detail information and improve structural restoration capabilities. Finally, the reconstructed features are projected back into the image space through convolutional mapping to output the restored image.

[0009] Two-stage semi-supervised learning strategy: This invention employs a staged optimization approach to train the underwater polarization image descattering model. In the first stage, paired supervised data is used to perform supervised training of the model under multi-scale constraints, enabling the model to learn the basic restoration mapping from the degraded polarization input to the target image and generate pseudo-labels required for subsequent training. In the second stage, real supervised labels and pseudo-labels are introduced into the training process to construct hybrid supervised constraints, further optimizing the model parameters. Through this two-stage training mechanism, while ensuring the model obtains a stable initial restoration capability, it can further utilize the potential information in unlabeled samples to enhance the model's ability to represent complex degradation patterns and improve the overall image restoration performance.

[0010] Fourier Modulation Refinement Module: To balance the global modeling requirements of complex underwater degradation with the need for local detail recovery, this invention designs a Fourier modulation refinement module. This module consists of a Fourier modulation attention module and a partially convolutional feedforward network module connected in series. The Fourier modulation attention module performs frequency modulation and adaptive recalibration on the input features at a global scale to enhance the network's ability to represent complex degradation patterns, global contextual relationships, and long-range dependencies. The partially convolutional feedforward network module performs local spatial refinement and structural compensation on the frequency-modulated features. By applying convolution operations on some channels, it introduces local neighborhood priors, enhancing the feature's ability to characterize edge contours, texture patterns, and fine-grained structural information.

[0011] Fourier Modulation Attention Mechanism: The Fourier modulation attention module includes a content branch and a frequency branch. The content branch extracts spatial semantic information and local structural responses from the input features; the frequency branch performs frequency transformation, frequency domain mapping, and inverse transformation on the input features to obtain frequency response features. Then, the content features and frequency response features are recombined into multi-head representations, and interactions occur within each head to achieve adaptive frequency modulation of the input features. This mechanism preserves the original spatial structural information while introducing frequency domain information to enhance the feature's perception of global degradation patterns, thereby improving the network's modeling effect on complex underwater scattering degradation.

[0012] Partial Convolutional Feedforward Network Module: This module further refines the features processed by the Fourier Modulation Attention Module. It first maps and nonlinearly transforms the input features, then divides them along the channel dimension into a preservation branch and a convolutional branch. The preservation branch maintains the original semantic information, while the convolutional branch extracts local neighborhood context information through convolution operations. Finally, the two branches are concatenated and mapped back to the original channel dimension. This design enhances local spatial awareness while controlling computational overhead, improving the restoration quality of edges, textures, and detailed structures.

[0013] Loss Functions: During model training, this invention designs corresponding loss functions for each of the two training stages. The first stage employs joint optimization of multi-scale reconstruction loss and frequency domain consistency loss to simultaneously constrain the consistency between the spatial domain restoration results and the frequency domain distribution, thereby learning a stable basic restoration mapping relationship. The second stage, building upon the multi-scale reconstruction loss and frequency domain consistency loss, further introduces gradient consistency loss to enhance the recovery ability of edge contours and texture details. Through the above loss design, the model's restoration accuracy and structure preservation ability in complex underwater environments can be effectively improved.

[0014] Compared with the prior art, the present invention has the following advantages: This invention provides an underwater polarization image descattering method based on two-stage semi-supervised learning. Compared to existing supervised learning methods that rely on large amounts of paired labeled data, this method can fully utilize the potential information in unlabeled polarization images under limited labeled data conditions, thereby effectively improving the model's training performance and practical application capabilities. Through a two-stage semi-supervised learning strategy, this invention enables the model to further mine additional supervisory information from pseudo-labels after achieving stable initial restoration capabilities, enhancing its adaptability to complex underwater degradation patterns.

[0015] Meanwhile, this invention achieves an organic combination of global frequency information modeling and local spatial refinement by introducing a Fourier modulation refinement module. Specifically, the Fourier modulation attention module enhances the network's ability to represent global contextual information and long-range dependencies, while the partially convolutional feedforward network module further strengthens the recovery of edge, texture, and fine-grained structural information. This results in improved restoration results in terms of structural integrity, texture clarity, and visual quality.

[0016] This invention, based on the use of polarization information to suppress background scattering interference, further combines a semi-supervised learning mechanism to improve the model's generalization ability, providing an effective, stable, and highly practical technical solution for polarization image restoration in complex underwater environments, which is of great significance to underwater imaging and underwater vision applications. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is an overall flowchart of the semi-supervised underwater polarization image restoration method of the present invention.

[0019] Figure 2 This is a network architecture diagram of the polarization image restoration model in an embodiment of the present invention.

[0020] Figure 3 This is an architecture diagram of the Fourier modulation refinement module in an embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram of the experimental setup for passive polarization imaging in an underwater environment, as described in an embodiment of the present invention.

[0022] Figure 5 These are partial images from the underwater polarization image dataset collected in an embodiment of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0024] like Figure 1-3 As shown, this invention discloses an underwater polarization image descattering method based on two-stage semi-supervised learning, characterized by the following steps: S1. Obtain an underwater polarization image dataset, which includes polarization images acquired in the same scene at four different polarization directions: 0°, 45°, 90°, and 135°. The underwater polarization image dataset also includes paired supervised data and unlabeled data.

[0025] Specifically, underwater passive polarization imaging experiments were conducted in an underwater environment. Using a Blackfly S BFS-U3-51S5PC-C polarization camera, four images of the same target or scene with different polarization directions were acquired in different underwater scenarios and grouped into a polarization image set. For paired supervised data, in addition to acquiring degraded underwater polarization images, corresponding clear reference images were also obtained for the first stage of supervised training. For unlabeled data, only degraded polarization images with different polarization directions were acquired for the second stage of semi-supervised training. This approach constructs an underwater polarization image dataset that includes both paired supervised samples and unlabeled samples to improve the model's adaptability to complex underwater environments.

[0026] S2. Construct an underwater polarization image restoration model. The method adopts a two-stage progressive semi-supervised learning framework. The underwater polarization image restoration model includes an encoder-decoder backbone network and a Fourier modulation refinement module set in the encoding process.

[0027] Specifically, a two-stage progressive semi-supervised learning structure is used. The input image is first constructed as a multi-scale representation of the same scene, and then fed into the corresponding layers of the encoder-decoder backbone network for feature extraction and image restoration. The input at each scale first passes through the convolutional input layer to complete feature mapping. At the encoding end, deep features are extracted step by step through downsampling operations, and a Fourier modulation thinning module is introduced during the encoding process to model and enhance the features. At the decoding end, the spatial resolution of the feature map is gradually restored through upsampling operations, and features at different scales are fused together with the stitching module. Finally, the restored result is obtained through the convolutional output layer.

[0028] Furthermore, the four input images with different polarization directions (0°, 45°, 90°, and 135°) are first organized into multi-polarization inputs within the same scene and then fed into the encoder-decoder backbone network. The input polarization images are first convolved and mapped to a high-dimensional feature space to complete shallow feature extraction. Subsequently, multi-scale feature extraction is performed through progressive downsampling to expand the receptive field and extract deep semantic information at different scales. During feature extraction, the extracted features are fed into a Fourier modulation thinning module for modeling and enhancement to improve the network's ability to represent complex degradation patterns and key structural information. In the decoding stage, the spatial resolution of the feature map is restored through progressive upsampling, and the corresponding layer features from the encoder and decoder are fused through skip connections and feature concatenation operations to reduce the loss of detail information and improve structural recovery capabilities. Finally, the reconstructed features are projected back into the image space through convolutional mapping to output the restored image.

[0029] Furthermore, the Fourier modulation refinement module is composed of a Fourier modulation attention module and a partially convolutional feedforward network module connected sequentially. It is used to perform staged enhancement and optimization of multi-scale features in the backbone network. The Fourier modulation attention module focuses on introducing frequency response information from the global scale and adaptively recalibrating the input features, while the partially convolutional feedforward network module focuses on refining and compensating the modulated features from the local scale. The two work together to enable the network to simultaneously take into account complex degradation modeling and local detail restoration.

[0030] Specifically, the Fourier modulation refinement module, through a cascaded setup of a Fourier modulation attention module and a partially convolutional feedforward network module, sequentially performs global frequency enhancement and local structural refinement on the input features. The Fourier modulation attention module primarily introduces frequency response information and adaptively recalibrates the features to improve the network's ability to perceive global degradation patterns and contextual relationships. The partially convolutional feedforward network module mainly performs local compensation and detail enhancement on the frequency-modulated features to improve the recovery of edges, textures, and local structures. Through the synergistic effect of these two modules, the network can simultaneously handle complex degradation modeling and fine detail reconstruction.

[0031] The Fourier modulation attention module scheme is as follows: for input features Perform layer normalization to obtain normalized features:

[0032] The normalized features Input the content branch and the frequency branch respectively, wherein the content branch generates content features through convolutional mapping:

[0033] The frequency branch generates frequency response features through frequency transformation, frequency domain linear mapping, and inverse transformation:

[0034] in, Indicates frequency transformation, Indicates inverse frequency transformation. This indicates a frequency domain mapping operation; the content features are... With the frequency response characteristics Reorganized into The number indicates that, in the... Element-wise multiplication is performed within each element to obtain the modulation result:

[0035] in, This represents element-wise multiplication. After aggregating and normalizing the modulation results of each head, the data is mapped back to the original channel dimension via convolution and then residually connected with the input features to obtain the output features of the Fourier modulation attention module.

[0036] in, This indicates a multi-head aggregation operation. This represents the output mapping convolution. Through the above process, the Fourier modulated attention module can effectively integrate frequency domain response information into the feature modeling process without destroying the original spatial structure representation, thereby enhancing the network's ability to perceive complex degradation patterns, global contextual relationships, and long-range dependency information.

[0037] The partial convolutional feedforward network module scheme is as follows: for input features Layer normalization is performed, where, As input features, This involves layer normalization; the normalized features are then processed through convolutional layers and Gaussian error linear unit activation functions to obtain intermediate features. , is represented as:

[0038] in, This represents the convolution operation, used to perform channel transformation and feature recombination on input features. Represents the activation function of the Gaussian error linear unit. It is an intermediate feature.

[0039] For the intermediate features along the channel dimension The division is represented as follows:

[0040] in, This indicates a channel partitioning operation; To preserve branches, which are used to retain the original semantic information; This is the convolution branch, used for local spatial modeling.

[0041] The convolution branch pass Convolution extracts local neighborhood context information to obtain local enhanced features, represented as follows:

[0042] in, express Convolution operation; This represents the local enhancement feature obtained after local convolution processing of the convolutional branch. The preserved branch and the local enhancement feature are concatenated along the channel dimension to obtain the fused feature, represented as:

[0043] in, This indicates a channel splicing operation; This is a fusion feature used to fuse original semantic information and local structural enhancement information.

[0044] The fused features are mapped back to the original channel dimension through a convolutional layer and then residually connected with the input features to obtain the output features of the partial convolutional feedforward network module, represented as:

[0045] in, The output features of some convolutional feedforward network modules, the This represents the convolution operation used to recover the original channel dimensions. Through the above process, the original semantic information without local convolution processing and the local structural information enhanced by convolution are effectively integrated, enabling the output features to simultaneously take into account the backbone representation after global modulation and fine-grained spatial details, thereby further improving the network's ability to recover target edges, texture details and local structures in complex underwater scenes.

[0046] In this embodiment, the polarization image descattering model is optimized using a staged loss function. Specifically, the first stage of supervised training is optimized using a first-stage loss function, which includes multi-scale reconstruction loss and frequency domain consistency loss; wherein, Set network output The descattering results at the i-th scale, the i-th The prediction results at each scale are denoted as and the corresponding supervised target is denoted as . The corresponding supervision target is denoted as The multi-scale reconstruction loss is expressed as:

[0047] in, Indicates the first Reconstruction loss weighting coefficients at each scale express The norm is used to constrain the reconstruction error between the downscattering results and the corresponding supervised target at each scale. The frequency domain consistency loss is expressed as:

[0048] in, Indicates the first Frequency domain consistency loss weighting coefficients at each scale This represents the Fourier transform, used to constrain the consistency between the downscattering results at each scale and the corresponding supervised target in the frequency domain. The first-stage loss function is expressed as:

[0049] in, This represents the weighting coefficient for the frequency domain consistency loss in the first stage. Through the above joint constraints, the model's recovery capability in both the spatial and frequency domains can be enhanced simultaneously.

[0050] In the second training phase, to further explore the potential information in the unlabeled data and improve the restoration of image edges and texture details, gradient consistency loss is introduced to optimize the model, building upon the multi-scale reconstruction loss and frequency domain consistency loss. Let the supervision objective for the second phase be denoted as... The supervision target is either a real label or a pseudo label. The prediction results at each scale are denoted as The multi-scale reconstruction loss is expressed as:

[0051] in, Indicates the first Reconstruction loss weighting coefficients at each scale express The norm is used to constrain the reconstruction error between the downscattering results and the corresponding supervised target at each scale. The frequency domain consistency loss is expressed as:

[0052] in, Indicates the first Frequency domain consistency loss weighting coefficients at each scale This represents the Fourier transform, used to constrain the consistency between the downscattering results at each scale and the corresponding supervised target in the frequency domain. At the highest resolution scale, a gradient consistency loss is applied, expressed as:

[0053] in, and These represent the gradient operators in the horizontal and vertical directions, respectively, used to constrain the consistency between the downscattering results at the highest resolution scale and the corresponding supervised target in terms of edge and texture details. The second-stage loss function is expressed as:

[0054] in, , and These represent the weight coefficients of the reconstruction loss, frequency domain consistency loss, and gradient consistency loss, respectively. Through these loss functions, the model can effectively restore complex underwater degraded images, improving the structural integrity, edge sharpness, and texture detail of the restored images.

[0055] S3. Train the polarization image restoration model based on the underwater polarization image dataset.

[0056] S4. Obtain the polarization image group to be restored, input the trained polarization image restoration model, and obtain the output of the polarization image restoration model as the restored image.

[0057] The solution and effects of the present invention will be further explained below based on specific application examples.

[0058] like Figure 4-5 As shown in the figure, this embodiment discloses an underwater polarization image restoration method based on semi-supervised learning, and the detailed steps are as follows: Step 1: Acquire underwater polarization image data Underwater passive polarization imaging experiments were conducted in an underwater environment. To diversify the types of objects photographed, sea cucumber models, scallop models, ink bottles, milk cartons, and aluminum cans were placed in a transparent glass tank. A schematic diagram of the underwater polarization imaging experiment is shown below. Figure 4 As shown, the glass water tank has a volume of 100×50×50 cm³. The object to be photographed was fixed at the bottom of the transparent water tank, and after pouring in clean water to submerge the object, a Blackfly S BFS-U3-51S5PC-C polarization camera was used to capture four images of the same scene with different polarization directions at one time, namely 0°, 45°, 90° and 135° polarization images.

[0059] Specifically, first, images of the scene are captured in clear water to serve as a reference clear image for that scene. Then, approximately 2 mL of milk is added to the water to simulate an underwater scattering environment, and a set of polarized images of the same scene is captured again, resulting in a set of degraded polarized images. This set includes four polarized images at 0°, 45°, 90°, and 135°. Milk is gradually added to the water, and the process is repeated until the water becomes so turbid that the target is difficult to discern. Through this method, reference clear images and a set of degraded polarized images of the same scene under different scattering conditions can be obtained.

[0060] Step 2: Construct an underwater polarization image dataset To meet the requirements of semi-supervised learning training, the collected data is divided into paired supervised data and unlabeled data. The paired supervised data includes degraded polarization image sets and their corresponding clear reference images, used for the first stage of supervised training. The unlabeled data only includes degraded polarization image sets and does not contain the corresponding clear reference images, used for the second stage of semi-supervised training. This data acquisition method allows for the simultaneous construction of an underwater polarization image data foundation combining limited labeled and unlabeled samples, providing data support for the subsequent two-stage progressive semi-supervised learning. Figure 5 The images captured are shown.

[0061] Step 3: Design an underwater polarization image descattering network, such as... Figure 2 As shown, an encoder-decoder backbone structure is adopted, combined with a two-stage progressive semi-supervised learning framework to achieve descattering of underwater polarized images. Specifically, four input images with different polarization directions are first organized into multi-polarization inputs within the same scene, and then further constructed into multi-scale representations before being fed into the network. Each scale input first passes through a convolutional input layer to complete shallow feature extraction, and then at the encoding end, downsampling operations are used to progressively extract deep features to expand the receptive field and obtain semantic information at different scales. At the decoding end, upsampling operations are used to gradually restore the spatial resolution of the feature maps, and a stitching module is used to fuse features at different scales. Finally, the reconstructed image is obtained through a convolutional output layer. Through this structure, the network can simultaneously extract global structural information and local detail information at different scales, thereby improving the ability to restore complex underwater degraded images.

[0062] A Fourier modulation thinning module is introduced during the feature extraction process of the backbone network to enhance and optimize multi-scale features. This module consists of a Fourier modulation attention module and a partially convolutional feedforward network module connected sequentially. The Fourier modulation attention module primarily introduces frequency response information from the global scale and adaptively recalibrates the input features to enhance the network's ability to represent complex degradation patterns, global contextual relationships, and long-range dependencies. The partially convolutional feedforward network module primarily performs local spatial thinning and structural compensation on the modulated features. By applying convolution operations to some channels, it introduces local neighborhood priors, thereby enhancing the feature's ability to express edge contours, texture patterns, and fine-grained structural information. Through the synergistic effect of these two modules, the network can simultaneously handle complex degradation modeling and local detail restoration.

[0063] The Fourier Modulation Attention Module (FAT) adaptively modulates input features by combining spatial content features and frequency response features. Specifically, the input features are first normalized layer-wise to reduce the impact of feature distribution fluctuations on subsequent mapping processes. The normalized features are then input to the content branch and the frequency branch, respectively. The content branch generates content features through convolutional mapping to preserve spatial semantic information and local structural responses in the input features. The frequency branch generates frequency response features through frequency transform, frequency domain mapping, and inverse transform to incorporate global spectral distribution information from the input features. Subsequently, the content features and frequency response features are reconstructed into multi-head representations and interact within each head to obtain the modulation result. The outputs of each head are then aggregated and normalized, and mapped back to the original channel dimension through convolution. Finally, a residual connection is formed with the input features to obtain the output features of the Fourier Modulation Attention Module. Through this process, the module can enhance the feature's ability to perceive complex degradation patterns, global contextual relationships, and long-range dependencies while preserving the original spatial structure representation and incorporating frequency domain response information.

[0064] The convolutional feedforward network module is used to perform local spatial refinement and structural compensation on the features processed by the Fourier modulated attention module. Specifically, the input features are first normalized by layers, and intermediate features are obtained by combining convolutional mapping with a nonlinear activation function. Then, the intermediate features are divided into a preservation branch and a convolutional branch along the channel dimension. The preservation branch preserves the original semantic information, while the convolutional branch extracts local neighborhood context information through 3×3 convolution. Next, the preservation branch and the convolutional branch are concatenated along the channel dimension to obtain fused features. These fused features are then mapped back to the original channel dimension through convolution and form residual connections with the input features to obtain the output features of part of the convolutional feedforward network module. Through this process, the semantic information without local convolution processing and the local structural information enhanced by convolution are effectively integrated, enabling the output features to simultaneously consider the globally modulated backbone representation and fine-grained spatial details, thereby improving the reconstruction quality of edge contours, texture patterns, and local details.

[0065] Step 4: Training and Testing of the Underwater Polarization Image Restoration Network In embodiments of the present invention, the designed underwater polarization image restoration network based on semi-supervised learning is used for training and testing. To verify the effectiveness of the proposed method in the underwater polarization image restoration task, an underwater polarization image dataset containing both reference samples and no reference samples was constructed. The reference dataset contains 1670 sets of underwater polarization images and their corresponding real reference images, used for the first stage of supervised training; the no-reference dataset contains 2000 sets of underwater polarization samples without real reference images, used to generate pseudo-labels after the first stage of training, and further participate in the second stage of semi-supervised training. The entire training process employs a two-stage optimization strategy: In the first stage, supervised learning of the network is performed using only reference samples for 80 epochs to obtain an initial model with strong basic reconstruction capabilities. Based on this, the model trained in the first stage is used to infer underwater polarization images without reference, generating corresponding pseudo-labels as weak supervision signals for subsequent training. In the second stage, after pseudo-label generation, both reference and unreference samples are incorporated into the training process. Reference samples are still supervised by real reference images, while unreference samples utilize the generated pseudo-labels to construct weak supervision constraints. These constraints, combined with consistency constraints, further optimize the network parameters, thereby fully exploring the potential information in unlabeled polarization images and improving the network's generalization ability and image reconstruction performance. The proposed semi-supervised underwater polarization image reconstruction network is implemented using PyTorch 2.4.1 and CUDA 12.1 and trained on a Linux platform. The Adam optimizer is used during training, with a batch size of 4. To unify the input scale and improve training efficiency, all input images are adjusted to a uniform spatial resolution before being input into the network for training. After training, the restoration effect of the method of the present invention can be verified using test samples, and the performance of the method can be comprehensively analyzed by combining qualitative visualization results and quantitative evaluation indicators.

[0066] The underwater polarization image descattering method proposed in this invention, based on two-stage semi-supervised learning, can effectively restore degraded polarization images in complex underwater scattering environments by fully utilizing limited labeled and unlabeled data. This method employs a two-stage progressive training strategy to further mine potential information in unlabeled samples while ensuring the model achieves stable basic restoration capabilities, thereby enhancing the network's adaptability to complex degradation patterns. Simultaneously, this invention combines polarization information with a Fourier modulation thinning mechanism to effectively suppress underwater scattering interference, improve the global structure preservation, edge contour recovery, and local texture detail expression capabilities of the restored image, making the restored result closer to a true, clear image. Compared with existing technologies, this invention exhibits superior performance in underwater polarization image descattering, structure preservation, and detail enhancement, providing an effective technical solution for underwater polarization imaging and underwater image descattering.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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; and these 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 method for underwater polarization image descattering based on two-stage semi-supervised learning, characterized in that, Includes the following steps: A color underwater polarization image dataset is obtained, which includes color polarization images acquired in the same scene at four different polarization directions: 0°, 45°, 90° and 135°. The color underwater polarization image dataset includes paired supervised data and unlabeled data. An underwater polarization image descattering model is constructed, which adopts a two-stage semi-supervised learning framework and includes a multi-scale encoder-decoder backbone network. Four color polarization images with different polarization directions from the same scene are constructed into multi-scale representations and then input into the respective scale branches of the multi-scale encoder-decoder backbone network. Each scale branch performs feature mapping through an input convolution module, and in the encoding stage, a downsampling convolution module extracts deep features step by step. In both the encoding and decoding stages, a Fourier modulation thinning module is used to jointly model and enhance the features in the spatial and frequency domains. In the decoding stage, an upsampling convolution module restores the spatial resolution of the feature map step by step, and a stitching module fuses the features from different scale branches. Finally, an output convolution module outputs the descattered image. The underwater polarization image descattering model is trained based on the aforementioned color underwater polarization image dataset. In the first stage, the underwater polarization image descattering model is trained under multi-scale constraints using paired supervised data to learn the basic descattering mapping from the degraded polarization image to the target image and generate pseudo-labels for subsequent training. In the second stage, the supervised labels and the pseudo-labels are introduced into the training process to construct hybrid supervised constraints and optimize the underwater polarization image descattering model. A set of underwater polarization images to be descattered is obtained, and the set of underwater polarization images to be descattered is input into a trained underwater polarization image descattering model. The output of the underwater polarization image descattering model is then used as the descattered image.

2. The underwater polarization image descattering method based on two-stage semi-supervised learning according to claim 1, characterized in that, The Fourier modulation thinning module includes a Fourier modulation attention module and a partially convolutional feedforward network module connected in series. The Fourier modulation attention module is used to perform frequency modulation and adaptive recalibration on the input features to enhance the network's ability to represent complex degradation patterns, global contextual information, and long-range dependencies. The partially convolutional feedforward network module is used to perform local thinning and structural compensation on the frequency-modulated features. By applying convolution operations to some channels, it enhances the local spatial perception ability to improve the recovery effect of edge, texture, and fine-grained structural information.

3. The underwater polarization image descattering method based on two-stage semi-supervised learning according to claim 2, characterized in that, The Fourier modulation attention module is used to adaptively modulate the input features by combining spatial content features and frequency response features, specifically including: For input features Perform layer normalization to obtain normalized features: The normalized features Input the content branch and the frequency branch respectively, wherein the content branch generates content features through convolutional mapping: The frequency branch generates frequency response characteristics through frequency transformation, frequency domain linear mapping, and inverse transformation: in, Indicates frequency transformation, Indicates inverse frequency transformation. This indicates a frequency domain mapping operation; the content features are... With the frequency response characteristics Reorganized into The number indicates that, in the... Element-wise multiplication is performed within each element to obtain the modulation result: in, This represents element-wise multiplication. After aggregating and normalizing the modulation results of each head, the result is mapped back to the original channel dimension via convolution and then residually connected with the input features to obtain the output features of the Fourier modulation attention module. in, This indicates a multi-head aggregation operation. This indicates the output mapping convolution.

4. The underwater polarization image descattering method based on two-stage semi-supervised learning according to claim 2, characterized in that, The convolutional feedforward network module is used to perform local spatial refinement and structural compensation on the features processed by the Fourier modulated attention module, specifically including: For input features Layer normalization is performed, where, For input features, This involves layer normalization; the normalized features are then processed through convolutional layers and Gaussian error linear unit activation functions to obtain intermediate features. , is represented as: in, This represents the convolution operation, used to perform channel transformation and feature recombination on input features. Represents the Gaussian error linear unit activation function. As an intermediate feature; For the intermediate features along the channel dimension The division is represented as follows: in, Indicates channel partitioning operation; To preserve branches, which are used to retain the original semantic information; This is the convolution branch, used for local spatial modeling; The convolution branch pass Convolution extracts local neighborhood context information to obtain local enhanced features, represented as follows: in, express Convolution operation; The local enhancement feature obtained after local convolution processing of the convolution branch is concatenated with the preserved branch and the local enhancement feature in the channel dimension to obtain the fused feature, represented as: in, Indicates a channel splicing operation; This is a fusion feature used to fuse original semantic information and local structural enhancement information; The fused features are mapped back to the original channel dimension through a convolutional layer and then residually connected with the input features to obtain the output features of the partial convolutional feedforward network module, represented as: in, The output features of some convolutional feedforward network modules, the This represents the convolution operation used to restore the original channel dimensions.

5. The underwater polarization image descattering method based on two-stage semi-supervised learning according to claim 1, characterized in that, Training the underwater polarization image descattering model based on the aforementioned color underwater polarization image dataset includes two stages: In the first stage, the underwater polarization image descattering model is trained under multi-scale constraints using paired supervised data to learn the basic descattering mapping from the degraded polarization image to the target image, and pseudo-labels are generated for the unlabeled data based on the trained model. In the second stage, the supervisory labels and the pseudo-labels are introduced into the training process to construct a hybrid supervisory constraint, which further optimizes the underwater polarization image descattering model.

6. The underwater polarization image descattering method based on two-stage semi-supervised learning according to claim 5, characterized in that, The first stage of supervised training is optimized using a first-stage loss function, which includes multi-scale reconstruction loss and frequency domain consistency loss; wherein, Set network output The descattering results at the i-th scale, the i-th The prediction results at each scale are denoted as and the corresponding supervised target is denoted as . The corresponding supervision target is denoted as The multi-scale reconstruction loss is expressed as: in, Indicates the first Reconstruction loss weighting coefficients at each scale express The norm is used to constrain the reconstruction error between the downscattering results and the corresponding supervised target at each scale. The frequency domain consistency loss is expressed as: in, Indicates the first Frequency domain consistency loss weighting coefficients at each scale The Fourier transform is used to constrain the consistency between the downscattering results at each scale and the corresponding supervised target in the frequency domain. The first-stage loss function is expressed as: in, This represents the weighting coefficient for frequency domain consistency loss in the first stage.

7. The underwater polarization image descattering method based on two-stage semi-supervised learning according to claim 6, characterized in that, The second stage of hybrid supervised training is optimized using a second-stage loss function, which includes multi-scale reconstruction loss, frequency domain consistency loss, and gradient consistency loss; where: Let the monitoring objective corresponding to the second stage be denoted as . The supervision target is either a real label or a pseudo label. The prediction results at each scale are denoted as The multi-scale reconstruction loss is expressed as: in, Indicates the first Reconstruction loss weighting coefficients at each scale express The norm is used to constrain the reconstruction error between the downscattering results and the corresponding supervised target at each scale. The frequency domain consistency loss is expressed as: in, Indicates the first Frequency domain consistency loss weighting coefficients at each scale This represents the Fourier transform, used to constrain the consistency between the downscattering results at each scale and the corresponding supervised target in the frequency domain. At the highest resolution scale, a gradient consistency loss is applied, expressed as: in, and Let represent the gradient operators in the horizontal and vertical directions, respectively, used to constrain the consistency between the downscattering results at the highest resolution scale and the corresponding supervised target in terms of edge and texture details. The second-stage loss function is expressed as: in, , and These represent the weight coefficients of reconstruction loss, frequency domain consistency loss, and gradient consistency loss, respectively.

Images