Underwater image enhancement method based on multichannel polarization information

Abstract

The invention belongs to the technical field of image processing, and particularly discloses an underwater image enhancement method based on multi-channel polarization information, which comprises the following steps: establishing a U-shaped convolutional neural network formed by a multi-stage encoder and a decoder, and introducing polarization degree-based space and channel double-branch attention at a jump joint; acquiring paired color images of the sample in non-scattering and scattering environments in a plurality of polarization directions; polarization channel and color channel cross recombination is carried out according to the scattering environment color image to generate multiple pseudo-color polarization images, a Stokes component is calculated in multiple polarization directions to obtain a single-channel polarization degree image, a training set is constructed, and a non-scattering image is used as a supervision label to train a network; inputting a to-be-restored underwater polarization image into the trained network, and outputting a restored image; the method has the following advantages that the sample and RGB polarization recombination are matched with polarization degree guide training, and the restoration stability and definition of the underwater image are improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to an underwater image enhancement method based on multi-channel polarization information. Background Technology

[0002] In underwater imaging, suspended particles cause strong backscattering and absorption of light, often resulting in blurred images, reduced contrast, and color distortion. Existing physical model-based underwater polarization imaging methods construct models containing parameters such as transmittance, scattering ratio, and background light intensity, and use images at different polarization angles for inversion to suppress scattering. However, these methods heavily rely on prior assumptions and parameter estimations, making them difficult to adapt to the complex and variable real-world scenarios in turbid waters, and exhibiting weak generalization ability. With the development of deep learning, data-driven polarization image restoration methods can learn the mapping relationship between degraded and clear images. However, these methods generally only use grayscale polarized images or simply stack polarization sub-channels, failing to fully utilize the synergistic characteristics of RGB color channels and polarization information. Due to the scarcity of real underwater data and the limited range of data augmentation methods, the stability and generalization ability of these methods under complex scattering conditions remain insufficient.

[0003] To address this issue, an underwater image enhancement method based on multi-channel polarization information is proposed. Summary of the Invention

[0004] The present invention aims to provide an underwater image enhancement method based on multi-channel polarization information, in order to solve or improve the problems mentioned above. Existing underwater polarization image restoration methods have obvious limitations in terms of strong dependence on parameter acquisition, insufficient utilization of polarization and color information, and weak generalization ability due to data scarcity, making it difficult to obtain stable and reliable clear images in complex scattering environments.

[0005] In view of this, a first aspect of the present invention is to provide an underwater image enhancement method based on multi-channel polarization information.

[0006] A second aspect of the present invention is to provide a system A third aspect of the present invention is to provide an electronic device.

[0007] A fourth aspect of the present invention is to provide a computer-readable storage medium.

[0008] The first aspect of the present invention provides an underwater image enhancement method based on multi-channel polarization information, comprising the following steps: constructing a convolutional neural network including multi-level encoders and decoders, and setting a dual-branch attention mechanism based on polarization degree at the jump connections between each level of encoder and decoder; acquiring a first color image of a sample in a non-scattering environment and a second color image in at least one scattering environment in multiple polarization directions; generating multiple pseudo-color polarization images by cross-recombining polarization channels and color channels based on the sub-images of the second color image in different polarization directions, and generating a single-channel polarization degree image by calculating Stokes components in multiple polarization directions; constructing a training set using the pseudo-color polarization images and the polarization degree image, inputting the training set into the convolutional neural network, and training the convolutional neural network using the first color image as the supervision label; inputting the underwater polarization image to be restored into the trained convolutional neural network to output an underwater restored image.

[0009] A second aspect of the present invention provides a system comprising: an image acquisition module for acquiring a first color image and a second color image under multiple polarization directions; a pseudo-color reconstruction module for generating a pseudo-color polarized image based on the cross-reconstruction of color channels and polarization channels of the second color image; a polarization degree generation module for calculating Stokes components according to multiple polarization directions and generating a single-channel polarization degree image; a training module for inputting the training set and the first color image into a convolutional neural network to perform network training; and an image restoration module for inputting an underwater polarized image to be restored into the trained convolutional neural network to output a restored image.

[0010] A third aspect of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the method described above.

[0011] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0012] The beneficial effects of this invention compared to the prior art are as follows: By constructing supervised samples using a first color image acquired in a non-scattering environment and a second color image acquired in at least one scattering environment, the convolutional neural network can learn the mapping relationship between degraded images and clear images under multi-turbidity and multi-polarization conditions. This effectively reduces the dependence on prior assumptions and artificial parameter estimation, thus giving it stronger adaptability and generalization ability to complex and variable scenarios in turbid water.

[0013] By cross-recombining polarization and color channels of sub-images in different polarization directions based on the second color image, various pseudo-color polarized images are generated. These are further combined with single-channel polarization degree images obtained by calculating Stokes components from multiple polarization directions to construct a training set. The pseudo-color polarized images are used as multi-dimensional feature inputs, and the polarization degree images are used as physical guidance signals. This allows the convolutional neural network to fully exploit the synergistic effect between RGB color channels and polarization information during encoding and decoding. Under the guidance of polarization degree, it can highlight high scattering regions and key structural regions, suppress noise and invalid background. As a result, it achieves underwater image restoration effects superior to existing technologies in terms of scattering suppression, contrast enhancement, detail restoration, and color correction. The combination of multiple pseudo-color polarization modes and multiple scattering conditions significantly expands the number of effective training samples, and it can still maintain high stability and generalization performance under the premise of limited real underwater data.

[0014] Additional aspects and advantages of embodiments of the invention will become apparent in the following description or may be learned by practice of embodiments of the invention. Attached Figure Description

[0015] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a flowchart of the method steps of the present invention; Figure 2 This is a schematic diagram of the RGB polarization feature recombination mode of the present invention; Figure 3 This is a schematic diagram of the network structure of the present invention; Figure 4 This is a flowchart of the dual-branch polarization degree attention mechanism algorithm of the present invention; Figure 5 This is a comparison image of the image restored by the present invention, the image to be restored, and the image restored by the prior art; Figure 6 This is a system logic block diagram of the present invention; Figure 7 This is a schematic diagram of the structure of an electronic device according to the present invention. Detailed Implementation

[0016] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments are combined with each other.

[0017] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention is also practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0018] Please see Figures 1-7 The following describes an underwater image enhancement method, system, electronic device, and computer-readable storage medium based on multi-channel polarization information according to some embodiments of the present invention.

[0019] An embodiment of the first aspect of the present invention proposes an underwater image enhancement method based on multi-channel polarization information. In some embodiments of the present invention, such as... Figures 1-5 As shown, the method includes the following steps: S101 constructs a convolutional neural network containing multi-level encoders and decoders, and sets a dual-branch attention mechanism based on polarization degree at the jump connections between each level of encoder and decoder.

[0020] Here, multiple encoders are set up sequentially from top to bottom on the side near the input end. Each encoder compresses the input features step by step through convolution and downsampling operations, transforming the detail and texture information in the original image spatial domain into deep feature representations with different receptive field scales. On the side near the output end, multiple decoders are set up sequentially from bottom to top. Each decoder restores the deep features step by step through upsampling and convolution operations, restoring the compressed feature mapping to a scale that matches the resolution of the input image, so as to output the resulting image for underwater image enhancement.

[0021] To preserve both shallow details and deep semantic information during feature transfer, skip connections are established between each encoder level and its corresponding decoder. These connections directly introduce features of different scales from the encoding path into the decoding path. A bi-branch attention mechanism based on polarization information is introduced at each skip connection. Specifically, in each skip connection, polarization-related information is used as a guide to generate spatial attention weights for modulating spatial location. This enhances features from the encoder in spatial regions with significant polarization response and suppresses features in regions with weak polarization response or dominated by noise. Simultaneously, channel attention weights are generated for modulating feature channels. Based on the contribution of each feature component to the change in polarization, feature channels with high correlation to polarization characteristics are preserved and enhanced, while channels with low correlation are weakened or suppressed.

[0022] As can be seen from the above, through the function of the dual-branch attention mechanism, the convolutional neural network can adaptively focus on multi-channel polarization features that are more related to polarization scattering suppression and target detail recovery during the feature transmission process of encoding and decoding, thereby providing a structural basis for subsequent underwater image enhancement effects.

[0023] Specifically, the architecture of a convolutional neural network is a U-shaped structure; and the steps for building a convolutional neural network containing multi-level encoders and decoders include: Multiple encoders are set up sequentially from top to bottom at the input end. Each encoder compresses the input features into deep features step by step through a downsampling structure.

[0024] A bottleneck layer is set between the bottom of the input and output ends to carry the deepest features and serve as a connection node between the encoding and decoding paths.

[0025] Multiple decoders are set up sequentially from bottom to top at the output end, and each decoder recovers deep features step by step through an upsampling structure.

[0026] Regarding the specific description above, multiple encoders are sequentially set up from top to bottom on the side near the input end. Each encoder receives the feature map from the previous layer in turn and compresses the input features step by step through a downsampling structure including convolution operations. This reduces the spatial resolution of the feature map layer by layer and increases the number of channels layer by layer, thereby gradually aggregating the information distributed in the fine-grained space of the original image into a deep feature representation with a larger receptive field. As the network progresses downwards from shallow to deep, the features extracted by the encoder gradually transition from low-level local information such as edges and textures to high-level semantic information such as contours, regional distribution, and even the overall structure of the object. This ensures that the entire encoding path maintains the main content of the input image and completes the differentiation and abstraction of degradation patterns and target structural patterns.

[0027] A bottleneck layer is placed at the very bottom of the encoding path, between the input and output ends. This bottleneck layer connects the output of the deepest encoder to the input of the most basic decoder. On one hand, it carries the deepest features compressed through multiple encoder stages, centrally representing key information in the current task, such as underwater scattering distribution, target structure contours, and multi-channel polarization response. On the other hand, it serves as a connection node between the encoding and decoding paths, uniformly transmitting the compressed high-dimensional features to the decoding path, allowing subsequent upsampling and reconstruction processes to unfold based on the global receptive field. Near the output end, multiple decoders are sequentially arranged from bottom to top. Each decoder receives feature maps from the previous layer, progressively restores the spatial resolution of the feature maps through upsampling structures, and refines and reconstructs the features using operations such as convolution, gradually supplementing details and texture information as the resolution is restored to near the input resolution.

[0028] As the network progresses step by step along the decoding path from bottom to top, each decoder retains the overall structure and color distribution trend by utilizing the deep features from the bottleneck layer and the previous decoder. On the other hand, it combines the coding-side features introduced at the jump connections to restore more detailed edges, local contrast, and details related to polarization characteristics layer by layer. As a result, the underwater image enhancement result generated at the output end inherits the global expressive power provided by the coding path and achieves targeted restoration at the spatial detail level.

[0029] As can be seen from the above, the U-shaped convolutional neural network achieves an organic combination of encoding and decoding, compression and restoration, and global and local functions in its structure, providing a stable and hierarchically clear network framework for underwater image enhancement driven by multi-channel polarization information.

[0030] Specifically, the steps for setting up a dual-branch attention mechanism based on polarization degree at the jump connections between encoders and decoders at each level include: At the jump connections between each level of encoder and decoder, spatial attention weights and channel attention weights are generated based on the polarization image.

[0031] Spatial attention weights are applied to features from the encoder to enhance the weighted spatial regions based on the degree of performance of the spatial response in the polarization distribution.

[0032] Channel attention weights are applied to features from the encoder to weaken the weighted channel components based on each channel's contribution to the polarization degree change.

[0033] Based on the specific description above, the polarization degree images are aligned according to a spatial size that matches the encoder output features. For each skip connection, the features are registered with the corresponding polarization degree image at that scale, ensuring that the feature response at the same spatial location corresponds to the polarization degree value at that location. Based on this, spatial attention branches and channel attention branches are constructed respectively: The spatial attention branch takes the polarization degree image as input and smooths, normalizes or simply transforms the spatial distribution of polarization degree to obtain an intermediate representation that reflects the difference in polarization response strength at different spatial locations. This intermediate representation is then mapped to a spatial attention weight map of the same size as the feature map. The channel attention branch is based on the encoder output features. By statistically converging the features in the spatial dimension and combining the statistics of the overall polarization degree distribution, the contribution of each channel in expressing polarization changes is obtained, and a channel attention weight vector for modulating the channel response is generated accordingly.

[0034] At each jump connection position, the feature map output by the encoder is multiplied with the corresponding spatial attention weight map on a pixel-by-pixel scale. This results in a larger attention weight for regions with high polarization degree and obvious response in the polarization image, thereby amplifying the response value of that region in the feature map. Conversely, the attention weight of pixels with low polarization degree or those corresponding to background noise or uniform water bodies is relatively reduced, thus being relatively suppressed in the feature map.

[0035] The encoder output features are globally averaged or otherwise statistically aggregated in the spatial dimension, compressing the overall response of each feature channel on the entire feature map into a scalar. This channel-level description is then combined with the polarization degree image or its statistics to obtain a comprehensive contribution index of each channel in terms of encoding polarization changes and reflecting scattering differences. Subsequently, the above index is converted into normalized channel attention weights through a set of nonlinear mappings. Channels with larger weights are retained and strengthened, while channels with smaller weights are suppressed or weakened.

[0036] As can be seen from the above, through the combined action of the spatial attention and channel attention branches at the jump connection positions at each level, the convolutional neural network in this invention can be guided by polarization information in both spatial and channel dimensions during feature transmission, thereby achieving fine selection and weighting of multi-channel polarization features. This is beneficial to improving the network's ability to identify scattering structures and target details in underwater image enhancement tasks.

[0037] Specifically, such as Figure 3 As shown, the network input consists of the image after RGB polarization feature extraction and the corresponding pixel-level DoP image. The backbone network is built based on U-net (U-shaped convolutional neural network). The network includes a downsampling encoding part and an upsampling decoding part, forming a symmetrical U-shaped structure, supplemented by skip connections between channels to achieve the combination of global features and local features.

[0038] This invention introduces a dual-branch attention mechanism on the backbone network and performs DOP modulation at skip connections. This part consists of two sub-modules: Spatial Attention (DSA) and Channel Attention (DCA). The DSA module generates pixel-wise attention weights by concatenating the DoP image with local features to highlight regions with high polarization information. The DCA module inputs the global average pooling description of the features and the statistics of the DoP into the network to generate channel weights to suppress task-irrelevant channel responses.

[0039] Suppose there are n hop connections, then the encoder output in a certain hop connection is E. (l), l ∈{1, 2, …,n}, which corresponds to the output D of the previous level decoder that needs to be connected. (l-1) Then, the final input of the decoder obtained after the skip connections is G. (l) .

[0040] Specifically, such as Figure 4 As shown, the network implementation based on the dual-branch polarization degree attention mechanism of DSA and DCA, the overall algorithm flow includes: ①DSA module design: For a given feature map E ∈ R C×H×W And the DoP graph P ∈ R C×H×W P∈[0,1], where C is the number of channels in the image, H is the number of pixels corresponding to the image height, and W is the number of pixels corresponding to the image width; firstly, a lightweight convolution is performed on the DoP image to achieve a smoothing effect, and the smoothed feature map is P′, represented as: In the formula, Smooth(•) is the Gaussian smoothing function. In its implementation, a Gaussian function with a standard deviation of σ is used to generate a 3×3 convolution kernel, and the image is then convolved. Next, the number of channels in P′ is expanded to align with the channel implementation of the given particular image E: The expanded P″ is concatenated with the original feature through channels to obtain a tensor E′ with 2C channels: Perform a 3×3 convolution on it to generate a spatial attention map M s : In the formula, σ(•) is the Sigmoid function, and Conv 3×3 (•) is a convolution operation using a 3×3 kernel. Finally, the spatial attention map is applied to the original features, i.e., the original features are modulated pixel by pixel, resulting in a spatially modulated feature map F′, which is represented as: In the formula, ⊙ represents the element-wise multiplication operation of matrices.

[0041] ②DCA Module Design: For the feature map F′ ∈ R obtained by the DSA module C×H×W And the DoP graph P ∈ R C×H×W For P ∈ [0, 1], first perform global average pooling (GAP) on the feature map F′: In the formula, ū can be regarded as the feature description of the overall image; then, the global average is taken for the DoP image P: In the formula, ā can be regarded as a polarization intensity indicator for the entire image; concatenating the feature description with the polarization statistics yields the concatenated vector Z, which is expressed as: The concatenated vector is then activated and mapped to a channel attention map M via a fully connected layer. c : The MLP(•) network mapping uses the Sigmoid function, resulting in the channel attention map M. c ∈ R C It needs to be broadcast to the spatial dimension, and the spatial modulation feature map F′ needs to be channel-modulated to obtain a feature map F″ containing both spatial modulation and channel modulation: ③ Features of splicing at jump connection points: Finally, the obtained F″ is concatenated with the output of the previous decoder to obtain the new feature G. (l) Feed into the subsequent decoding network: In the formula, concat(•) represents the concatenation operation of two or more arrays or strings. The above process makes the decoding stage prioritize high-value feature regions guided by DoP.

[0042] After completing the network structure design, a loss function design is also required to indicate the direction of network iteration. This invention employs a hybrid loss function combining Mean Squared Error (MSE) and Structural Similarity (SSIM), and calculates the total loss function based on a weighted hybrid approach. total : In the formula, ω1 and ω2 are the weights of the individual loss, and the loss... MSE For mean squared error loss, loss SSIM The structural similarity loss can be adjusted according to the actual effect of the algorithm; for example, ω1=60% and ω2=40%.

[0043] S102, acquire a first color image of the sample in a non-scattering environment and a second color image in at least one scattering environment in multiple polarization directions.

[0044] Here, the sample to be tested is placed inside an experimental water tank. First, clean water is injected into the tank to create a scatter-free environment. An active illumination source located outside the tank illuminates the underwater sample, and a polarizing device is placed in front of the light source to ensure the illumination light entering the water has a stable polarization state. Then, a color polarization camera is used to acquire color polarization images of the sample in multiple preset polarization directions. Each polarization direction corresponds to a first color image containing multiple color channels. These multiple polarization directions include several typical linear polarization angles achieved through polarizer rotation or the camera's internal beam splitting structure, exemplified by 0 degrees, 45 degrees, 90 degrees, and 135 degrees, to ensure that the color polarization response of the sample in the clean water environment can be acquired under different polarization observation directions. Through the above acquisition process, first color images with multiple polarization directions are obtained in a scatter-free environment surrounding the same target scene. These images maintain consistency in spatial composition, viewing angle, and imaging optical path, differing only from data acquired in subsequent scattering environments in terms of water scattering conditions, thus providing a foundation for subsequent construction of clear labels.

[0045] After data collection under a non-scattering environment, a pre-set concentration of skim milk solution was gradually added to the experimental water tank. This simulated scattering particles in natural water bodies by utilizing the tiny particles in the emulsion, allowing the water to transition from a clear state to scattering environments with different turbidity levels. During the addition of skim milk, the turbidity of the water was measured in real-time using a turbidimeter. The turbidity value was used as a quantitative indicator of scattering intensity, and the corresponding scattering conditions were recorded within the pre-set turbidity range. For each turbidity condition, while maintaining the sample position, imaging optical path, and polarization acquisition settings consistent with the non-scattering environment, a color polarization camera was used to acquire color polarized images of the sample in the aforementioned multiple polarization directions. This yielded a second color image that corresponded one-to-one with the first color image in the clear water environment in terms of scene content and spatial composition.

[0046] As described above, by repeating the acquisition process under different turbidity conditions and scattering environments, paired color polarization image data are obtained for the same target sample under multiple polarization directions and multiple scattering intensities. The multi-polarization direction color images under clear water conditions constitute the first color image set, and the corresponding multi-polarization direction color images under each scattering environment constitute the second color image set. This phased, paired acquisition method ensures a strict correspondence between the first and second color images in terms of target content, viewpoint, and geometric structure, facilitating the use of clear water images as supervisory labels for scattering images during subsequent training. Furthermore, by varying the concentration of skim milk to cover different scattering intensities, the second color images exhibit representativeness and hierarchy in turbidity distribution and scattering intensity, providing a rich and structured sample source for subsequent multi-channel polarization feature reconstruction and underwater image enhancement network training.

[0047] Specifically, the step of acquiring a first color image of a sample in a non-scattering environment and a second color image in at least one scattering environment in multiple polarization directions includes: First color images of the sample in multiple polarization directions were acquired in a non-scattering environment.

[0048] Color polarization images corresponding to the samples are acquired in at least one scattering environment with different simulated scattering characteristics.

[0049] The acquisition process is repeated based on the changes in the scattering intensity of the scattering medium, and all color polarization images are combined to form a second color image corresponding to the first color image.

[0050] Regarding the specific description above, the sample to be tested can be fixedly arranged inside an experimental water tank, filled only with clean water to make the water essentially transparent and negligible in terms of scattering. Based on this, an active light source is used to illuminate the sample, and polarization adjustment elements are configured along the light source path or imaging path, enabling the imaging system to acquire color polarized images of the same scene under multiple preset polarization directions. For example, by rotating a linear polarizer, switching polarizer groups, or using a split-focus plane polarization camera, color polarized images of the sample in the clean water environment are sequentially acquired under different polarization directions such as 0 degrees, 45 degrees, 90 degrees, and 135 degrees. For each polarization direction, the imaging device obtains a first color image containing red, green, and blue color channel information. During the acquisition process, the sample posture, camera position, and light source arrangement remain unchanged; only by changing the polarization direction, multiple images with highly consistent spatial content but different polarization states are obtained, thus forming a set of first color images surrounding the same target sample under scattering-free conditions.

[0051] After acquiring the first color image in multiple polarization directions under a non-scattering environment, color polarization images corresponding to the sample are acquired in at least one scattering environment with different simulated scattering characteristics. To this end, while keeping the sample position and imaging system configuration unchanged, a scattering medium solution to simulate the scattering characteristics of water can be gradually added to the experimental water tank. For example, a mixed solution containing tiny particles can be added to gradually change the water from a clear state to a scattering environment with a certain degree of turbidity.

[0052] After each adjustment of the scattering medium concentration, the turbidity of the water body is measured using a turbidimeter or similar measuring device. This turbidity value is used as a quantitative indicator of scattering intensity, and the process of multi-polarization direction color polarization imaging is repeated under the same scattering conditions. In other words, for the same sample and viewing angle as in a non-scattering environment, color polarization images are sequentially acquired in multiple polarization directions under the current scattering environment. This results in a set of color polarization images corresponding to the sample under each scattering condition. These images maintain a correspondence with the first color image in terms of spatial structure and target content. However, due to backscattering, attenuation, and blurring effects introduced by the scattering medium, their imaging quality is significantly degraded compared to the first color image.

[0053] Within a set scattering intensity range, the turbidity of the water body is gradually changed by adjusting the concentration of the scattering medium. Each time a new scattering intensity level is reached, a complete multi-polarization direction color polarization image acquisition process is executed, thereby obtaining a set of color polarization images under multiple scattering conditions for the same target sample and the same imaging structure. For any target sample, the first multi-polarization direction color image acquired in a non-scattering environment is considered the reference imaging result for that sample in clear water. All color polarization images acquired under different scattering intensity conditions are processed to form a second color image dataset that corresponds to the aforementioned reference image in terms of spatial location and scene content.

[0054] As described above, by acquiring images in pairs between no-scattering and multiple scattering environments, the consistency of the first and second color images in the scene geometry is ensured. This allows the first color image to be used as a supervisory label to guide the network in learning the mapping relationship from degraded imaging to clear imaging. On the other hand, by introducing multiple scattering intensity levels, the second color image has sufficient diversity in terms of degradation degree and scattering characteristics. This is beneficial for the network to fully engage with underwater imaging samples under different turbidity conditions during the training phase, thereby improving the adaptability and generalization performance of the underwater image enhancement method based on multi-channel polarization information in complex scattering environments.

[0055] Specifically, an underwater polarization imaging experimental platform based on active illumination was constructed. The experimental setup included a light source, a polarizing device, a polarizing camera, a lens, and an experimental water tank. A target object was placed in the water tank, and a polarizing device was placed in front of the light source to obtain polarized illumination. A color division of plane (DoFP) polarization camera was used to acquire RGB color polarimetric images of the target object at 0°, 45°, 90°, and 135°.

[0056] First, target images were acquired in a clean water environment. Then, skim milk solutions of different concentrations were gradually added. Since casein molecules in skim milk can effectively simulate Rayleigh scattering in seawater, a scattering medium environment was simulated in the experimental water tank. The turbidity of the scattering medium was measured using a turbidimeter, and data was collected in the scattering medium. Data on the target object under various solution turbidities were collected within the range of 5 NTU-50 NTU. After completing image acquisition for one set of target objects in both clean water and the scattering medium, the above acquisition steps were repeated for another set of target objects.

[0057] S103, based on the sub-images of the second color image in different polarization directions, multiple pseudo-color polarization images are generated by cross-recombining the polarization channel and color channel, and a single-channel polarization degree image is generated by calculating the Stokes component in multiple polarization directions.

[0058] Here, each second color image corresponds to multiple polarization directions during acquisition, resulting in a color polarimetric image containing red, green, and blue color channels under each polarization direction. For the same target scene, while ensuring a one-to-one correspondence in spatial position, the color channels and polarization channels of the aforementioned color polarimetric images are cross-recombined: A preset number of color channels are selected from the obtained sub-images, and a polarization channel with the same number of color channels is selected from all available polarization directions. On the other hand, a non-repeating correspondence is established between the selected color channels and the selected polarization channels, so that each color channel is matched with a different polarization direction.

[0059] By combining the above-mentioned cross-referencing methods, multiple pseudo-color polarization images are generated. Each pseudo-color polarization image carries light intensity information from different polarization directions through multiple color channels, thus reflecting the differentiated response of the same target scene under different polarization observation conditions. Compared with directly using the original color polarization sub-images, the pseudo-color polarization images obtained through recombination implicitly contain the re-encoding relationship of polarization angles between channels. This allows the network to perceive multiple polarization states within a single image, which is beneficial for amplifying the response differences between scattered light and target reflected light during the feature learning stage and enhancing the sensitivity to changes in polarization information. Since the same set of original sub-images generates multiple pseudo-color polarization images through different correspondences, the number of training samples and the diversity of feature combinations are effectively expanded without increasing the acquisition cost. This allows subsequent deep networks to learn in a richer multi-channel polarization feature space, thus providing a more sufficient input information foundation for underwater image enhancement.

[0060] After completing the pseudo-color polarization image reconstruction, in order to extract polarization features closely related to scattering suppression from a physical level, Stokes components are further calculated based on color polariton images in multiple polarization directions, and a single-channel polarization degree image is generated on this basis. Specifically, at each pixel location, the light intensity measurement values ​​of that pixel under multiple polarization directions are read respectively. The light intensity at different polarization angles is combined to construct the Stokes components describing the polarization state of the light field, so that each pixel corresponds to a set of physical quantities that can characterize the total light intensity and the linear polarization components.

[0061] Under linear polarization imaging conditions, the light field can be decomposed into unpolarized and linearly polarized components by measuring the light intensity at four typical polarization angles. This allows for the separation of the polarization-related component and its proportional relationship to the total light intensity at each pixel. Based on these Stokes components, the degree of polarization is calculated at the pixel level, numerically reflecting the proportion of the polarized component relative to the total light intensity at that pixel. Regions with high polarization typically correspond to areas with more pronounced polarization characteristics in the light field, such as a water background dominated by scattering components or the edge of a target with significant specular reflection characteristics.

[0062] Considering that each color polariton image in the original acquisition contains multiple color channels, the degree of polarization can be calculated separately on each color channel. Then, the degree of polarization of different color channels is weighted and fused according to the preset channel weights, and the polarization information of multiple channels is compressed into a single-channel grayscale polarization image. This polarization image retains the sensitivity to changes in polarization intensity as a whole, while also taking into account the channel contribution ratio consistent with human eye brightness perception.

[0063] As can be seen above, in addition to the pseudo-color polarization image corresponding to each second color image, an extra single-channel polarization degree image with clear physical meaning is obtained. This enables the subsequent network to learn multi-channel pseudo-color polarization features and use the polarization degree image as a guiding signal to highlight the scattering significant region and the key structural region from a physical perspective. This provides a fine and interpretable polarization prior for the subsequent polarization degree-based attention mechanism and underwater image enhancement process.

[0064] Specifically, the steps of generating multiple pseudo-color polarized images by cross-recombining polarization channels and color channels based on sub-images of the second color image in different polarization directions include: Extract m different color channels from the obtained sub-image, and select m polarization channels from all kinds of polarization directions.

[0065] A unique correspondence is established between color channels and polarization channels to form a pseudo-color polarized image obtained by combining m color channels with m distinct polarization channels.

[0066] Regarding the specific description above, for the acquired second color image, a corresponding color polarization sub-image is obtained under each polarization direction. Each sub-image contains multiple distinct color channels. Based on the acquired sub-images, in this embodiment, the number of color channels used for reconstruction is predetermined to be m. m distinct color channels are extracted from all sub-images, for example, any m channels corresponding to the red, green, and blue channels. From all available polarization directions, m polarization channels, the same number as the number of color channels, are selected, ensuring that the number of polarization directions participating in subsequent reconstruction is consistent with the number of color channels participating in reconstruction.

[0067] After selecting the color channels and polarization channels, a unique correspondence is established between them to form a pseudo-color polarized image obtained by combining m color channels with m distinct polarization channels. Specifically, in each reassembly, the aforementioned m selected color channels are considered as a set of output channels to be matched, and the selected m polarization channels are considered as a set of allocable polarization direction resources. A unique polarization channel is assigned to each color channel through a one-to-one pairing process, ensuring that any color channel extracts pixel information from only one sub-image of a polarization direction within the same pseudo-color polarized image, guaranteeing that all m polarization channels are not reused in a single reassembly. Thus, when constructing a single pseudo-color polarized image, each color channel in the image originates from a sub-image of a different polarization direction, thereby encoding responses from multiple polarization directions into the color dimension in a non-overlapping manner at the same spatial location.

[0068] As can be seen from the above, the cross-recombination strategy ensures that each pseudo-color polarization image carries information from multiple polarization directions, and introduces differentiated polarization response expressions at the sample level through the construction of multiple correspondences. This allows the same target scene to be expanded into multiple pseudo-color polarization images with different polarization encoding methods based on the second color image, which is beneficial to improving the diversity and discriminability of multi-channel polarization features in the subsequent training stage.

[0069] In any of the above embodiments, the pseudo-color polarization image is characterized by recombination of color channels and polarization channels to represent a polarization response different from that of the second color image, and to expand the number of samples in the training set.

[0070] In this embodiment, the second color image is a color polarized image obtained under a single polarization direction. Its three color channels reflect the light intensity distribution of the red, green, and blue bands respectively in the same polarization direction. The color channels are consistent in polarization attributes, reflecting more the spectral composition and scene brightness distribution. However, by cross-recombining the sub-images of the second color image in different polarization directions, in the pseudo-color polarized image, the color channels no longer come from the same polarization direction, but correspond to multiple different polarization directions, resulting in deliberately constructed differences in polarization response among the color channels within a pseudo-color polarized image.

[0071] In this way, at the same spatial location, the different color channels of the pseudo-color polarization image encode the response differences of that location under different polarization observation conditions. Compared with the second color image, which only reflects color information in a single polarization direction, the pseudo-color polarization image introduces a reorganized expression of polarization state in the channel dimension, thereby forming polarization response characteristics that are significantly different from those of the second color image at the image level. This is beneficial for the network to more sensitively capture the polarization differences between the scattered background and the target reflection during the learning process.

[0072] Furthermore, by employing multiple color and polarization channel recombination methods on the sub-images of the same set of original second-color images, a unique correspondence is established between the color and polarization channels, constructing multiple pseudo-color polarized images with distinct structures. Each pseudo-color polarized image corresponds to a different polarization information encoding mode. Although these pseudo-color polarized images correspond to the same target scene in spatial structure, they exhibit significant differences in channel combination and polarization response expression. From the perspective of training data, a set of second-color images originating from the same physical scene is expanded into multiple input sample sequences composed of different channels.

[0073] Specifically, the steps for generating a polarization degree image from Stokes components calculated using multiple polarization directions include: Based on multiple color polarimetric images acquired through m polarization channels, light intensity sampling is performed on the color channels under each polarization channel to determine the Stokes component used to characterize the polarization state.

[0074] The degree of polarization, which characterizes the ratio of polarization component to total light intensity, is calculated at the pixel level based on the Stokes component.

[0075] The polarization degrees corresponding to different color channels are fused to generate a polarization degree image.

[0076] Regarding the specific description above, for the same target scene, a color polarimetric image containing multiple color channels is acquired under each polarization channel. Each color channel corresponds to a light intensity distribution in spatial location. For any pixel location, the light intensity value under different polarization directions is read from the color polarimetric images corresponding to the pixel under m polarization channels, and sampling is performed channel by channel, so that each color channel forms a set of light intensity sampling sequences that vary with polarization direction. Based on the light intensity data acquired under multiple polarization channels, according to the physical model of polarization imaging, a Stokes component is constructed for each color channel to characterize the polarization state of the pixel's light field, so that it can numerically characterize the total light intensity and the amplitude and direction information of the linear polarization component, thereby providing basic parameters for subsequent polarization degree calculation.

[0077] After determining the Stokes components, a polarization degree value, representing the ratio of polarization components to total light intensity, is calculated at the pixel level based on these Stokes components. Specifically, for each pixel location, the polarization-related components in the light field are compared with the overall light intensity using the Stokes components obtained in the previous step to obtain the relative proportion of the polarization component to the total light intensity at that pixel. This forms a scalar value between zero and one, reflecting the degree of polarization significance of that pixel.

[0078] When the polarization degree value is close to zero, it indicates that the polarization component in the light field at that location is weak, and it is mostly unpolarized scattering or diffuse reflection background. When the polarization degree value is close to one, it indicates that the linear polarization component accounts for a higher proportion in the light field at that location, and it is more likely to correspond to a scattering region with obvious polarization directionality or a structural edge with a certain specular reflection component. Through the above-mentioned polarization degree calculation at the pixel level, each pixel obtains a polarization degree value with clear physical meaning, providing a pixel-level quantitative basis for subsequently introducing polarization priors in the spatial and feature dimensions.

[0079] After calculating the polarization degree of each color channel, the polarization degree values ​​of each channel are weighted and combined according to preset channel weights. This assigns greater weight to color channels with higher human eye sensitivity or better signal-to-noise ratios, while assigning less weight to color channels with relatively heavy noise interference or less contribution to the task. Through this fusion method, the polarization degree information from multiple color channels is compressed into a single-channel polarization degree image. This polarization degree image spatially reflects the distribution of the relative strength of polarization components in the entire image, retaining the comprehensive advantages of multi-channel polarization measurement while avoiding the complexity of directly processing multi-channel polarization degrees in the network.

[0080] As can be seen from the above, the final obtained polarization degree image, as a guiding feature with physical interpretation, is used in conjunction with the pseudo-color polarization image and input into the subsequent convolutional neural network to highlight the polarization-significant region in feature modulation and attention guidance, thereby improving the ability to identify and restore scattering structures and important target regions during underwater image enhancement.

[0081] Specifically, such as Figure 2 As shown, for the acquired polarization images at 0 degrees, 45 degrees, 90 degrees, and 135 degrees, each image contains three channels: R (Red), G (Green), and B (Blue). First, RGB polarization reconstruction features are calculated. For example... Figure 2 As shown, the color channels and polarization channels are cross-recombined, resulting in the following six pseudo-color polarization modes: Mode 1: R0°+G90°+B45°; Mode 2: R0°+G45°+B90°; Mode 3: R90° + G0° + B45°; Mode 4: R90° + G45° + B0°; Mode 5: R45° + G0° + B90°; Mode 6: R45° + G90° + B0°; This involves obtaining six sub-images of the same target scene and calculating the information entropy and variance of each image. The variance reflects the distribution of pixel values ​​in the image, while the information entropy reflects the randomness and complexity of the grayscale distribution, and is related to the ability of different polarization angles to suppress reflected and scattered light from the scene. When the variance shows consistency while the information entropy differs, the effectiveness of the feature extraction method is verified.

[0082] Next, polarization degree (DoP) image calculation is performed. Since DoP is essentially a physical quantity independent of color, measuring the ratio of polarization components to total light intensity in the light field, it can be calculated as a single-channel quantity. That is, the RGB three-channel polarization images at 0°, 45°, 90°, and 135° are calculated into a single-channel grayscale DoP image. The specific steps for pixel-wise calculation are given below: Let I0(x), I 45 (x), I 90 (x), I 135 (x) represents four polarizer images at pixel position x, further analyzed using I... θ (c) (x) represents the light intensity of the c-th color channel at angle θ, where θ∈{0°,45°,0°,35°}, c∈{R,G,B}, and S i (c)(x) represents the i-th Stokes vector component of the c-th color channel, i∈{0,1,2}, S (c) (x) represents the total Stokes vector of the c-th color channel.

[0083] In the case of linear polarization, Stokes components are constructed using measurements at four angles, and Stokes is calculated for each channel, expressed as follows: Then channel c contains only the polarization degree DoP of the linear polarization component. (c) (x) is represented as: In the formula, a very small constant ε is used to avoid the denominator being zero, and then channel fusion is performed on the three-channel DoP image to obtain the grayscale DoP image. gray (x), the calculation formula is: In the formula, ω (c) This represents the weight of each channel, usually ω. R =0.299, ω G =0.587, ω B =0.114, DoP R (x) represents the DoP image of the red channel, DoP G (x) represents the green channel DoP image, DoP B (x) represents the blue channel DoP image.

[0084] Finally, to meet the dataset requirements for deep learning, the obtained RGB polarization feature images and DoP images are simultaneously expanded, including steps such as cropping by step size, rotation, horizontal flipping, and vertical flipping. The images in the scattering medium are used to construct the sample space, and the clear water images corresponding to the same target scene are used to construct the label space.

[0085] S104: Construct a training set using a pseudo-color polarization image and a polarization degree image, input the training set into a convolutional neural network, and train the convolutional neural network using the first color image as the supervisory label.

[0086] Here, for the same target scene, the second color image obtained under at least one scattering environment is recombined with the polarization channel and color channel to form multiple pseudo-color polarization images. A single-channel polarization degree image is calculated based on the light intensity information of multiple polarization directions. For each scattering condition, the pseudo-color polarization image generated under that condition is spliced ​​or paired with the corresponding polarization degree image according to a preset organization method, making it a multi-source feature description of the same input sample. On the one hand, this allows the convolutional neural network to learn the correlation between multi-channel polarization response and scattering structure from the pseudo-color polarization images; on the other hand, by using the aforementioned single-channel physical quantities of the polarization degree image, the network can provide an explicit distinction between high-polarization and low-polarization regions.

[0087] For each set of target samples, the first color image with multiple polarization directions acquired in a clean water environment is preprocessed and used as the reference image of the sample in a non-scattering state; in a scattering environment, the image pair obtained by pseudo-color polarization reconstruction and polarization degree calculation is used as the input feature of the sample in a scattering degradation state.

[0088] In the specific training process, the samples in the training set are further subjected to geometric data enhancement in the form of cropping, rotation, flipping, etc., so that the polarization characteristics of the same physical scene in different fields of view and different directions can be fully presented, thereby increasing the number and diversity of training samples and reducing the network's tendency to overfit to specific scenes or specific compositions.

[0089] As can be seen above, by iteratively updating the network parameters in multiple rounds, the convolutional neural network gradually converges on the current training set, and finally obtains a training model that can effectively restore clear underwater images based on pseudo-color polarization images and polarization degree images, providing a foundation for subsequent enhancement and restoration when inputting underwater polarization images to be restored.

[0090] S105: Input the underwater polarization image to be restored into the trained convolutional neural network to output the underwater restored image.

[0091] Here, during the deployment phase, the underwater polarization imaging data acquired on-site is organized in the same manner as during the training phase. This ensures that the underwater polarization image to be restored remains compatible with the training samples in terms of resolution, channel arrangement, and polarization information representation, thereby guaranteeing that the trained convolutional neural network can directly process the underwater polarization image to be restored. In practical applications, the underwater polarization image to be restored is a color polarization image acquired in a scattering water environment and its multi-channel input form after channel reconstruction and polarization degree calculation. Through the same preprocessing procedure as in the training phase, it is converted into an input feature tensor for the convolutional neural network, and this feature tensor is fed into the trained network model without changing the network structure and parameters.

[0092] During forward inference, the trained convolutional neural network performs multi-level feature extraction and step-by-step reconstruction of the underwater polarization image to be restored along a predetermined encoding and decoding path: On the one hand, the encoder compresses and abstracts the multi-channel polarization features of the input, distinguishing and reorganizing the degradation information dominated by scattering and the intrinsic features related to the target structure in the feature space; on the other hand, the decoder continues to use the polarization-guided dual-branch attention mechanism at each level of jump connection to perform spatial and channel-level weighted modulation of the features from the encoder, so that the network focuses on the regions and feature components related to scattering suppression, edge contours and detail texture restoration during the inference stage, and weakens the response to components with low correlation to polarization changes.

[0093] As described above, through the feature transfer and reconstruction process, the underwater restored image generated by the network output is significantly improved in terms of contrast, clarity, and color reproduction compared to the original underwater polarized image to be restored. The turbidity and haze caused by backscattering in the water are effectively reduced, and the outline and surface texture of the target object are enhanced. Therefore, step S105, by inputting the underwater polarized image to be restored into the trained convolutional neural network, realizes the transfer of the image restoration mapping learned based on multi-channel polarization information to the inference application of the actual underwater scene. This allows the present invention to continuously output underwater restored images in online or batch processing scenarios after completing offline training, functionally closing the entire process from data acquisition, feature construction, network learning to image enhancement output.

[0094] This invention provides an underwater image enhancement method based on multi-channel polarization information. By cross-recombining RGB color channels with multiple polarization directions and introducing a polarization-degree-based dual-branch attention mechanism, it achieves deep fusion of multi-channel polarization information and effective injection of physical priors at both the network input and feature learning levels. On the one hand, by recombining RGB polarization channels, color channels are cross-combined with polarization directions such as 0 degrees, 45 degrees, and 90 degrees to generate multiple pseudo-color polarization modes. This re-encodes scene information that was originally expressed only in a single polarization direction and in the conventional RGB channels into a pseudo-color image that simultaneously contains differences in multiple polarization responses. Validated by information entropy and variance metrics, this not only breaks through the limitations of traditional grayscale polarization images but also forms a more discriminative multi-dimensional feature representation in terms of scattering suppression and target detail preservation. On the other hand, channel recombining expands the original data volume many times over without increasing additional acquisition costs. Combined with geometric transformations such as cropping, rotation, and flipping, the training samples are further expanded, effectively alleviating the training difficulties caused by the scarcity of real underwater data. This allows deep learning models to obtain more sufficient sample support under different turbidity and scattering intensities, thereby improving adaptability to complex scattering environments, reducing the risk of overfitting, and significantly enhancing the model's generalization ability in multi-condition underwater scenarios.

[0095] On the other hand, this invention utilizes polarization degree images calculated from multiple polarization directions to construct a dual-branch attention mechanism. Spatial attention branch and channel attention branch are introduced at the jump connection between the encoder and decoder. Guided by the polarization degree, a quantity with clear physical meaning, the invention highlights high scattering regions and structurally sensitive regions reflected by the polarization degree distribution in the spatial dimension. In the channel dimension, the feature response is weighted and modulated according to the contribution of each channel to the change in polarization degree. This enables the network to prioritize effective information closely related to polarization characteristics during feature transmission and reconstruction, while suppressing components that are unrelated to polarization or dominated by noise.

[0096] A second aspect of the present invention provides a system 2. In some embodiments of the present invention, such as... Figure 6 As shown, system 2 includes: Image acquisition module 201 is used to acquire a first color image and a second color image under multiple polarization directions.

[0097] The pseudo-color reconstruction module 202 is used to generate a pseudo-color polarized image based on the cross-reconstruction of the color channels and polarization channels of the second color image.

[0098] The polarization degree generation module 203 is used to calculate the Stokes component based on multiple polarization directions and generate a single-channel polarization degree image.

[0099] Training module 204 is used to input the training set and the first color image into the convolutional neural network to perform network training.

[0100] The image restoration module 205 is used to input the underwater polarization image to be restored into the trained convolutional neural network to output the restored image.

[0101] This invention provides a system that centralizes data acquisition and preprocessing in a front-end imaging device or edge computing unit, executes polarization degree generation and pseudo-color reconstruction on a local processing board with sufficient computing power, and runs the training module and image restoration module on a general-purpose computing platform with high parallelism. This facilitates flexible integration or expansion of the multi-channel polarization information processing capabilities of this invention into existing hardware systems, reducing significant modifications to the original imaging chain; it also allows for independent optimization and resource scheduling of each module under different hardware conditions, enabling the system to rationally utilize computing and storage resources while meeting real-time or offline high-precision requirements.

[0102] An embodiment of the third aspect of the present invention provides an electronic device. In some embodiments of the present invention, such as... Figure 7 As shown, an electronic device is provided, including desktop computers, laptops, handheld computers, and cloud servers. Electronic device 3 includes, but is not limited to, a processor 301 and a memory 302. Those skilled in the art will understand that... Figure 7 This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3, including more or fewer parts than shown, or different parts.

[0103] Processor 301 is a central processing unit (CPU), but can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0104] Memory 302 is an internal storage unit of electronic device 3, exemplified by a hard disk or RAM of electronic device 3. Memory 302 can also be an external storage device of electronic device 3, exemplified by a plug-in hard disk, Smart MediaCard (SMC), Secure Digital (SD) card, Flash Card, etc., installed on electronic device 3. Memory 302 includes both internal and external storage units of electronic device 3. Memory 302 is used to store computer programs and other programs and data required by electronic device 3.

[0105] An embodiment of the fourth aspect of the present invention provides a computer-readable storage medium. In some embodiments of the present invention, a computer-readable storage medium is provided that, when executed by processor 301, implements the steps of the above-described method. Therefore, the computer-readable storage medium provided in the fourth aspect of the present invention has all the technical effects of the above-described steps, which will not be repeated here.

[0106] The above embodiments are only used to illustrate the technical solutions of this disclosure, and are not intended to limit them. Although this disclosure 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 of the technical features. The above modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be included within the protection scope of this disclosure.

Claims

1. An underwater image enhancement method based on multi-channel polarization information, characterized in that, Includes the following steps: Convolutional neural networks with multi-level encoders and decoders are constructed, and a dual-branch attention mechanism is set at the jump connections between each level of encoder and decoder according to the degree of polarization. In multiple polarization directions, a first color image of the sample in a non-scattering environment and a second color image in at least one scattering environment are acquired respectively. Based on the sub-images of the second color image in different polarization directions, multiple pseudo-color polarization images are generated by cross-recombining the polarization channel and the color channel, and a single-channel polarization degree image is generated by calculating the Stokes components in multiple polarization directions. A training set is constructed using the pseudo-color polarization image and the polarization degree image. The training set is then input into the convolutional neural network, and the convolutional neural network is trained using the first color image as the supervision label. The underwater polarization image to be restored is input into the trained convolutional neural network to output the underwater restored image.

2. The underwater image enhancement method according to claim 1, characterized in that, The convolutional neural network has a U-shaped architecture; and the steps for constructing a convolutional neural network containing multi-level encoders and decoders include: Multiple encoders are set sequentially from top to bottom at the input end. Each encoder compresses the input features into deep features step by step through a downsampling structure. A bottleneck layer is set between the bottom of the input and output ends to carry the deepest features and serve as a connection node between the encoding and decoding paths; Multiple decoders are set sequentially from bottom to top at the output end, and each decoder recovers the deep features step by step through an upsampling structure.

3. The underwater image enhancement method according to claim 2, characterized in that, The step of setting up a dual-branch attention mechanism based on polarization degree at the jump connection between encoders and decoders at each level includes: At the jump connection between the encoder and the decoder at each level, spatial attention weights and channel attention weights are generated based on the polarization image, respectively. The spatial attention weights are applied to features from the encoder to enhance the weighted spatial region based on the degree of performance of the spatial response in the polarization distribution. The channel attention weights are applied to features from the encoder to weaken the weighted channel components based on each channel's contribution to the polarization degree change.

4. The underwater image enhancement method according to claim 1, characterized in that, The step of acquiring a first color image of a sample in a non-scattering environment and a second color image in at least one scattering environment in multiple polarization directions includes: The first color image of the sample in multiple polarization directions was acquired in the scatter-free environment. Color polarization images corresponding to the sample are acquired in at least one scattering environment with different simulated scattering characteristics; The acquisition process is repeated based on the change in scattering intensity of the scattering medium, and all the color polarization images are combined to form a second color image corresponding to the first color image.

5. The underwater image enhancement method according to claim 4, characterized in that, The step of generating multiple pseudo-color polarized images by cross-recombining the sub-images of the second color image in different polarization directions through polarization channels and color channels includes: Extract m different color channels from the obtained sub-images, and select m polarization channels from all types of polarization directions; A unique correspondence is established between color channels and polarization channels to form a pseudo-color polarized image obtained by combining m color channels with m distinct polarization channels.

6. The underwater image enhancement method according to claim 5, characterized in that, The pseudo-color polarized image is characterized by a polarization response different from that of the second color image through the recombination of color channels and polarization channels, and expands the number of samples in the training set.

7. The underwater image enhancement method according to claim 6, characterized in that, The step of generating a polarization degree image from Stokes components calculated through multiple polarization directions includes: Based on multiple color polarimetric images obtained from m polarization channels, light intensity sampling is performed on the color channels under each polarization channel to determine the Stokes component used to characterize the polarization state. The degree of polarization, representing the ratio of polarization component to total light intensity, is calculated at the pixel level based on the Stokes component. The polarization degrees corresponding to different color channels are fused to generate the polarization degree image.

8. A system for implementing the underwater image enhancement method according to any one of claims 1-7, characterized in that, include: The image acquisition module is used to acquire a first color image and a second color image under multiple polarization directions; A pseudo-color reconstruction module is used to generate a pseudo-color polarized image based on the cross-reconstruction of the color channels and polarization channels of the second color image; The polarization degree generation module is used to calculate the Stokes component based on multiple polarization directions and generate a single-channel polarization degree image. The training module is used to input the training set and the first color image into the convolutional neural network to perform network training; The image restoration module is used to input the underwater polarization image to be restored into the trained convolutional neural network to output the restored image.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the underwater image enhancement method as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the underwater image enhancement method as described in any one of claims 1 to 8.

Images