Underwater degraded image polarization recovery method based on neural network

Abstract

The invention provides an underwater degraded image polarization recovery method and system based on a neural network, and relates to the field of underwater image recovery. The method comprises the following steps: firstly, collecting polarization images in a real underwater environment through a polarization camera to form a training data set; then, a deep learning scattering removal network is designed, the network is based on a polarization image degradation model, and a plurality of parts including a channel attention mechanism module and a dense residual module are adopted, so that the capturing capacity of multi-scale features is enhanced, and the performance and the training efficiency of the network are improved; finally, a joint loss function is designed by combining polarization and intensity information so as to guide the training process of the network, and it is ensured that the recovered image is optimal in the aspects of polarization and intensity.

Description

Technical Field

[0001] The invention belongs to the technical field of underwater image restoration, and in particular relates to a method for restoring polarization of degraded underwater images based on a neural network. Background Art

[0002] Underwater image restoration technology plays a vital role in many key areas, especially for complex underwater operations such as underwater rescue, marine resource exploration and underwater survey. However, the scattering and absorption of light by particles suspended in the water will weaken the light signal received by the imaging device, which will cause serious degradation of underwater image quality, manifested as insufficient contrast, unclear edges and shortened viewing distance. Faced with the challenge of "unclear sight and short viewing distance" caused by turbid water and scattering, we urgently need to adopt innovative optical information capture technology and advanced image processing methods to solve the problem of image degradation in turbid water environments. This includes recovering the target image from the underwater scene, eliminating the influence of backscattered light, and improving the contrast and clarity of underwater images.

[0003] In recent years, people have proposed a variety of methods to improve the quality of underwater imaging, such as histogram stretching, image filtering, Retinex imaging, dark channel prior method, and underwater polarization imaging. Compared with other methods, the imaging effect is extremely limited when multiple degradation factors coexist. Underwater polarization imaging has the advantage of accurately estimating the difference and connection between target information light and backscattered light, which can achieve the purpose of scattered light suppression and improve imaging clarity, and stands out from a number of underwater imaging methods.

[0004] At present, various polarization-based underwater image restoration methods have been proposed, among which the most classic model is the polarization imaging model proposed by Schechner et al. This method extracts polarization parameters from multiple polarization images to restore clear underwater images, improves the quality of underwater images, and also proves the superiority of using polarization technology underwater. However, the traditional underwater polarization imaging physical model is simplified or idealized to some extent. For example, Treibitz et al. consider the polarization degree of backscattered light as a single value and use the background of the local area to estimate the polarization degree of backscattered light. However, the polarization degree of backscattered light varies across the entire image plane, and considering it as a fixed single value may have a large deviation from the actual situation. Hu et al. estimated the polarization degree of backscattered light in the image based on the parameter fitting method, and used the background area to perform polynomial fitting on the polarization degree of backscattered light to estimate the polarization degree of the target area. However, the polarization degree of the target area estimated by the plane fitted with the polarization degree of backscattered light still has a large error, the interpretation of the target information light is still inaccurate, the extraction of the target information light is still inaccurate, and the image clarity still needs to be further improved. Summary of the invention

[0005] To solve the above technical problems, the present invention proposes a method and system for polarimetric restoration of underwater degraded images based on a neural network.

[0006] The first aspect of the present invention discloses a method for polarimetric restoration of underwater degraded images based on a neural network; the method includes:

[0007] Step S1, collect polarimetric images in a real underwater environment, and form a training data set after preprocessing; the sample labels of each data sample in the training data set include an intensity image, a degree of polarization image, and a polarization angle image;

[0008] Step S2, construct a de-scattering network based on the polarimetric image degradation model; wherein, the de-scattering network includes a shallow feature extraction module, a channel attention mechanism module, a dense residual module, a global feature fusion module, and a reconstruction module connected in sequence;

[0009] Step S3, input the samples in the training data set into the de-scattering network in sequence to obtain an enhancement result;

[0010] Step S4, calculate the joint loss function value of polarization and intensity based on the enhancement result and the sample label, and then perform backpropagation calculation on the gradient of the de-scattering network according to the joint loss function, and then obtain the optimized de-scattering network;

[0011] Step S5, input the target image into the optimized de-scattering network to obtain the restored image.

[0012] In the step S1, all polarimetric images are cropped horizontally and vertically to obtain a training data set containing underwater polarimetric image pairs.

[0013] In the step S2, the shallow feature extraction module includes two convolutional layers with a kernel size of 3×3, which are used to extract the shallow feature information of the polarimetric image and obtain a feature map.

[0014] In the step S2, the channel attention mechanism module uses a global average pooling layer to reduce the feature map to a dimension of 1×1×C, where C represents the number of channels of the feature map; then, weight learning is performed on each channel feature to realize weight adjustment of each feature in the feature map.

[0015] In the step S2, the dense residual module includes multiple residual modules, and each residual module extracts features through a series of convolutional layers and activation layers, where the output of each convolutional layer will be concatenated with the input of the subsequent layer to realize dense connection of features; at the same time, each residual module contains a residual connection, and the input is directly added to the output to promote feature reuse and information accumulation.

[0016] The global feature fusion module first concatenates the outputs of all residual blocks in the channel dimension through a connection layer to merge feature maps of different levels. The concatenated feature maps then pass through two consecutive convolutional layers to further extract and fuse features.

[0017] The reconstruction module uses upsampling to reconstruct a high-resolution image of the target sample size.

[0018] In step S4, the joint loss function L is:

[0019] L=L cont +λ×L polar

[0020] Among them, λ represents the empirical value; L cont Indicates strength loss, L cont =||I HR -I HR-gt || 2 , I HR represents the underwater intensity image enhanced by the descattering network, I HR-gt represents the intensity image in clear water environment, ||.|| 2 represents the L2 norm; L polar Represents polarization loss, which consists of polarization degree loss and polarization angle loss, that is, L polar =||DOLP-DOLP gt || 2 +α||AOP-AOP gt || 2 , DOLP and DOLP gt They represent the polarization image enhanced by the descattering network and the polarization image in clear water, AOP and AOP respectively. gt They represent the polarization angle image enhanced by the descattering network and the polarization angle image in clear water respectively, and α is the weight factor.

[0021] The second aspect of the present invention discloses a polarization restoration system for underwater degraded images based on a neural network; the system comprises:

[0022] The first processing module is configured to collect polarization images in a real water environment and form a training data set after preprocessing: intensity image, polarization degree image and polarization angle image;

[0023] The second processing module is configured to construct a descattering network based on the polarization image degradation model; wherein the descattering network includes a shallow feature extraction module, a channel attention mechanism module, a dense residual module, a global feature fusion module and a reconstruction module connected in sequence;

[0024] A third processing module, configured to sequentially input samples in a training dataset into a de-scattering network to obtain an enhancement result;

[0025] A fourth processing module, configured to calculate a joint loss function value of polarization and intensity based on the enhancement result and the sample label, and then perform backpropagation based on the joint loss function to calculate the gradient of the de-scattering network, thereby obtaining an optimized de-scattering network;

[0026] A fifth processing module, configured to input a target image into the optimized de-scattering network to obtain a restored image.

[0027] A third aspect of the present invention discloses an electronic device. The electronic device includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, the steps in a method for restoring the polarization of an underwater degraded image based on a neural network according to any one of the first aspects of the present disclosure are implemented.

[0028] A fourth aspect of the present invention discloses a computer-readable storage medium. A computer program is stored on the computer-readable storage medium. When the computer program is executed by a processor, the steps in a method for restoring the polarization of an underwater degraded image based on a neural network according to any one of the first aspects of the present disclosure are implemented.

[0029] In summary, the solution proposed by the present invention has the following technical effects:

[0030] (1) By introducing a degradation model based on polarization images and a dense residual network, since this technical means combines an in-depth understanding of the physical imaging model and the feature extraction ability of deep learning, the present application solves the problems of inaccurate image restoration and insufficient clarity that easily occur in traditional underwater polarization imaging methods when dealing with complex underwater environments, and realizes the effective restoration of underwater targets and the significant improvement of image quality in a turbid water environment.

[0031] (2) By introducing a channel attention mechanism module into the network, since this technical means enables the network to identify and emphasize key regions in the image, and assigns different attention weights to different regions, the present application solves the problem that traditional networks may not be able to effectively distinguish targets and backgrounds when processing polarization images, realizes more accurate positioning of underwater targets and enhancement of image details, thereby improving the overall quality of underwater image restoration.

[0032] (3) The present application constructs a joint loss function of polarization and intensity. Since this technical means not only focuses on the image reconstruction quality at the pixel level, but also considers the reconstruction effect of polarization information at the feature level, it solves the problem of polarization information loss or inaccurate reconstruction that may occur in the underwater image restoration process of existing methods, achieves an effective balance between intensity information and polarization information, and improves the polarization restoration quality of degraded underwater images. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0034] Figure 1 It is a flowchart of a method for restoring polarization of degraded underwater images based on a neural network according to an embodiment of the present invention;

[0035] Figure 2 A schematic diagram of an experiment for photographing a polarization data set according to an embodiment of the present invention;

[0036] Figure 3 A schematic diagram of a de-scattering network structure according to an embodiment of the present invention;

[0037] Figure 4 Schematic diagram of the channel attention mechanism structure according to an embodiment of the present invention;

[0038] Figure 5 2 is a schematic diagram of the structure of a dense residual module according to an embodiment of the present invention;

[0039] Figure 6 The figure is a structural diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0041] It is understood that the terms "first", "second", etc. used in this application may be used herein to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish a first element from another element. For example, a first image may be referred to as a second image, and similarly, a second image may be referred to as a first image without departing from the scope of this application. Both the first image and the second image are images, but they are not the same image.

[0042] In view of the shortcomings of the existing technology in target interpretation accuracy and image restoration quality, this paper adopts a deep learning framework and relies on large-scale data sets to deeply mine target features. Through a multi-layer network structure, layer-by-layer abstraction from basic features to complex features is achieved, and these high-level features are used to parse new input images to achieve high-definition underwater imaging.

[0043] According to an embodiment of the present invention, in a first aspect, a method for restoring polarization of degraded underwater images based on a neural network is provided; see Figure 1 ,This method first collects polarization images in a real underwater environment through a polarization camera to form a training dataset. Then, a deep learning descattering network is designed. The network is based on the polarization image degradation model and uses multiple parts including a channel attention mechanism module and a dense residual module to enhance the ability to capture multi-scale features and improve the performance and training efficiency of the network. Finally, a joint loss function is designed by combining polarization and intensity information to guide the network training process, ensuring that the restored image is optimal in both polarization and intensity.

[0044] The method comprises:

[0045] Step S1, collect polarization images in a real water environment, and form a training data set after preprocessing; the sample label of each data sample in the training data set includes an intensity image, a polarization degree image, and a polarization angle image. The intensity image, the polarization degree image, and the polarization angle image can be obtained through simple calculations of polarization images at various angles.

[0046] It is difficult to obtain a real underwater image dataset. Therefore, most underwater image datasets are currently generated by simulation. These datasets are usually based on the classic underwater scattering model. By manually setting relevant parameters, the clear image is artificially converted into a simulated underwater turbid image. Although this method is easy to operate, the obtained simulation dataset is often of limited quality and significantly different from the actual underwater environment, which affects the practicality of the dataset. In order to solve the limitations of the simulation method, the corresponding real underwater image dataset obtained by the present invention includes the following four steps. The schematic diagram of the dataset experiment is shown in the figure. Figure 2 shown.

[0047] Place the target object firmly in a specific location and adjust the polarization camera, LED light source, and polarizer to the appropriate angles to ensure that their positions remain fixed.

[0048] Pour clean water into the water tank, and fix the target objects on the base in a predetermined order to take pictures in a clear water environment.

[0049] Add 1 ml of skim milk to the sink and stir constantly to mix the water evenly. Then readjust and fix the position of the target object and shoot again.

[0050] Continue adding milk and repeat step 3).

[0051] In order to ensure the generalization ability of the network at different turbidity levels and for targets of different materials, multiple sets of underwater target polarization images are collected and preprocessed to construct a dataset. The dataset is randomly divided into training set, validation set and test set according to a certain ratio, which can be selected as 8:1:1. In addition, in order to expand the scale of the dataset, all images are cropped horizontally and vertically, the image side length can be selected as 64 pixels, and the moving step can be selected as 32 pixels. Finally, a large-scale dataset containing underwater polarization image pairs is obtained.

[0052] Step S2, constructing a descattering network based on the polarization image degradation model; wherein the descattering network includes a shallow feature extraction module, a channel attention mechanism module, a dense residual module, a global feature fusion module and a reconstruction module connected in sequence;

[0053] This step proposes a method for polarization restoration of underwater degraded images based on a neural network. The network structure includes five parts: shallow feature extraction module (SFE), channel attention mechanism module (CAM), residual module (RM), global feature fusion module (GFF), reconstruction module (RM) and intensity and polarization joint loss function. The descattering network structure is shown in 3. The network used in the present invention has four-channel input, which consists of polarization images of 0°, 45°, 90° and 135°. The image size is H×W×4, where H and W represent the height and width of the input image respectively.

[0054] Shallow feature extraction module: In this step, a shallow feature extractor is designed to extract shallow feature information of polarization images. Specifically, the shallow feature extraction module contains two convolutional layers with a convolution kernel size of 3×3. Smaller convolution kernels can reduce the number of model parameters, thereby reducing the complexity of the model and speeding up training. At the same time, small convolution kernels can extract more detailed features, and through two 3×3 convolutional layers, the depth of the network can be increased, thereby improving the model's ability to learn complex features.

[0055] Channel attention mechanism module: This module first uses the Global Average Pooling (GAP) technique to reduce the feature map to a dimension of 1×1×C, where C represents the number of channels in the feature map. Then, two 1×1 convolution operations are deployed in succession to learn the weights of the features of each channel. For details, please refer to Figure 4 , the first 1×1 convolution adjusts the number of convolution kernels to promote the integration of features from different channels and reduce the feature dimension, thereby reducing the computational burden of the model. Then, the second 1×1 convolution operation restores the feature dimension to ensure that the number of weights corresponds exactly to the number of feature channels. Finally, these coefficients are multiplied by their corresponding original feature channels to complete the feature weight adjustment.

[0056] In this step, after adding the dense residual module to the channel attention mechanism module, features are extracted through a series of convolutional layers and activation layers. The output of each convolutional layer is concatenated with the input of the subsequent layer to achieve dense connection of features. At the same time, each RDB contains a residual connection, which directly adds the input to the output, thereby promoting feature reuse and information accumulation, and alleviating the gradient disappearance problem in deep network training. This design not only enhances the network's ability to capture multi-scale features, but also improves the performance and training efficiency of the network through the combination of residual learning and dense connections. Specific details are as follows Figure 5 shown.

[0057] Global feature fusion module: This step uses the global feature fusion module to fuse feature information at different levels to enhance the network's ability to reconstruct polarization image details, especially when processing low-resolution polarization images. The global feature fusion module is located after the dense residual module, and further fuses the features from different residual blocks so that the network can capture more global information, thereby improving the quality of image reconstruction. Specifically, the global feature fusion module first concatenates the outputs of all residual blocks in the channel dimension through a connection layer to merge feature maps at different levels so that the network can consider information at different scales at the same time. The concatenated feature map then passes through two consecutive convolutional layers, both of which use the ReLU activation function to further extract and fuse features. This process not only enhances the expressiveness of the features, but also helps the network learn more complex feature combinations.

[0058] Reconstruction module: This step uses an upsampling module to reconstruct a high-resolution image of the target image size. Specifically, the feature map is first further processed through a convolutional layer followed by a ReLU activation function. This helps extract a richer feature representation before upsampling. The upsampling is then achieved by rearranging the pixels of the feature map while maintaining the spatial consistency of the feature map. After upsampling, the feature map is processed through a final convolutional layer to generate the final high-resolution polarization image.

[0059] Step S3, inputting the samples in the training data set into the descattering network in sequence to obtain an enhanced result;

[0060] Step S4, based on the enhancement result and the sample label, calculate the joint loss function value of polarization and intensity, and then perform back propagation to calculate the gradient of the descattering network according to the joint loss function, so as to obtain the optimized descattering network;

[0061] Since directly calculating the loss of output and label cannot match the polarization image with the true value image well, this step combines the intensity loss and polarization loss as the final loss function for network training. By calculating the difference between the label and the output, the intensity loss is expressed as:

[0062] L cont =||I HR -I HR-gt || 2

[0063] Among them, I HR represents the underwater image enhanced by the network proposed in this patent, I HR-gt is the intensity image in clear water environment, and ||.||2 is the L2 norm.

[0064] Polarization loss consists of polarization degree loss and polarization angle loss, which are calculated as follows:

[0065] L polar =||DOLP-DOLP gt || 2 +α||AOP-AOP gt || 2

[0066] Among them, DOLP and DOLP gt They represent the polarization degree image enhanced by the proposed network and the polarization degree image in clear water, AOP and AOP respectively. gt They represent the polarization angle image enhanced by the proposed network and the polarization angle image in clear water respectively. α is the weight factor used to balance the parameters. Therefore, the final loss of the network is expressed as follows:

[0067] L=L cont +λ×Lpolar

[0068] Here, λ is an empirical value, which is used to balance the intensity loss and polarization-aware loss and constrain the two losses to be of the same order of magnitude.

[0069] Step S5: input the target image into the optimized descattering network to obtain a restored image.

[0070] The second aspect of the present invention discloses a polarization restoration system for underwater degraded images based on a neural network; the system comprises:

[0071] The first processing module is configured to collect polarization images in a real water environment and form a training data set after preprocessing; the sample label of each data sample in the training data set includes an intensity image, a polarization degree image, and a polarization angle image;

[0072] The second processing module is configured to construct a descattering network based on the polarization image degradation model; wherein the descattering network includes a shallow feature extraction module, a channel attention mechanism module, a dense residual module, a global feature fusion module and a reconstruction module connected in sequence;

[0073] A third processing module is configured to sequentially input samples in the training data set into the descattering network to obtain an enhancement result;

[0074] A fourth processing module is configured to calculate a joint loss function value of polarization and intensity based on the enhancement result and the sample label, and then perform back propagation to calculate the gradient of the descattering network according to the joint loss function, so as to obtain an optimized descattering network;

[0075] The fifth processing module is configured to input the target image into the optimized descattering network to obtain a restored image.

[0076] The third aspect of the present invention discloses an electronic device. The electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps in any one of the methods for restoring polarization of degraded underwater images based on a neural network in the first aspect of the present disclosure are implemented.

[0077] Figure 6 is a structural diagram of an electronic device according to an embodiment of the present invention, such as Figure 6As shown, the electronic device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. Among them, the processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the electronic device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, near field communication (NFC) or other technologies. The display screen of the electronic device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the electronic device can be a touch layer covered on the display screen, or a button, a trackball or a touch pad set on the housing of the electronic device, or an external keyboard, touch pad or mouse, etc.

[0078] Those skilled in the art will understand that Figure 6 The structure shown in the figure is only a structural diagram of the part related to the technical solution of the present disclosure, and does not constitute a limitation on the electronic device to which the technical solution of the present application is applied. The specific electronic device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0079] The fourth aspect of the present invention discloses a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in any one of the steps in the method for restoring polarization of underwater degraded images based on a neural network in the first aspect of the present disclosure are implemented.

[0080] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them. Although the present invention has been described in detail with reference to the above embodiments, it should be understood by those skilled in the art that the technical solutions described in the above embodiments can still be modified, or some or all of the technical features therein can be replaced by equivalents, and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for restoring polarization of degraded underwater images based on a neural network, characterized in that: The method comprises: Step S1, collecting polarization images and intensity images in a real water environment, and forming a training data set after preprocessing; the sample label of each data sample in the training data set includes an intensity image, a polarization degree image, and a polarization angle image; Step S2, constructing a descattering network based on the polarization image degradation model; wherein the descattering network includes a shallow feature extraction module, a channel attention mechanism module, a dense residual module, a global feature fusion module and a reconstruction module connected in sequence; Step S3, inputting the samples in the training data set into the descattering network in sequence to obtain an enhanced result; Step S4, based on the enhancement result and the sample label, calculate the joint loss function value of polarization and intensity, and then perform back propagation to calculate the gradient of the descattering network according to the joint loss function, so as to obtain the optimized descattering network; Step S5: input the target image into the optimized descattering network to obtain a restored image.

2. The method according to claim 1, characterized in that In step S1, all polarization images are cropped horizontally and vertically to obtain a training data set containing underwater polarization image pairs.

3. The method according to claim 2, characterized in that In step S2, the shallow feature extraction module includes two convolution layers, and the size of the convolution kernel is 3×3, which is used to extract the shallow feature information of the polarization image to obtain a feature map.

4. The method according to claim 3, characterized in that: In step S2, the channel attention mechanism module uses a global average pooling layer to reduce the feature map to a dimension of 1×1×C, where C represents the number of channels of the feature map; then, weight learning is performed on each channel feature to adjust the weight of each feature in the feature map.

5. The method according to claim 4, characterized in that In step S2, the dense residual module includes multiple residual modules, each residual module extracts features through a series of convolutional layers and activation layers, wherein the output of each convolutional layer is concatenated with the input of the subsequent layer to achieve dense connection of features; at the same time, each residual module contains a residual connection, which directly adds the input to the output, thereby promoting feature reuse and information accumulation.

6. The method according to claim 5, characterized in that The global feature fusion module first concatenates the outputs of all residual blocks in the channel dimension through a connection layer to merge feature maps of different levels. The concatenated feature maps then pass through two consecutive convolutional layers to further extract and fuse features. The reconstruction module uses upsampling to reconstruct a high-resolution image of the target sample size.

7. The method according to claim 1, characterized in that In step S4, the joint loss function L is: L=L cont +λ×L polar Among them, λ represents the empirical value; L cont Indicates strength loss, L cont =||I HR -I HR-gt ||2, I HR represents the underwater intensity image enhanced by the descattering network, I HR-gt represents the intensity image in clear water environment, ||.||2 represents the L2 norm; L polar Represents polarization loss, which consists of polarization degree loss and polarization angle loss, that is, L polar =||DOLP-DOLP gt ||2+α||AOP-AOP gt ||2, DOLP and DOLP gt They represent the polarization image enhanced by the descattering network and the polarization image in clear water, AOP and AOP respectively. gt They represent the polarization angle image enhanced by the descattering network and the polarization angle image in clear water respectively, and α is the weight factor.

8. A neural network-based underwater degraded image polarization restoration system, characterized in that: The system comprises: The first processing module is configured to collect polarization images in a real water environment and form a training data set after preprocessing; the sample label of each data sample in the training data set includes an intensity image, a polarization degree image, and a polarization angle image; The second processing module is configured to construct a descattering network based on the polarization image degradation model; wherein the descattering network includes a shallow feature extraction module, a channel attention mechanism module, a dense residual module, a global feature fusion module and a reconstruction module connected in sequence; A third processing module is configured to sequentially input samples in the training data set into the descattering network to obtain an enhancement result; A fourth processing module is configured to calculate a joint loss function value of polarization and intensity based on the enhancement result and the sample label, and then perform back propagation to calculate the gradient of the descattering network according to the joint loss function, so as to obtain an optimized descattering network; The fifth processing module is configured to input the target image into the optimized descattering network to obtain a restored image.

9. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps in the method for polarization restoration of degraded underwater images based on a neural network as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the steps of the method for polarization restoration of underwater degraded images based on a neural network according to any one of claims 1 to 7 are implemented.

Images