Polarization image defogging method and device based on deep learning and storage medium

Abstract

The invention discloses a polarization image defogging method based on deep learning. The method comprises the following steps: S1, collecting data and constructing a multi-mode water mist degradation data set; s2, analyzing the image data in the data set into four-channel polarization direction image data and RGB intensity image data; s3, inputting four-channel polarization direction image data, and constructing a network adaptive learning polarization parameter image to complete atmospheric transmittance estimation based on image polarization information; s4, connecting the polarization-transmittance output and the RGB-transmittance output along the channel direction, accessing a multimode information fusion network based on feature attention, and finally accessing a clear image generation module to output a clear image; and S5, optimizing network parameters by adopting an L1 loss function, and improving the generalization capability of the network. According to the method, through fusion of the polarization features and the RGB features, in combination with a feature attention mechanism, dominant feature information and polarization recessive feature information in the image are mined at the same time, and the robustness of a complex scene is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of polarization dehazing, and particularly relates to a polarization image dehazing method, device and storage medium based on deep learning. Background Art

[0002] The purpose of image dehazing technology is to improve the clarity and information accuracy of images by eliminating the degradation effect of haze on images. Traditional image dehazing methods are mainly based on physical models (such as dark channel prior, color attenuation prior) or image enhancement techniques (such as Retinex, wavelet transform), but there are problems such as strong scene dependence, complex manual adjustment of parameters, and limited dehazing effect. Therefore, the existing technology often combines the automatic feature learning ability of deep learning for use, and its core is to directly or indirectly estimate the atmospheric transmittance (scattering coefficient) through a neural network and restore the haze-free image.

[0003] The polarization imaging technology in the existing technology can obtain the polarization information of light (such as degree of polarization, polarization angle). By distinguishing the polarization characteristics of the target radiation light and the atmospheric scattered light, it can more accurately suppress the haze interference, thus providing a new physical basis for dehazing.

[0004] The existing polarization image dehazing methods based on deep learning mainly combine polarization information with a deep learning network to optimize the atmospheric transmittance estimation and thus achieve image restoration. In polarization image dehazing based on deep learning, the conventional operation is to use individual parameters of the polarization image (such as degree of polarization map, polarization angle map) as inputs and combine them with a convolutional neural network (CNN) for atmospheric transmittance estimation or image restoration. For example, by manually selecting the degree of polarization map or polarization angle map as feature inputs to construct an end-to-end dehazing model. However, such methods have significant limitations:

[0005] (1) The manually selected polarization parameters cannot comprehensively describe the polarization characteristics of complex scenes. Especially under complex conditions such as medium fog and thick fog, single or fixed polarization parameters are difficult to adapt to the dynamically changing haze concentration and scene structure;

[0006] (2) The manually selected polarization parameters are easily affected by noise, resulting in poor robustness of the network's estimation of the atmospheric transmittance. Finally, problems such as color distortion, insufficient contrast or detail loss may occur in the restored image;

[0007] (3) Existing technologies improve the utilization of polarization parameters by optimizing the network structure or combining multi-source prior information, or combine physical models (such as polarization difference models) with deep learning, and use polarization angle distribution or atmospheric light gradient prior to constrain the network output. However, they still rely on artificially designed polarization features or prior constraints and lack the ability to adaptively select polarization parameters, resulting in the network's difficulty in dynamically adjusting its dependence on key polarization parameters in complex scenarios (such as high-concentration haze and strong light changes), thus limiting the generalization performance of the algorithm.

[0008] To this end, the present application specifically proposes a polarization image dehazing method based on deep learning to solve the above technical problems. Summary of the Invention

[0009] The main objective of the present invention is to provide a polarization image dehazing method based on deep learning. By introducing an adaptive learning module, the network can autonomously select or fuse key polarization parameters (such as multi-angle polarization degree, polarization angle gradient, etc.) according to task requirements (such as dehazing effect, color fidelity), so as to more comprehensively describe the polarization characteristics of the scene, and finally improve the dehazing effect and algorithm generalization ability in complex haze scenes, so as to solve the technical problems proposed in the background art.

[0010] The present invention adopts the following technical solutions to solve the above technical problems:

[0011] S1. Collect the original color polarization image data and construct a multi-modal water mist degradation data set including scenes such as sand land and water surface.

[0012] S2. Parse the original color polarization image data in the data set into four-channel polarization direction image data and RGB intensity image data.

[0013] S3. Input the four-channel polarization direction image data, and through the polarization information parsing front-end network, make full use of the convolutional neural network to learn high-dimensional polarization information, obtain the polarization parameter image, and then use the polarization-transmittance mapping dehazing task network to input the polarization parameter image to obtain the atmospheric transmittance image estimated based on the image polarization information, so as to complete the atmospheric transmittance estimation based on the image polarization information.

[0014] The polarization information parsing front-end network includes multiple hidden layer units. The output of the previous hidden layer unit is the input of the next unit. Each hidden layer unit performs a non-linear pixel-by-pixel operation between the input images to fully extract the polarization information and output a new set of feature images. After passing through multiple hidden layer units, the original polarization direction image outputs a set of adaptive encoded polarization parameter images. The hidden layer unit is the basic component unit of the high-dimensional representation network and includes a convolutional layer, an activation layer, and a batch normalization layer (Batch Normalization, BN layer);

[0015] Among them, the polarization-transmittance mapping dehazing task network adopts the AOD-Net method, which includes 1 layer of 1×1 convolutional layer, 2 layers of 3×3 convolutional layers, 1 layer of 5×5 convolutional layer, 1 layer of 7×7 convolutional layer and ReLU activation layer, as well as three "concat" connection layers. Among them, the "concat1" layer connects the features from the "conv1" and "conv2" layers, "concat2" connects the features from the "conv2" and "conv3" layers, and "concat3" connects the features from the "conv1", "conv2", "conv3" and "conv4" layers, and finally outputs the atmospheric transmittance image estimated based on polarization information;

[0016] S4. Connect the output layer of the polarization-transmittance mapping dehazing task network and the output layer of the RGB-transmittance mapping network along the channel direction, connect them to the multi-modal information fusion network based on feature attention, and then connect them to the clear image generation module. After data calculation and processing, the corresponding clear image is finally output;

[0017] Among them, the multi-modal information fusion network based on feature attention is composed of N consecutive B modules (BasicBlockStructure). The B module is composed of local residual learning and feature attention FA module. Local residual learning allows the network to bypass less important information through multiple local residual connections to achieve effective feature extraction and attention allocation. By combining multiple B modules and using the feature attention mechanism, the network's ability to fuse dominant RGB features and implicit polarization information is enhanced. The number of B modules is a hyperparameter of network design. By adjusting this hyperparameter, the complexity and learning ability of the network can be controlled to better adapt to the requirements of the dehazing task;

[0018] At the same time, in order to obtain a 3-channel transmittance image, 2 layers of 3×3 convolutional layers are connected at the output end of the multi-modal information fusion network and connected to the clear image generation module. After data calculation and processing, the corresponding clear image is finally output;

[0019] S5. Use the L1 loss function to optimize the network parameters and improve the network generalization ability.

[0020] Preferably, in the S1 step, the data in the data set is collected by setting specified lighting conditions, imaging angles and fog concentration environments in the imaging simulation experimental environment, and using a split focal plane visible light color polarization camera to collect imaging data of various typical scaled targets with three concentrations of light fog, medium fog and heavy fog under different backgrounds.

[0021] Preferably, the specific operation process of data set parsing in the S2 step includes;

[0022] S21. Four polarization direction intensity maps are obtained through downsampling as the original polarization images;

[0023] S22. Perform color image analysis based on the Bayer arrangement on each of the four polarization direction intensity maps obtained by downsampling to obtain four RGB intensity maps in the RGB polarization directions;

[0024] S23. Fuse the four RGB intensity maps in the RGB polarization directions obtained in step S22 to finally obtain an RGB intensity map, where:

[0025]

[0026] Among them, I (R,G,B) (e) represents the RGB intensity map in the e -th angle polarization direction, and e is a specified angle value.

[0027] Preferably, the specific calculation expression for the clear image generation module to perform data calculation and processing in step S4 is:

[0028] J(x) = K(x)I(x) - K(x) + b

[0029] Among them, K(x) is the estimated transmittance, I(x) is the input foggy image, and J(x) is the generated clear image.

[0030] Preferably, the specific expression formula for the L1 loss function to optimize the network parameters in step S5 is:

[0031]

[0032] Among them, θΘ represents the parameters of the entire defogging network, I gt represents the real clear image, I haze represents the input foggy image, and G(I haze ) represents the clear image finally generated by the network.

[0033] On the other hand, the present invention also discloses a computer - readable storage medium storing a computer program, and when the computer program is executed by a processor, the processor is caused to execute the steps of the above - mentioned method.

[0034] On yet another hand, the present invention also discloses a computer device including a memory and a processor, the memory storing a computer program, and when the computer program is executed by the processor, the processor is caused to execute the steps of the above - mentioned method.

[0035] As can be seen from the above technical solutions, the present invention provides a polarization image defogging method based on deep learning. Compared with the prior art, the present invention has the following advantages:

[0036] 1. Based on the foggy day polarization image dataset, the system constructed in this application can, through the fusion of polarization features and RGB features, combined with the feature attention mechanism, mine the dominant feature information and polarization implicit feature information in the image, and enhance the robustness of complex scenes.

[0037] 2. The present invention constructs a polarization information parsing front-end network using N layers of 1×1 convolutional layers, ReLU activation layers, and batch normalization BN layers to adaptively extract multi-dimensional polarization features, breaking through the limitations of traditional artificial parameters. This network can automatically adjust the learning strategy according to task requirements to determine the polarization parameter image beneficial to the task, enabling the network to flexibly handle different application scenarios.

[0038] 3. The present invention provides the physical properties of fog (such as polarization angle, degree of polarization) based on polarization information to achieve image defogging, making up for the deficiencies of traditional RGB data under complex lighting / media. It can not only improve the adaptability to dynamic fog distribution using multi-scale and residual designs, but also enhance scene depth and material perception while retaining color information based on multi-modal feature complementarity.

[0039] It should be understood that the content described in this part is not intended to identify the key or important features of the embodiments of the present invention, nor is it used to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. Of course, any product implementing the present invention does not necessarily need to achieve all the above-mentioned advantages simultaneously. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The specification drawings constituting a part of this application are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0041] Figure 1 is the overall three-dimensional view of the present invention;

[0042] Figure 2 is the schematic diagram of the color polarization image parsing process of the present invention;

[0043] Figure 3 is the schematic diagram of the structure of the polarization image parsing front-end network of the present invention;

[0044] Figure 4 is the schematic diagram of the structure of the defogging task network of the present invention;

[0045] Figure 5 is the schematic diagram of the polarization image defogging network of the present invention;

[0046] Figure 6 is the schematic diagram of the structure of the feature attention mechanism of the present invention;

[0047] Figure 7 Schematic diagram of the multi-modal information fusion network structure of the present invention. Specific embodiments

[0048] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0049] In the embodiment, see in detail Figures 1 to 7 。

[0050] Such as Figure 1 shown, the polarization image dehazing method based on deep learning proposed in the embodiment of the present invention includes the following steps:

[0051] Step 1: Dataset construction

[0052] In the imaging simulation experiment environment, by setting various lighting conditions, imaging angles and fog concentration environments, use a split focal plane visible light color polarization camera to collect imaging data of various typical scaled targets under different backgrounds. The constructed multi-modal water fog degradation dataset (including scenes such as sand land and water surface) covers three concentrations of light fog, medium fog and heavy fog.

[0053] Step 2: Data parsing

[0054] The collected original polarization images cannot be directly used for training. Such as Figure 2 shown, according to the network input requirements, the original color polarization image dataset is parsed into a four-channel polarization direction image dataset and an RGB intensity image dataset. The parsing steps are as follows:

[0055] (1) The collected original polarization images are downsampled to obtain four polarization direction intensity maps (grayscale);

[0056] (2) Perform color image parsing based on the Bayer arrangement on the four polarization direction intensity maps obtained by downsampling to obtain four RGB intensity maps in the RGB polarization directions;

[0057] (3) The four RGB intensity maps in the RGB polarization directions obtained in (2) are fused according to the following formula (1) to finally obtain an RGB intensity map, and there is:

[0058]

[0059] Among them, I (R,G,B)(e) represents the RGB intensity map in the polarization direction of the e-th angle, where e is a specified angle value;

[0060] Step 3: Polarization information parsing and transmittance estimation

[0061] Input the polarization images in four directions (0°, 45°, 90°, 135°). Through the polarization information parsing front-end network as shown in Figure 3 , make full use of the convolutional neural network to learn the high-dimensional polarization information, and then use the polarization-transmittance mapping, such as the defogging task network shown in Figure 4 , to complete the atmospheric transmittance estimation based on the image polarization information.

[0062] Among them, the defogging task network adopts multi-scale convolution, which can extract the density distribution characteristics of fog with different receptive fields, adapt to thin fog and thick fog scenarios, and fuse multi-scale features (such as low-level details and high-level semantics) to enhance the robustness of transmittance estimation.

[0063] Step 4: Multi-modal information fusion

[0064] Connect the output layers of the polarization-transmittance mapping network and the RGB-transmittance mapping network along the channel direction, and connect them to the multi-modal information fusion network based on feature attention, as shown in Figure 5 , Figure 6 and Figure 7 . Then connect it to the clear image generation module, and finally output the corresponding clear image using the following formula:

[0065] J(x) = K(x)I(x) - K(x) + b

[0066] where K(x) is the estimated transmittance, I(x) is the input foggy image, and J(x) is the generated clear image.

[0067] Step 5: Network training and optimization

[0068] Use the L1 loss function to optimize the network parameters and improve the network generalization ability. The formula is:

[0069]

[0070] where θΘ represents the parameters of the entire defogging network, I gt represents the real clear image, I haze represents the input foggy image, and G(I haze ) represents the clear image finally generated by the network.

[0071] In a specific embodiment, for the target evaluation results of different algorithms, in actual experiments, there is the following table of data:

[0072] Method PSNR SSIM MSE VIF NIQE TIME HE 15.221 0.688 110.05 0.752 0.781 0.086 DCP 16.920 0.776 106.11 0.768 7.235 0.258 FFA-Net 14.023 0.648 112.07 0.777 8.071 2.992 POL 19.431 0.800 79.22 0.790 7.979 0.603 DehazeNet 13.395 0.428 108.21 0.565 0.225 3.736 AOD-Net 17.337 0.75 103.49 0.777 0.533 1.887 GridDehazeNet 10.884 0.622 112.09 0.774 8.201 3.304 PolNet 19.867 0.832 79.02 0.874 7.729 2.031

[0073] Therefore, through the fusion of polarization features and RGB features, combined with the feature attention mechanism, the present application simultaneously mines the dominant feature information and polarization implicit feature information in the image, improves the robustness of complex scenes, and constructs a polarization information parsing front-end network using N layers of 1×1 convolutional layers, ReLU activation layers, and batch normalization BN layers to adaptively extract multi-dimensional polarization features, breaking through the limitations of traditional artificial parameters. This network can automatically adjust the learning strategy according to task requirements to determine the polarization parameter image beneficial to the task. This adaptive learning ability is a major advantage of deep learning models, enabling the network to flexibly handle different application scenarios.

[0074] In summary, the system network adopted by this algorithm integrates polarization information parsing, multi-scale fusion, and attention mechanism. While retaining the advantages of physical models, it can further enhance the defogging ability in complex scenes using deep learning, especially suitable for various harsh environments.

[0075] On the other hand, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to execute the steps of the above method.

[0076] On yet another aspect, the present invention also discloses a computer device including a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to execute the steps of the above method.

[0077] In another embodiment provided by the present application, there is also provided a computer program product containing instructions, which, when running on a computer, causes the computer to execute any one of the polarization image defogging methods based on deep learning in the above embodiments.

[0078] It can be understood that the system provided by the embodiments of the present invention corresponds to the method provided by the embodiments of the present invention. The explanations, examples, and beneficial effects of related content can refer to the corresponding parts in the above method.

[0079] The embodiments of the present application also provide an electronic device including a processor, a communication interface, a memory, and a communication bus. Among them, the processor, the communication interface, and the memory complete communication with each other through the communication bus.

[0080] The memory is used to store a computer program.

[0081] The processor is used to implement the above polarization image defogging method based on deep learning when executing the program stored in the memory.

[0082] The communication bus mentioned in the above electronic device may be a Peripheral Component Interconnect Standard bus, an Extended Industry Standard Architecture bus, or the like. The communication bus may be divided into an address bus, a data bus, a control bus, and the like.

[0083] The communication interface is used for communication between the above electronic device and other devices.

[0084] The memory may include a random access memory and may also include a non-volatile memory, such as at least one disk memory. Optionally, the memory may also be at least one storage device located far from the aforementioned processor.

[0085] The above-mentioned processor may be a general-purpose processor, including a central processing unit, a network processor, etc.; it may also be a digital signal processor, an application-specific integrated circuit, a field programmable gate array, or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.

[0086] It should also be noted that the electronic device further includes a terminal device, which may also be referred to as a terminal, a user equipment, a mobile station, a mobile terminal, etc. The terminal device may be a mobile phone, a smart TV, a wearable device, a tablet computer, a computer with wireless transceiver function, a virtual reality terminal device, an augmented reality terminal device, a wireless terminal in industrial control, a wireless terminal in unmanned driving, a wireless terminal in remote surgery, a wireless terminal in smart grid, a wireless terminal in transportation safety, a wireless terminal in smart city, a wireless terminal in smart home, and so on. The embodiments of the present application do not limit the specific technologies and specific device forms adopted by the terminal device.

[0087] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wire (such as coaxial cable, optical fiber, digital subscriber line) or wirelessly (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer, or a data storage device such as a server or data center that includes one or more integrated available media. The available medium can be a magnetic medium (such as a floppy disk, hard disk, magnetic tape), an optical medium, or a semiconductor medium (such as a solid-state drive SSD), etc.

[0088] The above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

[0089] In addition, it should be noted that if there are directional indications (such as up, down, left, right, front, back...) involved in the embodiments of the present invention, the directional indications are only used to explain the relative positional relationship and movement conditions between components in a specific posture. If the specific posture changes, the directional indications will also change accordingly.

[0090] In addition, if there are descriptions involving "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only for descriptive purposes, and cannot be understood as indicating or implying their relative importance or implicitly indicating the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one such feature. In addition, the meaning of "and / or" appearing throughout the text includes three parallel scenarios. Taking "A and / or B" as an example, it includes scenario A, scenario B, or the scenario where A and B are satisfied simultaneously. In addition, in the embodiments of the present invention, "a plurality" means two or more. In addition, the technical solutions between the various embodiments can be combined with each other, but it must be based on the fact that those of ordinary skill in the art can implement it. When the combination of technical solutions results in contradictions or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the protection scope required by the present invention.

Claims

1. A polarization image dehazing method based on deep learning, characterized in that, It includes the following steps: S1. Collect the original color polarization image data and construct a multi-modal water fog degradation dataset including sandy land and water surface scenes; S2. Parse the original color polarization image data in the dataset into four-channel polarization direction image data and RGB intensity image data; S3. Input the four-channel polarization direction image data, pass it through the polarization information parsing front-end network, use the convolutional neural network to learn the high-dimensional polarization information, obtain the polarization parameter image, and then use the polarization-transmittance mapping defogging task network to input the polarization parameter image to obtain the atmospheric transmittance image estimated based on the polarization information, so as to complete the atmospheric transmittance estimation based on the image polarization information; S4. Connect the output layer of the polarization-transmittance mapping defogging task network and the output layer of the RGB-transmittance mapping network along the channel direction, connect them to the multi-modal information fusion network based on feature attention, and then connect them to the clear image generation module. After data calculation and processing, finally output the corresponding clear image; S5. Use the L1 loss function to optimize the network parameters and improve the network generalization ability.

2. The polarization image defogging method based on deep learning according to claim 1, wherein In the S1 step, the data in the dataset is collected by setting specified lighting conditions, imaging angles and fog concentration environments in the imaging simulation experiment environment, and using a split focal plane visible light color polarization camera to collect imaging data of multiple typical scaled targets with three concentrations of light fog, medium fog and heavy fog under different backgrounds.

3. The polarization image defogging method based on deep learning according to claim 1, wherein The specific operation process of the dataset parsing in the S2 step includes; S21. Obtain four polarization direction intensity maps as the original polarization images through downsampling; S22. Respectively perform color image parsing based on the Bayer arrangement on the four polarization direction intensity maps obtained by downsampling to obtain four RGB intensity maps in the RGB polarization directions; S23. Fuse the four RGB intensity maps in the RGB polarization directions obtained in the S22 step to finally obtain the RGB intensity map, where: Among them, I (R,G,B) (e) represents the RGB intensity map in the polarization direction of the e-th angle, where e is a specified angle value.

4. The polarization image defogging method based on deep learning according to claim 1, characterized in that In the S3 step, the polarization information parsing front-end network includes multiple hidden layer units. The output of the previous hidden layer unit is the input of the next unit. Each hidden layer unit performs a non-linear pixel-by-pixel operation between the input images to fully extract the polarization information and output a new set of feature images; The hidden layer unit is the basic building block of the high-dimensional representation network, including a convolutional layer, an activation layer and a batch normalization layer; After passing through multiple hidden layer units, the original polarization direction image outputs a set of adaptive encoded polarization parameter images.

5. The polarization image defogging method based on deep learning according to claim 1, wherein In the S3 step, the polarization-transmittance mapping defogging task network adopts the AOD-Net method, which includes 1 layer of 1×1 convolutional layer, 2 layers of 3×3 convolutional layers, 1 layer of 5×5 convolutional layer, 1 layer of 7×7 convolutional layer and a ReLU activation layer, as well as three concat connection layers.

6. The polarization image defogging method based on deep learning according to claim 1, characterized in that, In the S3 step, the multi-modal information fusion network based on feature attention is composed of N consecutive B basic block structures, where: The B basic block structure is composed of local residual learning and a feature attention FA module. Local residual learning allows the network to bypass less important information through multiple local residual connections to achieve effective feature extraction and attention distribution; Multiple B basic block structures are combined and the feature attention mechanism is utilized to enhance the network's fusion ability for explicit RGB features and implicit polarization information. The number of B basic block structures is a hyperparameter of the multi-modal information fusion network, which is used to adjust and control the complexity and learning ability of the network; A 2-layer 3×3 convolutional layer is connected to the output end of the multi-modal information fusion network and is connected to the clear image generation module, which is used to finally output the corresponding clear image after data calculation and processing.

7. The polarization image defogging method based on deep learning according to claim 1, characterized in that, The specific calculation expression for the clear image generation module to perform data calculation and processing in the S4 step is: J(x) = K(x)I(x) - K(x) + b where K(x) is the estimated transmittance, I(x) is the input foggy image, and J(x) is the generated clear image.

8. The polarization image defogging method based on deep learning according to claim 1, characterized in that The specific expression formula for the L1 loss function to optimize the network parameters in the S5 step is: Among them, θΘ represents the parameters of the entire defogging network, and I gt represents the real clear image, and I haze represents the input foggy image, and G(I haze ) represents the clear image finally generated by the network.

9. A computer-readable storage medium, characterized in that, A computer program is stored, and when the computer program is executed by a processor, the processor is caused to execute the steps of the method according to any one of claims 1 to 8.

10. A computer device, characterized in that, It includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor is caused to execute the steps of the method according to any one of claims 1 to 8.

Images