An unsupervised meter fuzzy image restoration method fusing physical degradation model

Abstract

This invention discloses an unsupervised method for restoring blurred meter images by incorporating a physical degradation model. This method addresses the problems of blurring, reduced contrast, and lost detail in meter images under complex environments due to factors such as low light and fog. The method includes: constructing an unpaired meter image dataset; constructing an unsupervised network for restoring blurred meter images, comprising a degradation category classification module, a physical parameter estimation module, and an image generation module; pre-training the degradation category classification module and the physical parameter estimation module; performing master training on the image generation module to establish a mapping relationship between the degraded image domain and the clear image domain; and inputting the blurred meter image to be restored into the trained restoration network to output a clear meter image. This invention integrates unsupervised learning, physical parameter constraints, and degradation category conditional mechanisms, effectively improving the quality of blurred meter image restoration even in the absence of paired training samples.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to an unsupervised restoration method for meter images degraded in low light and foggy weather. Background Technology

[0002] As a commonly used condition monitoring device in industrial scenarios such as chemical, power, energy and manufacturing, meters can display key parameters such as pressure, temperature, liquid level and flow rate in real time. The clear acquisition of their image information is of great significance for subsequent automatic identification and intelligent monitoring.

[0003] In real-world industrial environments, meter images are often affected by factors such as insufficient ambient lighting, fog interference, and limitations in imaging equipment performance, leading to degradation phenomena such as insufficient brightness, decreased contrast, loss of detail, and overall blurring. Specifically, low lighting conditions weaken visible information in the image, while foggy conditions increase light scattering during imaging, thus reducing meter image quality and affecting subsequent feature extraction and recognition results.

[0004] To address image degradation, existing technologies typically employ traditional image enhancement methods or deep learning-based image restoration methods. For image restoration in foggy conditions, traditional methods often rely on manually designed features or prior information to estimate transmittance and global atmospheric light. However, these methods suffer from poor prior adaptability and insufficient robustness in complex scenes. In contrast, while deep learning methods can automatically learn image features, most are designed for natural scenes and lack image physical degradation modeling for meter image restoration tasks. This makes it difficult to effectively constrain the image restoration process by incorporating the imaging mechanisms corresponding to different degradation categories.

[0005] Furthermore, existing supervised learning image restoration methods typically rely on a large number of paired degraded and clear images as training samples. However, in real-world industrial scenarios, obtaining strictly paired clear and blurred images of the same meter under different degradation conditions is difficult and costly, limiting the application of these methods in meter image restoration tasks. Additionally, existing methods lack the ability to specifically restore images of different degradation categories, making it difficult to balance the realism, clarity, and physical interpretability of the restored results.

[0006] Therefore, there is a need to provide an unsupervised method for restoring blurred meter images that can be trained on unpaired meter image datasets, combined with image degradation mechanisms, and restored for different degradation categories, so as to improve the authenticity, clarity and applicability of the restoration results and provide a reliable image foundation for subsequent meter image recognition and intelligent analysis. Summary of the Invention

[0007] The purpose of this invention is to provide an unsupervised method for restoring blurred meter images by incorporating a physical degradation model, in order to solve problems such as insufficient brightness, decreased contrast, loss of detail, and image blurring in meter images caused by factors such as low light and fog in complex industrial environments, thereby improving the authenticity, clarity, and applicability of the restored blurred meter images.

[0008] To achieve the objective of this invention, this invention provides an unsupervised method for restoring blurred meter images by incorporating a physical degradation model, comprising the following steps:

[0009] S1. Construct an unpaired meter image dataset, which includes clear meter images and degraded meter images, and the degraded meter images include low-light degraded images and foggy degraded images.

[0010] S11. Obtain clear meter images and degraded meter images;

[0011] S12. The clear meter images and degraded meter images are screened and sorted, and meter images with complete image content and meeting the training requirements are retained.

[0012] S13. Classify the screened degraded meter images according to the degradation category to form a subset of low-light degraded images and a subset of foggy-day degraded images, and organize the clear meter images into a clear image subset;

[0013] S14. The clear meter image and the degraded meter image are respectively used as clear image domain samples and degraded image domain samples, and divided according to the ratio of training set to test set of 8:2, for subsequent training and testing of unsupervised meter blur image recovery network. The ratio of low light degraded image to fog degraded image is set to 1:1, and the ratio of clear meter image, low light degraded image and fog degraded image is set to 2:1:1.

[0014] S2. Construct an unsupervised fuzzy meter image restoration network, which includes a degradation category classification module, a physical parameter estimation module, and an image generation module. The image generation module includes a degradation image generator, a clear image generator, and a discriminator.

[0015] S21. Construct a degradation category classification module to identify the degradation category of the input degraded meter image and output the corresponding degradation category to provide a category basis for subsequent physical degradation model selection and clear image restoration;

[0016] S22. Construct a physical parameter estimation module to estimate the physical parameters in the degradation process according to the physical degradation model corresponding to the degradation category. Specifically, it estimates the illumination component for low light degradation and the transmittance and global atmospheric light for fog degradation, so as to provide physical constraints for the subsequent degradation image generation process.

[0017] S23. Construct an image generation module, which includes a degraded image generator, a clear image generator, and a discriminator. The degraded image generator is used to receive a clear meter image and its corresponding physical parameters, and generate a degraded image that conforms to the physical degradation mechanism. The clear image generator is used to receive a degraded meter image and its corresponding degradation category, and restore degraded meter images of different degradation categories. The discriminator is used to determine the consistency between the generated image and the real image in the target image domain.

[0018] S3. The degradation category classification module is pre-trained using the degraded meter image so that the degradation category classification module can identify the degradation category of the input degraded meter image;

[0019] S31. Input the degraded meter image as a training sample into the degradation category classification module;

[0020] S32. The degradation category classification module is pre-trained using the training samples, so that the degradation category classification module learns the feature representation of meter images of different degradation categories;

[0021] S33. Use the pre-trained degradation category classification module to identify the degradation category of the input degradation meter image, so as to provide a category basis for the subsequent establishment of physical degradation model and conditional restoration of clear image generator.

[0022] S4. Establish a corresponding physical degradation model based on the degradation category, and pre-train the physical parameter estimation module using the degraded meter image, so that the physical parameter estimation module can estimate the physical parameters corresponding to its degradation process based on the input degraded meter image.

[0023] S41. Establish a corresponding physical degradation model based on the degradation category, wherein when the degradation category represents a low-light degradation image, a physical degradation model based on Retinex theory is established, and its expression is:

[0024] ;

[0025] in, The horizontal pixel coordinates of the image. These are the pixel coordinates in the vertical direction of the image. To input a low-light image, For the reflection component, For the illumination component, when the degradation category characterizes the input image as a foggy degradation image, a physical degradation model based on the atmospheric scattering model is established, and its expression is:

[0026] ;

[0027] in, For the observed foggy images, For fog-free images, For global atmospheric light, Let be the transmittance, and the transmittance satisfies the following relationship:

[0028] ;

[0029] in, Atmospheric scattering coefficient, For scene depth;

[0030] S42. Input the degraded meter image as a training sample into the physical parameter estimation module;

[0031] S43. The physical parameter estimation module is pre-trained using the training samples, so that the physical parameter estimation module learns the physical parameter feature representations corresponding to different degradation categories.

[0032] S44. The pre-trained physical parameter estimation module estimates the physical parameters corresponding to the degradation process based on the input degraded meter image. Specifically, it estimates the illumination component for low-light degradation and the transmittance and global atmospheric light for foggy degradation, so as to provide physical constraints for the degradation image generation process of the subsequent degradation image generator.

[0033] S5. The image generation module is trained using the clear meter image and the degraded meter image. The clear meter image is input into the degraded image generator, and the physical parameters are also input into the degraded image generator to generate a degraded image that conforms to the physical degradation model. The degraded meter image is input into the clear image generator, and the degradation category is also input into the clear image generator to generate a restored clear image. The generated image is compared with the real image in the target image domain based on the discriminator. Combined with adversarial loss, cycle consistency loss, identity transformation loss and physical consistency constraint, the image generation module is trained to perform generative adversarial training to establish a bidirectional mapping relationship between the degraded meter image domain and the clear meter image domain.

[0034] S51. Input the clear meter image into the degradation image generator, and input the physical parameters into the degradation image generator to generate a degradation image that conforms to the physical degradation model;

[0035] S52. Input the degraded meter image into the clear image generator, and introduce the degradation category into the clear image generator to generate a restored clear image;

[0036] The degradation category is introduced into the sharp image generator via category embedding, specifically including: mapping the discrete label corresponding to the degradation category to a continuous vector representation to obtain a category embedding vector; expanding the category embedding vector to the same spatial dimension as the image features to obtain expanded category features; fusing the expanded category features with the image features through channel concatenation; and applying a process to the fused features. Convolution is used for compression and feature reshaping to obtain conditional features for image restoration, thereby enabling targeted restoration of meter images with different degradation categories;

[0037] S53. Based on the discriminator, the generated image is distinguished from the real image in the target image domain, and adversarial training is performed on the image generation module by combining adversarial loss, cycle consistency loss, identity transformation loss and physical consistency constraint.

[0038] Specifically, the discriminator distinguishes between the generated image and the real image in the target image domain, calculates the adversarial loss, calculates the cycle consistency loss and identity transformation loss based on the bidirectional mapping between the degraded meter image domain and the clear meter image domain, constructs physical consistency constraints based on the physical degradation model and the physical parameter estimation results, and performs generative adversarial training on the image generation module by combining the adversarial loss, cycle consistency loss, identity transformation loss, and physical consistency constraints. The total loss function of the image generation module is... Including combating losses Cyclic consistency loss Identity transformation loss Physical model consistency loss loss of consistency with physical information Its expression is as follows:

[0039] ;

[0040] in, The weight coefficients for each loss term are used to optimize and train the image generation module using the total loss function, so as to establish a bidirectional mapping relationship between the degraded meter image domain and the clear meter image domain.

[0041] S6. Perform meter blur image restoration. Input the blurred meter image to be restored into the trained unsupervised meter blur image restoration network, and output the restored clear meter image.

[0042] S61. Input the blurred meter image to be recovered into the trained unsupervised blurred meter image recovery network;

[0043] S62. Use the degradation category classification module to identify the degradation category of the blurred image of the meter to be restored, and obtain the corresponding degradation category;

[0044] S63. Input the blurred image of the meter to be restored and the identified degradation category into the clear image generator, and use the clear image generator to perform targeted restoration of the blurred image of the meter to be restored.

[0045] S64. Output the restored clear meter image.

[0046] The significant advancement of this invention compared to existing technologies lies in:

[0047] (1) The present invention constructs an unsupervised meter blur image recovery network, which can realize meter blur image recovery without the need for a large number of paired clear images and degraded image training samples. This reduces the limitation of model training and application caused by the difficulty of paired data collection in industrial scenarios and improves the applicability of the method in complex industrial environments.

[0048] (2) In the unsupervised restoration process, the present invention integrates the image physical degradation model, establishes the corresponding physical constraint relationship according to the degradation category, and estimates the physical parameters in the degradation process through the physical parameter estimation module. Then, the physical parameters are introduced into the degradation image generator, thereby guiding the model to learn the degradation process that is more in line with the real imaging mechanism, improving the realism of the generated degradation image and the interpretability of the model.

[0049] (3) The present invention identifies the degradation category of the input meter image by setting a degradation category classification module, and introduces the degradation category into the clear image generator in a category embedding manner, so that the clear image generator can perform differentiated restoration for the blurred meter images of different degradation categories, thereby enhancing the pertinence of the restoration process and improving the clarity and detail expression of the restored image.

[0050] (4) In addition to adversarial loss and cycle consistency loss, this invention also introduces physical model consistency loss and physical information consistency loss during the training process, so that the generated image not only satisfies the inter-domain mapping relationship, but also further conforms to the corresponding physical degradation mechanism, thereby effectively improving the quality of meter blur image restoration.

[0051] (5) By integrating physical degradation model, category embedding mechanism and unsupervised learning mechanism, this invention can achieve better meter image restoration under low light degradation and fog degradation conditions, providing reliable image quality assurance for subsequent meter feature extraction and data recognition.

[0052] To make the objectives and technical solutions of this invention clearer, the following description is provided in conjunction with the accompanying drawings and specific embodiments. Attached Figure Description

[0053] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0054] Figure 1 This is a flowchart of the unsupervised meter fuzzy image restoration method based on the fusion of physical degradation model of the present invention;

[0055] Figure 2 This is a schematic diagram of the unsupervised meter blur image recovery network framework of the present invention;

[0056] Figure 3 This is a schematic diagram of the degradation category classification module structure of the present invention;

[0057] Figure 4 This is a schematic diagram of the physical parameter estimation module based on Retinex theory of the present invention;

[0058] Figure 5 This is a schematic diagram of the physical parameter estimation module based on the atmospheric scattering model of the present invention;

[0059] Figure 6 This is a schematic diagram of the degraded image generator structure of the present invention;

[0060] Figure 7 This is a schematic diagram of the clear image generator structure of the present invention;

[0061] Figure 8 This is a schematic diagram of the discriminator structure of the present invention. Detailed Implementation

[0062] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] In this embodiment, combined with Figure 1 The present invention provides an unsupervised method for restoring blurred meter images by incorporating a physical degradation model, comprising the following steps:

[0064] S1. Construct an unpaired meter image dataset, which includes clear meter images and degraded meter images, and the degraded meter images include low-light degraded images and foggy degraded images.

[0065] S11. Obtain clear meter images and degraded meter images. The clear meter images are used to form clear image domain samples, and the low-light degraded images and foggy degraded images are used to form degraded image domain samples. Since it is difficult to obtain strictly paired clear meter images and degraded meter images in actual industrial environments, unpaired meter image datasets are used for subsequent training and testing.

[0066] S12. The clear meter images and degraded meter images are screened and sorted, and meter images with complete image content and meeting the training requirements are retained. Preferably, meter images with complete dial areas, clear and distinguishable image content, and no serious occlusion or truncation are retained to ensure the effectiveness and stability of subsequent network training and testing.

[0067] S13. Classify the screened degraded meter images according to the degradation category to form a subset of low-light degraded images and a subset of foggy-day degraded images, and organize the clear meter images into a clear image subset;

[0068] S14. The clear meter image and the degraded meter image are respectively used as clear image domain samples and degraded image domain samples, and divided according to the ratio of training set to test set of 8:2, for subsequent training and testing of unsupervised meter blur image recovery network. The ratio of low light degraded image to fog degraded image is set to 1:1, and the ratio of clear meter image, low light degraded image and fog degraded image is set to 2:1:1.

[0069] S2, Combination Figure 2 An unsupervised network for restoring blurred images of meters is constructed. The unsupervised network for restoring blurred images of meters includes a degradation category classification module, a physical parameter estimation module, and an image generation module. The image generation module includes a degradation image generator, a clear image generator, and a discriminator.

[0070] Figure 2 middle, Represents the sharp image domain. Represents the degraded image domain. This represents a clear sample of the input meter image. The input represents a degraded meter image sample. The degradation category classification module is used to determine the degradation category of the input degraded meter image. The physical parameter estimation module is used to estimate the physical parameters corresponding to the image degradation process. The clear image generator is used to restore the samples in the degraded image domain to the clear image domain. The degraded image generator is used to map the samples in the clear image domain to the degraded image domain. The discriminator is used to determine the authenticity of the generated image and the real image in the target image domain.

[0071] S21. Construct a degradation category classification module to identify the degradation category of the input degraded meter image and output the corresponding degradation category to provide a category basis for subsequent physical degradation model selection and clear image restoration;

[0072] S22. Construct a physical parameter estimation module to estimate the physical parameters in the degradation process according to the physical degradation model corresponding to the degradation category. Specifically, it estimates the illumination component for low light degradation and the transmittance and global atmospheric light for fog degradation, so as to provide physical constraints for the subsequent degradation image generation process.

[0073] S23. Construct an image generation module, which includes a degraded image generator, a sharp image generator, and a discriminator, wherein the degraded image generator is used to implement the sharp image domain. To the degraded image domain The mapping, which is about to Mapped to ,Will Remapped to At the same time, Mapped to and will Remapped to ;

[0074] The sharp image generator is used to implement the degraded image domain. To the clear image domain The mapping, which is about to Mapped to and will Mapped to ;

[0075] The discriminator is used to determine the authenticity of the generated image and the corresponding real image in the image domain. Specifically, the discriminator... and To make a judgment, and To make a judgment, and To make a judgment, and at the same time, to and To make a judgment, and To make a judgment, and Perform the judgment;

[0076] in, and This indicates the recovered clear image. , , and This represents the generated degraded image.

[0077] S3, Combination Figure 3 The degradation category classification module is pre-trained using the degraded meter image, enabling the degradation category classification module to identify the degradation category of the input degraded meter image;

[0078] S31. The degraded meter image is used as a training sample and input into the degradation category classification module. The degradation category classification module consists of multiple convolutional layers, average pooling, flattening, and sigmoid activation functions. The front-end convolutional layer is used to perform initial feature extraction on the input degraded meter image. The subsequent convolutional layers consist of convolution, normalization, and activation functions, which are used to further extract image features. Subsequent convolutional layers are used to continue to extract and compress image features. Average pooling is used to reduce the dimensionality of the feature map. Flattening is used to convert the feature map into a one-dimensional feature vector. The sigmoid activation function is used to output the corresponding degradation category discrimination result.

[0079] S32. The degradation category classification module is pre-trained using the training samples, so that the degradation category classification module learns the feature representation of meter images of different degradation categories;

[0080] S33. Use the pre-trained degradation category classification module to identify the degradation category of the input degradation meter image, so as to provide a category basis for the subsequent establishment of physical degradation model and conditional restoration of clear image generator.

[0081] S4, Combination Figure 4 and Figure 5 A corresponding physical degradation model is established based on the degradation category, and the physical parameter estimation module is pre-trained using the degraded meter image, so that the physical parameter estimation module can estimate the physical parameters corresponding to its degradation process based on the input degraded meter image.

[0082] S41. Establish a corresponding physical degradation model based on the degradation category. When the degradation category represents a low-light degradation image, establish a physical degradation model based on Retinex theory (based on retina-cortex theory), whose expression is:

[0083] ;

[0084] in, The horizontal pixel coordinates of the image. These are the pixel coordinates in the vertical direction of the image. To input a low-light image, For the reflection component, For the illumination component, when the degradation category characterizes the input image as a foggy degradation image, a physical degradation model based on the atmospheric scattering model is established, and its expression is:

[0085] ;

[0086] in, For the observed foggy images, For fog-free images, For global atmospheric light, Let be the transmittance, and the transmittance satisfies the following relationship:

[0087] ;

[0088] in, Atmospheric scattering coefficient, For scene depth;

[0089] S42. The degraded meter image is used as a training sample and input into the physical parameter estimation module. When the input image is a low-light degraded image, the following steps are called: Figure 4 The physical parameter estimation module based on Retinex theory shown calls the following function when the input image is a degraded image due to fog: Figure 5 The physical parameter estimation module based on the atmospheric scattering model is shown.

[0090] S43. The physical parameter estimation module is pre-trained using the training samples, so that the physical parameter estimation module learns the physical parameter feature representations corresponding to different degradation categories.

[0091] The physical parameter estimation module based on Retinex theory is used to output the illumination component. After inputting a low-light degraded image, the system first performs convolutional feature extraction to obtain a feature map. Then, it passes through multiple feature extraction layers and residual blocks for feature learning. The residual block consists of two convolutional units to enhance feature representation and improve parameter estimation accuracy. Finally, the output is the illumination component corresponding to the input image. ;

[0092] The physical parameter estimation module based on the atmospheric scattering model is used to output the transmittance respectively. and global atmospheric light The upper branch is used to estimate the transmittance. After inputting a degraded image due to fog, the corresponding transmittance is output through multiple convolutional feature extraction layers. The lower branch is used to estimate global atmospheric light. After inputting a degraded image due to fog, features are extracted via convolution, then compressed via pooling, and finally... Convolution outputs global atmospheric light ;

[0093] S44. The pre-trained physical parameter estimation module estimates the physical parameters corresponding to the degradation process based on the input degraded meter image. Among them, the illumination component is estimated for low-light degradation. Estimating transmittance in response to foggy weather degradation and global atmospheric light This provides physical constraints for the degradation image generation process of the subsequent degradation image generator.

[0094] S5, Combination Figure 6 , Figure 7 and Figure 8 The image generation module is trained using the clear meter image and the degraded meter image. The clear meter image is input into the degraded image generator, and the physical parameters are also input into the degraded image generator to generate a degraded image that conforms to the physical degradation model. The degraded meter image is input into the clear image generator, and the degradation category is also input into the clear image generator to generate a restored clear image. The generated image is distinguished from the real image in the target image domain based on the discriminator. The image generation module is trained using adversarial loss, cycle consistency loss, identity transformation loss, and physical consistency constraints to establish a bidirectional mapping relationship between the degraded meter image domain and the clear meter image domain.

[0095] S51. Input the clear meter image into the degradation image generator, and input the physical parameters into the degradation image generator to generate a degradation image that conforms to the physical degradation model;

[0096] The degraded image generator employs an encoder-decoder structure. After a clear meter image is input, it first undergoes... Convolution is used for initial feature extraction, and then multiple convolutions are performed sequentially. Convolutional processing is used for deep feature extraction. Physical parameters output from the physical parameter estimation module are simultaneously input into the degraded image generator and fused with image features during the front-end feature extraction stage. After feature extraction, the features undergo feature transformation through residual blocks to enhance the generator's ability to model the physical degradation process. This is followed by multiple... Convolution and upsampling gradually restore spatial resolution, ultimately through Convolution outputs a degraded image;

[0097] S52. Input the degraded meter image into the clear image generator, and introduce the degradation category into the clear image generator to generate a restored clear image;

[0098] The clear image generator also employs an encoder-decoder structure. After the degraded meter image is input, it first passes through... Convolution is used for initial feature extraction, and then multiple convolutions are performed sequentially. Convolutional processing is used for deep feature extraction. The degraded categories are introduced into the sharp image generator via category embedding. Specifically, this includes: mapping the discrete labels corresponding to the degraded categories to continuous vector representations to obtain category embedding vectors; expanding the category embedding vectors to the same spatial dimension as the image features to obtain expanded category features; fusing the expanded category features with the image features through channel concatenation; and applying a process to the fused features. Convolutional processing compresses and reshapes features, which are then fed into subsequent networks. The category embedding features and image features are input into residual blocks, which are then processed through multiple... Convolution and upsampling gradually restore spatial resolution, ultimately through The convolution output restores a clear image of the meter.

[0099] S53. Based on the discriminator, the generated image is distinguished from the real image in the target image domain, and adversarial training is performed on the image generation module by combining adversarial loss, cycle consistency loss, identity transformation loss and physical consistency constraint.

[0100] The discriminator adopts a PatchGAN (Local Block Generative Adversarial Network) structure, which consists of multiple concatenated convolutional layers, the first four of which... Convolution, a 2-stride convolutional unit is used to progressively reduce the feature map spatial size and extract local discriminative features, the last one... Convolutional units with a stride of 1 are used to output the discrimination result;

[0101] Specifically, the discriminator distinguishes between the generated image and the real image in the target image domain, and calculates the adversarial loss. As shown in the following formula:

[0102] ;

[0103] in, Represents the sharp image domain. Represents the degraded image domain. This represents a degraded image generator. Represents a clear image generator. This represents a degraded image domain discriminator. This represents a sharp image domain discriminator. Represents clear image domain samples, This represents a degraded image domain sample. Represents the mathematical expectation. This represents the true data distribution of a clear image. This represents the true data distribution of the degraded image;

[0104] Calculate the cycle consistency loss based on the bidirectional mapping between the degraded meter image domain and the clear meter image domain. Its expression is as follows:

[0105] ;

[0106] in, Represents the L1 norm;

[0107] The loss from the identity transformation is calculated using the following expression:

[0108] ;

[0109] Based on the physical degradation model and the physical parameter estimation results, a physical model consistency loss is constructed. loss of consistency with physical information The physical model consistency loss The expression is as follows:

[0110] ;

[0111] in, This represents the total number of pixels in the input image. The degraded image generated by the degraded image generator is the first one. The pixel value of each pixel. For the degraded image calculated by the physical model, the first... The pixel value of each pixel;

[0112] The loss of physical information consistency The expression is as follows:

[0113] ;

[0114] in, The total number of dimensions of the physical parameters to be estimated. The physical parameters estimated from the generated degraded image, These are the physical parameters estimated from real degraded images;

[0115] Finally, construct the total loss function. Its expression is as follows:

[0116] ;

[0117] in, The weight coefficients for each loss term are used to optimize and train the image generation module using the total loss function, so as to establish a bidirectional mapping relationship between the degraded meter image domain and the clear meter image domain.

[0118] S6. Perform meter blur image restoration. Input the blurred meter image to be restored into the trained unsupervised meter blur image restoration network, and output the restored clear meter image.

[0119] S61. Input the blurred meter image to be recovered into the trained unsupervised blurred meter image recovery network;

[0120] S62. Use the degradation category classification module to identify the degradation category of the blurred image of the meter to be restored, and obtain the corresponding degradation category;

[0121] S63. Input the blurred image of the meter to be restored and the identified degradation category into the clear image generator, and use the clear image generator to perform targeted restoration of the blurred image of the meter to be restored.

[0122] S64. Output the restored clear meter image.

[0123] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0124] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An unsupervised method for restoring blurred meter images by incorporating a physical degradation model, characterized in that, Includes the following steps: S1. Construct an unpaired meter image dataset, which includes clear meter images and degraded meter images, and the degraded meter images include low-light degraded images and foggy degraded images. S2. Construct an unsupervised fuzzy meter image restoration network, which includes a degradation category classification module, a physical parameter estimation module, and an image generation module. The image generation module includes a degradation image generator, a clear image generator, and a discriminator. S3. The degradation category classification module is pre-trained using the degraded meter image so that the degradation category classification module can identify the degradation category of the input degraded meter image; S4. Establish a corresponding physical degradation model based on the degradation category, and pre-train the physical parameter estimation module using the degraded meter image, so that the physical parameter estimation module can estimate the physical parameters corresponding to its degradation process based on the input degraded meter image. S5. The image generation module is trained using the clear meter image and the degraded meter image. The clear meter image is input into the degraded image generator, and the physical parameters are also input into the degraded image generator to generate a degraded image that conforms to the physical degradation model. The degraded meter image is input into the clear image generator, and the degradation category is also input into the clear image generator to generate a restored clear image. The generated image is distinguished from the real image in the target image domain based on the discriminator. Combined with adversarial loss, cycle consistency loss, identity transformation loss, and physical consistency constraints, the image generation module is trained to perform generative adversarial training to establish a bidirectional mapping relationship between the degraded meter image domain and the clear meter image domain. S6. Perform meter blur image restoration. Input the blurred meter image to be restored into the trained unsupervised meter blur image restoration network, and output the restored clear meter image.

2. The unsupervised method for restoring blurred meter images based on a fused physical degradation model according to claim 1, characterized in that, Step S1 includes the following steps: S11. Obtain clear meter images and degraded meter images; S12. The clear meter images and degraded meter images are screened and sorted, and meter images with complete image content and meeting the training requirements are retained. S13. Classify the screened degraded meter images according to the degradation category to form a subset of low-light degraded images and a subset of foggy-day degraded images, and organize the clear meter images into a clear image subset; S14. The clear meter image and the degraded meter image are respectively used as clear image domain samples and degraded image domain samples, and divided according to the ratio of training set to test set of 8:2, for subsequent training and testing of unsupervised meter blur image recovery network. The ratio of low light degraded image to fog degraded image is set to 1:1, and the ratio of clear meter image, low light degraded image and fog degraded image is set to 2:1:

1.

3. The unsupervised method for restoring blurred meter images based on a fused physical degradation model according to claim 1, characterized in that, Step S2 includes the following steps: S21. Construct a degradation category classification module to identify the degradation category of the input degraded meter image and output the corresponding degradation category; S22. Construct a physical parameter estimation module to estimate the physical parameters in the degradation process according to the physical degradation model corresponding to the degradation category. Specifically, it estimates the illumination component for low light degradation and the transmittance and global atmospheric light for foggy degradation. S23. Construct an image generation module, which includes a degraded image generator, a clear image generator, and a discriminator. The degraded image generator is used to receive a clear meter image and its corresponding physical parameters, and generate a degraded image that conforms to the physical degradation mechanism. The clear image generator is used to receive a degraded meter image and its corresponding degradation category, and restore degraded meter images of different degradation categories. The discriminator is used to determine the consistency between the generated image and the real image in the target image domain.

4. The unsupervised method for restoring blurred meter images based on a fused physical degradation model according to claim 1, characterized in that, Step S3 includes the following steps: S31. Input the degraded meter image as a training sample into the degradation category classification module; S32. Pre-train the degradation category classification module using the training samples; S33. Use the pre-trained degradation category classification module to identify the degradation category of the input degradation meter image.

5. The unsupervised method for restoring blurred meter images based on a fused physical degradation model according to claim 1, characterized in that, Step S4 includes the following steps: S41. Establish the corresponding physical degradation model according to the degradation category; S42. Input the degraded meter image as a training sample into the physical parameter estimation module; S43. Pre-train the physical parameter estimation module using the training samples; S44. The pre-trained physical parameter estimation module estimates the physical parameters corresponding to the degradation process based on the input degraded meter image.

6. The unsupervised method for restoring blurred meter images based on a fused physical degradation model according to claim 5, characterized in that, Step S41 includes: when the degradation category characterizes the input image as a low-light degradation image, establishing a physical degradation model based on Retinex theory, the expression of which is: ； in, The horizontal pixel coordinates of the image. These are the pixel coordinates in the vertical direction of the image. To input a low-light image, For the reflection component, For the illumination component, when the degradation category characterizes the input image as a foggy degradation image, a physical degradation model based on the atmospheric scattering model is established, and its expression is: ； in, For the observed foggy images, For fog-free images, For global atmospheric light, Let be the transmittance, and the transmittance satisfies the following relationship: ； in, Atmospheric scattering coefficient, For scene depth.

7. The unsupervised method for restoring blurred meter images based on a fused physical degradation model according to claim 1, characterized in that, Step S5 includes the following steps: S51. Input the clear meter image into the degradation image generator, and input the physical parameters into the degradation image generator to generate a degradation image that conforms to the physical degradation model; S52. Input the degraded meter image into the clear image generator, and introduce the degradation category into the clear image generator to generate a restored clear image; S53. Based on the discriminator, the generated image and the real image in the target image domain are distinguished, and adversarial loss, cycle consistency loss, identity transformation loss and physical consistency constraint are combined to perform generative adversarial training on the image generation module to establish a bidirectional mapping relationship between the degraded meter image domain and the clear meter image domain.

8. The unsupervised method for restoring blurred meter images based on a fused physical degradation model according to claim 7, characterized in that, Step S52 includes the following steps: S521. The degradation category is introduced into the clear image generator in a category embedding manner, and the discrete label corresponding to the degradation category is mapped to a continuous vector representation to obtain the category embedding vector; S522. Extend the category embedding vector to the same spatial dimension as the image features to obtain the extended category features; S523. The expanded category features are fused with image features by channel splicing. S524, Apply the following to the fused features: Convolution is used for compression and feature reshaping to obtain conditional features for image restoration.

9. The unsupervised method for restoring blurred meter images based on a fused physical degradation model according to claim 7, characterized in that, Step S53 includes: discriminating between the generated image and the real image in the target image domain based on the discriminator; calculating adversarial loss; calculating cycle consistency loss and identity transformation loss based on the bidirectional mapping result between the degraded meter image domain and the clear meter image domain; constructing physical consistency constraints based on the physical degradation model and the physical parameter estimation results; and performing generative adversarial training on the image generation module by combining the adversarial loss, cycle consistency loss, identity transformation loss, and physical consistency constraints. The total loss function of the image generation module is... Including combating losses Cyclic consistency loss Identity transformation loss Physical model consistency loss loss of consistency with physical information Its expression is as follows: ； in, These are the weighting coefficients corresponding to each loss term.

10. The unsupervised method for restoring blurred meter images based on a fused physical degradation model according to claim 1, characterized in that, Step S6 includes the following steps: S61. Input the blurred meter image to be recovered into the trained unsupervised blurred meter image recovery network; S62. Use the degradation category classification module to identify the degradation category of the blurred image of the meter to be restored; S63. Input the blurred image of the meter to be restored and the identified degradation category into the clear image generator to restore the blurred image of the meter to be restored. S64. Output the restored clear meter image.

Images