Real-world image super-resolution system and method for joint learning of super-resolution and degradation

By introducing joint learning of degraded networks and multiple loss functions in the super-resolution method, the problem that the existing technology cannot be applied to real-world image super-scores is solved, and high-quality high-resolution image generation and real-world image pairs are achieved. Build.

CN114219712BActive Publication Date: 2024-11-19SUN YAT SEN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111545264.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-16
Publication Date
2024-11-19
Estimated Expiration
2041-12-16

AI Technical Summary

Technical Problem

The existing super-resolution method cannot be applied to real-world image super-scores and cannot effectively process the high and low resolution conversion of real-world images.

Method used

A real-world image super-segment system that combines super-resolution and degradation learning is proposed, including super-segment network module, degradation network module, discrimination module and loss function module, through weighting of counter-loss, cyclic consistent loss, information loss and perceived loss Summarize to achieve the generation and degradation of high-resolution images, ensure that the distribution of the output images is consistent with the real image and retain content information.

Benefits of technology

Effectively avoid artifacts caused by double-triple downsampling, maintain color components, remove non-real artifacts, generate high-resolution images with higher quality and less noise, suitable for super-scoring and building real-world images High and low resolution image pairs of the world.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114219712B_ABST
    Figure CN114219712B_ABST
Patent Text Reader

Abstract

The present invention provides a real-world image super-resolution system and method for jointly learning super-resolution and degradation. The system can generate real training pairs by training a degradation sub-network, thereby avoiding artifacts caused by bicubic downsampling, and also introduces information loss to avoid color changes, and introduces perceptual loss to remove non-real artifacts. Compared with other methods, the present invention has good performance, and the obtained high-resolution images have less noise and better visual quality. It solves the problem that existing super-resolution methods cannot be applied to real-world image super-resolution, and at the same time proposes a joint training framework, which can not only realize real-world image super-resolution, but also construct real-world high- and low-resolution image pairs, which is convenient for application in other super-resolution networks.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of computer vision, and more specifically, to a real-world image super-resolution system and method for jointly learning super-resolution and degradation. Background Art

[0002] Convolutional Neural Network is a deep neural network with convolution structure. The convolution structure can reduce the parameters of the network through three key modules: local receptive field, weight sharing, and pooling operation, thereby alleviating the overfitting problem of the model. The convolutional neural network structure includes: input layer, convolution layer, pooling layer, full connection layer, and output layer. There are multiple feature maps in each layer. Each feature map extracts the features of the input image through the convolution operation of the convolution kernel. The core modules for extracting features are the convolution layer and the pooling layer. The convolution layer consists of a series of neurons and activation functions. Neurons can be regarded as filters in image processing, and the weights connecting each neuron are fixed. The pooling layer is in the middle of the continuous convolution layer, which is used to compress data and parameters. It has the functions of ensuring the scale invariance of features, reducing the dimension of features, and reducing overfitting. The gradient descent method is a method used in convolutional neural networks to minimize the loss function. It can back-propagate the weight parameters in the network layer by layer, continuously iterate the training, and thus improve the accuracy of the network. The lower layers of the convolutional neural network are composed of alternating convolutional layers and maximum pooling layers, while the higher layers are composed of fully connected layers. The input of the corresponding fully connected layer is the feature map obtained by feature extraction from the lower layer, and the output is a classifier. Logistic regression, cross entropy regression, or support vector machine can be used to classify the input image. The fully connected layer of the convolutional neural network is the hidden layer and logistic regression classifier of the corresponding traditional multi-layer perceptron. Due to the characteristics of parameter sharing and sparse connection, the convolutional neural network has good self-learning ability, parallel processing ability and fault tolerance. Compared with other methods, its generalization ability is more prominent, and it has been widely used in pattern recognition, object detection, image super-resolution, etc.

[0003] Image super resolution refers to the restoration of a high-resolution image from a low-resolution image or image sequence. In a large number of electronic image application fields, people often expect to obtain high-resolution images. High resolution means that the pixel density in the image is high, which can provide more details, and these details are indispensable in many practical applications. Traditional image super resolution can be divided into three aspects: interpolation-based, reconstruction-based, and learning-based. Interpolation-based methods are mainly completed through interpolation functions such as linear interpolation and bicubic interpolation; reconstruction-based methods mainly solve super resolution problems through knowledge of probability theory and set theory; learning-based methods use some traditional machine learning algorithms, such as support vector machines. With the widespread use of convolutional neural networks, most super resolution methods now use machine learning-based methods, using paired high-resolution images and low-resolution images as supervision information to learn the reconstruction process. These methods have achieved relatively great success.

[0004] Generative Adversarial Networks (GANs) is a deep learning model and one of the most promising methods for unsupervised learning on complex distributions in recent years. The model produces more ideal outputs through the mutual game learning of (at least) two modules in the framework: the generative model and the discriminative model. Now most of the generative models and discriminative models are designed using neural networks. The input of the discriminative model is an image, and the output is a probability value, indicating the judgment of true or false; the input of the generative model is a set of random numbers, and the output is an image, the purpose of which is to generate samples. The two models compete with each other, the purpose of which is to make the generative model generate more realistic samples, so that the discriminator cannot distinguish them. When training this model, a separate alternating iterative training method is used, that is, the generator is trained once, the discriminator is trained once, and then the generator is trained, and so on. The power of the generative adversarial network lies in the ability to automatically learn the data distribution of the original real sample set.

[0005] The prior art discloses a patent for an efficient image super-resolution reconstruction method. The patent proposes a new image super-resolution network structure based on a convolutional neural network. The network structure is named SESRNet, in which a multi-scale convolution fusion module is used to efficiently obtain image features of different scales and the amount of calculation is greatly reduced. SESRNet is mainly composed of 5 Mixblocks, and the amount of calculation required is only 2.8GFLOPS, which is half the amount of calculation of the mainstream FSRCNN's 6GFLOPS. The image super-resolution network in this patent greatly reduces the amount of calculation while maintaining the effect of image restoration, so that the edge intelligent IoT terminal device can restore the image to the appropriate effect more smoothly, thereby improving the interactive effect of AIoT devices. However, the patent does not involve any solution to the problem that existing super-resolution methods cannot be applied to real-world image super-resolution. Summary of the invention

[0006] The present invention provides a real-world image super-resolution system that jointly learns super-resolution and degradation. The system solves the problem that existing super-resolution methods cannot be applied to real-world image super-resolution. At the same time, a joint training framework is proposed, which can not only realize real-world image super-resolution, but also construct real-world high- and low-resolution image pairs, which is convenient for application in other super-resolution networks.

[0007] Another object of the present invention is to provide a super-resolution method for the real-world image super-resolution system that jointly learns super-resolution and degradation.

[0008] In order to achieve the above technical effects, the technical solution of the present invention is as follows:

[0009] A real-world image super-resolution system that jointly learns super-resolution and degradation, including:

[0010] The super-resolution network module generates a high-resolution image from the input low-resolution image through the network;

[0011] The degradation network module degrades the high-resolution image obtained by the super-resolution network module into a blurred image with real-world features;

[0012] The discrimination module discriminates whether the image input to the degradation network module comes from the real image distribution or the generated image distribution;

[0013] The loss function module constrains both the super-resolution network module and the degradation network module to ensure the distribution of the output image and retain the content information. The loss function module includes an adversarial loss module, a cycle consistency loss module, an information loss module and a perceptual loss module. The weighted sum of them is used to obtain the total loss function module.

[0014] Furthermore, the super-resolution network module uses a generator to generate a high-resolution image from an input low-resolution image through a network: a structure with an iterative error feedback mechanism is used as a generator, and the projection error is back-propagated through alternating iterative upsampling and downsampling units, and the reconstruction results in the network are corrected multiple times; the upsampling unit generates more high-resolution spatial features, and the downsampling unit projects these features to a low-resolution space, retaining more high-resolution components, and more deep features can be effectively used.

[0015] Furthermore, the degradation network module is a generator network that degrades the input high-resolution image into a blurred image with real-world features: a deep linear network without an activation function is used to simulate the degradation process. The network is a 6-layer linear full convolutional network with a stride of 2: the sizes of the first three convolution kernels are 7x7, 5x5, and 3x3, and the size of the remaining layers is 1x1, which is equivalent to a convolution kernel with a receptive field of 17x17.

[0016] Furthermore, the discriminant module learns the distribution of image pixels to discriminate whether it is from the real image distribution or the generated image distribution: five 3x3 convolutional layers are used, and all but the first and last convolutional layers are LeakyReLU activation functions with BatchNornalization; a feature map is obtained after passing through the discriminator module, and each value in the feature map represents the confidence that the corresponding image area is judged to be true, and the confidence of the entire image can be obtained by averaging all the values ​​in the feature map; the discriminant module includes a high-resolution image discriminant module D HR and low-resolution discrimination module D LR .

[0017] Furthermore, the adversarial loss module is used to distinguish the super-resolution network module The distribution of the output image and the distribution of the original image, the adversarial loss module is described as:

[0018]

[0019] Among them, Y~P Y is the distribution of image data output by the super-resolution network module, X~P X is the distribution of the input low-resolution image training data.

[0020] Furthermore, since the content information in the image will be lost during the generation process, a cycle-consistent loss module is used to preserve the content information through self-supervised super-resolution and degradation sub-networks. The cycle-consistent loss module is described as:

[0021]

[0022] Furthermore, for the low-resolution image generated by the degradation network module, an information loss module is introduced to maintain the color components and avoid color changes. The description of the information loss module is:

[0023]

[0024] Where Bic is the bicubic downsampling operation.

[0025] Furthermore, since there are many artifacts in the generated high-resolution samples, a perceptual loss module is added between the high-resolution image and the corresponding original image, which is described as:

[0026]

[0027] where φ laver It is the feature of the layer of the pre-trained CNN, using the VGG-19 network pre-trained on ImagesNet.

[0028] Furthermore, the final loss function module is determined as:

[0029] Loss = λ adv L adv +λ cyc L cyc +λ Info L Info +λ per L per

[0030] where λ adv is the weight of the adversarial loss, λ cyc is the weight of the cycle consistency loss, λ Info is the weight of information loss, λ per are the weights of perceptual loss, which are 0.1, 10, 10, and 0.5 respectively.

[0031] A real-world image super-resolution method for jointly learning super-resolution and degradation includes the following steps:

[0032] S1: The super-resolution network module generates a high-resolution image from the input low-resolution image through the network;

[0033] S2: The degradation network module degrades the high-resolution image obtained by the super-resolution network module into a blurred image with real-world features;

[0034] S3: The discrimination module discriminates whether the image input by the degradation network module comes from the real image distribution or the generated image distribution;

[0035] S4: The loss function module constrains both the super-resolution network module and the degradation network module to ensure the distribution of the output image and preserve the content information.

[0036] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0037] The present invention can generate real training pairs by training a degenerate subnetwork, thereby avoiding artifacts caused by bicubic downsampling, and also introduces information loss to avoid color changes, and introduces perceptual loss to remove non-real artifacts; compared with other methods, the present invention has good performance, and the obtained high-resolution pictures have less noise and better visual quality; it solves the problem that the existing super-resolution methods cannot be applied to the super-resolution of real-world pictures, and at the same time proposes a joint training framework, which can not only realize the super-resolution of real-world pictures, but also construct real-world high- and low-resolution picture pairs, which is convenient for application in other super-resolution networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the system structure of the present invention;

[0039] Figure 2 is a flow chart of the method of the present invention;

[0040] Figure 3 It is a schematic diagram of the processing process of the super-resolution network module;

[0041] Figure 4 Schematic diagram of the processing process of the degraded network module. DETAILED DESCRIPTION

[0042] The drawings are for illustrative purposes only and should not be construed as limiting the present patent;

[0043] In order to better illustrate the present embodiment, some parts in the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product;

[0044] It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0045] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0046] Example 1

[0047] like Figure 1 As shown in FIG. 1 , a real-world image super-resolution system that jointly learns super-resolution and degradation includes:

[0048] The super-resolution network module generates a high-resolution image from the input low-resolution image through the network;

[0049] The degradation network module degrades the high-resolution image obtained by the super-resolution network module into a blurred image with real-world features;

[0050] The discrimination module discriminates whether the image input to the degradation network module comes from the real image distribution or the generated image distribution;

[0051] The loss function module constrains both the super-resolution network module and the degradation network module to ensure the distribution of the output image and retain the content information. The loss function module includes an adversarial loss module, a cycle consistency loss module, an information loss module and a perceptual loss module. The weighted sum of them is used to obtain the total loss function module.

[0052] Super-resolution network module This module uses a generator to generate a high-resolution image from an input low-resolution image through the network: a structure with an iterative error feedback mechanism is used as the generator, and the projection error is back-propagated through alternating iterative upsampling and downsampling units, and the reconstruction results in the network are corrected multiple times; the upsampling unit generates more high-resolution spatial features, while the downsampling unit projects these features to the low-resolution space, retaining more high-resolution components and effectively using more deep features.

[0053] The degradation network module is a generator network that degrades the input high-resolution image into a blurred image with real-world characteristics: a deep linear network without an activation function is used to simulate the degradation process. The network is a 6-layer linear full convolutional network with a stride of 2: the first three convolution kernels are 7x7, 5x5, and 3x3 in size, and the remaining layers are 1x1 in size, which is equivalent to a convolution kernel with a receptive field of 17x17.

[0054] The discriminant module learns the distribution of image pixels to distinguish whether it comes from the real image distribution or the generated image distribution: five 3x3 convolutional layers are used, except for the first and last convolutional layers, all of which are LeakyReLU activation functions with BatchNornalization; after passing through the discriminator module, a feature map is obtained, and each value in the feature map represents the confidence that the corresponding image area is judged to be true. The confidence of the entire image can be obtained by averaging all the values ​​in the feature map; the discriminant module includes a high-resolution image discriminant module D HR and low-resolution discrimination module D LR .

[0055] The adversarial loss module is used to distinguish the super-resolution network module G Up The distribution of the output image and the distribution of the original image, the adversarial loss module is described as:

[0056]

[0057] Among them, Y~P Y is the distribution of image data output by the super-resolution network module, X~P X is the distribution of the input low-resolution image training data.

[0058] Since the content information in the image will be lost during the generation process, a cycle-consistent loss module is used to preserve the content information through self-supervised super-resolution and degradation sub-networks. The cycle-consistent loss module is described as:

[0059]

[0060] For the low-resolution images generated by the degradation network module, an information loss module is introduced to maintain the color components and avoid color changes. The information loss module is described as follows:

[0061]

[0062] Where Bic is the bicubic downsampling operation.

[0063] Since there are many artifacts in the generated high-resolution samples, a perceptual loss module is added between the high-resolution image and the corresponding original image, which is described as:

[0064]

[0065] where φ laver It is the feature of the layer of the pre-trained CNN, using the VGG-19 network pre-trained on ImagesNet.

[0066] The final loss function module is determined as:

[0067] Loss = λ adv L adv +λ cyc L cyc +λ Info L Info +λ per L per

[0068] Among them, λ adv is the weight of the adversarial loss, λ cyc is the weight of the cycle consistency loss, λ Info is the weight of information loss, λ per are the weights of perceptual loss, which are 0.1, 10, 10, and 0.5 respectively.

[0069] Example 2

[0070] like Figure 2 As shown, a real-world image super-resolution method that jointly learns super-resolution and degradation includes the following steps:

[0071] S1: The super-resolution network module generates a high-resolution image from the input low-resolution image through the network;

[0072] S2: The degradation network module degrades the high-resolution image obtained by the super-resolution network module into a blurred image with real-world features;

[0073] S3: The discrimination module discriminates whether the image input by the degradation network module comes from the real image distribution or the generated image distribution;

[0074] S4: The loss function module constrains both the super-resolution network module and the degradation network module to ensure the distribution of the output image and preserve the content information.

[0075] Taking 4x super-resolution as an example, training is performed on 4 Titan Xp GPUs. The specific implementation steps are as follows:

[0076] Step 1: Collect observation data of a set of high-resolution image datasets X and a set of low-resolution image datasets Y;

[0077] Step 2: Input the low-resolution image X into the super-resolution network to obtain the corresponding high-resolution image X HR , and then input the obtained low-resolution image into the degradation network to obtain the corresponding low-resolution image X LR ;

[0078] Step 3: Input the high-resolution image Y into the degradation network to obtain the corresponding low-resolution image Y LR , and then input the obtained high-resolution image into the super-resolution network to obtain the corresponding low-resolution image Y HR ;

[0079] Step 4: Construct adversarial loss, cycle consistency loss, information loss and perceptual loss, and train the super-resolution network and the degradation network simultaneously according to steps 2 and 3, iterating about 300k times;

[0080] Step 5: Use the parameter model trained in step 4 to input the required low-resolution image into the super-resolution module to obtain the high-resolution image after super-resolution.

[0081] By jointly learning super-resolution and degradation, we can achieve super-resolution processing of natural images in the real world, which has practical significance for real life. At the same time, we can learn the degradation process of the real world, construct high- and low-resolution image pairs in the real world, and apply them to other SISR super-resolution networks.

[0082] Example 3

[0083] The super-resolution network module mainly uses a generator to generate a high-resolution image from the input low-resolution image through the network. Using a structure with an iterative error feedback mechanism as the generator, the projection error is back-propagated through alternating iterative upsampling and downsampling units, and the reconstruction results in the network can be corrected multiple times. The upsampling unit generates more high-resolution spatial features, while the downsampling unit projects these features to the low-resolution space. In this way, more high-resolution components can be retained and more deep features can be effectively used. A 7-layer structure is used, and each layer includes upsampling and downsampling sublayers.

[0084] The degradation network module needs to help the model learn the degradation distribution of the real world, which is the most significant difference from other SISR super-resolution methods. This module is also a generator network that degrades the input high-resolution image into a blurred image with real-world characteristics. Since nonlinear networks are inconsistent with actual theories, and deep linear networks are more suitable for optimization than single-layer networks, a deep linear network without an activation function is used to simulate the degradation process. The network is a 6-layer linear full convolutional network with a stride of 2: the first three convolution kernels are 7x7, 5x5, and 3x3 in size, and the remaining layers are 1x1 in size, which is equivalent to a convolution kernel with a receptive field of 17x17.

[0085] The discriminant module is to learn the distribution of regional pixels to discriminate whether it comes from the real image distribution or the generated image distribution. For the discriminator of high-resolution images, five 3x3 convolutional layers are used. The convolutional layers (except the first and last layers) are followed by the LeakyReLU activation function with BatchNornalization. In addition, a patch discriminator with a fixed receptive field is used. For example, a three-layer network corresponds to a 70×70 patch. That is, each output value of the discriminator is only related to a patch in a local fixed area. The patch loss will be fed back to the generator to optimize the gradient of local details. In general, an image is input and a feature map is obtained after passing through the discriminator. Each value in the feature map represents the confidence that the corresponding area is judged to be true. The confidence of the entire image is obtained by averaging all the values ​​in the feature map. The discriminator of low-resolution images has a similar structure, but has different step sizes in the initial layer.

[0086] Loss function module.

[0087] The loss function is used to guide the direction of model training. A suitable loss function can make model training more efficient. The following loss functions are used in this patent invention, and their weighted summation is used to obtain the overall loss function:

[0088] 1) Adversarial loss: Adversarial loss is used to distinguish the distribution of generated images from the distribution of original images.Up and the discriminator D for high-definition images HR For example, the adversarial loss can be described as:

[0089]

[0090] Among them, Y~P Y is the distribution of high-resolution data in the training set, X~P X is the distribution of low-resolution training data in the training set. Unlike the SISR super-resolution method, these two parts of data are unpaired. Since this patent invention has multiple discriminators, the overall adversarial loss L adv The sum of all losses.

[0091] 2) Cycle consistency loss: Although the adversarial loss can ensure that the generated image and the target image are in the same domain, the content information in the image may be lost during the generation process. Thanks to the closed loop in the framework, two cycle consistency losses are adopted, which can preserve the content information through self-supervised super-resolution and degradation sub-networks. Therefore, the cycle consistency loss is defined as:

[0092]

[0093] 3) Information loss: In addition to adversarial loss and cycle consistency loss, for low-resolution images generated by the degraded network, information loss is also introduced to maintain color components and avoid color changes:

[0094]

[0095] Where Bic is the bicubic downsampling operation.

[0096] 4) Perceptual loss: In our preliminary experiments, there are many artifacts in the generated high-resolution samples. In fact, the features extracted from the pre-trained deep network contain rich semantic information, and their distance can be used as a judgment of perceptual similarity. Therefore, a perceptual loss is added between the high-resolution image and the corresponding original image:

[0097]

[0098] where φ layer It is the feature of the layer of the pre-trained CNN. Here we use the VGG-19 network pre-trained on ImagesNet.

[0099] The final loss function can be determined as:

[0100] Loss = λ adv L adv +λ cyc L cyc +λInfo L Info +λ per L per

[0101] Among them, λ adv is the weight of the adversarial loss, λ cyc is the weight of the cycle consistency loss, λ Info is the weight of information loss, λ per is the weight of the perceptual loss. Experiments have shown that the weight combinations of 0.1, 10, 10, and 0.5 are better.

[0102] The same or similar reference numerals correspond to the same or similar components;

[0103] The positional relationships described in the drawings are only for illustrative purposes and should not be construed as limiting the present patent.

[0104] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A real-world image super-resolution system that jointly learns super-resolution and degradation. It is characterized in that include: The super-resolution network module generates a high-resolution image from the input low-resolution image through the network; The degradation network module degrades the high-resolution image obtained by the super-resolution network module into a blurred image with real-world features; The degradation network module is a generator network that degrades the input high-resolution image into a blurred image with real-world characteristics: a deep linear network without an activation function is used to simulate the degradation process. The network is a 6-layer linear full convolutional network with a stride of 2: the sizes of the first three convolution kernels are 7×7, 5×5, and 3×3, and the sizes of the remaining layers are 1×1, which is equivalent to a convolution kernel with a receptive field of 17×17. The discrimination module discriminates whether the image input to the degradation network module comes from the real image distribution or the generated image distribution; The loss function module constrains both the super-resolution network module and the degradation network module to ensure the distribution of the output image and retain the content information. The loss function module includes an adversarial loss module, a cycle-consistent loss module, an information loss module, and a perceptual loss module, and the weighted sum of them is used to obtain the total loss function module. For the low-resolution images generated by the degradation network module, an information loss module is introduced to maintain the color components and avoid color changes. The information loss module is described as follows: Where Bic is the bicubic downsampling operation.

2. The real-world image super-resolution system for jointly learning super-resolution and degradation according to claim 1, It is characterized in that The super-resolution network module uses a generator to generate a high-resolution image from an input low-resolution image through the network: a structure with an iterative error feedback mechanism is used as the generator, and the projection error is back-propagated through alternating iterative upsampling and downsampling units, and the reconstruction results in the network are corrected multiple times; the upsampling unit generates more high-resolution spatial features, and the downsampling unit projects these features to the low-resolution space, retaining more high-resolution components and effectively using more deep features.

3. The real-world image super-resolution system for jointly learning super-resolution and degradation according to claim 2, It is characterized in that The discriminant module learns the distribution of image pixels to discriminate whether it is from the real image distribution or the generated image distribution: five 3×3 convolutional layers are used, and all but the first and last convolutional layers use the LeakyReLU activation function with BatchNornalization; a feature map is obtained after passing through the discriminator module, and each value in the feature map represents the confidence that the corresponding image area is judged to be true. The confidence of the entire image can be obtained by averaging all the values ​​in the feature map; the discriminant module includes a high-resolution image discrimination module and low-resolution discrimination module .

4. The real-world image super-resolution system for jointly learning super-resolution and degradation according to claim 3, It is characterized in that The adversarial loss module is used to distinguish the super-resolution network module The distribution of the output image and the distribution of the original image, the adversarial loss module is described as: in is the distribution of image data output by the super-resolution network module, is the distribution of the input low-resolution image training data.

5. The real-world image super-resolution system for jointly learning super-resolution and degradation according to claim 4, It is characterized in that Since the content information in the image will be lost during the generation process, a cycle-consistent loss module is used to preserve the content information through self-supervised super-resolution and degradation sub-networks. The cycle-consistent loss module is described as: 。 6. The real-world image super-resolution system for jointly learning super-resolution and degradation according to claim 5, It is characterized in that Since there are many artifacts in the generated high-resolution samples, a perceptual loss module is added between the high-resolution image and the corresponding original image, which is described as: in It is the feature of the layer of the pre-trained CNN, using the VGG-19 network pre-trained on ImagesNet.

7. The real-world image super-resolution system for jointly learning super-resolution and degradation according to claim 6, It is characterized in that The final loss function module is determined as: in is the weight against the loss, is the weight of cycle consistency loss, is the weight of information loss, are the weights of perceptual loss, which are 0.1, 10, 10, and 0.5 respectively.

8. A super-resolution method, based on the real-world image super-resolution system for jointly learning super-resolution and degradation as claimed in claim 7, It is characterized in that The following steps are involved: S1: The super-resolution network module generates a high-resolution image from the input low-resolution image through the network; S2: The degradation network module degrades the high-resolution image obtained by the super-resolution network module into a blurred image with real-world features; S3: The discrimination module discriminates whether the image input by the degradation network module comes from the real image distribution or the generated image distribution; S4: The loss function module constrains both the super-resolution network module and the degradation network module to ensure the distribution of the output image and preserve the content information.