A method, device and storage medium for obtaining blind super-resolution images

Abstract

The present disclosure relates to a method, device and storage medium for obtaining blind super-resolution images. The method includes estimating image degradation model parameters in the natural image degradation process, wherein the image degradation model parameters include a convolution kernel pool and a noise pool; establishing a low-resolution-high-resolution paired image dataset according to the image degradation model parameters; using the paired image dataset to train and optimize an image generator; and training and optimizing an image discriminator based on the paired image dataset and the image generator to obtain a blind super-resolution image. The present disclosure estimates model parameters through a degradation model of a natural image so as to establish a paired dataset, which is then used to train and optimize an image generator and a discriminator, thereby being able to obtain a blind super-resolution image, greatly improving the quality of the super-resolution image.

Description

Technical Field

[0001] The present disclosure generally relates to the field of image processing technology. More specifically, the present disclosure relates to a method, device and computer-readable storage medium for acquiring blind super-resolution images. Background Art

[0002] This section is intended to provide background or context for the embodiments of the present disclosure set forth in the claims. The descriptions herein may include concepts that may be explored, but not necessarily concepts that have been previously thought of or explored. Therefore, unless otherwise noted herein, the content described in this section is not prior art with respect to the specification and claims of the present application, and is not admitted to be prior art by inclusion in this section.

[0003] With the development of information technology, people have higher and higher requirements for the quality of digital images, especially in the fields of computer vision such as medicine, autonomous driving, astronomy, and monitoring, which all require high-resolution, detailed, high-definition images. However, in actual image acquisition, the spatial resolution of the image obtained is not high, often affected by factors such as the spatial resolution of the imaging system itself, light or ray intensity, spatial distance, and system noise. The traditional way to improve the resolution of an image is usually designed based on system hardware devices or software algorithms, and improving the image resolution based on system hardware devices is a more direct way, but this method is costly and difficult to implement. In addition, the system hardware equipment is also susceptible to the characteristics of electromagnetic wave physics and component physics, which limits the improved spatial resolution.

[0004] Single Image Super Resolution (SISR) reconstruction technology is mainly based on the design of software algorithm level, which can achieve the improvement from low resolution (LR) observation image to high resolution (HR) observation image. Recently, due to the powerful function of deep convolutional neural network in processing image data with Euclidean structure, many convolutional neural network based super resolution techniques have solved the task of single image super resolution analysis from certain specific perspectives. However, these methods usually design with prior information to find a balance between detail recovery and noise suppression, which often leads to the loss of high-frequency detail information. In addition, there are many methods that are based on bicubic downsampling to obtain HR-LR images and use them for training. Therefore, when faced with the problem of natural image degradation caused by distance, electromagnetic wave signal strength, and system noise, the image degradation process does not match these natural degradation processes, resulting in the performance of super resolution models being limited. Summary of the invention

[0005] In order to at least solve the above problems, the present disclosure proposes a solution for obtaining a blind super-resolution image, that is, by estimating the degradation model parameters of a natural image, so as to obtain a paired image dataset based on the model parameters, so that an image generator and a discriminator can be trained and optimized to obtain a high-quality blind super-resolution image. In view of this, the present disclosure provides corresponding solutions in the following aspects.

[0006] In a first aspect, the present disclosure provides a method for acquiring a blind super-resolution image, comprising: estimating image degradation model parameters in a natural image degradation process, wherein the image degradation model parameters include a convolution kernel pool and a noise pool; establishing a low-resolution-high-resolution paired image dataset according to the image degradation model parameters; using the paired image dataset to train and optimize an image generator; and training and optimizing an image discriminator based on the paired image dataset and the image generator to obtain a blind super-resolution image.

[0007] In one embodiment, estimating image degradation model parameters in a natural image degradation process includes: constructing an image degradation model for degradation from a high-resolution image to a low-resolution image in the natural image; learning a blur kernel of each natural image in the image degradation model using a linear convolutional network to estimate a blur kernel pool; and extracting a noise distribution of the natural image in the image degradation model based on relative variance selection and sub-block partitioning to estimate a noise pool.

[0008] In another embodiment, establishing a low-resolution-high-resolution paired image data set includes: randomly sampling from the blur kernel pool and the noise pool respectively according to a random principle to obtain a target blur kernel and a target noise block; and simulating a degradation process for a known high-resolution image based on the target blur kernel, the target noise block and the image degradation model to establish a low-resolution image paired with the known high-resolution image.

[0009] In another embodiment, using the paired image dataset to train and optimize an image generator includes: under the constraint of quadratic Wasserstein distance, inputting a low-resolution image into the image generator to generate a pseudo high-resolution image to obtain a loss function of the image generator; and training the image generator based on the loss function.

[0010] In yet another embodiment, using the paired image dataset to train and optimize the image generator further includes: adding multiple target losses to the loss function to optimize the image generator.

[0011] In yet another embodiment, the multiple target losses include adversarial loss, boundary loss, pixel loss and / or perceptual loss.

[0012] In another embodiment, training and optimizing an image discriminator includes: inputting a pseudo high-resolution image and a true high-resolution image into the image discriminator to obtain a discrimination score; determining a numerical solution for the discrimination score using linear programming and a deep neural network; and training and optimizing the image discriminator based on the optimal transmission principle and the quadratic Wasserstein distance determined by the numerical solution.

[0013] In yet another embodiment, the above method further includes: adding an optimal transmission regularization term to the image discriminator to optimize the image discriminator.

[0014] In a second aspect, the present disclosure provides a device for acquiring a blind super-resolution image, comprising: a processor; and a memory connected to the processor, wherein the memory stores a computer program code, and when the computer program code is executed by the processor, the device executes the method described above and in its multiple embodiments.

[0015] In a third aspect, the present disclosure provides a computer-readable storage medium having stored thereon computer-readable instructions for acquiring a blind super-resolution image, wherein when the computer-readable instructions are executed by one or more processors, the method described above and in its multiple embodiments is implemented.

[0016] According to the scheme disclosed in the present invention, by estimating the degradation model parameters of natural images, a paired image data set is obtained based on the model parameters, so that the image generator and the discriminator can be trained and optimized to obtain high-quality blind super-resolution images. Furthermore, the embodiment disclosed in the present invention is based on the optimal transmission principle and uses the quadratic Wasserstein distance as a metric to promote mutual optimization between the image generator and the discriminator, so that the image is generated by the optimized generator and a high-resolution image that is closer to the real one. In addition, in the embodiment disclosed in the present invention, the image generator is trained by adding multiple target losses so that the generator can achieve the best in both edge high-frequency detail recovery and noise removal, thereby being able to obtain a blind super-resolution image with high fidelity of high-frequency details and significant noise reduction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] By reading the detailed description below with reference to the accompanying drawings, the above and other purposes, features and advantages of the exemplary embodiments of the present disclosure will become readily understood. In the accompanying drawings, several embodiments of the present disclosure are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0018] Figure 1 is a simplified flow chart illustrating a method for acquiring a blind super-resolution image according to an embodiment of the present disclosure;

[0019] Figure 2is a detailed flow chart showing a method for acquiring a blind super-resolution image according to an embodiment of the present disclosure;

[0020] Figure 3-Figure 6 is an exemplary schematic diagram showing generation of a blind super-resolution image according to an embodiment of the present disclosure; and

[0021] Figure 7 is an exemplary schematic diagram showing a blind super-resolution image generated according to an embodiment of the present disclosure and an image generated in other ways. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments of the present disclosure will be described clearly and completely below in conjunction with the accompanying drawings. It should be understood that the embodiments described in this specification are only some embodiments provided by the present disclosure to facilitate a clear understanding of the solutions and to comply with legal requirements, and are not all embodiments that can implement the present invention. Based on the embodiments disclosed in this specification, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present disclosure.

[0023] Figure 1 is a simplified flow chart illustrating a method 100 for acquiring a blind super-resolution image according to an embodiment of the present disclosure.

[0024] As shown in the figure, at step S102, image degradation model parameters in the natural image degradation process are estimated. In one embodiment, the aforementioned image degradation model parameters may include a convolution kernel pool and a noise pool. This step S102 can be understood as being able to estimate the convolution kernel and noise distribution that conform to the actual degradation process through learning (such as machine learning or deep learning) on ​​a low-quality image data set in a natural scene. Before estimating the image degradation model parameters, an image degradation model of the degradation from a high-resolution ("HR") image to a low-resolution ("LR") image in a natural image can be first constructed. In an implementation scenario, the image degradation model from a high-resolution image to a low-resolution image can be expressed as the following formula:

[0025]

[0026] Among them, LR represents low-resolution image, HR represents high-resolution image, k and n represent convolution kernel and noise respectively. represents the convolution operation, ↓ represents the downsampling operation, and s represents the downsampling multiple.

[0027] Based on the image degradation model constructed above, the image degradation model parameters (including the convolution kernel pool and the noise pool) can be estimated. In one embodiment, a linear convolution network (such as a KernelGAN network) can be used to learn the blur kernel of each natural image in the image degradation model, where all blur kernels can form a blur pool. In another embodiment, the noise distribution of the natural image in the image degradation model can be extracted according to relative variance selection and sub-block partitioning to form a noise pool. Specifically, the aforementioned noise distribution can be extracted using the following formula:

[0028]

[0029] Among them, p i represents an image block, represents the image block p i The local image sub-block in , mean(·) and var(·) represent mean operation and variance operation respectively, μ and γ can represent empirical parameters, and v max In some embodiments, μ and γ are usually parameters input by those skilled in the art based on empirical data. For example, in the solution of the present disclosure, μ and γ can be set to 0.1 and 0.25, respectively.

[0030] It can be understood that, for an image, when the variance of a local image sub-block is small, it can be said that the local image sub-block in the above image block belongs to a grayscale flat area, that is, there is no noise distribution or the noise is very small. When the variance of a local image sub-block is large, it means that the local image sub-block in the above image block belongs to an edge or high-frequency area (that is, the area contains noise or the noise is very large). Therefore, combined with the above formula (2), the variance of the local image sub-block is greater than the maximum variance threshold v max The image distribution is used as the noise distribution to form a noise pool. In the embodiment of the present disclosure, the maximum variance threshold v max Can be set to 40.

[0031] After obtaining the image degradation model parameters (including the blur kernel pool and the noise pool), at step S104, a low-resolution-high-resolution paired image dataset is established according to the image degradation model parameters. Regarding the establishment of the aforementioned paired image dataset, in one embodiment, the present disclosure proposes that first, random sampling can be performed from the blur kernel pool and the noise pool respectively according to the random principle, so as to obtain a target blur kernel and a target noise block. Then, based on the aforementioned extracted target blur kernel and target noise block, and substituting them into the image degradation model represented by the above formula (1), it is possible to simulate the degradation process for a known high-resolution image, so as to obtain a low-resolution image that conforms to the natural degradation process. Thus, the present disclosure establishes a low-resolution-high-resolution paired image dataset.

[0032] After establishing the above-mentioned paired image data set, the process proceeds to step S106. At this step, the paired image data set is used to train and optimize the image generator. In an implementation scenario, this step can be implemented according to the optimal transmission principle. As known to those skilled in the art, optimal transport can be used to solve the mapping problem between two distributions. For example, assume that two variables X and Y are given, and assume that the distributions of the two metrics X and Y in the corresponding space are μ and ν respectively. Under this assumption, a transmission transformation T:X→Y is found, which can transform the random variable X corresponding to the distribution μ into the random variable Y obeying the distribution ν, thereby obtaining the expectation of minimizing the transmission cost c(x,T(x)). Thus, the mapping between the above-mentioned two variables can be obtained. It should be understood that in the optimal transmission principle, its main transmission cost basis can be the quadratic Wasserstein distance.

[0033] Specifically, under the constraint of the quadratic Wasserstein distance, a low-resolution image can be input into an image generator to generate a pseudo high-resolution image. At this time, since there is a difference between the input low-resolution image and the generated pseudo high-resolution image, the loss function of the image generator can be obtained. Further, the image generator can be trained based on the obtained loss function. For example, the image generator can be trained by adjusting the weight coefficient of the loss function. Based on the trained image generator, a pseudo high-resolution image can be obtained for subsequent training and optimization of the image discriminator, thereby forming an optimal mapping from a low-resolution image to a high-resolution image. In one embodiment, the aforementioned quadratic Wasserstein distance can be specifically expressed as the following formula:

[0034]

[0035] Among them, G represents the image generator, D represents the image discriminator, and z i and i They represent the low-resolution images and high-resolution images in the paired image dataset respectively, and m represents the number of low-resolution images and high-resolution images in each iteration.

[0036] According to the foregoing description, image noise is usually distributed at the edge or high-frequency area. Therefore, in order to obtain high-fidelity images with high-frequency details and to obtain more paired image information, the disclosed embodiment may also add multiple target losses to the loss function to optimize the aforementioned image generator. In different implementation scenarios, the aforementioned multiple target losses may be, for example, adversarial loss, boundary loss, pixel loss and / or perceptual loss, and the adversarial loss may be the aforementioned quadratic Wasserstein distance. The boundary loss, pixel loss and perceptual loss in the aforementioned multiple target losses will be described in detail below.

[0037] In one embodiment, when obtaining an image with high-frequency details and high fidelity, a boundary loss can be added to the loss function. The boundary loss introduces an attention mechanism so that the image generator can pay more attention to the restoration of high-frequency details at the edge. edge It can be expressed as the following formula:

[0038]

[0039] in, represents a boundary operator, and G represents an image generator. Additionally, in the embodiment of the present disclosure, a deep neural network, for example, a RRDB model may be used to implement the aforementioned operations.

[0040] In one embodiment, in order to obtain more paired image information to optimize the image generator, a pixel loss representing the image texture content may be added to the loss function. In particular, the aforementioned pixel loss may be, for example, the pixel mean square error L 1 , and the pixel mean square error L 1 It can be expressed as the following formula:

[0041]

[0042] Where G represents the image generator, ||·|| 1 represents the 1-norm, h and w represent the number of rows and columns of the image respectively.

[0043] In one embodiment, in order to generate an image with better visual effects, a perceptual loss can be added to the loss function. per It can be expressed as the following formula:

[0044]

[0045] Where G represents the image generator, ||·|| 2represents the 2-norm, ψ represents the feature extractor, and h, w and c represent the number of rows, columns and channels of the image, respectively. In some implementation scenarios, the feature extractor may use a neural network model such as VGG-19 to achieve feature extraction.

[0046] Based on the multiple target losses added to the loss function, the multiple target losses can be subjected to weighted summation, for example, to obtain the final optimization target, so as to optimize the image generator. The optimization target is denoted as L G , it can be expressed as the following formula:

[0047] L G =λ adv L adv +λ edge L edge +λ 1 L 1 +λ per L per (7)

[0048] Among them, L adv , L edge , L 1 and L per They represent adversarial loss, boundary loss, pixel loss and perceptual loss respectively, and λ adv , edge , 1 and λ per They represent the weight coefficients corresponding to the aforementioned various losses. In the embodiment of the present disclosure, the weight coefficient λ adv , edge , 1 and λ per Set to 0.05, 0.01, 0.01 and 0.7 respectively.

[0049] According to the paired image dataset and image generator obtained above, the process of method 100 proceeds to step S108. At step S108, the image discriminator is trained and optimized based on the paired image dataset and the image generator to obtain a blind super-resolution image. As mentioned above, the paired image dataset contains low-resolution images and high-resolution images, and the low-resolution images in the paired image dataset are input into the image generator to generate pseudo-high-resolution images. Furthermore, the true high-resolution images and pseudo-high-resolution images in the paired image dataset can be simultaneously input into the image discriminator to train and optimize the image discriminator. More specifically, linear programming and deep neural networks such as VGG-128 can be used to solve the Monge-Kantorovich dual problem in a discrete case, so as to obtain a numerical solution of the discriminant score after the image discriminator, and then train and optimize the image discriminator.

[0050] In one implementation scenario, the Monge-Kantorovich dual problem in the discrete case can be expressed as follows:

[0051]

[0052] Where SR and HR represent the pseudo high-resolution image set and the true high-resolution image set, respectively, and m represents the number of pseudo high-resolution images and true high-resolution images in each iteration. i, x j ) represents the secondary transmission cost function, which is defined as follows:

[0053]

[0054] In formula (9), ||·|| 2 represents the 2-norm, x j and i The distribution represents the pseudo high-resolution image and the true high-resolution image. Combining the above formula (8) and formula (9), the discrimination score φ(y i ) and τ(x j ), for example, they can be expressed as and

[0055] Based on the obtained numerical solution of the discriminant score, the image discriminator is trained and optimized again according to the optimal transmission principle. Similar to the above training and optimization of the image generator, when training and optimizing the image discriminator, the quadratic Wasserstein distance based on the quadratic transmission cost can be used as the optimization target. Further, the quadratic Wasserstein distance can be determined by the numerical solution of the discriminant score to obtain the optimal image discriminator. At this time, the quadratic Wasserstein distance is related to the numerical solution of the discriminant score, and its specific expression is as follows:

[0056]

[0057] In one embodiment, in order to improve the stability of the image discriminator, the embodiment of the present disclosure also adds an optimal transmission regularization term. The optimal transmission regularization term L OTR It can be expressed by the following formula:

[0058]

[0059] In addition, for the quadratic Wasserstein distance L wass and the optimal transmission regularization term L OTR For example, weighted summation is used to obtain the final optimization target L of the image discriminator. D, so that a more stable optimal image discriminator can be obtained. The above optimization goal L D It can be expressed as the following formula:

[0060] L D =1×L wass +0.41×L OTR (12)

[0061] It should be understood that L in formula (12) wass The coefficients 1 and L OTR The coefficient of 0.41 is merely illustrative and not restrictive, and those skilled in the art may also select other suitable coefficients to obtain the optimization goal according to the teachings of the present disclosure.

[0062] In combination with the above description, it can be seen that the embodiment of the present disclosure first estimates the real degradation process of the image in the natural scene and the corresponding parameters, and further obtains the low-resolution image that conforms to the natural degradation process according to the parameters, thereby establishing a low-resolution-high-resolution paired image data set. Based on the obtained paired data set and the optimal transmission principle, the image generator and the image discriminator are trained and optimized respectively. For example, the quadratic Wasserstein distance based on the quadratic transmission cost can be used as a measure of the distance between the pseudo high-resolution image space and the true high-resolution image space, and the linear programming is used to solve the Monge-Kantorovich dual problem to continuously optimize the quadratic Wasserstein distance, so that the generated pseudo-high-resolution image is closer and closer to the true high-resolution image. In addition, the embodiment of the present disclosure adds multiple target losses in the optimization process so that a blind super-resolution image with high fidelity of high-frequency details and obvious noise reduction effect can be obtained. With the scheme of the present disclosure, the restoration of high-frequency details of the image and the removal of noise can be optimized at the same time.

[0063] Figure 2 Detailed flowchart of the method for obtaining a blind super-resolution image according to an embodiment of the present disclosure. Figure 1 As can be seen from the description, method 200 is Figure 1 The method 100 is shown as a specific implementation manner, so the above description of the method 100 is also applicable to the method 200.

[0064] As shown in the figure, at step S202, an unpaired natural image data set is input into the constructed image degradation model. In one embodiment, the image degradation model can be constructed based on the above formula (1). Then, at step S204, the natural image degradation process is estimated based on the input natural image data set to obtain image degradation model parameters, such as convolution pool K and noise pool N. Among them, the blur pool can be obtained by learning the convolution kernel of the natural image through a linear convolution network, and the noise pool N can be obtained based on the above formula (2). As mentioned above, the image distribution whose local image sub-block variance is greater than the maximum variance threshold is usually selected as the noise distribution to form the noise pool.

[0065] According to the obtained image degradation model parameters (including the convolution pool K and the noise pool N), at step S206, a convolution kernel k and a noise n can be randomly extracted from K and N according to the random allocation principle, and applied to the high-resolution image to obtain a low-resolution image that conforms to the true degradation process. Repeatedly perform step S206 until the low-resolution-high-resolution (LR-HR) paired image dataset of the present disclosure is obtained. Further, at step S208, the low-resolution image in the paired image dataset obtained in the above step S206 is input into the image generator, and under the constraint of the quadratic Wasserstein distance, the image generator is continuously trained and optimized so that the image generator can gradually approach an optimal mapping from low resolution to high resolution. In other words, based on the aforementioned image generator, the input low-resolution image can generate a pseudo high-resolution image. At step S210, the pseudo high-resolution image and the true high-resolution image are input into the image discriminator, and the image discriminator is used to distinguish the pseudo high-resolution image from the true high-resolution image (for example, using the discrimination score described above to distinguish), and the image discriminator is made to approach the optimal discriminator by continuously optimizing the quadratic Wasserstein distance. Finally, by repeatedly executing the above steps S208 and S210, the image generator and the image discriminator can promote and optimize each other, and finally generate a high-quality blind super-resolution image, such as Figure 3-Figure 6 shown.

[0066] Figure 3-Figure 6 is an exemplary schematic diagram showing the generation of a blind super-resolution image according to an embodiment of the present disclosure. As shown in the figure, Figure 3-Figure 6 The left images in the figure are all natural images before processing. Figure 3-Figure 6 It can be seen that natural images are all blurred, that is, the resolution is low. Figure 3-Figure 6 The images on the right in the figure all represent images processed using the disclosed solution. It can be seen that the processed images are clearer and have better display effects.

[0067] Figure 7It is an exemplary schematic diagram showing a blind super-resolution image generated according to an embodiment of the present disclosure and an image generated by other methods. As shown in the figure, the first column on the left side of the figure shows multiple images, and they are all low-resolution ("LR") images. The multiple columns immediately following the first column are images generated based on different processing methods for the local images in the first column of images on the left side (shown in the rectangular box in the figure). The figure further shows that the second column to the last column on the left are the result images generated based on the scheme of the present disclosure, the paired data set realSR, the DAN algorithm, the IKC algorithm, the DPSR algorithm, the ESRGAN algorithm and the ZSSR algorithm. By comparing the results of processing by different algorithms, it can be seen that the result image obtained by the method of the embodiment of the present disclosure is clearer (that is, the resolution is higher).

[0068] According to the above description in conjunction with the accompanying drawings, those skilled in the art can also understand that the embodiments of the present disclosure can also be implemented by a software program. Therefore, the present disclosure also provides a computer program product. The computer program product can be used to implement the method for obtaining a blind super-resolution image described in conjunction with Figures 1 and 2 of the present disclosure.

[0069] It should be noted that although the operations of the disclosed method are described in a particular order in the accompanying drawings, this does not require or imply that the operations must be performed in this particular order, or that all the operations shown must be performed to achieve the desired results. On the contrary, the steps depicted in the flow chart can be performed in a different order. Additionally or alternatively, some steps can be omitted, multiple steps can be combined into one step, and / or one step can be decomposed into multiple steps.

[0070] It should be understood that when the terms "first", "second", "third" and "fourth" are used in the claims, specifications and drawings of the present disclosure, they are only used to distinguish different objects, rather than to describe a specific order. The terms "include" and "comprise" used in the specification and claims of the present disclosure indicate the presence of the described features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or their collections.

[0071] It should also be understood that the terms used in this disclosure are only for the purpose of describing specific embodiments and are not intended to limit the present disclosure. As used in this disclosure and claims, the singular forms of "a", "an", and "the" are intended to include the plural forms unless the context clearly indicates otherwise. It should also be further understood that the term "and / or" used in this disclosure and claims refers to any combination of one or more of the associated listed items and all possible combinations, including these combinations.

[0072] Although the embodiments of the present invention are as described above, the contents are only examples used to facilitate understanding of the present invention, and are not intended to limit the scope and application scenarios of the present invention. Any technician in the technical field of the present invention can make any modifications and changes in the form and details of the implementation without departing from the spirit and scope disclosed by the present invention, but the scope of patent protection of the present invention shall still be based on the scope defined by the attached claims.

Claims

1. A method for obtaining a blind super-resolution image, characterized in that: include: estimating image degradation model parameters in a natural image degradation process, wherein the image degradation model parameters include a convolution kernel pool and a noise pool; Establishing a low-resolution-high-resolution paired image dataset according to the image degradation model parameters; Under the constraint of the quadratic Wasserstein distance, the low-resolution image is input into the image generator to generate a pseudo high-resolution image, so as to obtain a loss function of the image generator; training an image generator based on the loss function; Inputting the pseudo high-resolution image and the true high-resolution image into the image discriminator to obtain a discriminant score; Determining a numerical solution for the discriminant score using linear programming and a deep neural network; as well as Based on the optimal transmission principle and the quadratic Wasserstein distance determined by numerical solution, the image discriminator is trained and optimized to obtain blind super-resolution images. Among them, the image degradation model parameters in the natural image degradation process are estimated including: Constructing an image degradation model for degradation from a high-resolution image to a low-resolution image in the natural image; Using a linear convolutional network to learn the blur kernel of each of the natural images in the image degradation model to estimate a blur kernel pool; and extracting the noise distribution of the natural image in the image degradation model according to relative variance selection and sub-block partitioning to estimate a noise pool; Among them, the extraction formula of noise distribution includes: in, is the image block, is selected from the image block The local image sub-block in and are mean operation and variance operation respectively, μ and γ represent empirical parameters, v max Represents the maximum variance threshold.

2. The method according to claim 1, characterized in that The low-resolution-high-resolution paired image datasets include: Randomly sampling from the blur kernel pool and the noise pool respectively according to a random principle to obtain a target blur kernel and a target noise block; and A degradation process is simulated for a known high-resolution image based on the target blur kernel, the target noise block and the image degradation model to establish a paired low-resolution image.

3. The method according to claim 1, characterized in that Wherein using the paired image dataset to train and optimize the image generator further comprises: A variety of target losses are added to the loss function to optimize the image generator.

4. The method according to claim 3, characterized in that The multiple target losses include adversarial loss, boundary loss, pixel loss and / or perceptual loss.

5. The method according to claim 1, characterized in that: Also includes: An optimal transmission regularization term is added to the image discriminator to optimize the image discriminator.

6. A device for obtaining a blind super-resolution image, comprising: processor; as well as A memory connected to the processor, wherein the memory stores computer program code, and when the computer program code is executed by the processor, the device executes the method according to any one of claims 1 to 5.

7. A computer-readable storage medium storing computer-readable instructions for acquiring a blind super-resolution image, wherein when the computer-readable instructions are executed by one or more processors, the method according to any one of claims 1 to 5 is implemented.

Images