A Blind Image Super-Resolution Method Based on Learning Degradation Information

Abstract

This invention discloses a blind image super-resolution method based on learned degradation information, belonging to the field of image processing technology. The specific steps are as follows: preparing a training set; training an unsupervised degradation representation network, which provides degradation information to assist a supervised global context information attention super-resolution network; training the overall network using the degradation information; taking the loss function value and determining the network model; inputting the test dataset into the determined network model to obtain the super-resolution result. This blind image super-resolution method based on learned degradation information solves the problem of existing super-resolution methods failing to effectively remove unknown degradation factors in real images due to the assumption of known degradation. For both synthetic and real natural images, it improves image resolution while removing unknown degradation factors such as noise and blur.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a blind super-resolution method for images based on learned degradation information. Background Technology

[0002] In recent years, with the widespread use of various visual acquisition devices such as cameras, camcorders, and mobile phones, receiving and using images has become an integral part of people's lives. Image resolution is a set of performance parameters used to evaluate the richness of detail information contained in an image; it reflects the actual ability of an imaging system to reflect the detailed information of an object. Compared to low-resolution images, high-resolution images typically contain higher pixel density, richer texture details, and higher reliability, providing more effective assistance to people. However, due to constraints from many factors such as acquisition equipment, environmental influences, transmission media, and bandwidth, it is usually difficult to directly obtain ideal high-resolution images with sharpened edges, no obvious noise, or blocky blurring. The most direct way to improve image resolution is to improve the optical hardware in the acquisition system, but this requires a large investment and the improvement is relatively small. Therefore, researchers have proposed another solution: using super-resolution reconstruction methods, starting from the software and algorithm perspective, to perform calculations on existing low-quality images to simulate high-quality images. Image super-resolution reconstruction technology has a wide range of applications and research significance in many fields.

[0003] With the continuous development of deep learning technology in image processing, image super-resolution reconstruction has become a hot research topic. Most existing super-resolution methods tend to assume that the degradation model is known, for example, by using simple bicubic downsampling from high-quality images to construct low-resolution and high-resolution image pairs for training. However, when the assumed degradation model deviates from the degradation model in real images, it is difficult to achieve good performance. Therefore, accurately learning and removing degradation information in images is essential and more in line with real-world needs. Summary of the Invention

[0004] The purpose of this invention is to solve the aforementioned problems existing in the prior art.

[0005] To achieve the above objectives, this invention provides a blind super-resolution method for images based on learned degradation information, the specific steps of which are as follows:

[0006] Step S1: Prepare the training set;

[0007] Step S2: Train an unsupervised degenerate representation network. The unsupervised degenerate representation network provides degenerate information to assist the supervised global context information attention super-resolution network.

[0008] Step S3: Train the overall network using the degradation information from step S2;

[0009] Step S4: Calculate the loss function value and determine the network model;

[0010] Step S5: Input the test dataset into the network model determined in step S4 to obtain the super-resolution result.

[0011] Preferably, step S1 is as follows:

[0012] Step S11: Select images from the DIV2K and Flickr2K datasets to form the DF2K_HR dataset and assign them numbers;

[0013] Step S12: Downsample the images in the DF2K_HR dataset using bicubic interpolation to obtain the DF2Kx4_Bicubic dataset, and number the images in the DF2Kx4_Bicubic dataset.

[0014] Step S13: Add isotropic Gaussian blur to the images in the DF2Kx4_Bicubic dataset to obtain a low-resolution blurred dataset, denoted as DF2Kx4_GTBlur dataset, and number the images in the DF2Kx4_GTBlur dataset;

[0015] Step S14: Add anisotropic Gaussian blur to the images in the DF2Kx4_Bicubic dataset to obtain a low-resolution blurred dataset, denoted as DF2Kx4_GYBlur dataset, and number the images in the DF2Kx4_GYBlur dataset.

[0016] Step S15: Add Gaussian noise to the DF2Kx4_GYBlur dataset to obtain a low-resolution noisy blurred dataset, denoted as DF2Kx4_BN dataset, and number the images in the DF2Kx4_BN dataset;

[0017] Step S16: Combine the images in the DF2K_HR dataset and the DF2Kx4_GYBlur dataset into image pairs according to their numbering order. Use the dataset composed of these image pairs as the training set for the synthetic natural image experiment. Combine the images in the DF2K_HR dataset and the DF2Kx4_BN dataset into image pairs according to their corresponding numbering order. Use these pairs as the training set for the real natural image experiment.

[0018] Step S17: Randomly select a set number of images from the low-resolution dataset and crop out a set number and size of image patches from each image.

[0019] Specifically:

[0020] When conducting experiments to synthesize natural images, images from the DF2Kx4_GTBlur dataset were selected and sliced ​​into blocks; the resulting image blocks are denoted as follows:

[0021] When conducting experiments with real natural images, images from the DF2Kx4_BN dataset were selected and sliced ​​into blocks, and the resulting image blocks were denoted as follows:

[0022] Preferably, step S2 is as follows:

[0023] Step S21: Input the low-resolution image patch into the unsupervised degradation representation network.

[0024] When conducting experiments to synthesize natural images, Image patches are input into an unsupervised degradation representation network;

[0025] When conducting experiments with real natural images, Image patches are input into an unsupervised degradation representation network;

[0026] Step S22: Start training, set the number of iterations parameter for the unsupervised degenerate representation network, and initialize the number of iterations parameter to zero;

[0027] Step S23: The degenerate representation is obtained by sequentially passing through a convolutional structure and a global average pooling layer;

[0028] Step S24: Calculate the synthetic contrast loss of the synthetic image experiment, perform a set number of iterations of training until the synthetic contrast loss reaches the convergence requirement, and stop training; otherwise, backpropagate the synthetic contrast loss value, update the parameters again through the ADAM optimization algorithm, and enter steps S22 and S23 again to repeat the training.

[0029] Calculate the true contrast loss of the real image experiment, perform a set number of iterations of training until the true contrast loss reaches the convergence requirement, and stop training; otherwise, backpropagate the true contrast loss value, update the parameters again through the ADAM optimization algorithm, and enter steps S22 and S23 again to repeat the training.

[0030] Preferably, step S3 specifically includes:

[0031] Step S31: Input the low-resolution image patch simultaneously into the unsupervised degradation representation network and the supervised global context information attention network.

[0032] For experiments involving synthesizing natural images, low-resolution blurred image patches are used. Input into the network;

[0033] For experiments with real natural images, low-resolution, noisy, blurred image patches are used. Input into the network;

[0034] Step S32: Start training, set the network loop training number parameters for the synthetic image experiment and the network loop training number parameters for the real image experiment respectively, and the initial values ​​are all zero;

[0035] Step S33: Pass the low-resolution image patch through a convolutional layer and a global average pooling layer to obtain a degenerate representation, and extract shallow features; at the same time, input the low-resolution image patch into another convolutional layer to obtain a shallow feature map;

[0036] Step S34: Perform deep feature extraction by simultaneously inputting the shallow feature map and the degenerate representation into the deep feature extraction network to obtain the deep feature map;

[0037] Step S35: Use the upsampling module to obtain the reconstructed high-resolution image SR.

[0038] Preferably, in step S34, the deep feature extraction network includes several global context information attention residual groups and a merging layer. The several global context information attention residual groups are connected in sequence and connected to the merging layer through skip connections. The merging layer performs element-wise summation and outputs the result.

[0039] Each global context information attention residual group includes several degenerate information attention blocks, a first convolutional layer, and a selectable convolutional kernel block. Several degenerate information attention blocks are connected in sequence and then connected in sequence with the first convolutional layer and the selectable convolutional kernel block. The output of the first convolutional layer is summed element-wise with the input of the global context information attention residual group and then input into the selectable convolutional kernel block.

[0040] Each degenerate information attention block includes two degenerate information attention layers and two second convolutional layers, with the degenerate information attention layers and second convolutional layers alternately connected in series;

[0041] Each selectable convolutional kernel block includes a third convolutional layer, a fourth convolutional layer, a global average pooling layer, two fully connected layers, a batch normalization layer, a linear rectified layer, and an activation function layer. The third and fourth convolutional layers are set up side by side. The outputs of the third and fourth convolutional layers are summed element by element and then fed into the global average pooling layer. The global average pooling layer, the first fully connected layer, the batch normalization layer, the linear rectified layer, the second fully connected layer, and the activation function layer are connected in sequence. The outputs of the third and fourth convolutional layers are multiplied element by element with the output of the activation function layer and then summed element by element.

[0042] Preferably, step S4 is as follows:

[0043] Step S41: In the synthesized image experiment, calculate the synthesis contrast loss of the unsupervised degradation network and the synthesis loss of the supervised global context information attention super-resolution network.

[0044] In real-image experiments, the true contrast loss of the unsupervised degradation network and the true loss of the supervised global context information attention super-resolution network were calculated.

[0045] Step S42: Calculate the total loss function value of the synthetic image experiment and the total loss function value of the real image experiment, respectively;

[0046] The total loss function value of the synthesized image experiment is calculated using the following formula:

[0047] Loss_1=Loss_degard_1+Loss_sr_1

[0048] Where Loss_1 is the total loss function value of the synthetic image experiment, Loss_degrad_1 is the synthetic contrast loss of the unsupervised degradation network, and Loss_sr_1 is the synthetic loss of the supervised global context information attention super-resolution network.

[0049] Meanwhile, if the set number of updates is reached, the network stops training and saves the training model obtained from the last round of updates as the network model; otherwise, it enters step S3 again and repeats the training loop.

[0050] The total loss function value for the real-image experiment is calculated using the following formula:

[0051] Loss_2=Loss_degard_2+Loss_sr_2

[0052] Where Loss_2 is the total loss function value in the real image experiment, Loss_degrad_2 is the true contrast loss of the unsupervised degradation network, and Loss_sr_2 is the true loss of the supervised global context information attention super-resolution network.

[0053] Meanwhile, if the set number of updates is reached, the network stops training and saves the training model obtained from the last round of updates as the network model; otherwise, it enters step S3 again and repeats the training.

[0054] Therefore, this invention employs a blind super-resolution method for images based on learning degradation information, which mainly comprises two stages: the first stage is an unsupervised degradation representation network that learns an abstract representation of the feature space to distinguish various degradations in the image; the second stage is a supervised global context information attention super-resolution network that can flexibly adapt to various degradations and generate high-resolution results from low-resolution input images. We use skip connections to merge the outputs of each residual group to better integrate shallow and deep information. At the end of each residual group, selectable convolutional kernel blocks that apply self-attention to kernel size selection are added to more effectively model global context information. Whether for synthetic or real natural images, this method not only improves image resolution but also removes unknown degradation factors such as noise and blur, demonstrating high application value.

[0055] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0056] Figure 1 Schematic diagram of unsupervised degenerate representation and supervised global context information attention super-resolution network;

[0057] Figure 2 Global context information, note the residual group diagram;

[0058] Figure 3 Optional convolutional layer block diagram;

[0059] Figure 4 Visual contrast images with 4x super-resolution on a synthetic natural image dataset;

[0060] Figure 5 Visual contrast images with 4x super-resolution on a real natural image dataset. Detailed Implementation

[0061] Example

[0062] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0063] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0064] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0065] A blind image super-resolution method based on learning degradation information, the specific steps of which are as follows:

[0066] Step S1: Prepare the training set.

[0067] Step S11: Import 900 high-resolution natural images from the DIV2K dataset and 2650 high-resolution natural images from the Flickr2K dataset as the training set. The dataset consisting of these 3550 high-resolution images is denoted as DF2K_HR, and each image is denoted as DF2K_HR1, DF2K_HR2, ..., DF2K_HR3550. Import 219 pairs of high / low resolution images from the Set5, Set14, BSD100, and Urban100 datasets as the test set for the synthetic natural image experiment. The high-resolution images are denoted as Benchmark_HR, and each image is denoted as Benchmark_H. R1, Benchmark_HR2, ..., Benchmark_HR219; denoted as Benchmarkx4_Bicubic, each image is denoted as Benchmarkx4_Bicubic1, Benchmarkx4_Bicubic2, ..., Benchmarkx4_Bicubic219; 20 real low-resolution images from the RealSR dataset are imported as the test set for the real natural image experiment, denoted as RealSR_BNLR, each image is denoted as RealSR_BNLR1, RealSR_BNLR2, ..., RealSR_BNLR20.

[0068] Step S12: Downsampling. The dataset DF2K_HR is downsampled by 4 times through bicubic interpolation to obtain the dataset DF2Kx4_Bicubic. Each image is denoted as DF2Kx4_Bicubic1, DF2Kx4_Bicubic2, ..., DF2Kx4_Bicubic3550.

[0069] Step S13: Add isotropic Gaussian blur. Add Gaussian blur with a kernel size of 21×21 and a kernel width of (0.2, 4.0) to the dataset DF2Kx4_Bicubic after downsampling by 4 times to obtain a low-resolution blurred dataset DF2Kx4_GTBlur. Each image is denoted as DF2Kx4_GTBlur1, DF2Kx4_GTBlur2, ..., DF2Kx4_GTBlur3550. Perform the same processing on the dataset Benchmarkx4_Bicubic to obtain the dataset Benchmarkx4_GTBlur. Each image is denoted as Benchmarkx4_GTBlur1, Benchmarkx4_GTBlur2, ..., Benchmarkx4_GTBlur219.

[0070] Step S14: Add anisotropic Gaussian blur. Add Gaussian blur with a blur kernel size of 21×21, random eigenvalues ​​λ1, λ2∈(0.2,4), and random rotation angle θ∈(0,π) to the dataset DF2Kx4_Bicubic after downsampling by 4 times, to obtain a low-resolution blurred dataset DF2Kx4_GYBlur. Each image is denoted as DF2Kx4_GYBlur1, DF2Kx4_GYBlur2, ..., DF2Kx4_GYBlur3550.

[0071] Step S15: Add Gaussian noise. Add Gaussian blur with a noise level of (0,25) to the dataset DF2Kx4_GYBlur to obtain a low-resolution noisy blur dataset DF2Kx4_BN. Each image is denoted as DF2Kx4_BN1, DF2Kx4_BN2, ..., DF2Kx4_BN3550.

[0072] Step S16: Arrange the images in the datasets DF2K_HR and DF2Kx4_GTBlur into image pairs in corresponding order, denoted as... As the training set for the synthetic natural image experiment; images from the datasets DF2K_HR and DF2Kx4_BN are paired in corresponding order, denoted as . As the training set for real-world image experiments; images from the Benchmark_HR and Benchmarkx4_GTBlur datasets are paired in corresponding order, denoted as... The RealSR_BNLR dataset serves as the test set for experiments on synthesized natural images; denoted as RealSR_BNLR. As a test set for experiments with real natural images.

[0073] Step S17: Inputting into the network, 32 images from the low-resolution dataset are randomly selected each time, and two 48×48 image patches are cropped from each image and input into the first network; for the synthetic natural image experiment, images from the DF2Kx4_GTBlur dataset are selected and cut into patches, denoted as image patches. For experiments using real natural images, images from the DF2Kx4_BN dataset are selected and sliced ​​into blocks, denoted as image blocks.

[0074] Step S2: Train an unsupervised degenerate representation network. The unsupervised degenerate representation network provides degenerate information to assist the supervised global context information attention super-resolution network.

[0075] The specific steps are as follows:

[0076] Step S21: Input the low-resolution image patch into the unsupervised degradation representation network. For the experiment of synthesizing natural images, input the low-resolution blurred image patch. Input to the network; for experiments with real natural images, low-resolution, noisy, blurred image patches are used. Input into the network.

[0077] Step S22: Start training. Define the variable epoch_degrad_1 as the number of times the unsupervised degradation representation network is trained in the synthetic image experiment, and initialize it to 0. Define the variable epoch_degrad_2 as the number of times the unsupervised degradation representation network is trained in the real image experiment, and initialize it to 0.

[0078] Step S23: First, after 6 convolutional layers, Patch_Conv is obtained. Then, after a global average pooling layer, the degenerate representation LR_Degradation is obtained.

[0079] Step S24: Calculate the synthetic contrast loss of the synthetic image experiment, denoted as loss_degrad_1, and set epoch_degrad_1 = epoch_degrad_1 + 1; when epoch_degrad_1 reaches 100 iterations, loss_degrad_1 is close to convergence, the network stops training, and the training model file model_1_100.pt is saved; otherwise, the loss_degrad_1 value is backpropagated, the parameters are updated again through the ADAM optimization algorithm, and steps S22 and S23 are entered again to repeat the training.

[0080] Calculate the true contrast loss of the real image experiment, denoted as loss_degrad_2, and set epoch_degrad_2 = epoch_degrad_2 + 1; when epoch_degrad_2 reaches 100 times, loss_degrad_2 is close to convergence, the network stops training, and the training model file model_2_100.pt is saved; otherwise, the loss_degrad_2 value is backpropagated, the parameters are updated again through the ADAM optimization algorithm, and steps S22 and S23 are entered again to repeat the training.

[0081] Step S3: Train the overall network using the degradation information from step S2.

[0082] Step S31: Input the low-resolution image patch simultaneously into the unsupervised degradation representation network and the supervised global context information attention network; for the synthetic natural image experiment, input the low-resolution blurred image patch... Input to the network; for experiments with real natural images, low-resolution, noisy, blurred image patches are used. Input into the network.

[0083] Step S32: Start training the entire network. Define the variable epoch_1 as the number of times the network will be trained in the synthetic image experiment, and initialize it to 0; define the variable epoch_2 as the number of times the network will be trained in the real image experiment, and initialize it to 0.

[0084] Step S33: Feed the low-resolution image patch into 6 convolutional layers and one global average pooling layer to obtain the degraded representation LR_Degradation; at the same time, perform shallow feature extraction by feeding the low-resolution image patch into a 3×3 convolutional layer to obtain the shallow feature map Feature_Shallow.

[0085] Step S34: Deep feature extraction, comprising 5 global context information attention residual groups, 1 merging layer, 5 skip connections, and an element-wise summation operation; each residual group contains 5 degenerate information attention blocks, a 3×3 convolutional layer, and a selectable convolutional kernel block; each degenerate information attention block contains two degenerate information attention layers and two 3×3 convolutional layers; each selectable convolutional kernel block contains a 5×5 convolutional layer, a 3×3 convolutional layer, a global average pooling layer, two fully connected layers, a batch normalization layer, a linear rectified layer, and a softmax activation function layer; the shallow feature map Feature_Shallow and the degenerate representation LR_Degradation are simultaneously fed into the above structure for deep feature extraction to obtain the deep feature map Feature_Deep.

[0086] like Figure 1-3As shown, the deep feature extraction network includes several global context information attention residual groups and a merging layer. The several global context information attention residual groups are connected sequentially and then connected to the merging layer via skip connections. The merging layer performs element-wise summation and outputs the result.

[0087] Each global context information attention residual group includes several degenerate information attention blocks, a first convolutional layer, and a selectable convolutional kernel block. The degenerate information attention blocks are sequentially connected and then sequentially connected to the first convolutional layer and the selectable convolutional kernel block. The output of the first convolutional layer is summed element-wise with the input of the global context information attention residual group and then input to the selectable convolutional kernel block.

[0088] Each degenerate information attention block includes two degenerate information attention layers and two second convolutional layers, with the degenerate information attention layers and second convolutional layers alternately connected in series;

[0089] Each selectable convolutional kernel block includes a third convolutional layer, a fourth convolutional layer, a global average pooling layer, two fully connected layers, a batch normalization layer, a linear rectified layer, and an activation function layer. The third and fourth convolutional layers are set up side by side. The outputs of the third and fourth convolutional layers are summed element by element and then fed into the global average pooling layer. The global average pooling layer, the first fully connected layer, the batch normalization layer, the linear rectified layer, the second fully connected layer, and the activation function layer are connected in sequence. The outputs of the third and fourth convolutional layers are multiplied element by element with the output of the activation function layer and then summed element by element.

[0090] Step S35: Reconstruction, which includes an upsampling module. The deep feature map Feature_Deep is fed into this module to obtain the reconstructed high-resolution image SR.

[0091] Step S4: Calculate the loss function value and determine the network model.

[0092] Step S41: In the synthetic image experiment, calculate the synthetic contrast loss for the unsupervised degradation network, denoted as Loss_degrad_1, and calculate the synthetic loss for the supervised global context information attention super-resolution network, denoted as Loss_sr_1; in the real image experiment, calculate the real contrast loss for the unsupervised degradation network, denoted as Loss_degrad_2, and calculate the real loss for the supervised global context information attention super-resolution network, denoted as Loss_sr_2.

[0093] Step S42: Calculate the total loss function value for the synthetic image experiment, denoted as Loss_1, i.e., Loss_1 = Loss_degard_1 + Loss_sr_1; simultaneously, if the epoch reaches 500, the network stops training and saves the training model model_1_600.pt obtained from the last epoch as the network model; otherwise, proceed to step 3 again and repeat the training loop. Calculate the total loss function value for the real image experiment, denoted as Loss_2, i.e., Loss_2 = Loss_degard_2 + Loss_sr_2; simultaneously, if the epoch reaches 500, the network stops training and saves the training model model_2_600.pt obtained from the last epoch as the network model; otherwise, proceed to step 3 again and repeat the training loop.

[0094] Step S5: Input the test dataset into the network model determined in step S4 to obtain the super-resolution result.

[0095] Step C051: Import the model into the test code; for the synthetic image experiment, import the network model file model_1_600.pt obtained in Part 04 into the test code; for the real image experiment, import the network model file model_2_600.pt obtained in Part 04 into the test code.

[0096] Step C052: Input the natural image used for testing into the model, obtain the result and save it; for synthetic image experiments, input the synthetic image dataset used for testing. In the input model, the peak signal-to-noise ratio (PSNR) and image after super-resolution reconstruction are obtained and saved, and compared with existing super-resolution algorithms; for real-image experiments, the real-image dataset used for testing is used. Input the model to obtain and save the super-resolution reconstructed image.

[0097] The peak signal-to-noise ratio (PSNR) of the embodiments of the present invention is compared with that of existing technology algorithms in the following table.

[0098]

[0099] In the comparison algorithms above, we trained the DASR network separately using the same parameters and dataset as our network, while the other algorithms used the test data from the DASR paper.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A blind image super-resolution method based on learning degradation information, characterized in that, The specific steps are as follows: Step S1: preparing a training set; The specific steps of step S1 are as follows: Step S11: selecting images in the DIV2K dataset and the Flickr2K dataset to combine into a DF2K_HR dataset and numbering; Step S12: down-sampling the images in the DF2K_HR dataset through a bicubic interpolation operation to obtain a DF2Kx4_Bicubic dataset, and numbering the images in the DF2Kx4_Bicubic dataset; Step S13: adding isotropic Gaussian blur to the images in the DF2Kx4_Bicubic dataset to obtain a low-resolution blur dataset, denoted as a DF2Kx4_GTBlur dataset, and numbering the images in the DF2Kx4_GTBlur dataset; Step S14: adding anisotropic Gaussian blur to the images in the DF2Kx4_Bicubic dataset to obtain a low-resolution blur dataset, denoted as a DF2Kx4_GYBlur dataset, and numbering the images in the DF2Kx4_GYBlur dataset; Step S15: adding Gaussian noise to the DF2Kx4_GYBlur dataset to obtain a low-resolution noise blur dataset, denoted as a DF2Kx4_BN dataset, and numbering the images in the DF2Kx4_BN dataset; Step S16: grouping the images in the DF2K_HR dataset and the DF2Kx4_GYBlur dataset into image pairs in order of numbering, and grouping the images in the DF2K_HR dataset and the DF2Kx4_BN dataset into image pairs in order of corresponding numbering, to obtain a training set for a synthetic natural image experiment and a training set for a real natural image experiment; Step S17: randomly selecting a certain number of low-resolution dataset images and cropping a certain number and size of image blocks in each image, Specifically: When performing the synthetic natural image experiment, images from the DF2Kx4_GTBlur dataset are selected and tiled, resulting in image tiles denoted as When the experiment of real natural images is performed, images from the DF2Kx4_BN dataset are selected and cropped, and the obtained image blocks are denoted as Step S2: training an unsupervised degradation representation network, which provides auxiliary degradation information for a supervised global context information attention super-resolution network; The specific steps of step S2 are as follows: Step S21: inputting the low-resolution image block into the unsupervised degradation representation network, When performing the synthetic natural image experiment, the the image patch is input to the unsupervised degradation representation network; When the experiment of real natural images is performed, the the image patch is input to the unsupervised degradation representation network; Step S22: starting training, setting a cycle training number parameter of the unsupervised degradation representation network, and setting the initial value of the cycle training number parameter to zero; Step S23: sequentially passing through a convolution structure and a global average pooling layer to obtain a degradation representation; Step S24: calculating a synthetic contrast loss of a synthetic image experiment, performing a certain number of cycle training, until the synthetic contrast loss reaches a convergence requirement, and stopping training; otherwise, the synthetic contrast loss value is back-propagated, the parameters are updated again through an ADAM optimization algorithm, step S22 and step S23 are entered again, and the training is repeated; calculating a real contrast loss of a real image experiment, performing a certain number of cycle training, until the real contrast loss reaches a convergence requirement, and stopping training; otherwise, the real contrast loss value is back-propagated, the parameters are updated again through an ADAM optimization algorithm, step S22 and step S23 are entered again, and the training is repeated; Step S3: training the whole network through the degradation information in step S2; Step S3 specifically is: Step S31: inputting the low-resolution image block into the unsupervised degradation representation network and the supervised global context information attention network at the same time, For the experiment of synthesizing natural images, the low resolution blurred image blocks are fed into the network; For the real natural image experiment, the low resolution noisy blurred image blocks are fed into the network; Step S32: starting training, setting the network cycle training number parameters in the synthetic image experiment and the network cycle training number parameters in the real image experiment respectively, and the initial values are both zero; Step S33: obtaining the degradation representation by passing the low-resolution image block through the convolution layer and the global average pooling layer, and extracting the shallow layer features; meanwhile, inputting the low-resolution image block into another convolution layer to obtain the shallow feature map; Step S34: performing deep layer feature extraction, inputting the shallow feature map and the degradation representation into the deep layer feature extraction network to obtain the deep layer feature map; Step S35: obtaining the reconstructed high-resolution image SR by using the up-sampling module; Step S4: taking the loss function value and determining the network model; Step S5: inputting the test data set into the network model determined in step S4 to obtain the super-resolution result. 2.The image blind super-resolution method based on learning degradation information according to claim 1, characterized in that: In step S34, the deep layer feature extraction network includes a plurality of global context information attention residual groups and a merging layer, the plurality of global context information attention residual groups are connected in sequence and connected with the merging layer through a jump connection, and the merging layer outputs after performing element-wise summation operation; Each global context information attention residual group includes a plurality of degradation information attention blocks, a first convolution layer, and a selectable convolution kernel block, the plurality of degradation information attention blocks are connected in sequence and then connected with the first convolution layer and the selectable convolution kernel block in sequence, the output of the first convolution layer is input into the selectable convolution kernel block after performing element-wise summation operation with the input of the global context information attention residual group; Each degradation information attention block includes two degradation information attention layers and two second convolution layers, and the degradation information attention layers and the second convolution layers are connected in series alternately; Each selectable convolution kernel block includes a third convolution layer, a fourth convolution layer, a global average pooling layer, two fully connected layers, a batch normalization layer, a linear rectification layer, and an activation function layer, the third convolution layer and the fourth convolution layer are arranged side by side, the outputs of the third convolution layer and the fourth convolution layer are input into the global average pooling layer after performing element-wise summation, the global average pooling layer, the first fully connected layer, the batch normalization layer, the linear rectification layer, the second fully connected layer, and the activation function layer are connected in sequence, and the outputs of the third convolution layer and the fourth convolution layer are multiplied with the output of the activation function layer respectively after performing element-wise multiplication operation and then performing element-wise summation operation. 3.The image blind super-resolution method based on learning degradation information according to claim 2, characterized in that: Step S4 specifically is as follows: Step S41: in the synthetic image experiment, calculating the synthetic contrast loss of the unsupervised degradation network and the synthetic loss of the supervised global context information attention super-resolution network; In the real image experiment, calculating the real contrast loss of the unsupervised degradation network and the real loss of the supervised global context information attention super-resolution network; Step S42: calculating the total loss function value of the synthetic image experiment and the total loss function value of the real image experiment respectively; The calculation formula of the total loss function value of the synthetic image experiment is as follows: Loss_1=Loss_degard_1+Loss_sr_1 Wherein, Loss_1 is the total loss function value of the synthetic image experiment, Loss_degrad_1 is the synthetic contrast loss of the unsupervised degradation network, and Loss_sr_1 is the synthetic loss of the supervised global context information attention super-resolution network. At the same time, if the update reaches a set number of times, the network stops training, and the training model obtained in the last round of update is saved as the network model; otherwise, step S3 is entered again, and the training is repeated cyclically. The total loss function value of the real image experiment is calculated, and the calculation formula is as follows: Loss_2=Loss_degard_2+Loss_sr_2 Wherein, Loss_2 is the total loss function value of the real image experiment, Loss_degrad_2 is the real contrast loss of the unsupervised degradation network, and Loss_sr_2 is the real loss of the supervised global context information attention super-resolution network, At the same time, if the update reaches a set number of times, the network stops training, and the training model obtained in the last round of update is saved as the network model; otherwise, step S3 is entered again, and the training is repeated cyclically.

Images