CN115526777A - Blind over-separation network establishing method, blind over-separation method and storage medium - Google Patents

Abstract

The invention discloses a blind super-resolution network establishment method, a blind super-resolution method and a storage medium, belonging to the field of computer vision, including: constructing training samples from high-resolution images and corresponding degraded images, and dividing training sets, test sets and Validation set; build a blind super-resolution network, including a degradation estimation network and a generative network; the generative network includes an upsampling network and a feature extraction network with alternately connected multiple deformable convolutional layers and feature extraction modules; the degradation estimation network estimates the input image The degradation information of each pixel position in the image is input to each deformable convolutional layer; after the feature extraction network extracts the feature map of the input image, the upsampling module reconstructs it to the specified magnification of the input image size to obtain a super-resolution image; to train The degraded image in the sample is the input image, and the blind super-resolution network is trained, tested and verified to obtain a blind super-resolution network for super-resolution reconstruction of the image. The present invention can improve the effect of super-resolution reconstruction.

Description

A method for establishing a blind super-resolution network, a blind super-resolution method, and a storage medium

technical field

The invention belongs to the field of computer vision, and more specifically relates to a blind super-resolution network establishment method, a blind super-resolution method and a storage medium.

Background technique

In today's society, with the popularization of smartphones, the rise of webcasting, and surveillance equipment all over the streets, images have become an indispensable part of daily life. However, due to the limitation of shooting equipment, the influence of complex shooting environment, and the compression loss of network transmission, there are always various problems in the image, such as noise, compression artifacts, low resolution, etc. These defects greatly reduce the visual experience , and have adverse effects on tasks such as target detection and face recognition. Therefore, how to improve the image resolution on the basis of the existing hardware level has become an urgent problem to be solved.

Image super-resolution technology can improve the image resolution only through the corresponding algorithm without improving the hardware level, so it has attracted widespread attention. However, the existing super-resolution technology is mainly aimed at ideal images. It assumes that low-resolution images are obtained from high-resolution images through Bicubic downsampling, and constructs data sets in this way to train super-resolution networks. However, when faced with real scene images, the performance of existing super-resolution techniques is greatly reduced due to the fact that real scene images often have defects such as noise and artifacts. Therefore, the blind super-resolution method for real scene images has extremely high practical application value, and is also the development trend of super-resolution technology. In recent years, with the development of deep learning represented by convolutional neural networks, researchers have begun to apply it to super-resolution technology, allowing the network to automatically extract features from low-resolution images, and then construct high-resolution images.

However, due to the wide variety of defects and complex and diverse backgrounds in real scene images, the existing blind super-resolution methods based on deep learning cannot pay attention to the texture-rich and severely degraded areas in the image, resulting in unsatisfactory super-resolution reconstruction results.

Contents of the invention

Aiming at the defects and improvement needs of the prior art, the present invention provides a method for establishing a blind super-resolution network, a blind super-resolution method and a storage medium, the purpose of which is to improve the effect of super-resolution reconstruction.

In order to achieve the above object, according to one aspect of the present invention, a method for establishing a blind super-resolution network is provided, including:

Perform degeneration operations on each high-resolution image in the high-resolution image dataset to obtain the corresponding degraded image; construct training samples from the high-resolution image and the corresponding degraded image, and divide all training samples into training set and test set and validation set;

Construct the blind super-resolution network to be trained; the blind super-resolution network includes a degradation estimation network and a generation network; the degradation estimation network is used to estimate the degradation information of each pixel position in the input image, and the generation network is used to use the degradation information to perform super-resolution on the input image The generation network includes a feature extraction network and an upsampling network; the feature extraction network includes multiple deformable convolution layers and multiple feature extraction modules connected alternately, and the degradation information output by the degradation estimation network is input to each deformable convolution layer, the feature extraction network is used to extract features from the input image to obtain a feature map; the up-sampling module is used to reconstruct the feature map to a specified magnification of the input image size to obtain a super-resolution image;

Take the degraded image in the training sample as the input image, use the training set, test set and verification set to train, test and verify the blind super-resolution network to be trained respectively, and obtain the blind super-resolution network for super-resolution reconstruction of the image .

The blind super-resolution network established by the present invention includes a degradation estimation network for estimating the degradation information of each pixel position in an input image and a generation network for generating a super-resolution image. The feature extraction network of the generation network uses a deformable convolution module to convert The degradation information of each pixel position predicted by the degradation estimation network is introduced into the generation network, and the offset of the deformable convolution will be generated after the introduction. Since the degradation information of different pixel positions may be different, correspondingly, the offset also exists in different positions. The difference enables the deformable convolution to extract more useful information for different positions of the input degraded image and improve the effect of super-resolution reconstruction.

Furthermore, the degradation estimation network is a UNet network, and a spatial attention module is inserted into the encoding module.

The present invention uses the UNet network as the backbone network of the degradation estimation network, and introduces a spatial attention module into the encoding module, so that the degradation estimation network can pay attention to the texture-rich positions in the degradation image, and enhance the estimation of the degradation information of each pixel position Accuracy, which in turn assists the generative network to generate better high-resolution images.

Further, a channel attention module is inserted into the decoding module of the UNet network.

Different degrees of degradation have different sizes of fuzzy kernels. In the UNet network, the input of the decoding module comes from the previous decoding module and encoding module, and its receptive field is also different. Therefore, choosing a suitable receptive field is crucial to the estimation accuracy of degraded information. Important; the present invention uses the UNet network as the backbone network of the degradation estimation network, and introduces a channel attention module in the decoding module, so that the network can adaptively select different degrees of receptive fields for different degrees of degraded images, thereby effectively improving the corresponding The estimated accuracy of the degradation information can then assist the generation network to generate better high-resolution images.

Further, the blind super-resolution network to be trained is trained, including:

Pre-training stage: use the training set to train the degradation estimation network to obtain the trained degradation estimation network;

Joint training phase: use the training set to jointly train the trained degradation estimation network and the generation network.

The blind super-resolution network established by the present invention includes the degradation estimation network and the generation network at the same time. Due to the complex network structure, the end-to-end training method is directly adopted, and the training is difficult; the present invention adopts a two-stage training method. In the first stage, that is, in the pre-training stage, the degradation estimation network is first pre-trained to make it have better degradation estimation performance; then in the second stage, that is, the joint training stage, the pre-trained degradation estimation network and the generation network are used Joint training can effectively reduce the training difficulty and improve the training efficiency on the basis of ensuring the super-resolution reconstruction effect of the overall network.

Further, the degradation operation includes: using the spatial variation blur kernel to perform the blurring operation on the high-resolution image and then downsampling, and the training sample also includes the spatial variation blur kernel corresponding to the degraded image, and the degradation information estimated by the degradation estimation network is spatially varying blur kernel;

The loss function in the pre-training phase is:

loss _p ＝||k _p -k _g || ₁ +||k _p -k _g || ₁ *g(I _LR )

The loss function of the joint training phase is:

loss＝ω ₁ *loss _p +ω ₂ *loss _g

Among them, loss represents the loss of the blind super-resolution network; loss _p and loss _g represent the losses of the degradation estimation network and the generation network, respectively, and ω ₁ and ω ₂ represent the corresponding weights; k _p represents the spatially varying fuzzy kernel estimated by the degradation estimation network , k _g represents the spatially varying blur kernel used in the degradation operation; I _LR represents the input degraded image, g() represents the gradient operator; |||| ₁ represents the mean absolute error.

When the present invention degrades the high-resolution image, it will use the spatial variation blur kernel to perform the fuzzy operation on the high-resolution image, which ensures that the degradation information of different pixel positions in the degraded image is different, and effectively improves the accuracy of the blind super-resolution network for real The ability to perform super-resolution reconstruction of images in the scene; when training the network, the combination of degraded gradient loss and degraded pixel loss is used as the loss function of the degraded estimation network. The degraded gradient loss makes the degraded estimation network pay more attention to the degraded image. Texture-rich locations, enhancing the estimation accuracy of texture-rich locations.

Further, loss _g =||I _SR -I _HR || ₁ +||g(I _SR )-g(I _HR )|| ₁

Among them, I _SR represents the super-resolution image output by the generation network, and I _HR represents the high-resolution image.

The present invention uses the combination of pixel loss and gradient loss of the super-resolution image as the loss function of the generation network. The gradient loss of the super-resolution image can make the generation network pay more attention to the texture-rich positions in the degraded image, and enhance the reconstruction effect of the super-resolution image.

Further, the gradient operator is a Scharr operator.

When the present invention constructs the loss function for the degraded estimation network and the generation network, it specifically uses the Scharr operator to calculate the gradient, so that better training effect can be obtained.

Furthermore, in the process of training the blind super-scoring network to be trained, stochastic gradient descent with momentum term is used as the optimizer.

According to another aspect of the present invention, there is provided a blind super-resolution method for real scene images, comprising: inputting the real scene images into the blind super-resolution network established by the blind super-resolution network establishment method provided by the present invention, by blind super-resolution The sub-network performs super-resolution reconstruction on real scene images to obtain high-resolution images.

According to another aspect of the present invention, a computer-readable storage medium is provided, including: a stored computer program; when the computer program is executed by a processor, it controls the device where the computer-readable storage medium is located to perform the above-mentioned blind super-resolution method provided by the present invention. A network establishment method, and/or the above-mentioned blind super-resolution method for real scene images provided by the present invention.

Generally speaking, through the above technical solutions conceived by the present invention, the following beneficial effects can be obtained:

(1) The present invention uses deformable convolution to introduce the degradation information of each pixel position predicted by the degradation estimation network into the generation network. Since there may be differences in the degradation information of different positions, the bias of the deformable convolution is generated by using the degradation information. There are also differences in the offset and offset at different positions, so that the deformable convolution can extract more useful information for different positions of the degraded image, and improve the effect of super-resolution reconstruction.

(2) The present invention uses UNet as the backbone network of the degradation prediction network, and introduces a spatial attention module in the encoding module and a channel attention module in the decoding module. Among them, the spatial attention module is introduced in the encoding module, so that the degradation estimation network can pay attention to the texture-rich position in the degraded image, enhance the estimation accuracy of the degradation information of the texture-rich position, and then enhance the super-resolution reconstruction effect of the texture-rich region ;Adding a channel attention module to the decoding module can make the network adaptively select the appropriate receptive field for different degrees of degradation, thereby improving the estimation accuracy of the corresponding degradation information, and then assisting the generation network to generate better high-resolution images. image.

(3) The present invention uses the combination of degraded gradient loss and degraded pixel loss as the loss function of the degraded estimation network. The degraded gradient loss makes the degraded estimation network pay attention to the texture-rich position in the degraded image, enhances the estimation accuracy of the texture-rich position, and then assists Generative networks for better high-resolution images.

Description of drawings

Fig. 1 is a flow chart of a method for establishing a blind super-resolution network provided by an embodiment of the present invention;

FIG. 2 is a schematic structural diagram of a blind super-resolution network provided by an embodiment of the present invention;

Fig. 3 is a schematic diagram of a deformable convolution structure provided by an embodiment of the present invention.

detailed description

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not constitute a conflict with each other.

In the present invention, the terms "first", "second" and the like (if any) in the present invention and drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

In order to solve the technical problem that the super-resolution reconstruction effect of the existing super-resolution method is not ideal, the present invention provides a method for establishing a blind super-resolution network, a blind super-resolution method and a storage medium. There are many types of image defects and complex and diverse backgrounds. When performing blind super-resolution, the degradation estimation network in the blind super-resolution network is used to estimate the degradation information for each pixel, and then introduced into the generation network, and the generation network is used to perform super-resolution. During resolution reconstruction, more useful information is extracted for different locations, thereby improving the effect of super-resolution reconstruction; on this basis, a spatial attention mechanism and a channel attention mechanism are introduced in the degradation estimation network to make the network more efficient. Pay attention to the texture-rich positions in the degraded image, improve the estimation accuracy of the degraded information, and assist the generation network to generate better high-resolution images.

The following are examples.

Example 1:

A method for establishing a blind super-resolution network, as shown in Figure 1, comprising:

Firstly, the degradation operation is performed on each high-resolution image in the high-resolution image dataset to obtain the corresponding degraded image;

Optionally, in this embodiment, the selected high-resolution image data set is specifically the DIV2K data set; in some other embodiments of the present invention, other data sets composed of high-resolution images may also be used.

In this embodiment, for the high-resolution image in the high-resolution image data set, the specific way of degrading it is: use the spatial variation blur kernel to perform the blurring operation on the high-resolution image and then perform down-sampling; use _IHR and I _LR represent high-resolution images and low-resolution images (degraded images) respectively, then the above degradation operations can be expressed by the following expressions:

I _LR ＝(k*I _HR )↓

C、H、W分别代表图像通道数、图像高、图像宽，并且d代表空间变化模糊核在每个位置的模糊核宽度，↓表示下采样操作；Among them, k represents the spatial variation blur kernel, the so-called spatial variation blur kernel means that the blur kernels of different pixel positions in the image are different;

C, H, and W represent the number of image channels, image height, and image width, respectively, and d represents the blur kernel width of the spatially varying blur kernel at each position, and ↓ represents the downsampling operation;

其中s为放大因子，旋转角为

核的大小d为21；需要说明的是，此处参数描述仅为示例性描述，不应理解为对本发明的唯一限定，在实际应用中，可以根据需要设定为其他大小；Optionally, the spatially varying blur kernel used in this embodiment is composed of an anisotropic Gaussian blur kernel, and the anisotropic Gaussian blur kernel is randomly selected; then the selected blur kernel is used to perform convolution on the input image to complete the blurring operation ; The embodiment of the present invention adopts the DIV2K data set as the input image, and the input image is randomly cropped to a size of 320×320, and the kernel width of the anisotropic Gaussian blur kernel is

where s is the magnification factor and the rotation angle is

The size d of the core is 21; it should be noted that the parameter description here is only an exemplary description, and should not be understood as the only limitation to the present invention. In practical applications, it can be set to other sizes as needed;

The image after the blur operation is down-sampled by 1/s, that is, at the pixel position of s×s, the pixel in the upper left corner is selected to complete the down-sampling;

Afterwards, training samples are composed of high-resolution images, degraded images, and corresponding spatially varying blur kernels; as an optional implementation, this embodiment will further perform all The combination of training samples is subjected to data enhancement operations of horizontal and vertical flipping and random rotation (90°, 180°, 270°), and finally divides the training set, verification set and test set according to the ratio of 8:1:1.

As shown in Figure 1, this embodiment also includes: constructing a blind super-resolution network to be trained, the blind super-resolution network includes a degradation estimation network and a generation network, and the degradation estimation network is used to estimate the degradation information of each pixel position in the input image. In the embodiment, the degradation information specifically refers to generating a spatially varying blur kernel corresponding to the degraded image, and the generation network is used to perform super-resolution reconstruction of the input image using the degradation information; in this embodiment, the network structure is specifically shown in FIG. 2 ;

In this embodiment, the degradation estimation network uses the UNet network as the backbone network. The traditional UNet network includes an encoding module (EncBlock), a decoding module (DecBlock) and an intermediate connection module between the encoding module and the decoding module. In this embodiment, the traditional On the basis of the UNet network, a spatial attention module (SAM) is added to the encoding module to extract the spatial information of the degraded image, so that the network can focus on texture-rich positions, and a channel attention module (CAM) is added to the decoding module , so that the network adaptively selects different degrees of receptive fields for different degrees of degraded images; referring to Figure 2, in this embodiment, the encoding module includes a convolutional layer (Conv), an activation function (Relu), a maximum pooling layer (MaxPool ) and spatial attention module (SAM), the decoding module includes channel attention module (CAM), convolutional layer (Conv) and activation function (Relu), the intermediate connection module consists of two convolutional layers and activation function; it should be explained What is worth is that the specific structure of the spatial attention module (SAM) and its position in the encoding module, as well as the specific structure of the channel attention module (CAM) and its position in the decoding module, can be flexibly adjusted according to actual needs; Optionally, in this embodiment, both the spatial attention module and the channel attention module adopt the structural form in the CBAM network, the spatial attention module is introduced after the last convolution layer, and the channel attention module is introduced after the cascade layer;

In this embodiment, the generation network includes a feature extraction network and an upsampling network; referring to Fig. 2, the feature extraction network includes a plurality of deformable convolutional layers (DCN) and a plurality of feature extraction modules connected alternately, and the space of the degradation estimation network output The change blur kernel is input to each deformable convolutional layer respectively, and the feature extraction network is used to extract features from the input image to obtain a feature map; the up-sampling module is used to reconstruct the feature map to a specified magnification of the input image size to obtain super-resolution image; it should be noted that the specific structure of the deformable convolution, feature extraction network and upper module can be flexibly selected according to actual needs; between the input image and the feature extraction network, there is also a deformable convolution, which can Optionally, in this embodiment, the first deformable convolution (that is, the deformable convolution between the input image and the feature extraction network) adopts the existing DCNv2 structure, and the subsequent deformable convolution in the feature extraction network is as follows As shown in Figure 3, the offsets (offsets) are obtained by performing convolution operations on the spatially variable fuzzy kernel. The feature extraction network adopts the structural form in the ESRGAN network, that is, the RRDB module, and the upsampling module is composed of Pixelshuffle;

The spatially varying blur kernel predicted by the degradation estimation network represents the degradation information of each pixel position in the degraded image. The generation network of this embodiment uses a deformable convolution layer to introduce the degradation information estimated by the degradation estimation network into the generation network, which can make full use of the degradation The spatial information of the fuzzy kernel in the information generates the offset of the deformable convolution, so that the deformable convolution can extract more useful information from the degraded image, and obtain a higher-quality super-resolution image;

Referring to Fig. 2, in this embodiment, a connection module composed of a convolutional layer and an activation function is also included between the degradation estimation network and the generation network, which is used to adjust the spatial variation blur kernel to make the spatial variation of the degradation estimation network output The blur can be adapted as input to a deformable convolutional layer in a generative network.

Referring to Fig. 1, in this embodiment, after constructing the training set, test set and verification set, and establishing the blind super-resolution network to be trained, the following steps will be performed: take the degraded image in the training sample as the input of the blind super-resolution network Image, use the training set, test set and verification set to train, test and verify the blind super-resolution network respectively, and obtain the blind super-resolution network for super-resolution reconstruction of the image.

Considering that the network structure is relatively complex, it is difficult to directly adopt the end-to-end training method. Therefore, this embodiment adopts a two-stage training method to train the blind super-resolution network to be trained, which specifically includes:

Pre-training stage: use the training set to train the degradation estimation network to obtain the trained degradation estimation network;

Joint training stage: use the training set to jointly train the trained degradation estimation network and the generation network;

In the above two-stage training method, in the first stage, that is, the pre-training stage, the degradation estimation network is pre-trained to make it have better degradation estimation performance; then in the second stage, that is, the joint training stage, use The joint training of the pre-trained degradation estimation network and the generation network can effectively reduce the training difficulty and improve the training efficiency on the basis of ensuring the super-resolution reconstruction effect of the overall network.

In order to make the network pay more attention to the information of the texture-rich position in the image, as a preferred implementation mode, this embodiment uses the combination of degraded gradient loss and degraded pixel loss as the loss function of the degraded estimation network, and adopts the super-resolution image The combination of pixel loss and gradient loss is used as the loss function of the generation network; the degradation gradient loss makes the degradation estimation network pay more attention to the texture-rich position in the degraded image, and enhances the estimation accuracy of the texture-rich position. The gradient loss of the super-resolution image can make the generation The network pays more attention to the texture-rich positions in the degraded image, and enhances the reconstruction effect of the super-resolution image. Specifically, the loss function of the degradation estimation network is:

loss _p ＝||k _p -k _g || ₁ +||k _p -k _g || ₁ *g(I _LR )

The loss function of the generative network is:

loss _g ＝||I _SR -I _HR || ₁ +||g(I _SR )-g(I _HR )|| ₁

Correspondingly, the loss function in the pre-training stage is: loss _p ;

The loss function of the joint training phase is:

loss＝ω ₁ *loss _p +ω ₂ *loss _g

Among them, loss represents the loss of the blind super-resolution network; loss _p and loss _g represent the losses of the degradation estimation network and the generation network, respectively, and ω ₁ and ω ₂ represent the corresponding weights; k _p represents the spatially varying fuzzy kernel estimated by the degradation estimation network , k _g represents the spatial variation blur kernel used in the degradation operation; I _LR represents the input degraded image; g() represents the gradient operator. In this embodiment, the gradient operator is specifically the Scharr operator. Experiments show that, in In the above loss function, the Scharr operator is used to calculate the gradient, which can effectively improve the training effect of the network; ‖ ‖ ₁ means the average absolute error.

In order to further improve the training effect of the network, in this embodiment, in the process of training the blind super-resolution network to be trained, stochastic gradient descent (SGD) containing momentum is used as the optimizer, the momentum (Momentum) is 0.9, and the weight penalty The coefficient (Weight Decay) is 5×10 ^-4 , the batch size (Batch Size) is 8, the initial learning rate is 10 ^-3 , every 50 rounds (Epoch) is reduced by 10 times, and the number of training rounds is 200 rounds.

In general, the blind super-resolution network established in this embodiment can accurately estimate the degradation information of the image, so that the network can better focus on the texture-rich regions in the image, and effectively improve the effect of super-resolution reconstruction.

Example 2:

A blind super-resolution method for real scene images, comprising: inputting the real scene images into the blind super-resolution network established by the blind super-resolution network establishment method provided by the present invention, and performing super-resolution on the real scene images by the blind super-resolution network Reconstructed to obtain a high-resolution image.

Example 3:

A computer-readable storage medium, comprising: a stored computer program; when the computer program is executed by a processor, the device where the computer-readable storage medium is located is controlled to execute the blind super-resolution network establishment method provided in Embodiment 1 above, and/or the above implementation Example 2 provides a blind super-resolution method for real scene images.

It is easy for those skilled in the art to understand that the above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention, All should be included within the protection scope of the present invention.

Claims (10)

1. A method for establishing a blind super-resolution network, comprising: Perform degeneration operations on each high-resolution image in the high-resolution image dataset to obtain the corresponding degraded image; construct training samples from the high-resolution image and the corresponding degraded image, and divide all training samples into training set and test set and validation set; Build a blind super-resolution network to be trained; the blind super-resolution network includes a degradation estimation network and a generation network; the degradation estimation network is used to estimate the degradation information of each pixel position in an input image, and the generation network is used to utilize the Degradation information performs super-resolution reconstruction on the input image; the generation network includes a feature extraction network and an upsampling network; the feature extraction network includes a plurality of deformable convolutional layers and a plurality of feature extraction modules connected alternately, so The degradation information output by the degradation estimation network is respectively input to each deformable convolution layer, and the feature extraction network is used to perform feature extraction on the input image to obtain a feature map; the up-sampling module is used to convert the feature map Reconstruction is a specified magnification of the input image size to obtain a super-resolution image; Taking the degraded image in the training sample as the input image, using the training set, the test set and the verification set to train, test and verify the blind super-resolution network to be trained respectively, to obtain Blind super-resolution network for super-resolution reconstruction. 2. The method for establishing a blind super-resolution network according to claim 1, wherein the degradation estimation network is a UNet network, and a spatial attention module is inserted in the coding module therein. 3. The method for establishing a blind super-resolution network according to claim 2, wherein a channel attention module is inserted in the decoding module of the UNet network. 4. The method for establishing a blind super-resolution network according to any one of claims 1 to 3, wherein training the blind super-resolution network to be trained includes: Pre-training stage: using the training set to train the degradation estimation network to obtain a trained degradation estimation network; Joint training stage: using the training set to jointly train the trained degradation estimation network and the generation network. 5. The method for establishing a blind super-resolution network according to claim 4, wherein the degeneration operation comprises: using a spatially variable blur kernel to perform a fuzzy operation on a high-resolution image and then downsampling, and the training samples It also includes a spatially varying blur kernel corresponding to the degraded image, and the degradation information estimated by the degradation estimation network is a spatially varying blur kernel; The loss function of the pre-training stage is: loss _p ＝||k _p -k _g || ₁ +||k _p -k _g || ₁ *g(I _LR ) The loss function of the joint training phase is: loss＝ω ₁ *loss _p +ω ₂ *loss _g Among them, loss represents the loss of the blind super-resolution network; loss _p and loss _g represent the loss of the degradation estimation network and the generation network respectively, and ω ₁ and ω ₂ represent the corresponding weights respectively; k _p represents the spatial variation estimated by the degradation estimation network The blur kernel, k _g represents the spatially varying blur kernel used in the degradation operation; I _LR represents the input degraded image, g() represents the gradient operator; ‖‖ ₁ represents the mean absolute error. 6. blind super-resolution network establishment method as claimed in claim 5, is characterized in that, loss _g = ‖I _SR -I _HR ‖ ₁ + ‖g(I _SR )-g(I _HR )‖ ₁ Among them, I _SR represents the super-resolution image output by the generation network, and I _HR represents the high-resolution image. 7. The method for establishing a blind super-resolution network according to claim 5, wherein the gradient operator is a Scharr operator. 8. The method for establishing a blind super-resolution network as claimed in claim 4, wherein, in the process of training the blind super-resolution network to be trained, stochastic gradient descent containing a momentum term is used as an optimizer. 9. A blind super-resolution method for real scene images, comprising: inputting real scene images into the blind super-resolution network established by the blind super-resolution network establishment method described in any one of claims 1 to 8 , performing super-resolution reconstruction on the real scene image by the blind super-resolution network to obtain a high-resolution image. 10. A computer-readable storage medium, comprising: a stored computer program; when the computer program is executed by a processor, it controls the device where the computer-readable storage medium is located to perform any one of claims 1-8 The method for establishing a blind super-resolution network, and/or the blind super-resolution method for real scene images provided in claim 9.

Images