Light-weight image blind super-resolution reconstruction method based on near-end gradient descent

Abstract

The invention discloses a lightweight image blind super-resolution reconstruction method based on near-end gradient descent, and the method comprises the steps: obtaining a high-resolution image data set, constructing a training set and a test set, and carrying out the image preprocessing; an image blind super-resolution reconstruction problem is decomposed into degradation estimation and image reconstruction, a blurring kernel updating module and a high-resolution feature map updating module are constructed, and the blurring kernel updating module updates a blurring kernel by combining a low-resolution feature map and a high-resolution feature map optimized in an alternating optimization process. The high-resolution feature map updating module uses a fuzzy kernel to assist in image reconstruction through a near-end gradient descent method; total loss is constructed based on the L1 loss between the fuzzy kernel updated in each iteration stage and the real fuzzy kernel and the L1 loss between the super-resolution image and the original high-resolution image; and performing model training based on the total loss, testing the model based on the test set, and outputting a reconstructed super-resolution image. According to the method, high-quality image blind super-resolution reconstruction can be realized with less model parameters.

Description

Technical Field

[0001] The present invention relates to the technical field of image reconstruction, and in particular to a lightweight image blind super-resolution reconstruction method based on proximal gradient descent. Background Art

[0002] The rapid development of computer and communication technologies has led to increasingly higher demands for image quality. Higher image quality often translates to higher resolution, richer texture and detail, and better visual effects. Image resolution depends on the precision of the image acquisition device. However, due to inherent limitations of image acquisition hardware, such as sensor pixel density and analog-to-digital conversion bit depth, recovering high-resolution images with rich textures and sharp details from existing low-resolution images has become a major challenge in image processing.

[0003] Image super-resolution (SR) technology generates high-resolution (HR) images with richer textures and sharper details by appropriately interpolating low-resolution (LR) images. Current mainstream super-resolution reconstruction methods are generally based on the assumption that low-resolution images are generated from high-resolution images via a single, predefined degradation process. Training sets constructed under this paradigm typically contain only image pairs with a specific degradation type, resulting in the trained model essentially being able to learn only the inverse transformation of that specific degradation process. However, in the real world, image degradation is complex and varied. At the imaging system level, physical constraints such as optical component defocus, mechanical jitter errors, the sensor's spatial sampling rate limitations, photoelectric conversion characteristics, and noise suppression capabilities directly restrict raw image quality. At the information transmission level, channel attenuation, random noise interference, and lossy compression encoding mechanisms under low bandwidth constraints can cause high-frequency information loss and compression artifacts in image data. Furthermore, the natural environment directly impacts image quality. Harsh lighting conditions caused by fog, wind, rain, and heavy snow can cause optical sensors to over- or under-expose, failing to accurately capture sufficient detail and resulting in image distortion. When there is a significant domain discrepancy between the predefined single degradation model and the complex degradation distribution of the actual low-resolution image, the reconstruction performance of existing mainstream super-resolution algorithms often degrades significantly. Therefore, the real world urgently requires a method that can handle more complex degraded images and reasonably reconstruct low-resolution images without knowing their degradation, generating high-resolution images rich in texture and detail. Summary of the Invention

[0004] In order to overcome the defects and shortcomings of the existing technology, the present invention provides a lightweight image blind super-resolution reconstruction method based on proximal gradient descent. The present invention decomposes the image blind super-resolution reconstruction problem into two steps: degradation estimation and image reconstruction, and constructs a blur kernel update module and a high-resolution feature map update module. The blur kernel update module realizes more accurate degradation estimation of low-resolution images by combining low-resolution feature maps with high-resolution feature maps reconstructed in an alternating optimization process. The high-resolution feature map update module utilizes the estimated blur kernel to assist image reconstruction through extraction descent, avoiding the destruction of the internal spatial relationship of the blur kernel and obtaining a more accurate high-resolution feature map. At the same time, the parameters of each module are shared, so that a better image blind super-resolution effect is achieved with a smaller number of parameters.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The present invention provides a lightweight image blind super-resolution reconstruction method based on proximal gradient descent, comprising the following steps:

[0007] Obtain high-resolution image datasets, construct training and test sets, and perform image preprocessing;

[0008] The high-resolution images in the training set are converted into low-resolution images through the degradation model;

[0009] The blind super-resolution image reconstruction problem is decomposed into degradation estimation and image reconstruction steps, which are performed alternately in multiple stages. The high-resolution feature map is optimized in the alternating optimization process, and the reconstructed super-resolution image is output.

[0010] In the degradation estimation step, a blur kernel update module is constructed. The blur kernel update module iteratively updates the blur kernel by combining the low-resolution feature map with the high-resolution feature map optimized in the alternating optimization process.

[0011] In the image reconstruction step, a high-resolution feature map update module is constructed. The high-resolution feature map update module decomposes the image reconstruction problem into gradient descent and proximal mapping, and uses a blur kernel to assist image reconstruction.

[0012] The total loss is constructed based on the L1 loss between the updated blur kernel and the true blur kernel at each iterative stage, as well as the L1 loss between the super-resolved image and the original high-resolution image;

[0013] The image blind super-resolution reconstruction model is trained based on the total loss to obtain the trained image blind super-resolution reconstruction model;

[0014] The blind super-resolution reconstruction model of the test image is tested based on the low-resolution images of the test set, and the reconstructed super-resolution image is output.

[0015] As a preferred technical solution, constructing training sets and test sets specifically includes:

[0016] Obtain the DF2K dataset as a data sample. The DF2K dataset includes the DIV2K dataset and the Flickr2K dataset. Part of the DIV2K dataset and high-resolution images of the Flickr2K dataset are used as training sets.

[0017] Another part of the DIV2K dataset is processed with different blur kernels to obtain 2x downsampled low-resolution images and 4x downsampled low-resolution images, which are used as the DIV2KRK test set.

[0018] As a preferred technical solution, image preprocessing is performed, specifically including:

[0019] The high-resolution images in the training set are randomly flipped horizontally and vertically and rotated 90° for data augmentation. As a preferred technical solution, the blur kernel update module iteratively updates the blur kernel by combining the low-resolution feature map with the high-resolution feature map optimized in the alternating optimization process. Specifically, the following steps are performed:

[0020] The blur kernel update module downsamples the high-resolution feature map through strided convolution, concatenates the downsampled high-resolution feature map with the low-resolution feature map channel, interacts with information and extracts features through two layers of residual blocks and average pooling, enhances the nonlinear expression ability of the module through multi-layer convolution and ReLU activation function, compresses the feature size into vector form through global average pooling and expands the channel to the blur kernel size through convolution, and adds it to the blur kernel in the previous iteration process to realize blur kernel update. The final blur kernel is obtained through normalization processing to assist in updating the high-resolution feature map information in the next stage of image reconstruction.

[0021] As a preferred technical solution, the high-resolution feature map update module decomposes the image reconstruction problem into gradient descent and proximal mapping, which can be specifically expressed as follows:

[0022]

[0023] Among them, I LR is a low-resolution image, I SR To reconstruct a high-resolution image, k is the blur kernel, Represents the convolution operation, ↓ s Indicates downsampling with a scaling factor of s, ↑ s Indicates upsampling with a scaling factor of s, i represents the index of different iteration stages, prox λ Represents the proximal mapping operator.

[0024] As a preferred technical solution, the downsampling operation is implemented through 3×3 strided convolution, and the upsampling is implemented using the bicubic operator.

[0025] As a preferred technical solution, the proximal mapping operator is completed using the residual feature distillation module RFDM.

[0026] As a preferred technical solution, the total loss is constructed based on the L1 loss between the updated blur kernel and the true blur kernel in each iterative stage, as well as the L1 loss between the super-resolved image and the original high-resolution image. It is specifically expressed as:

[0027]

[0028] Among them, I HR For high-resolution images, I SR To reconstruct a high-resolution image, k represents the true blur kernel, k i Represents the blur kernel estimated in each iterative stage.

[0029] The present invention also provides a computer-readable storage medium storing a program, which, when executed by a processor, implements the above-mentioned lightweight image blind super-resolution reconstruction method based on proximal gradient descent.

[0030] The present invention also provides a computer device, comprising a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the above-mentioned lightweight image blind super-resolution reconstruction method based on proximal gradient descent is implemented.

[0031] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0032] The present invention decomposes the image blind super-resolution reconstruction problem into two steps: degradation estimation and image reconstruction. It also constructs a blur kernel update module and a high-resolution feature map update module. The blur kernel update module realizes a more accurate degradation estimation of the low-resolution image by combining the low-resolution feature map with the high-resolution feature map reconstructed in the alternating optimization process. The high-resolution feature map update module assists in image reconstruction by extracting and descenting the estimated blur kernel, avoiding the destruction of the internal spatial relationship of the blur kernel and obtaining a more accurate high-resolution feature map. At the same time, the parameters of each module are shared, so that a better image blind super-resolution effect is achieved with a smaller number of parameters.

[0033] The present invention conducts quantitative comparison on the public test set DIV2KRK. In the blind super-resolution reconstruction task with a scaling factor of 2, a PSNR value of 32.67dB and an SSIM value of 0.9058 are achieved with a parameter amount of 0.3M. In the blind super-resolution reconstruction task with a scaling factor of 4, a PSNR value of 28.70dB and an SSIM value of 0.7852 are achieved with a parameter amount of 0.37M. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1Schematic diagram of the overall implementation framework of the lightweight image blind super-resolution reconstruction method based on proximal gradient descent of the present invention;

[0035] Figure 2 This is a schematic diagram of the network architecture of the fuzzy kernel update module of the present invention;

[0036] Figure 3 Schematic diagram of the network architecture of the high-resolution feature map update module of the present invention;

[0037] Figure 4 This is a comparison chart of experimental results of the present invention and the existing image blind super-resolution reconstruction method on synthetic images;

[0038] Figure 5 This is a comparison chart of experimental results of the present invention and the existing image blind super-resolution reconstruction method on real images. DETAILED DESCRIPTION

[0039] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0040] Example 1

[0041] like Figure 1 As shown, this embodiment provides a lightweight image blind super-resolution reconstruction method based on proximal gradient descent, comprising the following steps:

[0042] S1. Obtain image data set and perform image preprocessing;

[0043] In this embodiment, the DF2K dataset is obtained as a data sample. The DF2K dataset is obtained by merging the DIV2K dataset and the Flickr2K dataset. The DIV2K dataset has a total of 1,000 high-resolution images, but only 900 of them are publicly available. The first 800 high-resolution images are used for training, and the 801st to 900th high-resolution images are used for testing. The 801st to 900th high-resolution images are processed with different blur kernels to obtain 2x downsampled low-resolution images and 4x downsampled low-resolution images, which are used as the DIV2KRK test set to verify the model performance.

[0044] The Flickr2K dataset contains 2,650 high-resolution images, including people, animals, landscapes, and other types, covering a wide range of natural scenes and everyday objects.

[0045] This example uses 800 high-resolution images from the DIV2K dataset and 2,650 high-resolution images from the Flickr2K dataset as training sets. During the training process, the DIV2K dataset is subjected to a degradation model to obtain low-resolution images (LRs) with a size of 48×48. The high-resolution images in the training set are then randomly flipped horizontally and vertically and rotated 90° for data augmentation.

[0046] S2. Multiple low-resolution feature maps are obtained from low-resolution images through 3×3 convolution, and image information is transferred from the image domain to the feature domain. The image blind super-resolution reconstruction problem is decomposed into two steps: degradation estimation and image reconstruction. A blur kernel update module and a high-resolution feature map update module are constructed. These steps are performed alternately in multiple stages to achieve image super-resolution reconstruction.

[0047] S21: Constructing a blur kernel update module in the degradation estimation step. The blur kernel update module accurately estimates the blur kernel by combining the low-resolution feature map with the high-resolution feature map reconstructed during the alternating optimization process, providing more accurate blur kernel guidance for the high-resolution feature map update module in the next stage.

[0048] like Figure 2 As shown in the figure, the blur kernel update module first downsamples the high-resolution feature map through 3×3 strided convolution, and then splices it with the low-resolution feature map channel. Then, two layers of residual blocks and average pooling are used to realize information interaction between the two and further extract effective information. Then, nine layers of 1×1 convolution and ReLU activation function are used to further enhance the nonlinear expression ability of the module. Then, the feature size is compressed into vector form through global average pooling and the channel is expanded to the blur kernel size through 1×1 convolution. Then, the blur kernel is added to the blur kernel in the previous iteration process to realize the blur kernel update, and the final blur kernel is obtained through normalization to assist in the update of high-resolution feature map information in the next stage of image reconstruction.

[0049] S22: Construct a high-resolution feature map update module in the image reconstruction step. The high-resolution feature map update module uses gradient descent and proximal mapping to efficiently utilize the blur kernel to assist image reconstruction, avoiding the destruction of the internal spatial information of the blur kernel. Under the guidance of a more accurate blur kernel, better image reconstruction is achieved, providing high-quality high-resolution feature maps for the blur kernel update module in the next stage.

[0050] like Figure 3As shown in the figure, the high-resolution feature map update module decomposes the image reconstruction problem into gradient descent and proximal mapping. The downsampling in gradient descent is implemented by 3×3 strided convolution, the upsampling is implemented by the bicubic operator, and the proximal mapping operator is implemented by the residual distillation block. The entire module does not perform channel dimensionality reduction, keeping the information flowing in the feature domain during the reconstruction process, avoiding the loss of rich information in the feature domain during the reconstruction process.

[0051] In this embodiment, gradient descent and proximal mapping are specifically expressed as follows:

[0052]

[0053] Among them, I LR is a low-resolution image, I SR To reconstruct a high-resolution image, k is the blur kernel, Represents the convolution operation, ↓ s Indicates downsampling with a scaling factor of s, ↑ s Indicates upsampling with a scaling factor of s, i represents the index of different iteration stages, prox λ The proximal mapping operator is represented by the 3×3 strided convolution, and the upsampling is implemented using the bicubic operator. The proximal mapping operator in the formula is often considered a denoising problem in other image restoration tasks and is usually completed using two residual blocks. This embodiment uses a more efficient residual feature distillation module (RFDM) to complete the task. Finally, under the guidance of a more accurate blur kernel, the blur kernel is efficiently used to assist image reconstruction through gradient descent, avoiding the destruction of the internal spatial relationship of the blur kernel and obtaining a more accurate high-resolution feature map.

[0054] In this embodiment, the calculation formula of the blur kernel is expressed as:

[0055]

[0056] Among them, the standard deviation of the blur kernel is uniformly sampled between [0.6, 5.0], the rotation angle θ is randomly sampled between [-π, π], for 2x (scale factor s = 2) reconstruction, the blur kernel uses an anisotropic Gaussian kernel of size 11×11 (i∈[1,11],i∈Z; j∈[1,11],j∈Z); for 4x (scale factor s = 4) reconstruction, the blur kernel uses an anisotropic Gaussian kernel of size 31×31 (i∈[1,31],i∈Z; j∈[1,31],j∈Z), which can be expressed as:

[0057]

[0058] In this embodiment, the standard deviation is 2.6 (σ x =2.6,σ y=2.6) as the initial blur kernel. The complete blur kernel is used during training instead of the blur kernel features after principal component analysis dimensionality reduction. Stronger supervision helps to better optimize the blur kernel update module.

[0059] S23: Alternately perform degradation estimation and image reconstruction steps in multiple stages to optimize, gradually obtaining more reliable degradation estimation and more accurate high-resolution feature maps. Preferably, the two modules are alternately performed four times, and finally the high-resolution feature map is reduced from the feature domain to the image domain through 3×3 convolution to obtain a reconstructed super-resolution image SR. At the same time, the parameters of each module are shared during the alternating process, ultimately achieving high-quality image blind super-resolution reconstruction with a smaller number of model parameters;

[0060] S3. Combine the L1 loss between the reconstructed blur kernel and the true blur kernel at each iterative stage and the L1 loss between the super-resolved image and the original high-resolution image to optimize the training of the blur kernel update module and the high-resolution feature map update module. The total loss is expressed as:

[0061]

[0062] Among them, I HR For high-resolution images, I SR To reconstruct a high-resolution image, k represents the true blur kernel, k i Represents the blur kernel estimated in each iterative stage;

[0063] In this example, the initial learning rate is set to 2×10 -4 , the number of epochs is set to 2700, and the learning rate is halved at 1 / 2 and 3 / 4 of the total epoch number respectively;

[0064] S4. Training an image blind super-resolution reconstruction model based on the total loss to obtain a trained image blind super-resolution reconstruction model;

[0065] S5. Test the blind super-resolution reconstruction model of the image based on the low-resolution image of the test set, and output the reconstructed super-resolution image.

[0066] In this embodiment, the performance of the designed model PGDSR is evaluated by calculating the peak signal-to-noise ratio (PSNR) and the result similarity (SSIM) between the super-reconstructed image and the original high-resolution image. As shown in Table 1 below, the lightweight image blind super-resolution reconstruction method based on proximal gradient descent (denoted by PGDSR) of the present invention is compared with various existing image blind super-resolution methods. The quantitative comparative effect evaluation is performed on the DIV2KRK dataset. The optimal indicators in the table are in bold:

[0067] Table 1 Comparison of reconstruction results of various image blind super-resolution reconstruction methods on the test set DIV2KRK

[0068]

[0069]

[0070] Through Table 1 and Figure 4 It was found that the PGDSR of the present invention has the least number of parameters and the best reconstructed visual effect. In terms of objective evaluation indicators, the lightweight image blind super-resolution reconstruction method based on proximal gradient descent (denoted by PGDSR) of the present invention performs significantly better than the two-step separation optimization method that combines blur kernel estimation with non-blind super-resolution reconstruction method. Compared with the image blind super-resolution method based on blur kernel estimation, it is also better than most methods such as IKC, DANv1, IDENet and DASR. Compared with DCLS, which has the highest PSNR, the model parameters of the proposed PGDSR are only 2.7% of DCLS, and the PSNR is only reduced by 0.08dB in the super-resolution reconstruction task with a scale factor of 2, and by 0.29dB in the super-resolution reconstruction task with a scale factor of 4. Compared with the current optimal model DANv3, the model parameters of the proposed PGDSR are only 19.0% of the DANv3 model, and the PSNR is only reduced by 0.29dB in the SR task with a scale factor of 2, and by 0.20dB in the SR task with a scale factor of 4;

[0071] pass Figure 4 Comparing the visual effects of the proposed method with the proposed method, we can see that while the PSNR is not as high as that of DCLS, it uses fewer parameters and restores high-frequency information that is more similar to the original image, resulting in better visual reconstruction. However, the image reconstructed by DANv1 is too blurry. The image reconstructed by DANv2 is also slightly blurry. While the image reconstructed by DANv3 is clearer, it also over-smoothes the burrs on the butterfly's edge. The image reconstructed by DCLS restores more detail in the butterfly pattern, but also restores a large number of artifacts.

[0072] pass Figure 5 It can be seen from the real image super-resolution reconstruction visual effect diagram that among the real image visual effect diagrams of various methods, the reconstructed image of the present invention has no obvious scratch artifacts, while the reconstructed images of IKC, DANv1, and DCLS have obvious scratch artifacts at the upper right of the English letter X, which indicates that PGDSR has better performance in the real-world image super-resolution task, and the method of the present invention has significant advantages in the number of parameters.

[0073] Example 2

[0074] This embodiment provides a storage medium, which may be a ROM, RAM, disk, optical disk or other storage medium. The storage medium stores one or more programs. When the program is executed by the processor, the lightweight image blind super-resolution reconstruction method based on proximal gradient descent of embodiment 1 is implemented.

[0075] Example 3

[0076] This embodiment provides a computing device, which can be a desktop computer, a laptop computer, a smart phone, a PDA handheld terminal, a tablet computer or other terminal device with a display function. The computing device includes a processor and a memory, and the memory stores one or more programs. When the processor executes the program stored in the memory, the lightweight image blind super-resolution reconstruction method based on proximal gradient descent of Example 1 is implemented.

[0077] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A lightweight image blind super-resolution reconstruction method based on proximal gradient descent, characterized by: The steps include: Obtain high-resolution image datasets, construct training and test sets, and perform image preprocessing; The high-resolution images in the training set are converted into low-resolution images through the degradation model; The blind super-resolution image reconstruction problem is decomposed into degradation estimation and image reconstruction steps, which are performed alternately in multiple stages. The high-resolution feature map is optimized in the alternating optimization process, and the reconstructed super-resolution image is output. In the degradation estimation step, a blur kernel update module is constructed. The blur kernel update module iteratively updates the blur kernel by combining the low-resolution feature map with the high-resolution feature map optimized in the alternating optimization process. In the image reconstruction step, a high-resolution feature map update module is constructed. The high-resolution feature map update module decomposes the image reconstruction problem into gradient descent and proximal mapping, and uses a blur kernel to assist image reconstruction. The total loss is constructed based on the L1 loss between the updated blur kernel and the true blur kernel at each iterative stage, as well as the L1 loss between the super-resolved image and the original high-resolution image; The image blind super-resolution reconstruction model is trained based on the total loss to obtain the trained image blind super-resolution reconstruction model; The blind super-resolution reconstruction model of the test image is tested based on the low-resolution images of the test set, and the reconstructed super-resolution image is output.

2. The lightweight image blind super-resolution reconstruction method based on proximal gradient descent according to claim 1, characterized in that: Construct training and test sets, including: Obtain the DF2K dataset as a data sample. The DF2K dataset includes the DIV2K dataset and the Flickr2K dataset. Part of the DIV2K dataset and high-resolution images of the Flickr2K dataset are used as training sets. Another part of the DIV2K dataset is processed with different blur kernels to obtain 2x downsampled low-resolution images and 4x downsampled low-resolution images, which are used as the DIV2KRK test set.

3. The lightweight image blind super-resolution reconstruction method based on proximal gradient descent according to claim 1, characterized in that: Perform image preprocessing, including: The high-resolution images in the training set are randomly flipped horizontally and vertically and rotated 90° for data augmentation.

4. The lightweight image blind super-resolution reconstruction method based on proximal gradient descent according to claim 1, characterized in that: The blur kernel update module iteratively updates the blur kernel by combining the low-resolution feature map with the high-resolution feature map optimized in the alternating optimization process. Specifically, The blur kernel update module downsamples the high-resolution feature map through strided convolution, concatenates the downsampled high-resolution feature map with the low-resolution feature map channel, interacts with information and extracts features through two layers of residual blocks and average pooling, enhances the nonlinear expression ability of the module through multi-layer convolution and ReLU activation function, compresses the feature size into vector form through global average pooling and expands the channel to the blur kernel size through convolution, and adds it to the blur kernel in the previous iteration process to realize blur kernel update. The final blur kernel is obtained through normalization processing to assist in updating the high-resolution feature map information in the next stage of image reconstruction.

5. The lightweight image blind super-resolution reconstruction method based on proximal gradient descent according to claim 1, characterized in that: The high-resolution feature map update module decomposes the image reconstruction problem into gradient descent and proximal mapping, which can be expressed as follows: Among them, I LR is a low-resolution image, I SR To reconstruct a high-resolution image, k is the blur kernel, Represents the convolution operation, ↓ s Indicates downsampling with a scaling factor of s, ↑ s Indicates upsampling with a scaling factor of s, i represents the index of different iteration stages, prox λ Represents the proximal mapping operator.

6. The lightweight image blind super-resolution reconstruction method based on proximal gradient descent according to claim 5, characterized in that: The downsampling operation is implemented by 3×3 strided convolution, and the upsampling is implemented by the bicubic operator.

7. The lightweight image blind super-resolution reconstruction method based on proximal gradient descent according to claim 5, characterized in that: The proximal mapping operator is completed using the residual feature distillation module RFDM.

8. The lightweight image blind super-resolution reconstruction method based on proximal gradient descent according to claim 1, characterized in that: Based on the L1 loss between the updated blur kernel and the true blur kernel at each iterative stage, and the L1 loss between the super-resolved image and the original high-resolution image, the total loss is constructed, which is specifically expressed as: Among them, I HR For high-resolution images, I SR To reconstruct a high-resolution image, k represents the true blur kernel, k i Represents the blur kernel estimated in each iterative stage.

9. A computer-readable storage medium storing a program, characterized in that: When the program is executed by a processor, the lightweight image blind super-resolution reconstruction method based on proximal gradient descent as described in any one of claims 1 to 8 is implemented.

10. A computer device comprising a processor and a memory for storing a program executable by the processor, characterized in that: When the processor executes the program stored in the memory, it implements the lightweight image blind super-resolution reconstruction method based on proximal gradient descent as described in any one of claims 1 to 8.

Images