Image inpainting method, device, equipment, storage medium and program product
By using encoder and decoder modules in the neural network model and training with a reconstruction loss function of real images, the problem of poor image restoration effect is solved, achieving more efficient image restoration and accelerated inference speed.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-07
- Publication Date
- 2026-04-07
AI Technical Summary
In existing technologies, the image restoration algorithms based on deep learning models generate unsatisfactory coarse results in the first stage, resulting in poor image restoration effects, and the model training and inference speed is slow.
The trained neural network model, including an encoder module and a decoder module, is used to obtain semantic and style encoding information of the image to be repaired and the region mask image through constraint training based on real images and reconstruction loss functions. The repaired image is then output through the decoder module.
It improves image restoration results, accelerates model training inference speed, outputs restored images that are closer to real images, and makes the training process more efficient.
Smart Images
Figure CN115908188B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of image restoration technology, and in particular to an image restoration method, apparatus, device, storage medium, and program product. Background Technology
[0002] Image inpainting, an important technique in image processing, is widely used in image content generation and object removal, such as watermark and logo removal. Image inpainting algorithms need to infer the area to be repaired based on the context information of the image. The challenge lies in making the repaired image appear harmonious and realistic.
[0003] In related technologies, image restoration algorithms are mainly divided into two categories: one is based on traditional mathematical and physical theories, and the other is based on deep learning models. Deep learning-based algorithms mostly consist of two stages. The first stage performs structural restoration of the image to generate a restored image with a coarse result. The second stage uses the result of the first stage to generate texture. This algorithm relies heavily on the coarse result generated in the first stage; if the coarse result generated in the first stage is not ideal, the restored image effect will be relatively unsatisfactory and needs improvement. Summary of the Invention
[0004] This application provides an image restoration method, apparatus, device, storage medium, and program product, which solves the problem of unsatisfactory image restoration effect in related technologies, resulting in better image restoration effect and faster inference speed during model training.
[0005] In a first aspect, embodiments of this application provide an image restoration method, the method comprising:
[0006] Obtain the image to be repaired and the region mask image corresponding to the image to be repaired;
[0007] The image to be repaired and the region mask image are input into a trained neural network model, which includes an encoder module and a decoder module. During the training of the neural network model, constraints are imposed based on the real image and the set reconstruction loss function.
[0008] The encoder module obtains semantic and style encoding information corresponding to the image to be repaired and the region mask image.
[0009] The decoder module outputs the semantic encoding information and the corresponding repaired image based on the style encoding information.
[0010] Secondly, embodiments of this application also provide an image restoration apparatus, comprising:
[0011] The image acquisition module is configured to acquire the image to be repaired and the region mask image corresponding to the image to be repaired;
[0012] The image input module is configured to input the image to be repaired and the region mask image into a trained neural network model. The neural network model includes an encoder module and a decoder module. During the training of the neural network model, constraints are imposed based on the real image and a set reconstruction loss function.
[0013] The image output module is configured to obtain semantic encoding information and style encoding information corresponding to the image to be repaired and the region mask image through the encoder module, and to output the repaired image corresponding to the semantic encoding information and the style encoding information through the decoder module.
[0014] Thirdly, embodiments of this application also provide an image restoration device, the device comprising:
[0015] One or more processors;
[0016] Storage device for storing one or more programs.
[0017] When the one or more programs are executed by the one or more processors, the one or more processors implement the image restoration method described in the embodiments of this application.
[0018] Fourthly, embodiments of this application also provide a non-volatile storage medium for storing computer-executable instructions, which, when executed by a computer processor, are used to perform the image restoration method described in embodiments of this application.
[0019] Fifthly, embodiments of this application also provide a computer program product, which includes a computer program stored in a computer-readable storage medium. At least one processor of the device reads from the computer-readable storage medium and executes the computer program, causing the device to perform the image restoration method described in embodiments of this application.
[0020] In this embodiment, the image to be repaired and its corresponding region mask image are obtained. These images are then input into a trained neural network model. The neural network model includes an encoder module and a decoder module. During training, constraints are applied based on the real image and a set reconstruction loss function. The encoder module obtains semantic and style encoding information corresponding to the image to be repaired and the region mask image. The decoder module outputs the repaired image corresponding to the semantic and style encoding information. This approach, by applying constraints based on the real image and the set reconstruction loss function during neural network model training, results in better repaired image performance during model use and accelerates inference speed during model training. Attached Figure Description
[0021] Figure 1 A flowchart illustrating an image restoration method provided in this application embodiment;
[0022] Figure 2 A schematic diagram of an image to be repaired provided in an embodiment of this application;
[0023] Figure 3 A schematic diagram of a region mask image provided in an embodiment of this application;
[0024] Figure 4 A schematic diagram of a real image provided for an embodiment of this application;
[0025] Figure 5 A schematic diagram of an encoder module provided in an embodiment of this application;
[0026] Figure 6 A flowchart illustrating the training process of a neural network model provided in an embodiment of this application;
[0027] Figure 7 A structural block diagram of an image restoration device provided in an embodiment of this application;
[0028] Figure 8 This is a schematic diagram of the structure of an image restoration device provided in an embodiment of this application. Detailed Implementation
[0029] The embodiments of this application will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the embodiments of this application and are not intended to limit the scope of the embodiments. Furthermore, it should be noted that, for ease of description, only the parts relevant to the embodiments of this application are shown in the accompanying drawings, not the entire structure.
[0030] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.
[0031] The image restoration method provided in this application embodiment can be applied to image processing scenarios such as image content generation and removal of specific targets in an image. Given an image that needs to be restored and a corresponding image containing the restoration area, a trained neural network model is used to output the restored image corresponding to the image that needs to be restored, thereby realizing the restoration processing of a specific area in the image that needs to be restored.
[0032] Figure 1 A flowchart of an image restoration method provided in this application embodiment specifically includes the following steps:
[0033] Step S101: Obtain the image to be repaired and the region mask image corresponding to the image to be repaired.
[0034] The image to be repaired is the image that needs to be repaired. This may include images containing watermarks, text, or other specified target content that needs to be removed. The region mask image is an image corresponding to the image to be repaired, marking the areas that need to be repaired. Optionally, it can be an image within the original image to be repaired that marks the target areas. An example of the image to be repaired and the corresponding region mask image is shown below. Figure 2 and Figure 3 As shown, where, Figure 2 This is a schematic diagram of an image to be repaired provided in an embodiment of this application. Figure 3 This is a schematic diagram of a region mask image provided in an embodiment of this application. Figure 2 As shown, assuming that image 10 is a landscape image, there is a text watermark 11 in the lower right corner of image 10. Image 10 is the image to be repaired, and it needs to be processed to remove the text watermark 11. Figure 3 for Figure 2The region mask image 20 corresponding to image 10 in the image is used to mark the target region 21 that needs to be repaired. This scheme uses a trained neural network model to repair the region 21, thereby removing the watermark text 11. At the same time, it ensures that the original watermark text in image 20 after the removal of watermark text 11 matches the background of image 10.
[0035] Step S102: Input the image to be repaired and the region mask image into the trained neural network model. The neural network model includes an encoder module and a decoder module. During the training of the neural network model, constraints are imposed based on the real image and the set reconstruction loss function.
[0036] After obtaining the image to be repaired and the corresponding region mask image, they are input into a trained neural network model. Optionally, this neural network model includes an encoder module and a decoder module, and adopts an encoder-decoder network architecture as the generative model, that is, the generator includes an encoder module and a decoder module.
[0037] Optionally, during the training of this neural network model, constraints are applied based on real images and a set reconstruction loss function. The constraints imposed during model training by the set reconstruction loss function enable the generative model to learn the ability to repair images through training. In one embodiment, the neural network model training uses a reconstruction loss function based on real images. This reconstruction loss function calculates the similarity between the repaired image output by the generative model and the real image. The real image is the actual image that should correspond to the image to be repaired after correction, as set during training; that is, the image before the content in the target region to be removed was added. Figure 4 As shown, Figure 4 This is a schematic diagram of a real image provided in an embodiment of this application. As shown in the figure, the lower right corner of image 30 does not contain the target area that needs to be repaired. Taking the target area as the area containing the watermark as an example, it is the image before the watermark is added.
[0038] Step S103: Obtain the semantic encoding information and style encoding information corresponding to the image to be repaired and the region mask image through the encoder module.
[0039] In one embodiment, after the image to be repaired and the corresponding region mask image are input into a trained neural network model, the encoder module of the neural network model outputs corresponding semantic encoding information and style encoding information. Optionally, the encoder module can be configured using a residual module, such as... Figure 5 As shown, Figure 5 This is a schematic diagram of an encoder module provided in an embodiment of this application. The residual module receives input x, passes it through a convolutional layer, and adds it to the input x. The residual module learns the residual f(x) between the input and output. The residual module can directly propagate gradients through the residual path, effectively preventing gradient vanishing and gradient exploding, and reducing the difficulty of model learning. Taking the image to be repaired and the corresponding region mask image as the input of the encoder module as an example, the output is the corresponding semantic encoding information and style encoding information.
[0040] Step S104: Output the semantic encoding information and the repaired image corresponding to the style encoding information through the decoder module.
[0041] In one embodiment, after obtaining the semantic and style encoding information output by the encoder module, the decoder module of the neural network model outputs the restored image corresponding to the semantic and style encoding information. Optionally, the decoder module uses a skip connection approach to fuse the features of each resolution with the corresponding resolution features in the encoder to better preserve the information of the image to be restored, thus outputting the final restored image. Optionally, the decoder module uses the modulation module in the StyleGAN2 model as the base module, and uses `style` to inject style into the weights `w` in the convolutional layer. Optionally, the modulation module can be represented by the following formula:
[0042] w′ ijk =s i ·w ijk
[0043]
[0044] Where w i ′ jk For the convolution of the convolutional layer, s i This is the i-th channel element for injecting style. i is the channel coordinate, j is the horizontal coordinate, k is the vertical coordinate, and ∈ is a preset constant to prevent errors where the denominator is 0.
[0045] In one embodiment, the training process of the neural network model is further given, such as... Figure 6 As shown, Figure 6 A flowchart illustrating the training process of a neural network model provided in this application embodiment specifically includes:
[0046] Step S201: Input the first image and the corresponding region mask image into the neural network model to obtain the corresponding first encoding information and first style information.
[0047] Wherein, the first image and the corresponding region mask image are sample images set during the training of the neural network model. The first image is the image that needs to be repaired, and the region mask image is the corresponding image with the target region that needs to be repaired set. Optionally, after the first image and the corresponding region mask image are input into the neural network model, the encoder module of the neural network model outputs first encoding information and first style information, exemplarily denoted as code and style.
[0048] Step S202: Input the real image corresponding to the first image into the trained reconstruction model to obtain the corresponding second encoding information and second style information.
[0049] In one embodiment, during the training of the neural network model, the real image corresponding to the first image is input into the trained reconstruction model to obtain the corresponding second encoding information and second style information. Optionally, the reconstruction model has the same structure as the neural network model set during training, and the real image is the original real image corresponding to the first image that does not contain the repaired region. Optionally, the reconstruction model also outputs the second encoding information and second style information through its encoder module, exemplarily denoted as code rec and style rec. By inputting the real image into the trained reconstruction model, code rec and style rec are obtained. These two features almost perfectly encode the real image, possessing the correct encoding of the real image that the image to be repaired should be repaired into.
[0050] Step S203: Calculate the distillation loss based on the first encoding information, the first style information, the second encoding information, and the second style information, and update the parameters of the neural network model based on the distillation loss.
[0051] In one embodiment, during the training of the neural network model, after obtaining the first encoding information, the first style information, the second encoding information, and the second style information, the parameters of the neural network model are updated by calculating the distillation loss. Optionally, the formula for the distillation loss function is:
[0052] L distill =(code-code rec) 2 +(style-style rec) 2
[0053] Among them, L distillThe distillation loss is represented by `code` (first encoded information), `coderec` (second encoded information), `style` (first style information), and `stylerec` (second style information). In this embodiment, unlike existing methods that use image reconstruction loss functions via backpropagation to indirectly enable the target generation model to learn intermediate features to make the generated image more similar to the real image, this method is more challenging to train. In this embodiment, the teacher network used for distillation is a reconstruction model. Its input is a real image, and its goal is to obtain correct features close to the real image to directly constrain the generation model, thereby making it easier for the model to learn the correct features for image restoration.
[0054] In one embodiment, the process includes training a reconstruction model. Before inputting the real image corresponding to the first image into the trained reconstruction model, the method further includes training the reconstruction model based on the set real image to obtain a trained reconstruction model. Because the task is simple, this reconstruction model can be trained very well, meaning that the intermediate features of the reconstruction model can almost perfectly reconstruct the image to be repaired. During training, the image to be repaired is input into the training neural network model, while the undamaged real image from the image to be repaired is input into the reconstruction model. At this point, two types of features are obtained: features from the neural network model and features from the reconstruction model. The intermediate features of the reconstruction model are then used to distill the intermediate features of the neural network model, thereby achieving direct supervision at the feature level and significantly reducing the learning difficulty of the generative network.
[0055] In one embodiment, during the training of the neural network model, three functions are used as reconstruction loss functions. Besides the aforementioned distillation loss function, the L1 loss function and the perceptual loss function are also included. Specifically, the training process further includes: inputting a first image and its corresponding region mask image into the neural network model to obtain a corresponding second image; obtaining the real image corresponding to the first image; calculating the L1 loss and perceptual loss of the second image and the real image respectively using the set L1 loss function and perceptual loss function; and updating the parameters of the neural network model based on the L1 loss and perceptual loss.
[0056] Optionally, the formula for the L1 loss function can be exemplarily written as:
[0057] L1(y,y ′ )=|yy ′ |
[0058] Where y, y ′ The images generated are the real image and the output image of the neural network model, which is also the second image. The gradient of the L1 loss function is constant, and it has a stable gradient for any input, thus having a robust solution.
[0059] Optionally, the perceptual loss function uses a VGG neural network on y,y ′ Features are extracted separately, and the L2 loss function is used to calculate the loss function. An example formula for calculating the perceptual loss function is as follows:
[0060] L perceptual (y,y ′ )=(VGG(y)-VGG(y ′ )) 2
[0061] Where y,y ′ These are the generated real image and the output image of the neural network model, which is also the second image.
[0062] In one embodiment, the training process of the neural network model further includes a training process based on a set adversarial loss function. Optionally, this includes: calculating the difference between the second image and the real image using the set adversarial loss function, and updating the parameters of the neural network model based on the calculation result. Here, the second image is the repaired image output by the neural network model during training, using the image to be repaired as input.
[0063] Optionally, the adversarial loss function can be the least squares adversarial loss function, LSGAN Loss. Through the adversarial loss function, the discriminator learns the difference between the real image and the second image. During training, the neural network model needs to output images that are as realistic as possible, making it impossible for the discriminator to distinguish the difference between the real image and the second image. The formula for LSGAN is as follows:
[0064] Loss_G=(D(G(x))-1) 2
[0065] Loss_D=(D(y)-1) 2 +D(G(x)) 2
[0066] Where D is the discriminator, G is the neural network model, and x and y are the input image to be repaired and the corresponding real image, respectively.
[0067] In one embodiment, during the training process of the above-mentioned reconstruction model, a dataset with as many samples as possible and as much diversity as possible is used to train the reconstruction model.
[0068] In one embodiment, during the training process of the neural network model described above, after the loss function converges, the weights of the neural network model are selected based on the loss value of the validation set to obtain the trained neural network model, which is then used to output the corresponding repaired image after inputting the image to be repaired and the corresponding region mask image.
[0069] As described above, by acquiring the image to be repaired and its corresponding region mask image, these images are input into a trained neural network model. The neural network model includes an encoder module and a decoder module. During training, constraints are applied based on the real image and a set reconstruction loss function. The encoder module obtains the semantic and style encoding information corresponding to the image to be repaired and the region mask image. The decoder module outputs the repaired image corresponding to the semantic and style encoding information. In this scheme, the constraints applied to the neural network model training based on the real image and the set reconstruction loss function result in better repaired image performance during model use, while also accelerating inference speed during model training.
[0070] Figure 7 This is a structural block diagram of an image restoration apparatus provided in an embodiment of this application. The apparatus is used to execute the image restoration method provided in the above embodiments, and has corresponding functional modules and beneficial effects for executing the method. Figure 7 As shown, the device specifically includes: an image acquisition module 101, an image input module 102, and an image output module 103, wherein,
[0071] Image acquisition module 101 is configured to acquire the image to be repaired and the region mask image corresponding to the image to be repaired;
[0072] The image input module 102 is configured to input the image to be repaired and the region mask image into a trained neural network model. The neural network model includes an encoder module and a decoder module. During the training of the neural network model, constraints are imposed based on the real image and the set reconstruction loss function.
[0073] The image output module 103 is configured to obtain semantic encoding information and style encoding information corresponding to the image to be repaired and the region mask image through the encoder module, and to output the repaired image corresponding to the semantic encoding information and the style encoding information through the decoder module.
[0074] As described above, the method involves acquiring the image to be repaired and its corresponding region mask image, then inputting these images into a trained neural network model. The neural network model includes an encoder module and a decoder module. During training, constraints are applied based on the real image and a set reconstruction loss function. The encoder module obtains the semantic and style encoding information corresponding to the image to be repaired and the region mask image. The decoder module outputs the repaired image corresponding to the semantic and style encoding information. This method, by applying constraints based on the real image and the set reconstruction loss function during neural network model training, results in better repaired image performance during model use and accelerates inference speed during training.
[0075] In one possible embodiment, the device further includes a model training module configured as follows:
[0076] The first image and the corresponding region mask image are input into the neural network model to obtain the corresponding first encoding information and first style information.
[0077] The real image corresponding to the first image is input into the trained reconstruction model to obtain the corresponding second encoding information and second style information;
[0078] The distillation loss is calculated based on the first encoding information, the first style information, the second encoding information, and the second style information, and the parameters of the neural network model are updated based on the distillation loss.
[0079] In one possible embodiment, the reconstructed model and the neural network model have the same structure, and the model training module is further configured with:
[0080] Before inputting the real image corresponding to the first image into the trained reconstruction model, the reconstruction model is trained based on the set real image to obtain the trained reconstruction model.
[0081] In one possible embodiment, the model training module is configured as follows:
[0082] According to formula L distill =(code-code rec) 2 +(style-style rec) 2 Calculate the distillation loss, where L distill This represents the distillation loss, where code is the first encoding information, code rec is the second encoding information, style is the first style information, and style rec is the second style information.
[0083] In one possible embodiment, the model training module is configured as follows:
[0084] The first image and the corresponding region mask image are input into the neural network model to obtain the corresponding second image;
[0085] Obtain the real image corresponding to the first image, calculate the L1 loss and perceptual loss of the second image and the real image respectively using the set L1 loss function and perceptual loss function, and update the parameters of the neural network model based on the L1 loss and the perceptual loss.
[0086] In one possible embodiment, the model training module is configured as follows:
[0087] The difference between the second image and the real image is calculated using a set adversarial loss function;
[0088] The parameters of the neural network model are updated based on the calculation results.
[0089] Figure 8 This is a schematic diagram of the structure of an image restoration device provided in an embodiment of this application, as shown below. Figure 8 As shown, the device includes a processor 201, a memory 202, an input device 203, and an output device 204; the number of processors 201 in the device can be one or more. Figure 8 Taking a processor 201 as an example; the processor 201, memory 202, input device 203, and output device 204 in the device can be connected via a bus or other means. Figure 8 Taking a bus connection as an example, the memory 202, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the image restoration method in this embodiment. The processor 201 executes various functional applications and data processing of the device by running the software programs, instructions, and modules stored in the memory 302, thereby realizing the above-mentioned image restoration method. The input device 203 can be used to receive input digital or character information and generate key signal inputs related to user settings and function control of the device. The output device 204 may include a display device such as a screen.
[0090] This application also provides a non-volatile storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform an image restoration method described in the above embodiments, wherein the method includes:
[0091] Obtain the image to be repaired and the region mask image corresponding to the image to be repaired;
[0092] The image to be repaired and the region mask image are input into a trained neural network model, which includes an encoder module and a decoder module. During the training of the neural network model, constraints are imposed based on the real image and the set reconstruction loss function.
[0093] The encoder module obtains semantic and style encoding information corresponding to the image to be repaired and the region mask image.
[0094] The decoder module outputs the semantic encoding information and the corresponding repaired image based on the style encoding information.
[0095] It is worth noting that in the above-described embodiments of the image restoration device, the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy differentiation and are not used to limit the protection scope of the embodiments of this application.
[0096] In some possible implementations, various aspects of the methods provided in this application can also be implemented as a program product comprising program code that, when run on a computer device, causes the computer device to perform the steps of the methods according to the various exemplary embodiments of this application described above. For example, the computer device can perform the image restoration method described in the embodiments of this application. The program product can be implemented using any combination of one or more readable media.
Claims
1. An image restoration method, characterized in that, include: Obtain the image to be repaired and the region mask image corresponding to the image to be repaired; The image to be repaired and the region mask image are input into a trained neural network model. The neural network model includes an encoder module and a decoder module. During the training of the neural network model, constraints are imposed based on the real image and a set reconstruction loss function. The training process of the neural network model includes: inputting the first image and the corresponding region mask image into the neural network model to obtain the corresponding first encoding information and first style information; inputting the real image corresponding to the first image into the trained reconstruction model to obtain the corresponding second encoding information and second style information; calculating the distillation loss based on the first encoding information, the first style information, the second encoding information, and the second style information; and updating the parameters of the neural network model based on the distillation loss. The encoder module obtains semantic and style encoding information corresponding to the image to be repaired and the region mask image. The decoder module outputs the semantic encoding information and the corresponding repaired image based on the style encoding information.
2. The image restoration method according to claim 1, characterized in that, The reconstruction model and the neural network model have the same structure. Before inputting the real image corresponding to the first image into the trained reconstruction model, the method further includes: The reconstruction model is trained based on the set real images to obtain the trained reconstruction model.
3. The image restoration method according to claim 1, characterized in that, The step of calculating distillation loss based on the first encoding information, the first style information, the second encoding information, and the second style information includes: According to the formula Calculate the distillation loss, where, Indicates distillation loss, For the first encoded information, For the second encoded information, For primary style information, This is information for the second style.
4. The image restoration method according to any one of claims 1-3, characterized in that, The training process of the neural network model includes: The first image and the corresponding region mask image are input into the neural network model to obtain the corresponding second image; Obtain the real image corresponding to the first image, calculate the L1 loss and perceptual loss of the second image and the real image respectively using the set L1 loss function and perceptual loss function, and update the parameters of the neural network model based on the L1 loss and the perceptual loss.
5. The image restoration method according to claim 4, characterized in that, The training process of the neural network model also includes: The difference between the second image and the real image is calculated using a set adversarial loss function; The parameters of the neural network model are updated based on the calculation results.
6. An image restoration device, characterized in that, include: The image acquisition module is configured to acquire the image to be repaired and the region mask image corresponding to the image to be repaired; The image input module is configured to input the image to be repaired and the region mask image into a trained neural network model. The neural network model includes an encoder module and a decoder module. During the training of the neural network model, constraints are imposed based on the real image and a set reconstruction loss function. The model training module is configured to input a first image and a corresponding region mask image into a neural network model to obtain corresponding first encoding information and first style information; input the real image corresponding to the first image into the trained reconstruction model to obtain corresponding second encoding information and second style information; calculate distillation loss based on the first encoding information, the first style information, the second encoding information, and the second style information; and update the parameters of the neural network model based on the distillation loss. The image output module is configured to obtain semantic encoding information and style encoding information corresponding to the image to be repaired and the region mask image through the encoder module, and to output the repaired image corresponding to the semantic encoding information and the style encoding information through the decoder module.
7. An image restoration device, the device comprising: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the image restoration method according to any one of claims 1-5.
8. A non-volatile storage medium storing computer-executable instructions, which, when executed by a computer processor, are used to perform the image restoration method according to any one of claims 1-5.
9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the image restoration method according to any one of claims 1-5.
Citation Information
Patent Citations
Generative image restoration method based on attention cross-layer transfer mechanism
CN114266711A