Image super-resolution reconstruction method based on cascade feature distillation network

Abstract

The invention provides an image super-resolution reconstruction method based on a cascade feature distillation network, and relates to the technical field of image super-resolution reconstruction, and the method comprises the steps: carrying out the convolution operation of a low-resolution image, and obtaining an initial feature; a symmetric cascade feature distillation module is adopted as a network backbone, nonlinear mapping is performed on extracted initial features, and effective representation of deep features is learned; splicing the learned deep features with the initial features through global jump connection, and constructing a global information dependency relationship; a high resolution image is generated using convolutional layers and sub-pixel convolution operations based on the global information dependencies. The network model constructed by adopting the scheme has moderate parameter quantity and calculation quantity, and can meet the application requirement of deploying the model on the edge equipment.

Description

Technical Field

[0001] The present application relates to the technical field of image super-resolution reconstruction, and particularly to an image super-resolution reconstruction method and device based on a cascaded feature distillation network. Background Art

[0002] In daily life, images can convey information visually and are a powerful medium for transmitting information. However, due to factors such as image acquisition and weather, the obtained images often suffer from the loss of details and edge information, and even appear blurred. For example, in the application scenario of drone aerial photography, if the image is blurred, it may lead to incorrect identification of obstacles or traffic signs. Therefore, improving image quality has become an urgent problem to be solved in current technological development and practical applications.

[0003] To improve the quality of images, on the one hand, image sensors with higher pixels can be used to capture delicate and rich image details. However, due to hardware limitations, the obtained image resolution is not high. In addition, devices such as high-resolution sensors are costly. In this case, the image super-resolution reconstruction technology has unique advantages. This technology can obtain high-quality images without increasing hardware costs, with higher flexibility, lower economic costs, and is easy to update and upgrade.

[0004] With the development of deep learning technology, convolutional neural networks can recover high-resolution details from low-resolution images, thereby improving image quality. However, due to the complexity of deep learning models themselves, especially the problems of large computational volume and high storage cost, it is difficult to practically apply advanced super-resolution deep learning models on resource-constrained edge devices such as smartphones and drones. Summary of the Invention

[0005] The present application aims to at least solve one of the technical problems in the related art to some extent.

[0006] To this end, the first object of the present application is to propose an image super-resolution reconstruction method based on a cascaded feature distillation network. Aiming at the technical problem that the existing convolutional neural network model has a large number of parameters and a large computational volume, resulting in difficult model deployment, an image super-resolution reconstruction model is proposed, which has a moderate number of parameters and a moderate computational volume, and can meet the application requirements of deploying the model on edge devices.

[0007] The second object of the present application is to propose a computer device.

[0008] To achieve the above object, the first aspect embodiment of the present application proposes an image super-resolution reconstruction method based on a cascaded feature distillation network, including:

[0009] Perform a convolution operation on the low-resolution image to obtain initial features;

[0010] Use a symmetric cascaded feature distillation module as the network backbone to perform a non-linear mapping on the extracted initial features and learn an effective representation of the deep features;

[0011] Concatenate the learned deep features with the initial features through global skip connections to construct a global information dependency relationship;

[0012] Generate a high-resolution image based on the global information dependency relationship using a convolutional layer and a sub-pixel convolution operation.

[0013] Optionally, in an embodiment of the present application, performing a convolution operation on the low-resolution image to obtain initial features, expressed as:

[0014] F 0 = Conv 3×3 (I LR )

[0015] where I LR is the low-resolution image, and Conv 3×3 (·) is a 3×3 convolution operation;

[0016] The learned deep features are:

[0017] F = H SCFDM (F 0 )

[0018] where SCFDM is a symmetric cascaded feature distillation module;

[0019] Concatenate the learned deep features with the initial features through global skip connections to construct a global information dependency relationship, expressed as:

[0020] F fusion = F 0 + F

[0021] Generate a high-resolution image using a 3×3 convolutional layer and a sub-pixel convolution operation, expressed as:

[0022] I SR = H Pixel (Conv 3×3 (F fusion ))

[0023] where H Pixel (·) represents a pixel reconstruction operation.

[0024] Optionally, in an embodiment of the present application, the symmetric cascaded feature distillation module includes a group of feature distillation modules. The extracted initial features are input, and through a group of feature distillation modules, the learned deep features are output by the information distillation branch of each feature distillation module.

[0025] Optionally, in an embodiment of the present application, the feature distillation module includes a first twin feature extraction module, a second twin feature extraction module, and a third twin feature extraction module. The information distillation branch of the feature distillation module includes a first pixel recalibration attention mechanism, a second pixel recalibration attention mechanism, a third pixel recalibration attention mechanism, and a cascaded feature aggregation module. Among them, the third twin feature extraction module of the first feature distillation module is connected to the first twin feature extraction module of the second feature distillation module. Inside each feature distillation module, the twin feature extraction modules are connected through a Split module. The first Split module connects the first and second twin feature extraction modules and the first pixel recalibration attention mechanism. The second Split module connects the second and third twin feature extraction modules and the second pixel recalibration attention mechanism. The third twin feature extraction module is connected to the third pixel recalibration attention mechanism. The cascaded feature aggregation module connects all pixel recalibration attention mechanisms.

[0026] Optionally, in an embodiment of the present application, the twin feature extraction module includes a twin enhanced residual weighted module GERWB and a twin convolutional residual weighted module GCRWB. After the feature F passes through GERWB and GCRWB, the output feature F GCRW is:

[0027] F GCRW′ = λ x1 H ESA [H CR (F)] + λ res1 F

[0028] F GCRW = λ x2 H ESA [H CR (F GCRW′ )] + λ res2 Conv 3×3 (F GCRW′ )

[0029] Among them, λ x1 , λ res1 are the adaptive weights on different path branches in GERWB. H ESA [·] represents the enhanced spatial attention operation. H CR (·) represents the convolution and ReLU activation function operations before the attention mechanism. λ x2 , λ res2is the adaptive weight on different path branches in GCRWB.

[0030] Optionally, in an embodiment of the present application, after the previous twin feature extraction module in the connection, the Split module uses an information distillation mechanism to separate the channel features. The first part is transmitted along the backbone network to the next twin feature extraction module in the connection, and the second part is input into the connected pixel recalibration attention mechanism as the distilled branch for pixel recalibration, which is expressed as:

[0031] {F remaink ,F distillk} = Split k (F GCRW )

[0032] where Split k (·) represents the channel separation operation performed by the k-th Split module, and F remaink , F distillk are the first part and the second part respectively.

[0033] Optionally, in an embodiment of the present application, the pixel recalibration attention mechanism is expressed as:

[0034]

[0035] where Conv 1×1 (·) represents a 1×1 convolution operation, σ[·] is the ReLU activation function, δ{·} is the sigmoid function, represents element-wise multiplication.

[0036] Optionally, in an embodiment of the present application, the cascaded feature aggregation module integrates the features output by all pixel recalibration mechanisms through concatenation and convolution operations to obtain deep features, which is expressed as:

[0037] F = Conv 1×1 (Concat(F PRA1 , F PRA2 , F PRA3 ))

[0038] where Concat(·) represents the concatenation operation.

[0039] Optionally, in an embodiment of the present application, given a training set containing N pairs of low-resolution images and their corresponding high-resolution images I LR is the low-resolution image, I HR is the original high-resolution image, and the loss function L 1 is used to optimize the network model, and the optimization process is expressed as:

[0040]

[0041] Among them, I SR is the high-resolution image output by the model, and θ represents the set of model parameters.

[0042] To achieve the above object, an embodiment of the second aspect of the present invention proposes a computer device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the above image super-resolution reconstruction method based on the cascaded feature distillation network is implemented.

[0043] The image super-resolution reconstruction method based on the cascaded feature distillation network in the embodiment of the present application has symmetry in the network backbone, combines the information distillation mechanism and the cascaded extraction method to enhance the information expression ability, enhances the feature interaction and reduces the model parameter quantity through the channel separation operation and the skip connection method, so that the model can restore the high-resolution effect while minimizing the unnecessary computational burden as much as possible; optimizes the local feature extraction ability, introduces the twin enhanced residual weighted module and the twin convolutional residual weighted module, which can effectively improve the reuse rate of local information in the network and accelerate the convergence speed of the network; designs a new pixel recalibration attention mechanism, which can weight important pixels on the information distillation branch and dynamically adjust the weight according to the image content, ensuring that the important area can be more accurately focused on during image reconstruction, and realizing the optimized integration of features.

[0044] Additional aspects and advantages of the present application will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The above and / or additional aspects and advantages of the present application will become apparent and be readily understood from the following description of the embodiments in conjunction with the drawings, wherein:

[0046] Figure 1 is a schematic flowchart of an image super-resolution reconstruction method based on a cascaded feature distillation network provided by Embodiment 1 of the present application;

[0047] Figure 2 is the network structure diagram of the embodiment of the present application;

[0048] Figure 3 is the structure diagram of GERWB of the embodiment of the present application;

[0049] Figure 4 is the structure diagram of GCRWB of the embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] Embodiments of the present application will be described in detail below. Examples of the embodiments are shown in the accompanying drawings, where like or similar reference numerals denote like or similar elements or elements having like or similar functions throughout. The embodiments described below by referring to the accompanying drawings are exemplary and are intended to explain the present application, and should not be construed as limiting the present application.

[0051] The following describes a method and apparatus for image super-resolution reconstruction based on a cascaded feature distillation network according to an embodiment of the present application with reference to the accompanying drawings.

[0052] Figure 1 It is a schematic flowchart of a method for image super-resolution reconstruction based on a cascaded feature distillation network provided by Embodiment 1 of the present application.

[0053] As Figure 1 shown, the method for image super-resolution reconstruction based on a cascaded feature distillation network includes the following steps:

[0054] Step 101: Perform a convolution operation on the low-resolution image to obtain initial features;

[0055] Step 102: Use a symmetric cascaded feature distillation module as the network backbone to perform a non-linear mapping on the extracted initial features to learn an effective representation of deep features;

[0056] Step 103: Concatenate the learned deep features with the initial features through a global skip connection to construct a global information dependence relationship;

[0057] Step 104: Based on the global information dependence relationship, use a convolutional layer and a sub-pixel convolutional operation to generate a high-resolution image.

[0058] In the method for image super-resolution reconstruction based on a cascaded feature distillation network according to an embodiment of the present application, the network backbone has symmetry, combines an information distillation mechanism and a cascaded extraction method to enhance the information expression ability, enhances feature interaction and reduces the model parameter quantity through a channel separation operation and a skip connection method, so that while the model can restore the high-resolution effect, it can minimize unnecessary computational burden as much as possible; optimizes the local feature extraction ability, introduces a twin enhanced residual weighting module and a twin convolutional residual weighting module, can effectively improve the reuse rate of local information in the network, and speeds up the convergence rate of the network; designs a new pixel recalibration attention mechanism, can weight important pixels on the information distillation branch, and dynamically adjusts the weight according to the image content, ensuring that it can more accurately focus on important regions during image reconstruction and realizing the optimal integration of features.

[0059] Optionally, in an embodiment of the present application, the network structure is as Figure 2 shown, where I LR 、I SRrespectively represent the input low-resolution (LR) image and the reconstructed super-resolution (SR) image.

[0060] First, perform a 3×3 convolution operation on the low-resolution image to obtain the initial feature F 0 , and this process can be expressed as:

[0061] F 0 = Conv 3×3 (I LR )

[0062] where I LR is the low-resolution image, and Conv 3×3 (·) is the 3×3 convolution operation;

[0063] Then, send the extracted initial feature into the symmetrical cascaded feature distillation module (SCFDM), which performs a non-linear mapping to obtain an effective representation of the deep feature. This process can be expressed as:

[0064] F = H SCFDM (F 0 )

[0065] where SCFDM is the symmetrical cascaded feature distillation module;

[0066] The learned deep feature is concatenated with the initial feature through a global skip connection to construct a global information dependency relationship. This process can be expressed as:

[0067] F fusion = F 0 + F

[0068] Finally, use a 3×3 convolutional layer and a sub-pixel convolution operation to generate a high-resolution image. This process can be expressed as:

[0069] I SR = H Pixel (Conv 3×3 (F fusion ))

[0070] where H Pixel (·) represents the pixel reconstruction operation.

[0071] Optionally, in an embodiment of the present application, as Figure 2As shown, the symmetric cascaded feature distillation module includes a set of feature distillation modules. The initially extracted features are input, and through the set of feature distillation modules, the learned deep features are output by the information distillation branches of each feature distillation module.

[0072] Optionally, in an embodiment of the present application, as Figure 2 shown, the feature distillation module includes a first twin feature extraction module, a second twin feature extraction module, and a third twin feature extraction module. The information distillation branch of the feature distillation module includes a first pixel recalibration attention mechanism, a second pixel recalibration attention mechanism, a third pixel recalibration attention mechanism, and a cascaded feature aggregation module. Among them, the third twin feature extraction module of the first feature distillation module is connected to the first twin feature extraction module of the second feature distillation module. Inside each feature distillation module, the twin feature extraction modules are connected through a Split module. The first Split module connects the first and second twin feature extraction modules and the first pixel recalibration attention mechanism. The second Split module connects the second and third twin feature extraction modules and the second pixel recalibration attention mechanism. The third twin feature extraction module is connected to the third pixel recalibration attention mechanism. The cascaded feature aggregation module connects all the pixel recalibration attention mechanisms.

[0073] Optionally, in an embodiment of the present application, the twin feature enhancement module includes a Gemel Enhanced Residual Weighting Block (GERWB) and a Gemel Convolutional Residual Weighting Block (GCRWB). Both jointly perform local feature extraction and introduce an adaptive weight feature operation; aiming to achieve better performance while maintaining fewer parameters and computational loads.

[0074] Figure 3 is the structural diagram of GERWB, Figure 4 is the structural diagram of GCRWB, as Figure 3 、 Figure 4 shown. In order to effectively increase the proportion of important features in GERWB and GCRWB, this embodiment combines it with an Enhanced Spatial Attention (ESA) mechanism without introducing any additional parameters. Specifically, a 3×3 convolutional layer and a ReLU activation function are used before ESA, and at the same time, a basic module with an adaptive weight operation is combined to keep the model lightweight and have a multi-scale feature extraction function. Compared with GERWB, GCRWB adds a 3×3 convolutional layer on the skip connection branch, which can increase the number of output channels to match the original input size and realize the association of different channel dimensions.

[0075] As shown in Figure 2 、 3 、4, assuming that the initial feature f 0 is input into GERWB and GCRWB, the output F GCRW and F GCRW can be expressed as:

[0076] F GCRW = λ x1 H ESA [H CR (F 0 )]+λ res1 F 0

[0077] F GCRW = λ x2 H ESA [H CR (F 0 )]+λ res2 Conv 3×3 (F 0 )

[0078] Among them, λ xl , λ resl respectively represent the adaptive weight operations on the k-th path branch, l = 1, 2, H ESA [·] represents the enhanced spatial attention (ESA) operation, while H CR (·) represents the convolution and ReLU activation function operations before the attention mechanism.

[0079] Optionally, in an embodiment of the present application, after the twin feature extraction module, an information distillation mechanism is used to separate the channel features. In this way, the feature information is divided into two parts: one part continues to be transmitted along the backbone network, and the other part is used as the distilled branch for pixel recalibration. This process can be expressed as:

[0080] {F remaink , F distillk} = Split k (F GCRW )

[0081] Among them, Split k (·) represents the k-th channel separation operation, and F remaink , F distillk respectively represent the backbone branch and the distilled branch retained after each separation.

[0082] Optionally, in an embodiment of the present application, the present invention proposes a Pixel-Recalibrated Attention (PRA) mechanism, which includes a 1×1 convolutional layer, a ReLU activation function, and a sigmoid function. The PRA structure is simple and has a very strong non-linear reconstruction ability. Its formula is expressed as:

[0083]

[0084] Among them, Conv 1×1 (·) represents a 1×1 convolutional operation, σ[·] is the ReLU activation function, δ{·} is the sigmoid function, represents element-wise multiplication.

[0085] Optionally, in an embodiment of the present application, the cascaded feature aggregation module integrates multiple sets of efficient and accurate features extracted by the pixel recalibration mechanism to obtain deep features. This process can be expressed as:

[0086] F = Conv 1×1 (Concat(F PRA1 , F PRA2 , F PRA3 ))

[0087] Among them, Concat(·) represents the concatenation operation, and F is the output result of the cascaded feature aggregation block. {F PRA1 , F PRA2 , F PRA3} respectively represent the output results of multiple sets of extracted features.

[0088] Optionally, in an embodiment of the present application, given a training set containing N pairs of low-resolution images and their corresponding high-resolution images I HR is the original high-resolution (HR) image. The loss function L 1 is used to optimize the network model. The optimization process is expressed as:

[0089]

[0090] Among them, θ represents the set of model parameters.

[0091] To implement the above embodiments, the present invention also proposes a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the methods described in the above embodiments are implemented.

[0092] In the description of this specification, the descriptions referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of this application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, without contradiction, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0093] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include at least one of such features. In the description of this application, "a plurality of" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0094] Any process or method description shown in a flowchart or described in other ways herein can be understood to represent a module, segment, or portion of code including one or more executable instructions for implementing a customized logic function or process, and the scope of the preferred embodiments of this application includes additional implementations, where the functions may be executed in a substantially simultaneous manner or in a reverse order according to the functions involved, rather than in the order shown or discussed, which should be understood by those skilled in the art to which the embodiments of this application belong.

[0095] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a definable sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device), or used in combination with these instruction execution systems, apparatus, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. More specific examples (non-exhaustive list) of computer-readable media include the following: an electrical connection portion having one or more wirings (electronic device), a portable computer diskette (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or otherwise processing as appropriate, and then storing it in a computer memory.

[0096] It should be understood that various parts of the present application can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0097] Those of ordinary skill in the art of this technology can understand that all or part of the steps carried by the methods of the above embodiments can be completed by a program instructing relevant hardware, and the program can be stored in a computer-readable storage medium. When the program is executed, it includes one or a combination of the steps of the method embodiments.

[0098] In addition, each functional unit in various embodiments of the present application may be integrated into one processing module, may exist separately as individual physical units, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. When the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.

[0099] The above-mentioned storage medium may be a read-only memory, a magnetic disk, an optical disc, etc. Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present application. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present application.

Claims

1. An image super-resolution reconstruction method based on a cascaded feature distillation network, characterized in that: include: Perform convolution operation on the low-resolution image to obtain initial features; A symmetrical cascade feature distillation module is used as the network backbone to perform nonlinear mapping on the extracted initial features and learn effective representations of deep features. The learned deep features are concatenated with the initial features through global skip connections to build global information dependencies; Based on the global information dependency, a high-resolution image is generated using a convolutional layer and a sub-pixel convolution operation.

2. The method according to claim 1, characterized in that Perform convolution operation on the low-resolution image to obtain the initial features, which are expressed as: F0=Conv 3×3 (I LR ) Among them, I LR For low-resolution images, Conv 3×3 (·) is a 3×3 convolution operation; The deep features learned are: F=H SCFDM (F0) Among them, SCFDM is a symmetrical cascade feature distillation module; The learned deep features are concatenated with the initial features through global skip connections to construct global information dependencies, which can be expressed as: F fusion =F0+F A 3×3 convolution layer and a sub-pixel convolution operation are used to generate a high-resolution image, expressed as: I SR =H Pixel (Conv 3×3 (F fusion )) Among them, H Pixel (·) denotes a pixel reconstruction operation.

3. The method according to claim 2, characterized in that The symmetrical cascade feature distillation module includes a group of feature distillation modules, which input the extracted initial features, pass through the group of feature distillation modules, and output the learned deep features by the information distillation branch of each feature distillation module.

4. The method according to claim 3, characterized in that The feature distillation module includes a first twin feature extraction module, a second twin feature extraction module, and a third twin feature extraction module. The information distillation branch of the feature distillation module includes a first pixel recalibration attention mechanism, a second pixel recalibration attention mechanism, a third pixel recalibration attention mechanism, and a cascade feature aggregation module. The third twin feature extraction module of the first feature distillation module is connected to the first twin feature extraction module of the second feature distillation module. In each feature distillation module, the twin feature extraction modules are connected through a Split module. The first Split module connects the first and second twin feature extraction modules and the first pixel recalibration attention mechanism. The second Split module connects the second and third twin feature extraction modules and the second pixel recalibration attention mechanism. The third twin feature extraction module is connected to the third pixel recalibration attention mechanism. The cascade feature aggregation module connects all pixel recalibration attention mechanisms.

5. The method according to claim 4, characterized in that The twin feature extraction module includes the twin enhanced residual weighting module GERWB and the twin convolution residual weighting module GCRWB. The feature F is passed through GERWB and GCRWB, and the output feature F GCRW for: F GCRW ′=λ x1 H ESA [H CR (F)]+λ res1 F F GCRW =λ x2 H ESA [H CR (F GCRW′ )]+λ res2 Conv 3×3 (F GCRW′ ) Among them, λ x1 , res1 is the adaptive weight on different path branches in GERWB, H ESA [·] represents the enhanced spatial attention operation, H CR (·) represents the convolution and ReLU activation function operations before the attention mechanism, λ x2 , res2 are the adaptive weights on different path branches in GCRWB.

6. The method according to claim 5, characterized in that After the previous twin feature extraction module is connected, the Split module uses the information distillation mechanism to separate the channel features. The first part is transmitted along the backbone network to the next twin feature extraction module connected, and the second part is used as the distilled branch input to the connected pixel recalibration attention mechanism for pixel recalibration, which is expressed as: {F remaink ,F distillk }=Split k (F GCRW ) Among them, Split k (·) represents the channel separation operation performed by the kth Split module, F remaink 、F distillk They are Part One and Part Two respectively.

7. The method according to claim 6, characterized in that The pixel recalibration attention mechanism is expressed as: Among them, Conv 1×1 (·) represents a 1×1 convolution operation, σ[·] is the ReLU activation function, δ{·} is the sigmoid function, Represents element-wise multiplication.

8. The method according to claim 7, characterized in that The cascade feature aggregation module integrates the features output by all pixel recalibration mechanisms through splicing and convolution operations to obtain deep features, which are expressed as: F=Conv 1×1 (Concat(F PRA1 ,F PRA2 ,F PRA3 )) Among them, Concat(·) represents the concatenation operation.

9. The method according to claim 1, characterized in that Given a training set containing N pairs of low-resolution images and their corresponding high-resolution images I LR is a low-resolution image, I HR For the original high-resolution image, the loss function L1 is used to optimize the network model. The optimization process is expressed as: Among them, I SR is the high-resolution image output by the model, and θ represents the parameter set of the model.

10. A computer device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method according to any one of claims 1 to 9 is implemented.

Images