A method for image super-resolution reconstruction based on dual-domain splicing attention mechanism

Abstract

This invention discloses an image super-resolution reconstruction method based on a dual-domain embedded attention mechanism, belonging to the field of image super-resolution reconstruction technology. The invention includes: acquiring and preprocessing images; constructing an image processing model; training the image processing model using the preprocessed images, employing an adaptive attention allocation strategy during training to dynamically adjust local attention weights, using an edge-aware multi-constraint loss function, and dynamically coordinating overall reconstruction with edge quality to improve the detail clarity and contour naturalness of the reconstructed image; and performing image super-resolution reconstruction using the trained image processing model. This invention embeds a ViT module based on a dual-domain embedded attention mechanism into the bottleneck layer of a U-Net network and employs an edge-aware multi-constraint loss function based on a gradient-aligned curriculum evolution strategy. The resulting model exhibits excellent image super-resolution reconstruction performance.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to image super-resolution reconstruction technology, specifically an image super-resolution reconstruction method based on a dual-domain chimeric attention mechanism. Background Technology

[0002] Image super-resolution reconstruction aims to recover richly detailed high-resolution images from low-resolution images, addressing the problem of detail loss due to imaging system limitations. Early methods relied primarily on interpolation and hand-designed prior models, but with limited effectiveness. A significant performance leap was achieved only with the advent of deep learning-based approaches. Current mainstream techniques utilize complex convolutional neural networks or Vision Transformer (ViT) structures to learn complex mapping relationships from low-resolution to high-resolution images end-to-end, generating realistically detailed information.

[0003] ViT's revolutionary idea lies in completely abandoning the local receptive field limitations of traditional convolution. It segments the feature map into fixed-size sequential blocks, projects them linearly, and then inputs them into the standard Transformer architecture. Its core multi-head self-attention mechanism allows the model to break through locality constraints while capturing global dependencies between all image blocks, achieving efficient integration of contextual information. However, this mechanism is essentially a global attention mechanism, which tends to over-focus on low-frequency redundant information in large areas of the background, crowding out modeling resources for high-frequency key details such as edges and contours. This results in problems such as blurred boundary contours and insufficient detail sharpness in the model output. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides an image super-resolution reconstruction method based on a dual-domain embedded attention mechanism. This invention embeds the ViT module based on the dual-domain embedded attention mechanism into the bottleneck layer of the U-Net network and employs an edge-aware multi-constraint loss function based on a gradient-aligned curriculum evolution strategy. The resulting model exhibits excellent image super-resolution reconstruction performance.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] An image super-resolution reconstruction method based on a dual-domain splicing attention mechanism includes the following steps:

[0007] S1. Acquire images and perform preprocessing;

[0008] S2. Construct an image processing model. The image processing model uses a multi-level U-Net as the main network structure. The bottleneck layer of U-Net adopts the ViT module. The ViT module first converts the feature map output by the U-Net encoder into sequence data through image block embedding. Then, it is input into the Transformer encoder with a dual-domain chiral attention mechanism for feature enhancement. Finally, the enhanced sequence data is decoded by the Transformer decoder to restore the feature map and used as the input of the U-Net decoder.

[0009] S3. The image processing model is trained using the preprocessed image. An adaptive attention allocation strategy is used during training to dynamically adjust the relative importance of global attention and local attention. An edge-aware multi-constraint loss function is used to dynamically coordinate the overall reconstruction and edge quality, which significantly improves the detail clarity and contour naturalness of the reconstructed image.

[0010] S4. Input the image to be processed into the trained image processing model to perform image super-resolution reconstruction.

[0011] Beneficial effects:

[0012] (1) A dual-domain chimeric attention mechanism is used in ViT. It adopts a dual-domain branching approach. The global perception domain uses multi-head self-attention to capture the long-range dependence and macroscopic structure of the image as a whole, while the local perception domain uses masked attention to limit the receptive field, so that the network only focuses on the attention within the nearby window and extracts high-frequency details, edges and textures within the window. At the same time, the two branches are combined using an attention transfer scheme based on the degree of reconstruction: when the degree of reconstruction is low, attention is paid to the global aspect, and when the degree of reconstruction is high, attention is paid to the local aspect.

[0013] (2) The present invention adopts a joint loss function - edge-aware multi-constraint loss, which is divided into three parts: basic reconstruction loss, pixel-level edge loss and structural edge loss. The basic reconstruction loss is responsible for ensuring the quality of basic reconstruction. The pixel-level edge loss ensures the correctness of edge intensity distribution from the pixel level. The structural edge loss pays more attention to edge contour information in the reconstructed image and forces the model to prioritize edge structure quality. At the same time, the combination of the three losses adopts a gradient alignment course evolution strategy during training: the basic reconstruction loss has a higher weight in the early stage of training, the pixel-level edge loss has a higher weight in the middle stage, and the structural edge loss has a higher weight in the later stage. Attached Figure Description

[0014] Figure 1 This is a network structure diagram in an embodiment of the present invention;

[0015] Figure 2 This is a schematic diagram of the dual-domain embedded attention mechanism algorithm in an embodiment of the present invention;

[0016] Figure 3 The image to be predicted in the embodiments of the present invention is the real image;

[0017] Figure 4 The image to be predicted in this embodiment of the invention is reconstructed by a model trained with edge-aware multi-constraint loss;

[0018] Figure 5 The image to be predicted in this embodiment of the invention is reconstructed by a model trained with mean squared error loss. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0020] An image super-resolution reconstruction method based on a dual-domain chimeric attention mechanism, such as Figure 1 As shown, it includes the following steps:

[0021] S1. Acquire images and perform preprocessing.

[0022] The images in this embodiment are from the Urban100 dataset. The preprocessing method can be selected as needed. For example, this embodiment uses conventional operations such as data normalization (Min-Max normalization), random rotation, and adding Gaussian noise. These are existing technologies and will not be described in detail here.

[0023] S2. Construct an image processing model, such as Figure 1 As shown, the image processing model uses U-Net as the main network structure. The U-Net encoder employs four cascaded downsampling modules. Each downsampling module sequentially performs convolution, batch normalization, LeakyReLU activation, and max pooling operations to progressively compress the image size and extract multi-scale features. The U-Net bottleneck layer uses the ViT module. The ViT module first converts the feature map output by the U-Net encoder into sequence data through image patch embedding, and then inputs it into the Transformer encoder, which employs a dual-domain chiral attention mechanism, for feature enhancement. Finally, the enhanced sequence data is decoded by the Transformer decoder to restore the feature map, which serves as the input to the U-Net decoder. The U-Net decoder employs four upsampling modules. Each upsampling module performs deconvolution and skip connection operations to progressively restore high-resolution image details.

[0024] Specifically:

[0025] In downsampling, convolution is the core of convolutional neural networks. It uses learnable convolutional kernels to perform sliding window-style local correlation calculations on the input data, effectively extracting spatial or temporal features. In this embodiment, the convolutional kernel size is 3×3. Pooling is used to compress data. In this embodiment, max pooling with a stride of 2 and a window size of 2×2 is used to compress the image.

[0026] In the bottleneck layer, the output of the U-Net encoder does not match the input of the Transformer encoder in the ViT module in terms of dimension. Therefore, it is necessary to convert the feature map output by the U-Net encoder into a sequence vector through image patch embedding, which includes the following steps:

[0027] S211. Slice and serialize the output of the U-Net encoder;

[0028] The purpose of slicing and partitioning is to transform low-resolution images ( H , W , C The image is divided into segments representing its height, width, and number of channels (where R is the set of real numbers). The size is Small pieces, including:

[0029]

[0030] The output data is a flattened sequence of image patches. ;

[0031] S212, Linear Mapping: Mapping the vector of each block to a linear layer of length . vector , It is the feature dimension of ViT;

[0032]

[0033] It is a linear transformation matrix;

[0034] S213. Embed position information for each sequence data;

[0035]

[0036] For the position embedding matrix;

[0037] In this embodiment, the Transformer encoder is a four-layer cascaded structure. Each layer processes the input data sequentially through layer normalization, dual-domain chiral attention mechanism enhancement, residual, layer normalization, MLP feedforward network layer and residual processing, and then outputs the data. The Transformer decoder uses mathematical rearrangement to piece together the image block sequence back into a complete image in its original position, that is, to convert the sequence vector into a feature map.

[0038] like Figure 2 As shown, the dual-domain chimeric attention mechanism includes two parallel branches: a global branch and a local branch. The global branch employs a multi-head self-attention mechanism, where the outputs of each attention head are concatenated to obtain the output of the global branch. subscript This indicates the current layer number of the Transformer encoder; the local branches employ multi-head self-attention and masked self-attention mechanisms. The masked self-attention mechanism uses a mask matrix to ensure that each position only focuses on its five neighboring positions, thus obtaining the local attention weights of each attention head. Then, the output of each attention head is obtained by combining the value vectors of each attention head, and the outputs of each attention head are merged to obtain the local branch output. Finally, the global branch outputs. With local branch output Weighted fusion yields fused output ;

[0039]

[0040] The mask matrix is ​​as follows:

[0041]

[0042] Represents the nth element in the mask matrix i Line 1 j The element values ​​of the column; The weight of a local branch is an adaptive parameter.

[0043] Specifically, the dual-domain embedded attention mechanism includes the following steps:

[0044] S221, Input data Perform a linear transformation to obtain three sets of vectors: query, key, and value.

[0045]

[0046]

[0047]

[0048] In the formula, , , are linear transformation matrices for query, key, and value, respectively, and , are learnable parameters; Q , K , V These are respectively query, key, and value vectors; D The dimension of the input sequence;

[0049] S222. Divide the query, key, and value vectors along the feature dimension. h Size;

[0050]

[0051]

[0052]

[0053] In the formula, Indicates the first A single-headed query, key, and value vector; It is a single-head dimension. ; Represents three sets of vectors: query, key, and value;

[0054] S223. Calculate the global attention weights of each attention head when using a multi-head attention mechanism. ;

[0055]

[0056] In the formula, , indicating the first Global attention weights for each attention head;

[0057] S223, Calculate the global branch output;

[0058] =

[0059]

[0060] In the formula, Indicates the first branch in the global branch The output of each attention head is indexed by its sequence number. , represents a linear layer used to achieve linear changes; This indicates the output of the global branch; Indicates a splicing operation; Indicates row normalization;

[0061] S224. Calculate the output of the local branch using the masked self-attention mechanism;

[0062] The local branch employs a local mask self-attention mechanism, which uses a preset mask matrix to shield irrelevant regions outside the window, strictly limiting attention calculations to the local neighborhood window of each image patch; that is, using the global attention weights of each attention head. As the initial local attention weights for each attention head Next, the initial local attention weights were adjusted. Perform a masking operation to obtain the local attention weights for each attention head. Then combine the value vector Get the first in the local branch The output of each attention head The output of the local branch is obtained by concatenating the outputs of all attention heads in the local branch. ;

[0063] S225. The weighted fusion of global branch output and local branch output yields the fused output.

[0064] .

[0065] S3. The image processing model is trained using the preprocessed image. An adaptive attention allocation strategy is adopted during training to dynamically adjust the relative importance of global attention and local attention. An edge-aware multi-constraint loss function is used to dynamically coordinate the overall reconstruction and edge quality, which significantly improves the detail clarity and contour naturalness of the reconstructed image.

[0066] This embodiment employs an adaptive attention allocation strategy during training, using the Structural Similarity Index (SSIM) as an indicator to evaluate the reconstruction quality level at the current training stage. When the reconstruction quality level is low, the weights of local branches are adjusted. By selecting smaller values, the weight of global attention is amplified, accelerating the improvement of reconstruction capabilities; when the reconstruction quality level is high, the weight of local branches is increased as the reconstruction quality level improves. To improve the acquisition of local information and enhance the processing of fine edges and texture information, the specific steps are as follows:

[0067] S331. Set the structural similarity index as an indicator for evaluating the reconstruction quality level;

[0068] The formula for calculating the structural similarity index is as follows:

[0069]

[0070] These represent the super-resolution reconstructed image and the real image, respectively. Indices representing the structural similarity index between two images; These represent the local means of the two images, respectively. These represent the local variances of the two images, respectively. This represents the local covariance of two images; while It is a constant;

[0071] S332. Set the weight of local branches The initial value is 0 and training is performed. During the training process, the value of the structural similarity index is monitored in real time and its trend is plotted. When the convergence state reaches the preset first convergence judgment condition (the difference between two adjacent samples is less than 0.001), it means that the reconstruction quality of the model is relatively high. The number of training rounds at this time is recorded as the critical training round.

[0072] S333. For each training round after the critical training round, the weights of the local branches are dynamically adjusted.

[0073] The formula for dynamically adjusting the weight of local branches is:

[0074]

[0075] Indicates the first The weights of local attention during training rounds; Indicates the critical training round; This is a constant whose value can be adjusted according to different datasets; in this example, it is set to 0.3. With this setting, when... hour, The value is very small, which ensures that the global branch weight is high when the reconstruction quality of the model is low; when > hour, It will grow faster, thus increasing the proportion of local branches when the reconstruction quality of the model is high.

[0076] In this example, an edge-aware multi-constraint loss function is used, which consists of three losses: basic reconstruction loss, pixel-level edge loss, and structural edge loss. The basic reconstruction loss uses the mean squared error loss function to calculate the overall loss of the entire image. The pixel-level edge loss uses the mean squared error loss function to calculate the loss of the edges of the entire image. The mean squared error loss function is a common loss function that imposes a higher penalty on larger errors, which can effectively drive the model to optimize in the direction of reducing prediction errors. The structural edge loss assigns higher weights to edge regions, forcing the model to prioritize edge structure quality.

[0077] Specifically:

[0078] The edge-aware multi-constraint loss function is as follows:

[0079]

[0080] In the formula, This represents the joint loss, with the number of iterations in parentheses; Indicates the basic reconstruction loss MSE; Indicates pixel-level edge loss; Indicates structural-level edge loss; The weights represent the pixel-level edge loss, with the number of iterations in parentheses; The weights represent the structural edge loss, with the number of iterations in parentheses;

[0081]

[0082] In the formula, N The total number of samples; m This represents the total number of pixels in a single data set. Representing the location of the real image and the super-resolution reconstructed image, respectively. The values ​​at these locations represent the actual value and the predicted value, respectively.

[0083] Edge intensity images are obtained from both the real image and the super-resolution reconstructed image using a pixel-level edge intensity constraint strategy: First, to preserve high-frequency details, the Sobel operator is directly applied to calculate the gradient magnitude. Then, to avoid the gradient vanishing problem caused by hard thresholding in traditional edge detection operators and to preserve the strength gradient information of texture edges, a differentiable nonlinear activation function, Sigmoid, is introduced. This operation smoothly and continuously maps the gradient magnitude to a probability range of 0 to 1. Within this range, the numerical value no longer represents a binary "yes or no" classification, but rather the confidence level that the pixel belongs to the edge texture, thus generating an edge intensity map with gradient continuity. Finally, The loss is obtained by calculating the mean square error between the edge intensity maps of the super-resolution reconstructed image and the real image. By regressing the edge intensity distribution, the network is forced to learn the correct gradient change trend, while ensuring the effective backpropagation of the gradient.

[0084] Specifically,

[0085] Sobel gradient calculation yields the gradient magnitude. :

[0086]

[0087]

[0088]

[0089] In the formula, Indicates input data; , Represents the Sobel operator convolution kernel; , These represent the gradients in the horizontal and vertical directions, respectively. Indicates the gradient magnitude; Indicates the gradient direction; This represents the convolution operation;

[0090] Introducing the non-linear activation function Sigmoid will reduce the unbounded gradient magnitude. Smooth and continuous mapping to The final edge intensity map is obtained from the interval. ;

[0091]

[0092]

[0093] These represent the edge intensity images corresponding to the super-resolution reconstructed image and the ground truth image at the location, respectively. The values ​​at these locations represent the actual value and the predicted value, respectively.

[0094] The edge intensity images of both the real image and the super-resolution reconstructed image are divided into two categories. M There are 10 windows, each with a size of 1. w × w ;

[0095]

[0096]

[0097] M The number of windows to divide the edge intensity image; For the first Edge weights of each window; The edge intensity images of the real image and the super-resolution reconstructed image are in the first... The structural similarity index of each window; Indicates the first The sum of the edge strength values ​​of each window; Indicates the first The edge pixel density of each window; It is a constant, and a very small non-zero value is chosen to ensure that the weight is not zero.

[0098] In addition, this embodiment adopts a gradient-aligned course evolution strategy for training, which divides the training into three ordered course stages, namely the first to the third stage.

[0099] The first stage focuses on basic reconstruction. In this stage, the joint loss function only uses the basic reconstruction loss. The purpose is to establish low-frequency content anchors using the basic reconstruction loss. During training, the gradient change of the joint loss is monitored. When the convergence state reaches the preset second convergence criterion (continuous convergence), the training continues. The loss decreases less than the threshold in the next iteration. Then proceed to the second stage;

[0100]

[0101] This is the window size for determining the convergence plateau period; in this embodiment, it is set to 3. This is the threshold for determining convergence; in this embodiment, it is set to a value of [value to be filled in]. ; This represents the joint loss during the iteration process, with the number of iterations indicated in parentheses;

[0102] The loss function for this stage is:

[0103]

[0104] Second stage injection of pixel-level edge loss This process corrects blurred boundaries, thereby promoting edge sharpening. In this stage, the joint loss uses basic reconstruction loss and pixel-level edge loss. In order to align gradients and ensure that injecting new loss does not cause gradient imbalance, a reasonable initial weight needs to be set.

[0105] The loss function for this stage is:

[0106]

[0107] Initial weights ( ):

[0108]

[0109] In the formula, This represents the gradient of the basic reconstruction loss with respect to the training parameters, and the number of iterations is in parentheses; Indicates the number of iterations at the start of the second phase; express Norm; Represents the numerical stability constant;

[0110] Dynamic updates ( ): Every time passing F A smooth alignment is performed in the next iteration.

[0111]

[0112] In the formula, This represents the weight smoothing coefficient, which is a constant and can be a value close to 1, such as 0.99. The purpose is to ensure that the weights do not change drastically and cause gradient imbalance. These represent the gradients of the pixel-level edge loss with respect to the training parameters, and the numbers in parentheses indicate the number of iterations. Represents the numerical stability constant;

[0113] When the convergence state reaches the preset second convergence criterion, it enters the third stage;

[0114] Third-stage injection of structural edge loss This stage corrects the edge contour structure and promotes structural texture synthesis. In this stage, the joint loss uses basic reconstruction loss, pixel-level edge loss and structural edge loss.

[0115] The loss function for this stage is:

[0116]

[0117] Initial weights ( ):

[0118]

[0119] In the formula, This represents the gradient of the structural edge loss with respect to the training parameters, and the number of iterations is in parentheses. Indicates the number of training iterations at the start of the third phase;

[0120] Dynamic updates ( ):

[0121]

[0122]

[0123] In the formula, This represents the numerical stability constant.

[0124] S4. Input the image to be processed into the trained image processing model to perform super-resolution reconstruction of the image to be processed.

[0125] Figure 3 To use a real image, first reduce the image to half its original size (both width and height become half), then add Gaussian noise. Figure 4 The image shown is the result of processing by the image processing model in this embodiment, which demonstrates its excellent image reconstruction capabilities.

[0126] In addition, this embodiment also sets up a model that uses mean squared error loss, which only modifies the loss function compared to the above model, and cancels the gradient alignment course evolution strategy.

[0127] The modified loss function is as follows:

[0128]

[0129] Figure 5 For images reconstructed using this model, compare Figure 4 and Figure 5 It can be seen that the edge-aware multi-constraint loss has a clearer edge contour and better restores the real information of the image in the details.

[0130] Table 1 shows the average training results of the two models using mean squared error and edge-aware multi-constraint loss, respectively, on the Urban100 dataset. In the table, PSNR is Peak Signal-to-Noise Ratio, SSIM is Structural Similarity, PSNR_Y is the Peak Signal-to-Noise Ratio of the Luminance Component, and SSIM_Y is the Structural Similarity of the Luminance Component. As shown in Table 1, compared to the commonly used mean squared error, the edge-aware multi-constraint loss slightly reduces the PSNR, while SSIM shows a significant improvement. This is because the model no longer simply pursues pixel alignment, sacrificing a tiny bit of pixel accuracy, and shifts its focus to restoring structure, texture, and edges, significantly improving structural quality.

[0131] Table 1 shows the training results metrics for the model under different loss functions.

[0132]

[0133] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the embodiments of the present invention should be included within the scope of protection of the present invention.

Claims

1. An image super-resolution reconstruction method based on a dual-domain chimeric attention mechanism, characterized in that, Includes the following steps: S1. Acquire images and perform preprocessing; S2. Construct an image processing model. The image processing model uses a multi-level U-Net as the main network structure. The bottleneck layer of U-Net adopts the ViT module. The ViT module first converts the feature map output by the U-Net encoder into sequence data through image block embedding. Then, it is input into the Transformer encoder with a dual-domain chiral attention mechanism for feature enhancement. Finally, the enhanced sequence data is decoded by the Transformer decoder to restore the feature map and used as the input of the U-Net decoder. S3. The image processing model is trained using the preprocessed image. During training, an adaptive attention allocation strategy is adopted to dynamically adjust the relative importance of global attention and local attention. An edge-aware multi-constraint loss function is used to dynamically coordinate the overall reconstruction and edge quality, thereby improving the detail clarity and contour naturalness of the reconstructed image. S4. Input the image to be processed into the trained image processing model to perform image super-resolution reconstruction; The Transformer encoder is a multi-layered cascaded structure, and each layer employs a dual-domain splicing attention mechanism. This mechanism includes two parallel branches: a global branch and a local branch. The global branch uses a multi-head self-attention mechanism, where the outputs of each attention head are concatenated to obtain the global branch output. The local branch employs a masked self-attention mechanism. This mechanism uses a mask matrix to ensure that each position only focuses on its multiple neighboring positions, thereby obtaining the local attention weights of each attention head. These weights are then combined with the value vectors of each attention head to obtain its output. Finally, the outputs of all attention heads are merged to obtain the local branch output. Finally, the global branch output... With the local branch output Weighted fusion yields fused output ; In the formula, Indicates the weight of a local branch; The mask matrix is ​​as follows: Represents the mask matrix of the first i Line 1 j The element values ​​of the column; This represents the neighborhood of interest for the target location in the mask matrix; In step S3, the training employs an adaptive attention allocation strategy, specifically including the following steps: S331. Set the structural similarity index as an indicator for evaluating the reconstruction quality level; The formula for calculating the structural similarity index is as follows: These represent the super-resolution reconstructed image and the real image, respectively. Indices representing the structural similarity index between two images; These represent the local means of the two images, respectively. These represent the local variances of the two images, respectively. Represents the local covariance of two images; It is a constant; S332. Set the weight of local branches The initial value is 0 and training is performed. During the training process, the structural similarity index is monitored in real time. When the convergence state of the structural similarity index reaches the preset first convergence judgment condition, the number of training rounds at this time is recorded as the critical training round. S333. For each training round after the critical training round, dynamically adjust the weight of the local branch; The formula for dynamically adjusting the weight of local branches is: Indicates the first The weights of local branches during training epochs; Indicates the critical training round; It is a constant.

2. The image super-resolution reconstruction method based on a dual-domain chimeric attention mechanism according to claim 1, characterized in that, The edge-aware multi-constraint loss function is as follows: In the formula, This represents the joint loss, with the number of iterations in parentheses; Indicates losses from basic reconstruction; Indicates pixel-level edge loss; Indicates structural-level edge loss; The weights represent the pixel-level edge loss, with the number of iterations in parentheses; The weights represent the structural edge loss, with the number of iterations in parentheses; In the formula, N The total number of samples; m This represents the total number of pixels in a single data set. , Representing the location of the real image and the super-resolution reconstructed image, respectively. The value at; Both the real image and the super-resolution reconstructed image are used to obtain edge intensity images through a pixel-level edge intensity constraint strategy; These represent the edge intensity images corresponding to the super-resolution reconstructed image and the ground truth image at the location, respectively. The value at; The edge intensity images of both the real image and the super-resolution reconstructed image are divided into two categories. M There are 10 windows, each with a size of 1. w × w ; M The number of windows to divide the edge intensity image; For the first Edge weights of each window; The edge intensity images of the real image and the super-resolution reconstructed image are in the first... The structural similarity index of each window; Indicates the first The sum of the edge strength values ​​of each window; Indicates the first Average edge strength of each window; It is a constant.

3. The image super-resolution reconstruction method based on a dual-domain chimeric attention mechanism according to claim 2, characterized in that, The process of obtaining the edge intensity image through a pixel-level edge intensity constraint strategy includes the following steps: S341. Calculate the gradient magnitude of the image using the Sobel operator; In the formula, Indicates input data; , Represents the Sobel operator convolution kernel; , These represent the gradients in the horizontal and vertical directions, respectively. Indicates the gradient magnitude; This represents the convolution operation; S342. Introduce the non-linear activation function Sigmoid to adjust the gradient magnitude. Mapped to The final edge intensity map is obtained from the interval. ; 。 4. The image super-resolution reconstruction method based on a dual-domain chimeric attention mechanism according to claim 2, characterized in that, In step S3, a gradient-aligned course evolution strategy is used for training. The gradient-aligned course evolution strategy divides the training process into three ordered course stages, namely the first to the third stages. During training, observe the gradient change of the joint loss. When the gradient of the joint loss reaches the preset second convergence criterion, proceed from the previous stage to the next stage. In the first phase, the joint loss only uses the basic reconstruction loss to establish low-frequency content anchors; In the second stage, the joint loss uses basic reconstruction loss and pixel-level edge loss to correct blurred boundaries, thereby promoting edge sharpening evolution; The joint loss function for the second stage is as follows: Initial weights, In the formula, Indicates the current iteration number; This represents the gradient of the basic reconstruction loss with respect to the training parameters, and the number of iterations is in parentheses; This represents the gradient of the pixel-level edge loss with respect to the training parameters, and the number of iterations is in parentheses. Indicates the number of iterations at the start of the second phase; express Norm; Represents the numerical stability constant; Dynamic updates Every time you pass F Perform a smooth alignment in each iteration; In the formula, Indicates the weighted smoothing coefficient; Represents the numerical stability constant; In the third stage, the joint loss employs basic reconstruction loss, pixel-level edge loss, and structural edge loss to correct edge contour structure and promote structural texture synthesis. The joint loss function for the third stage is as follows: Initial weights, : In the formula, This represents the gradient of the structural edge loss with respect to the training parameters, and the number of iterations is in parentheses. Indicates the number of training iterations at the start of the third phase; Dynamic updates : In the formula, This represents the numerical stability constant.

Images