An image super-resolution reconstruction method and system based on an optical cavity transition layer

Abstract

The application relates to the field of low-level computer vision, and discloses an image super-resolution reconstruction method and system based on a light hole transition layer, which comprises the following steps: S1. constructing and training a reconstruction network used for image super-resolution reconstruction; S3. inputting an image to be reconstructed into the reconstruction network to extract shallow layer features; S4. extracting several level features of the shallow layer features; S5. extracting attention features of the several level features through a channel attention layer extraction block; S6. fusing the several level features through a first 1*1 convolution layer, combining the fused several level features with attention features obtained through a second 1*1 convolution layer, and obtaining combined features; and convolving the combined features through a second 3*3 convolution layer to obtain reconstruction features; and S7. combining the reconstruction features with the image to be reconstructed, inputting the combination into an up-sampling block to perform up-sampling, and obtaining a final reconstructed image. The application solves the problem of low efficiency in the prior art, and has the characteristics of convenient calculation and simple structure.

Description

Technical Field

[0001] This invention relates to the field of low-level computer vision technology, and more specifically, to an image super-resolution reconstruction method and system based on an optical void transition layer. Background Technology

[0002] Image super-resolution reconstruction technology has wide applications in medical image reconstruction, facial / license plate reconstruction and recognition in surveillance cameras, satellite remote sensing image reconstruction, and video stream super-resolution. However, most existing image super-resolution networks improve performance by deepening or widening the network structure, leading to complex network structures, increased model parameters, and slow reconstruction speeds. Furthermore, most image super-resolution models are built by continuously stacking residual blocks, resulting in the loss of local residual information within these convolutional blocks, leading to the loss of texture details in the reconstructed image. Additionally, deepening the network significantly increases computational demands and memory consumption, limiting the application of image super-resolution algorithms on terminal devices, especially customer-facing devices such as mobile phones and drones. To better deploy image super-resolution technology on terminal devices, a lightweight method capable of rapidly reconstructing images is needed.

[0003] An existing technology provides an image super-resolution reconstruction method based on an attention mechanism-based dilated convolutional neural network. This method designs an attention mechanism-based dilated convolutional module, a multi-channel attention mechanism-based dilated convolutional module, an attention residual block, and a multi-channel feature extraction module. Based on this, an attention mechanism-based dilated convolutional neural network is designed.

[0004] However, existing technologies suffer from low reconstruction efficiency. Therefore, how to invent a highly efficient image super-resolution reconstruction method is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] To address the problem of low efficiency in existing technologies, this invention provides an image super-resolution reconstruction method and system based on an optical void transition layer, which features convenient computation and simple structure.

[0006] To achieve the above-mentioned objectives of this invention, the technical solution adopted is as follows:

[0007] The first aspect of this invention provides an image super-resolution reconstruction method based on an optical void transition layer, comprising the following steps:

[0008] S1. Construct and train a reconstruction network for image super-resolution reconstruction; the reconstruction network includes several optically dilated convolutional extraction blocks, residual attention extraction blocks, upsampling blocks, a first 3*3 convolutional layer, a second 3*3 convolutional layer, a first 1*1 convolutional layer, and a second 1*1 convolutional layer.

[0009] S2. Obtain the image to be reconstructed;

[0010] S3. Input the image to be reconstructed into the reconstruction network and extract shallow features through the first 3*3 convolutional layer;

[0011] S4. Extract several hierarchical features of shallow features through several optically dilated convolutional extraction blocks;

[0012] S5. The residual attention extraction block includes several channel attention layers, and attention features of several levels of features are extracted through the channel attention layer extraction block;

[0013] S6. Several hierarchical features are fused through the first 1*1 convolutional layer, and the fused hierarchical features are combined with the attention features passed through the second 1*1 convolutional layer to obtain the combined features; the combined features are convolved through the second 3*3 convolutional layer to obtain the reconstructed features.

[0014] S7. Combine the reconstructed features with the image to be reconstructed, input the upsampling block for upsampling, and obtain the final reconstructed image.

[0015] Preferably, in step S4, several levels of features of shallow features are extracted through several optically dilated convolutional extraction blocks. Specifically, the shallow features are input into the first optically dilated convolutional extraction block to obtain the first level features; the first level features are input into the second optically dilated convolutional extraction block to obtain the second level features; and the level features output by each level of the optically dilated convolutional extraction block are input into the next level of the optically dilated convolutional extraction block to obtain the features of each level step by step.

[0016] Furthermore, the optically voided convolutional extraction block includes several optically voided transition layers; the several optically voided transition layers are connected in series and used to extract features sequentially.

[0017] Furthermore, more specifically, the optically voided convolutional extraction block includes 3 optically voided transition layers, 5 1*1 convolutional layers, 4 channel attention layers, and 1 enhanced spatial attention layer.

[0018] Furthermore, optically dilated convolution extracts hierarchical features from blocks, specifically:

[0019] A1. During feature extraction, the feature to be extracted is input into the first optical void transition layer to obtain the first feature. The feature to be extracted is combined with the first feature to obtain the first combined feature. The first combined feature is input into the second optical void transition layer to obtain the second feature. The first combined feature is combined with the second feature to obtain the second combined feature. The second combined feature is input into the third optical void transition layer to obtain the third feature. The second combined feature is combined with the third feature to obtain the third combined feature.

[0020] A2. The features to be extracted, the first combined feature, and the second combined feature are each convolved through a 1*1 convolutional layer and then fused with the third combined feature. The fused features are then passed through a 1*1 convolutional layer and an enhanced spatial attention layer in sequence and output as the first-level features. The first-level features are used as input to the next optical hole extraction block to further extract the next-level features or to the first 1*1 convolutional layer for feature fusion.

[0021] A3. The first feature, the second feature, and the third feature are extracted into attention features through the channel attention layer and then fused. The fused attention features are then passed through a 1*1 convolutional layer and a channel attention layer in sequence and output as the second-level features. The second-level features are used to extract attention features from the input residual attention extraction block.

[0022] Furthermore, the optical void transition layer includes several convolutional kernels; the specific steps for extracting features from the optical void transition layer are as follows;

[0023] B1. Set the compression ratio K, the number of features to be extracted N, and the branch factor n. b M features are input into M / K 1×1 convolution kernels for point convolution, resulting in M / K point convolution features;

[0024] B2. Perform parallel convolutions on the M / K convolutional features using 1×1 and 3×3 convolutional kernels respectively. The 1×1 convolutional kernel outputs N / n. b Extracting features from channels, and outputting NN / n via 3×3 convolution. b Channel extraction features;

[0025] B3. N / n b Channel extraction features and NN / n b The extracted features of the channels are fused to obtain N extracted features.

[0026] Furthermore, the channel attention layer includes a pooling layer, a 1*1 convolutional layer, a signal layer, and a 1D channel convolutional layer. When the channel attention layer extracts features, the input features of the channel attention layer are sequentially processed by the pooling layer, the 1*1 convolutional layer, the signal layer, and the 1D channel convolutional layer to extract features. The extracted features are then fused with the original input features, and the fused features are used as the output features of the channel attention layer.

[0027] Furthermore, the enhanced spatial attention layer can adaptively select a 1D convolution kernel.

[0028] Furthermore, the number of optically voided convolutional extraction blocks is four.

[0029] The second aspect of this invention provides an image super-resolution reconstruction system based on an optical void transition layer, comprising a model building module, an image acquisition module, a shallow feature extraction module, a hierarchical feature extraction module, a residual attention extraction module, a reconstruction feature module, and an image reconstruction module;

[0030] The model building module is used to build and train a reconstruction network for image super-resolution reconstruction; the reconstruction network includes several optically dilated convolution extraction blocks, residual attention extraction blocks, upsampling blocks, a first 3*3 convolutional layer, a second 3*3 convolutional layer, a first 1*1 convolutional layer, and a second 1*1 convolutional layer.

[0031] The image acquisition module is used to acquire the image to be reconstructed;

[0032] The shallow feature extraction module is used to input the image to be reconstructed into the reconstruction network and extract shallow features through the first 3*3 convolutional layer;

[0033] The aforementioned hierarchical feature extraction module is used to extract several hierarchical features of shallow features through several optically dilated convolution extraction blocks;

[0034] The residual attention extraction module is used to extract attention features of several hierarchical features;

[0035] The reconstructed feature module is used to fuse several hierarchical features through a first 1*1 convolutional layer, and combine the fused hierarchical features with the attention features passed through a second 1*1 convolutional layer to obtain combined features; the combined features are then convolved through a second 3*3 convolutional layer to obtain reconstructed features.

[0036] The image reconstruction module is used to combine the reconstruction features with the image to be reconstructed, input them into the upsampling block for upsampling, and obtain the final reconstructed image.

[0037] The beneficial effects of this invention are as follows:

[0038] This invention constructs and trains a reconstruction network for image super-resolution reconstruction. Several hierarchical features of shallow features are extracted through several optically dilated convolutional extraction blocks of the reconstruction network. Attention features of the shallow features and several hierarchical features are extracted through residual attention extraction blocks of the reconstruction network. Finally, several hierarchical features are fused through a first 1*1 convolutional layer, and the fused hierarchical features are combined with the attention features passed through a second 1*1 convolutional layer to obtain combined features. These combined features are then convolved through a second 3*3 convolutional layer to obtain reconstructed features, resulting in a reconstructed image. This reduces the parameters of the image super-resolution reconstruction model, decreases the parameters and computational requirements for reconstruction, and improves the efficiency of the reconstruction model. Therefore, this invention solves the problem of low efficiency in existing technologies and is computationally convenient and structurally simple. Attached Figure Description

[0039] Figure 1 This is a schematic flowchart of an image super-resolution reconstruction method based on an optical void transition layer according to the present invention.

[0040] Figure 2 This is a schematic diagram of the overall framework of an image super-resolution reconstruction method based on an optical void transition layer according to the present invention.

[0041] Figure 3 This is a schematic diagram of the framework of the optically voided convolution extraction block of an image super-resolution reconstruction method based on an optically voided transition layer according to the present invention.

[0042] Figure 4 This is a schematic diagram of the structure of the optical void transition layer in the image super-resolution reconstruction method based on the optical void transition layer of the present invention.

[0043] Figure 5 This is a schematic diagram of the channel attention layer structure in an image super-resolution reconstruction method based on an optical void transition layer according to the present invention.

[0044] Figure 6 This invention relates to an enhanced spatial attention layer in an image super-resolution reconstruction method based on an optical void transition layer. Detailed Implementation

[0045] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0046] Example 1

[0047] like Figure 1 As shown, an image super-resolution reconstruction method based on an optical void transition layer includes the following steps:

[0048] S1. Construct and train a reconstruction network for image super-resolution reconstruction; the reconstruction network includes several optically dilated convolutional extraction blocks, residual attention extraction blocks, upsampling blocks, a first 3*3 convolutional layer, a second 3*3 convolutional layer, a first 1*1 convolutional layer, and a second 1*1 convolutional layer.

[0049] S2. Obtain the image to be reconstructed;

[0050] S3. Input the image to be reconstructed into the reconstruction network and extract shallow features through the first 3*3 convolutional layer;

[0051] S4. Extract several hierarchical features of shallow features through several optically dilated convolutional extraction blocks;

[0052] S5. The residual attention extraction block includes several channel attention layers, and attention features of several levels of features are extracted through the channel attention layer extraction block;

[0053] S6. Several hierarchical features are fused through the first 1*1 convolutional layer, and the fused hierarchical features are combined with the attention features passed through the second 1*1 convolutional layer to obtain the combined features; the combined features are convolved through the second 3*3 convolutional layer to obtain the reconstructed features.

[0054] S7. Combine the reconstructed features with the image to be reconstructed, input the upsampling block for upsampling, and obtain the final reconstructed image.

[0055] Example 2

[0056] More specifically, in this embodiment, the publicly available DIV2K dataset is used as the training set to train the reconstruction network; a total of 800 images are used, with 100 images for the validation set and 100 images for the test set. Since the test dataset has not yet been released, this embodiment uses Set5, Set14, BSD100, and Urban100 as the test sets. Set5, Set14, BSD100, and Urban10 are all real images of natural landscapes, people, etc., with 5, 14, 100, and 100 images respectively.

[0057] In this embodiment, before training the reconstruction network, the DIVK2K images are horizontally flipped, vertically flipped, and horizontally and vertically flipped to expand the training set. The images are also cropped into blocks of sizes 256×256, 255×255, and 256×256, respectively, to train the reconstruction networks at magnifications of 2, 3, and 4 times. This embodiment also simultaneously trains pre-trained super-resolution reconstruction network models at magnifications of 2, 3, and 4 times from scratch.

[0058] In this embodiment, the reconstructed features are reconstructed to obtain a high-resolution image through subpixel convolution. After obtaining the reconstructed image, the quality of the reconstructed image is evaluated by two evaluation metrics: peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) between the reconstructed image and the original high-resolution image.

[0059] In step S4, several levels of features of shallow features are extracted through several optically dilated convolutional extraction blocks. Specifically, the shallow features are input into the first optically dilated convolutional extraction block to obtain the first level features; the shallow features are input into the second optically dilated convolutional extraction block to obtain the second level features. By inputting the level features output by each level of the optically dilated convolutional extraction block into the next level of the optically dilated convolutional extraction block, the features of each level are obtained step by step.

[0060] In one specific embodiment, the optically voided convolutional extraction block includes several optically voided transition layers; the several optically voided transition layers are connected in series and used to extract features sequentially.

[0061] In one specific embodiment, such as Figure 3As shown, more specifically, the optically voided convolutional extraction block includes 3 optically voided transition layers, 5 1*1 convolutional layers, 4 channel attention layers, and 1 enhanced spatial attention layer.

[0062] In one specific embodiment, such as Figure 3 As shown, optically dilated convolution extracts hierarchical features from blocks, specifically:

[0063] A1. During feature extraction, the feature to be extracted is input into the first optical void transition layer to obtain the first feature. The feature to be extracted is combined with the first feature to obtain the first combined feature. The first combined feature is input into the second optical void transition layer to obtain the second feature. The first combined feature is combined with the second feature to obtain the second combined feature. The second combined feature is input into the third optical void transition layer to obtain the third feature. The second combined feature is combined with the third feature to obtain the third combined feature.

[0064] A2. The features to be extracted, the first combined feature, and the second combined feature are each convolved through a 1*1 convolutional layer and then fused with the third combined feature. The fused features are then passed through a 1*1 convolutional layer and an enhanced spatial attention layer in sequence and output as the first-level features. The first-level features are used as input to the next optical hole extraction block to further extract the next-level features or to the first 1*1 convolutional layer for feature fusion.

[0065] A3. The first feature, the second feature, and the third feature are extracted into attention features through the channel attention layer and then fused. The fused attention features are then passed through a 1*1 convolutional layer and a channel attention layer in sequence and output as the second-level features. The second-level features are used to extract attention features from the input residual attention extraction block.

[0066] In one specific embodiment, such as Figure 4 As shown, the optical void transition layer includes several convolutional kernels; the specific steps for extracting features from the optical void transition layer are as follows;

[0067] B1. Set the compression ratio K, the number of features to be extracted N, and the branch factor n. b M features are input into M / K 1×1 convolution kernels for point convolution, resulting in M / K point convolution features;

[0068] B2. Perform parallel convolutions on the M / K convolutional features using 1×1 and 3×3 convolutional kernels respectively. The 1×1 convolutional kernel outputs N / n. b Extracting features from channels, and outputting NN / n via 3×3 convolution. b Channel extraction features;

[0069] B3. N / n b Channel extraction features and NN / n bThe extracted features of the channels are fused to obtain N extracted features.

[0070] In one specific embodiment, such as Figure 5 As shown, the channel attention layer includes a pooling layer, a 1*1 convolutional layer, a signal layer, and a 1D channel convolutional layer. When the channel attention layer extracts features, the input features of the channel attention layer are sequentially processed by the pooling layer, the 1*1 convolutional layer, the signal layer, and the 1D channel convolutional layer to extract features. The extracted features are then fused with the original input features, and the fused features are used as the output features of the channel attention layer.

[0071] In one specific embodiment, such as Figure 6 As shown, the enhanced spatial attention layer can adaptively select a 1D convolution kernel.

[0072] In one specific embodiment, the number of optically voided convolution extraction blocks is four.

[0073] In one specific embodiment, the upsampling is to enlarge the original image.

[0074] This invention utilizes a distillation structure based on the RFDN network, employing an optically hollow transition layer to replace the 3×3 convolutions in the RFDN for extracting image feature information. This significantly reduces the number of network parameters and accelerates reconstruction. The image reconstruction effect is particularly noticeable for tasks with large magnification factors. Furthermore, this invention uses a lightweight model for training, achieving good reconstruction results even with a relatively shallow network, thus reducing the overall model complexity.

[0075] This invention constructs and trains a reconstruction network for image super-resolution reconstruction. Several hierarchical features of shallow features are extracted through several optically dilated convolutional extraction blocks of the reconstruction network. Attention features of the shallow features and several hierarchical features are extracted through residual attention extraction blocks of the reconstruction network. Finally, several hierarchical features are fused through a first 1*1 convolutional layer, and the fused hierarchical features are combined with the attention features passed through a second 1*1 convolutional layer to obtain combined features. These combined features are then convolved through a second 3*3 convolutional layer to obtain reconstructed features, resulting in a reconstructed image. This reduces the parameters of the image super-resolution reconstruction model, decreases the parameters and computational requirements for reconstruction, and improves the efficiency of the reconstruction model. Therefore, this invention solves the problem of low efficiency in existing technologies and is computationally convenient and structurally simple.

[0076] Example 3

[0077] An image super-resolution reconstruction system based on an optical void transition layer includes a model building module, an image acquisition module, a shallow feature extraction module, a hierarchical feature extraction module, a residual attention extraction module, a reconstruction feature module, and an image reconstruction module.

[0078] The model building module is used to build and train a reconstruction network for image super-resolution reconstruction; the reconstruction network includes several optically dilated convolution extraction blocks, residual attention extraction blocks, upsampling blocks, a first 3*3 convolutional layer, a second 3*3 convolutional layer, a first 1*1 convolutional layer, and a second 1*1 convolutional layer.

[0079] The image acquisition module is used to acquire the image to be reconstructed;

[0080] The shallow feature extraction module is used to input the image to be reconstructed into the reconstruction network and extract shallow features through the first 3*3 convolutional layer;

[0081] The aforementioned hierarchical feature extraction module is used to extract several hierarchical features of shallow features through several optically dilated convolution extraction blocks;

[0082] The residual attention extraction module is used to extract attention features of several hierarchical features;

[0083] The reconstructed feature module is used to fuse several hierarchical features through a first 1*1 convolutional layer, and combine the fused hierarchical features with the attention features passed through a second 1*1 convolutional layer to obtain combined features; the combined features are then convolved through a second 3*3 convolutional layer to obtain reconstructed features.

[0084] The image reconstruction module is used to combine the reconstruction features with the image to be reconstructed, input them into the upsampling block for upsampling, and obtain the final reconstructed image.

[0085] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. An image super-resolution reconstruction method based on an optical void transition layer, characterized in that: Includes the following steps: S1. Construct and train a reconstruction network for image super-resolution reconstruction; the reconstruction network includes several optically dilated convolutional extraction blocks, residual attention extraction blocks, upsampling blocks, a first 3*3 convolutional layer, a second 3*3 convolutional layer, a first 1*1 convolutional layer, and a second 1*1 convolutional layer; the optically dilated convolutional extraction block includes several optically dilated transition layers; the several optically dilated transition layers are connected in series to extract features sequentially. The optical void transition layer includes several convolutional kernels; the specific steps for extracting features from the optical void transition layer are as follows; B1. Set the compression ratio K, the number of features to be extracted N, and the branching factor. M features are input into M / K 1×1 convolution kernels for point convolution, resulting in M / K point convolution features; B2. Perform parallel convolutions on the M / K point convolutional features using 1×1 and 3×3 convolutional kernels respectively. The output of the 1×1 convolutional kernel is... Extracting features from channels, outputting via 3×3 convolution. Channel extraction features; B3. Will Channel extraction features and The extracted features of the channels are fused to obtain N extracted features; S2. Obtain the image to be reconstructed; S3. Input the image to be reconstructed into the reconstruction network and extract shallow features through the first 3*3 convolutional layer; S4. Extract several hierarchical features of shallow features through several optically dilated convolutional extraction blocks; S5. The residual attention extraction block includes several channel attention layers, through which attention features of several hierarchical features are extracted; S6. Several hierarchical features are fused through the first 1*1 convolutional layer, and the fused hierarchical features are combined with the attention features passed through the second 1*1 convolutional layer to obtain the combined features; the combined features are convolved through the second 3*3 convolutional layer to obtain the reconstructed features. S7. Combine the reconstructed features with the image to be reconstructed, input the upsampling block for upsampling, and obtain the final reconstructed image.

2. The image super-resolution reconstruction method based on an optical void transition layer according to claim 1, characterized in that: In step S4, several levels of features of shallow features are extracted through several optically dilated convolutional extraction blocks. Specifically, the shallow features are input into the first optically dilated convolutional extraction block to obtain the first level features; the first level features are input into the second optically dilated convolutional extraction block to obtain the second level features. By inputting the level features output by each level of the optically dilated convolutional extraction block into the next level of the optically dilated convolutional extraction block, the features of each level are obtained step by step.

3. The image super-resolution reconstruction method based on an optical void transition layer according to claim 1, characterized in that: The optically voided convolutional extraction block includes 3 optically voided transition layers, 5 1*1 convolutional layers, 4 channel attention layers, and 1 enhanced spatial attention layer.

4. The image super-resolution reconstruction method based on an optical void transition layer according to claim 3, characterized in that: Optical dilated convolution extracts hierarchical features from blocks, specifically: A1. During feature extraction, the feature to be extracted is input into the first optical void transition layer to obtain the first feature. The feature to be extracted is combined with the first feature to obtain the first combined feature. The first combined feature is input into the second optical void transition layer to obtain the second feature. The first combined feature is combined with the second feature to obtain the second combined feature. The second combined feature is input into the third optical void transition layer to obtain the third feature. The second combined feature is combined with the third feature to obtain the third combined feature. A2. The features to be extracted, the first combined feature, and the second combined feature are each convolved through a 1*1 convolutional layer and then fused with the third combined feature. The fused features are then passed through a 1*1 convolutional layer and an enhanced spatial attention layer in sequence and output as the first-level features. The first-level features are used as input to the next optical hole extraction block to further extract the next-level features or to the first 1*1 convolutional layer for feature fusion. A3. The first feature, the second feature, and the third feature are extracted into attention features through the channel attention layer and then fused. The fused attention features are then passed through a 1*1 convolutional layer and a channel attention layer in sequence and output as the second-level features. The second-level features are used to extract attention features from the input residual attention extraction block.

5. The image super-resolution reconstruction method based on an optical void transition layer according to claim 3, characterized in that: The channel attention layer includes a pooling layer, a 1*1 convolutional layer, a signal layer, and a 1D channel convolutional layer. When the channel attention layer extracts features, the input features of the channel attention layer are sequentially processed by the pooling layer, the 1*1 convolutional layer, the signal layer, and the 1D channel convolutional layer to extract features. The extracted features are then fused with the original input features, and the fused features are used as the output features of the channel attention layer.

6. The image super-resolution reconstruction method based on an optical void transition layer according to claim 3, characterized in that: The enhanced spatial attention layer adaptively selects a 1D convolutional kernel.

7. The image super-resolution reconstruction method based on an optical void transition layer according to claim 1, characterized in that: The number of optically voided convolution extraction blocks is 4.

8. An image super-resolution reconstruction system based on an optical void transition layer, characterized in that: The method for implementing the method as described in any one of claims 1 to 7 includes a model building module, an image acquisition module, a shallow feature extraction module, a hierarchical feature extraction module, a residual attention extraction module, a feature reconstruction module, and an image reconstruction module. The model building module is used to build and train a reconstruction network for image super-resolution reconstruction; the reconstruction network includes several optically dilated convolution extraction blocks, residual attention extraction blocks, upsampling blocks, a first 3*3 convolutional layer, a second 3*3 convolutional layer, a first 1*1 convolutional layer, and a second 1*1 convolutional layer. The image acquisition module is used to acquire the image to be reconstructed; The shallow feature extraction module is used to input the image to be reconstructed into the reconstruction network and extract shallow features through the first 3*3 convolutional layer; The aforementioned hierarchical feature extraction module is used to extract several hierarchical features of shallow features through several optically dilated convolution extraction blocks; The residual attention extraction module is used to extract attention features of several hierarchical features; The reconstructed feature module is used to fuse several hierarchical features through a first 1*1 convolutional layer, and combine the fused hierarchical features with the attention features passed through a second 1*1 convolutional layer to obtain combined features; the combined features are then convolved through a second 3*3 convolutional layer to obtain reconstructed features. The image reconstruction module is used to combine the reconstruction features with the image to be reconstructed, input them into the upsampling block for upsampling, and obtain the final reconstructed image.

Images