An image super-resolution method based on pyramid attention mechanism and symmetric network
By introducing a pyramid attention mechanism and symmetric network into the image super-resolution algorithm, combining L1 and MSE loss functions to optimize the model weight, the problem of insufficient image detail processing in the existing technology is solved, and high-quality high-resolution image generation is achieved and improve model generalization capabilities.
Patent Information
- Application Number
- CN202111364295.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-17
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2041-11-17
AI Technical Summary
Existing image super-resolution algorithms have problems of smoothness, ringing and low clarity in detail processing, and deep learning-based methods have challenges in model training deduction speed and model size, making it difficult to apply on finite resource platforms.
Using the image super-resolution method based on the pyramid attention mechanism and symmetric network, high and low resolution image pairs are generated through the bicubital interpolation function, combined with the pyramid attention module and symmetric convolution network for feature extraction and reconstruction, and L1 and MSE are used. The loss function is optimized in reverse gradient and adjusts the model weights to improve the quality of the generated results.
Without reducing the training speed of the model, the quality and effect of high-resolution images are significantly improved, and the generation and generalization capabilities of the network are improved.
Smart Images

Figure CN114022362B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer vision and visual image reconstruction, and mainly relates to an image super-resolution method based on a pyramid attention mechanism and a symmetric network. Background Art
[0002] Image super-resolution algorithms are a current branch of computer vision research, finding widespread application in urban surveillance, medical imaging, remote sensing, and other fields. Their primary task is to input low-resolution images into a network, which is then trained and inferred to ultimately generate high-resolution images with greater detail and texture features. Although image super-resolution algorithms have been extensively studied, the high-resolution images produced by most traditional methods suffer from issues such as smoothing, ringing, and low clarity, and their processing of detail often falls short of expectations.
[0003] Current image super-resolution algorithms can be divided into three categories based on the basic framework: traditional methods based on interpolation processing, methods based on image reconstruction, and methods based on deep learning. The traditional method based on interpolation processing is a process of estimating new pixels through given pixels. This method believes that the newly added pixels in the image are only related to the values of the surrounding pixels. This is the simplest way to improve low-resolution images. The advantage is that the algorithm is simple and fast, but its disadvantage is also obvious. The quality of the generated high-resolution images is very limited. The method based on image reconstruction uses the degradation model of the image to study how the original visual scene evolves to obtain the observed image, among which motion estimation and prior knowledge extraction are key factors. The advantage of this method is that the quality of the generated image is better than that of the traditional method based on interpolation. However, in scenarios with high upsampling ratios, the low-resolution image contains very little information, resulting in poor reconstructed image quality.
[0004] Currently, the most widely researched image super-resolution methods are those based on deep learning. Most methods study the mapping relationship between low-resolution images and their corresponding high-resolution counterparts, using deep networks to learn and reconstruct the input image. Typically, deep learning-based image super-resolution algorithms first partition the image into blocks and construct pairs of low-resolution and high-resolution images. They then learn the relationship between the two images, and finally reconstruct the high-resolution image based on the input low-resolution image using this learned relationship. In 2014, during the rise of convolutional neural networks, Dong C et al. first applied convolutional neural networks to image super-resolution. They employed a three-layer pure convolutional network and used large-scale training to directly learn the mapping relationship between low-resolution and high-resolution images. The subsequent advent of the FSRCNN algorithm significantly improved the algorithm's inference speed. The use of sub-pixel convolutional layers improved upsampling efficiency compared to deconvolution while maintaining constant network parameters. The application of deep learning methods in super-resolution has rapidly advanced, leading to the development of blind super-resolution algorithms and attention-based super-resolution algorithms. However, image super-resolution algorithms based on deep learning networks have made a qualitative leap in image accuracy, but the speed of model training and deduction and the model size have also increased accordingly. More and more convolutional layers are stacked into the super-resolution network, making the algorithm difficult to apply on platforms with limited resources. Moreover, most image super-resolution algorithms based on deep convolutional network structures still have problems such as blurred details, smoothness, and pseudo-details in the generated results, and the quality of high-resolution images still needs to be improved. Summary of the Invention
[0005] The technical problem to be solved by this invention is to overcome the shortcomings of the existing technology and provide an image super-resolution method based on the pyramid attention mechanism and symmetric network. The method uses two symmetric convolutional networks with pyramid attention modules to process the global distribution of the input high-resolution image, reconstruct the input low-resolution image, and further adjust and optimize the model weights in combination with reverse gradient optimization of three loss functions. This method improves the quality and effectiveness of the generated results without reducing the model training speed.
[0006] An image super-resolution method based on a pyramid attention mechanism and a symmetric network includes the following steps:
[0007] Step 1: For the dataset to be processed, use the bicubic interpolation function to create a high-resolution image (x HR ) and its corresponding low-resolution image (x LR ), the low resolutions are 2.0, 3.0, and 4.0, which are 2 times, 3 times, and 4 times smaller;
[0008] Step 2: Convert the high-resolution image (x HR ) is input into the first part of the symmetric network for feature extraction. It first passes through the pyramid attention module to integrate the multi-scale information of the real image. Then, it passes through the convolution layer, the ReLU activation layer, and the maximum pooling layer to generate an intermediate feature map. The entire process obtains the multi-level information distribution of the image and enhances the high-frequency details.
[0009] Step 3: The intermediate feature map generated after convolution is compared with the low-resolution image to calculate the error loss. The loss is used to update the network parameters and gradients. The specific loss formula is as follows:
[0010]
[0011] Where x t Represents the intermediate feature map generated by the tth layer of the first network, y t represents the intermediate feature map generated by the tth layer of the second network, and N represents the number of pixels on the feature map. When calculating the intermediate feature map (x3) generated by the first network and the input low-resolution image (x LR ) when the loss value is between:
[0012]
[0013] Step 4: Input the intermediate feature map generated by the first network into the second part of the network for reconstruction to obtain a reconstructed high-resolution image. With the input high-resolution image Calculate the loss and update the network parameters and gradients in reverse. The specific formula for calculating the loss is as follows:
[0014]
[0015] In the formula Represents the high-resolution image reconstructed from the intermediate feature map generated by the first network.
[0016] Step 5: Convert the low-resolution image ( x L R ) is input into the second part of the network, and the low-resolution image is enlarged resolution by resolution, and finally the generated high-resolution image is obtained. and with high-resolution images (x HR ) Calculate the loss and update the network parameters again. The specific formula for calculating the loss is as follows:
[0017]
[0018] In the formula Represents the low-resolution image (x) that the second network will input LR) reconstructed into a high-resolution image.
[0019] The loss functions used to calculate the difference in steps 4 and 5 are because the intermediate feature maps reconstructed from the high-resolution image should be as close as possible to the original image to ensure that the network learns the true distribution of image information; while the reconstruction from the low-resolution image should ultimately preserve the image's high-frequency information and global distribution. The two loss functions have different effects on these two aspects: the generated results trained with the L1 loss function tend to be sharp, while the results trained with the MSE loss function tend to be smooth.
[0020] Furthermore, the symmetric network in step 2 is composed of two networks connected in series with exactly the same number of network layers, structure, convolution kernel size, pooling method, padding step size, and attention module and shared weights.
[0021] Furthermore, the first network structure and corresponding parameters are: Conv1+ReLU layer, size and number of channels 3*3*96, Padding: valid; MaxPool1 layer, size and number of channels 2*2, Padding: valid; Conv2+ReLU layer, size and number of channels 3*3*256, Padding: same; MaxPool2 layer, size and number of channels 2*2, Padding: same; Conv3+ReLU layer, size and number of channels 3*3*512, Padding: valid; the filter step size of each layer is 1.
[0022] Furthermore, the second network structure and corresponding parameters are: Conv1+ReLU layer, size and number of channels 3*3*512, Padding: valid; Conv2+ReLU layer, size and number of channels 3*3*256, Padding: same; UpConv1 layer, size and number of channels 2*2, Padding: same; Conv3+ReLU layer, size and number of channels 3*3*96, Padding: valid; UpConv2 layer, size and number of channels 2*2, Padding: valid; the filter step size of each layer is 1.
[0023] Furthermore, when high-resolution images from the dataset are input into the first network, the high-resolution images are reduced in size by a factor of 2 layer by layer; when the intermediate feature maps are input into the second network, the low-resolution images are enlarged in size by a factor of 2 layer by layer.
[0024] Furthermore, the feature dimensions and sizes after calculation and output of each layer of the first network are as follows: Conv1+ReLU layer outputs 480*480*96, MaxPool1 layer outputs 240*240*96, Conv2+ReLU layer outputs 240*240*256, MaxPool2 layer outputs 120*120*256, and Conv3+ReLU layer outputs 120*120*512.
[0025] Furthermore, the feature dimensions and sizes after calculation and output of each layer of the second network are as follows: Conv1+ReLU layer outputs 120*120*512, Conv2+ReLU layer outputs 120*120*256, UpConv1 layer outputs 240*240*256, Conv3+ReLU layer outputs 240*240*96, and UpConv2 layer outputs 480*480*96.
[0026] Furthermore, the detailed feature of step 3 is that the training data set and the test data set used for training include:
[0027] The DIV2K dataset contains 800 natural scene training images and 100 validation images; the Set5 dataset contains 5 high-resolution images of people, animals, and cartoons; the Set14 dataset contains 14 high-resolution images of people, animals, cartoons, black and white images, and scenes; the BSD100 dataset contains 100 high-resolution images of people, animals, scenes, buildings, etc.; the Urban100 dataset contains 100 high-resolution images of buildings;
[0028] Among them, the low-resolution images used in training and testing are all produced using the bicubic downsampling function.
[0029] Furthermore, the derivation process of the pyramid attention module is as follows:
[0030]
[0031] Among them, K i Represents the size of the i-th convolution kernel, K i =2×(i+1)+1; x t Represents the intermediate feature map of the tth layer; Represents the convolution layer using grouped convolution, G i represents the group size of the i-th layer, F i Represents grouped feature maps of different scales; Z iRepresents the generated feature descriptor, δ represents the ReLU layer operation, S represents the number of groups divided according to the number of channels of the feature map, which is set to 4 here, and σ represents the Sigmoid activation function. represents the full convolution layer, H and W represent the length and width of the input image, and m and n represent F i The row and column indices of the feature map; H i Represents the feature map of recalibrated weights, exp(·) represents the exponential function with e as the base; ⊙ is the dot multiplication operation; τ(·) represents the scalar function used for normalization, represents the function used to calculate the correlation of regional features, θ(·) represents a feature conversion function that generates a new feature representation; Concat(·) is a fusion operation that fuses the recalibrated feature blocks of each group, and Out represents the final output of the attention module.
[0032] Furthermore, during the training phase, feature loss is calculated between the feature maps output by the module groups of corresponding sizes of the first network and the second network to update the network parameters. The training process is divided into two parts:
[0033] The high-resolution image is fed into the first network, where it is gradually scaled down to generate intermediate feature maps. This image is then fed into the second network, where it is gradually scaled up to generate intermediate feature maps. Feature maps of the same scale are computed using the L1 norm loss, and the network gradients are updated inversely. The resulting super-resolution image is also compared to the input high-resolution image using loss calculations (L1 loss) and metrics (PSNR / SSIM) to verify the network's generalization capabilities.
[0034] The low-resolution image corresponding to the high-resolution image is input into the second network, and the size is increased layer by layer to generate an intermediate feature map. The loss is calculated with the intermediate feature map of the same scale generated by the first network. Here, the mean square error loss (MSE) is used for calculation. The network gradient is updated in reverse to obtain the final reconstructed super-resolution image to verify the generalization ability of the network.
[0035] During the test and deduction phase, only low-resolution images are input into the second network to generate the final super-resolution images, and the generated results are evaluated using the full-reference image quality evaluation index (PSNR / SSIM).
[0036] The beneficial effects achieved by the present invention are as follows: this method updates and iterates the network parameters through the attention mechanism and loss function, greatly improving the network's generation and generalization capabilities, and improving the quality and effect of the final high-resolution image while ensuring the model training speed. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] Figure 12 is a network structure diagram of an image super-resolution method based on a pyramid attention mechanism and a symmetric network in an embodiment of the present invention.
[0038] Figure 2 This is a flow chart of an image super-resolution method based on a pyramid attention mechanism and a symmetric network in an embodiment of the present invention. DETAILED DESCRIPTION
[0039] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.
[0040] The following is combined with Figure 2 The technical solution of the present invention is further described in detail:
[0041] like Figure 2 As shown in the figure, when performing image super-resolution reconstruction based on the pyramid attention mechanism and symmetric network, a deep neural network is first built to perform the image reconstruction task, which mainly includes two parts: a pyramid attention module for enhancing the network feature extraction capability and an end-to-end symmetric network part for performing high-resolution and low-resolution image training and reconstruction tasks.
[0042] Specifically, the network as a whole can be divided into the first network and the second network. The first network contains three module groups from top to bottom. The first two module groups are composed of a pyramid attention module, a convolutional layer, a ReLU activation layer, and a maximum pooling layer. The image downsampling is implemented by the maximum pooling layer. The third module group is composed of a pyramid attention mechanism module, a convolutional layer, and a ReLU activation layer. The second network contains three module groups from bottom to top. The first module group is composed of a pyramid attention module, a ReLU activation layer, and a convolutional layer. The second two module groups are composed of a pyramid attention module, a ReLU activation layer, a convolutional layer, and an upper convolutional layer. The image upsampling is implemented by the upper convolutional layer. The network structure adopts a symmetrical series dual network. The network reduces the feature size and increases the number of feature channels at each stage.
[0043] The symmetric network consists of two networks connected in series, with the same number of network layers, convolution kernel size, pooling method, and padding step size, and with shared weights. The specific network structure and parameters are shown in Tables 1 and 2:
[0044] Table 1. First network structure and parameters
[0045]
[0046] Table 2 Second network structure and parameters
[0047]
[0048] In the table, Layer represents all network layers from the first convolutional layer Conv1 that receives the original high-resolution image input to the last pooling layer;
[0049] Size indicates the convolution kernel or filter size and number of channels of the current convolution layer or pooling layer; Stride indicates the filter step size of the current network layer; Padding indicates the Padding method used by the current network layer: Same method or Valid method.
[0050] When high-resolution images from the dataset are input into the first network, the high-resolution images are reduced in size by a factor of 2 layer by layer. When the intermediate feature maps are input into the second network, the low-resolution images are increased in size by a factor of 2 layer by layer. The feature dimensions and sizes of the outputs calculated at each layer are shown in Tables 3 and 4.
[0051] Table 3 Feature dimensions and sizes of each layer of the first network
[0052]
[0053] Table 4 Feature dimensions and sizes of each layer of the second network
[0054]
[0055] Among them, Layer represents all network layers in the symmetric network; Size represents the feature size and number of channels after the current convolutional layer or pooling layer is processed.
[0056] The pyramid attention module first performs group convolution on the input features according to the number of channels. After obtaining the grouped features, it uses the Squeeze-and-Excitation network to extract and activate the feature weights. After calculation through the Softmax normalization function, the attention weight of each group is obtained. After multiplying it with the feature value, a feature map with richer details can be obtained.
[0057] The pyramid attention mechanism can help the network more accurately obtain rich details of the image, improve the network's feature extraction ability and generative generalization ability. The derivation process of the pyramid attention module is as follows:
[0058]
[0059] Among them, K i Represents the size of the i-th convolution kernel, K i =2×(i+1)+1; x t Represents the intermediate feature map of the tth layer; Represents the convolution layer using grouped convolution, G i represents the group size of the i-th layer, Fi Represents grouped feature maps of different scales; Z i Represents the generated feature descriptor, δ represents the ReLU layer operation, S represents the number of groups divided according to the number of channels of the feature map, which is set to 4 here, and σ represents the Sigmoid activation function. represents the full convolution layer, H and W represent the length and width of the input image, and m and n represent F i The row and column indices of the feature map; H i Represents the feature map of recalibrated weights, exp(·) represents the exponential function with e as the base; ⊙ is the dot multiplication operation; τ(·) represents the scalar function used for normalization, represents the function used to calculate the correlation of regional features, θ(·) represents a feature conversion function that generates a new feature representation; Concat(·) is a fusion operation that fuses the recalibrated feature blocks of each group, and Out represents the final output of the attention module.
[0060] The following is a brief description of the network training process and test inference process. During the training phase, feature loss is calculated between the feature maps output by the module groups of corresponding sizes of the first network and the second network to update the network parameters. The training process is divided into two parts:
[0061] 1. A high-resolution image is fed into the first network, where it is gradually downsized to generate intermediate feature maps. This image is then fed into the second network, where it is gradually upsized to generate intermediate feature maps. Feature maps of the same scale are computed using the L1 norm loss, and the network gradients are updated inversely. The resulting super-resolution image is compared to the input high-resolution image using loss calculations (L1 loss) and metrics (PSNR / SSIM) to verify the network's generalization capabilities.
[0062] 2. Input the low-resolution image corresponding to the high-resolution image into the second network, increase the size layer by layer to generate intermediate feature maps, and calculate the loss with the intermediate feature maps of the same scale generated by the first network. Here, the mean squared error loss (MSE) is used for calculation, and the network gradient is updated in reverse to obtain the final reconstructed super-resolution image to verify the generalization ability of the network.
[0063] During the test and deduction phase, only low-resolution images are input into the second network to generate the final super-resolution images, and the generated results are evaluated using the full-reference image quality evaluation index (PSNR / SSIM).
[0064] After building a symmetric network for image super-resolution reconstruction, we began training the network to obtain good model weights. We used an end-to-end training method, and the image super-resolution training and test datasets used for training included:
[0065] The DIV2K dataset contains 800 natural scene training images and 100 validation images.
[0066] The Set5 dataset contains 5 high-resolution images of people, animals, and cartoons.
[0067] The Set14 dataset contains 14 high-resolution images including human figures, animals, cartoons, black and white images, and scene images.
[0068] The BSD100 dataset contains 100 high-resolution images of people, animals, scenes, buildings, etc.
[0069] The Urban100 dataset contains 100 high-resolution building images.
[0070] The following test experiments are conducted. From the methods proposed in recent years, a variety of image super-resolution methods based on deep features are selected. The PSNR and SSIM indicators of traditional methods, deep methods and the proposed method are compared on the Set5, Set14, BSD100 and Urban100 datasets. The experimental results of 4x super-resolution are shown in Table 5:
[0071] Table 5 Test set experimental results
[0072]
[0073]
[0074] Among them, the first column shows several typical super-resolution methods, and the first row shows four commonly used super-resolution test datasets. The peak signal-to-noise ratio (PSNR) is a full-reference image quality evaluation metric, that is, an error-sensitive image quality evaluation. The larger its value, the better the generated quality. Structural similarity (SSIM) is also a full-reference image quality evaluation metric, which measures image similarity from three aspects: brightness, contrast, and structure. The larger its value, the less distorted the generated image. As can be seen from the data in the table, the present invention outperforms most experimental comparison objects in terms of image generation quality, i.e., indicators, and the network structure complexity is much lower than other comparison methods, thus proving the effectiveness of the present invention.
[0075] The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiment. Any equivalent modifications or changes made by ordinary technicians in this field based on the contents disclosed in the present invention should be included in the protection scope recorded in the claims.
Claims
1. An image super-resolution method based on pyramid attention mechanism and symmetric network, characterized by: The following steps are involved: Step 1: For the data set to be processed, use the bicubic interpolation function to create a high-resolution image x HR Its corresponding low-resolution image x LR , the low resolutions are 2.0, 3.0, and 4.0, which are 2 times, 3 times, and 4 times smaller; Step 2: Transform the high-resolution image x HR The input is sent to the first part of the symmetric network for feature extraction. It first passes through the pyramid attention module to fuse the multi-scale information of the real image, and then passes through the convolution layer-ReLU activation layer-max pooling layer operation to generate an intermediate feature map. The whole process obtains the multi-level information distribution of the image and enhances the high-frequency detail information; The symmetric network in step 2 is composed of two networks connected in series with the same number of network layers, structure, convolution kernel size, pooling method, padding step size, and attention module and sharing weights; The first network structure and corresponding parameters are: Conv1+ReLU layer, size and number of channels 3*3*96, Padding: valid; MaxPool1 layer, size and number of channels 2*2, Padding: valid; Conv2+ReLU layer, size and number of channels 3*3*256, Padding: same; MaxPool2 layer, size and number of channels 2*2, Padding: same; Conv3+ReLU layer, size and number of channels 3*3*512, Padding: valid; the filter step size of each layer is 1; The second network structure and corresponding parameters are: Conv1+ReLU layer, size and number of channels 3*3*512, Padding: valid; Conv2+ReLU layer, size and number of channels 3*3*256, Padding: same; UpConv1 layer, size and number of channels 2*2, Padding: same; Conv3+ReLU layer, size and number of channels 3*3*96, Padding: valid; UpConv2 layer, size and number of channels 2*2, Padding: valid; the filter step size of each layer is 1; The derivation process of the pyramid attention module is as follows: Among them, K i represents the size of the i-th convolution kernel, K i =2×(i+1)+1; x t Represents the intermediate feature map of the tth layer; Represents a convolutional layer using grouped convolution, G i represents the group size of the i-th layer, F i Represents grouped feature maps of different scales; Z i represents the generated feature descriptor, δ represents the ReLU layer operation, S represents the number of groups divided according to the number of channels of the feature map, which is set to 4 here, and σ represents the Sigmoid activation function. represents the full convolutional layer, H and W represent the length and width of the input image, and m and n represent F i The row and column indices of the feature map; H i represents the feature map of recalibrated weights, exp(·) represents the exponential function with base e; ⊙ is the dot multiplication operation; τ(·) represents the scalar function used for normalization, represents a function used to calculate the correlation of regional features, θ(·) represents a feature conversion function that generates a new feature representation; Concat(·) is a fusion operation that fuses the recalibrated feature blocks of each group, and Out represents the final output of the attention module; Step 3: The intermediate feature map generated after convolution is compared with the low-resolution image to obtain the error loss. The loss is used to update the network parameters and gradients. The specific loss formula is as follows: In the formula, x t represents the intermediate feature map generated by the tth layer of the first network, y t represents the intermediate feature map generated by the tth layer of the second network, and N represents the number of pixels on the feature map; when calculating the intermediate feature map x3 generated by the first network and the input low-resolution image x LR When the loss value is between: Step 4: Input the intermediate feature map generated by the first network into the second network for reconstruction to obtain a reconstructed high-resolution image. With the input high-resolution image x HR Calculate the loss and update the network parameters and gradients in reverse. The specific formula for calculating the loss is as follows: In the formula Represents the high-resolution image reconstructed from the intermediate feature map generated by the first network; Step 5: Convert the low-resolution image x LR Input into the second part of the network, enlarge the low-resolution image resolution by resolution, and finally get the generated high-resolution image and with the high-resolution image x HR Calculate the loss and update the network parameters again; the specific formula for calculating the loss is as follows: In the formula Represents the low-resolution image x that the second network will input LR Reconstructed high-resolution image.
2. The image super-resolution method based on pyramid attention mechanism and symmetric network according to claim 1, characterized in that: When high-resolution images from the dataset are input into the first network, the size of the high-resolution images is reduced by a factor of 2 layer by layer; when the intermediate feature maps are input into the second network, the size of the low-resolution images is increased by a factor of 2 layer by layer.
3. The image super-resolution method based on pyramid attention mechanism and symmetric network according to claim 2, characterized in that: The feature dimensions and sizes after calculation and output of each layer of the first network are as follows: Conv1+ReLU layer outputs 480*480*96, MaxPool1 layer outputs 240*240*96, Conv2+ReLU layer outputs 240*240*256, MaxPool2 layer outputs 120*120*256, and Conv3+ReLU layer outputs 120*120*512.
4. The image super-resolution method based on pyramid attention mechanism and symmetric network according to claim 2, characterized in that: The feature dimensions and sizes after calculation and output of each layer of the second network are as follows: Conv1+ReLU layer outputs 120*120*512, Conv2+ReLU layer outputs 120*120*256, UpConv1 layer outputs 240*240*256, Conv3+ReLU layer outputs 240*240*96, and UpConv2 layer outputs 480*480*96.
5. The image super-resolution method based on pyramid attention mechanism and symmetric network according to claim 1, characterized in that: The detailed feature of step 3 is that the training data set and the test data set used for training include: The DIV2K dataset contains 800 natural scene training images and 100 validation images; the Set5 dataset contains 5 high-resolution images of people, animals, and cartoons; the Set14 dataset contains 14 high-resolution images of people, animals, cartoons, black and white, and scenes; the BSD100 dataset contains 100 high-resolution images of people, animals, scenes, and buildings; the Urban100 dataset contains 100 high-resolution images of buildings; Among them, the low-resolution images used in training and testing are all produced using the bicubic downsampling function.
6. The image super-resolution method based on pyramid attention mechanism and symmetric network according to claim 1, characterized in that: During the training phase, feature loss is calculated between the feature maps output by the module groups of corresponding sizes of the first network and the second network to update the network parameters. The training process is divided into two parts: Input the high-resolution image into the first network, reduce the size layer by layer to generate an intermediate feature map, and input it into the second network, increase the size layer by layer to generate an intermediate feature map; feature maps of the same scale are calculated using the L1 norm loss, and the network gradient is updated in reverse; the final super-resolution image and the input high-resolution image are also subjected to loss calculation and indicator evaluation to verify the generalization ability of the network; Input the low-resolution image corresponding to the high-resolution image into the second network, increase the size layer by layer to generate an intermediate feature map, and calculate the loss with the intermediate feature map of the same scale generated by the first network. Here, the mean square error loss is used for calculation, and the network gradient is updated in reverse to obtain the final reconstructed super-resolution image to verify the generalization ability of the network; During the test and deduction phase, only the low-resolution image is input into the second network to generate the final super-resolution image, and the generated result is evaluated using the full reference image quality evaluation index.
