A single image super-resolution method based on multi-angle knowledge distillation

Abstract

The application discloses a single image super-resolution method based on multi-angle knowledge distillation, which comprises the following steps: preprocessing a data set; training a teacher network by using the data set, and obtaining a teacher super-resolution image; inputting a low-resolution image into a student network to obtain a student super-resolution image; extracting knowledge of the teacher network and the student network from two angles of inter-channel and intra-channel; calculating pixel L1 loss between the student super-resolution image and a real high-resolution image, pixel L1 loss between the student super-resolution image and the teacher super-resolution image, and L1 loss between the knowledge extracted by the teacher network and the student network, updating parameters through back propagation to obtain a final student network; and inputting the low-resolution image into the final student network to reconstruct a high-resolution image. The application can more efficiently extract knowledge from the teacher network, further improve the performance of the student network, and reconstruct a clearer image.

Description

Technical Field

[0001] This invention relates to a single-image super-resolution method based on multi-angle knowledge distillation, belonging to the field of computer vision image super-resolution. Background Technology

[0002] Single-image super-resolution (SR) technology refers to reconstructing a high-resolution (HR) image from a single low-resolution (LR) image. It is a fundamental task in computer vision and plays a crucial role in fields such as medical imaging and video surveillance. With the rise of deep learning, the continuous development of hardware resources such as GPUs, and the emergence of massive datasets, Convolutional Neural Networks (CNNs) have been widely applied to image super-resolution, achieving significant performance improvements compared to traditional super-resolution methods. Most super-resolution networks tend to explore complex and sophisticated network structures to improve network performance; however, complex networks are often difficult to deploy on hardware devices due to their large number of parameters and high computational complexity.

[0003] Knowledge distillation is an effective model compression technique, typically consisting of four parts: a teacher network, knowledge, a distillation algorithm, and a student network. It transfers knowledge from a heavyweight, high-performance teacher network to a lightweight student network to guide student learning, improving the performance of the lightweight student network without introducing additional parameters. Research shows that in the knowledge distillation framework, the performance of the teacher network and the representation of knowledge are crucial to the performance of the lightweight student network. The better the performance of the teacher network, the richer the information contained in its intermediate feature maps, thus ensuring the quality of the knowledge source. Knowledge representation refers to what knowledge is extracted from the intermediate feature maps of the teacher network, which is crucial for knowledge transfer and the student network. Based on the above description, this invention proposes a single-image super-resolution algorithm based on multi-angle knowledge distillation. Summary of the Invention

[0004] Purpose of the invention: To address the problem of inefficient knowledge extraction from teacher networks in knowledge distillation, this invention proposes a single-image super-resolution method based on multi-angle knowledge distillation. Specifically, it extracts knowledge from the teacher network from two perspectives: inter-channel and intra-channel. Inter-channel extraction extracts attention information from the intermediate feature maps of the teacher network, while intra-channel extraction extracts activation information from the intermediate feature maps of the teacher network. This is known as Attention and Activation Information Knowledge Distillation (AAKD).

[0005] Technical solution: To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A single-image super-resolution method based on multi-angle knowledge distillation includes the following steps:

[0007] Preprocess the dataset;

[0008] The teacher network was trained using the dataset, and super-resolution images of the teachers were obtained.

[0009] Inputting a low-resolution image into the student network yields a high-resolution image of the student.

[0010] Knowledge of the teacher network and student network is extracted from two perspectives: between channels and within channels.

[0011] Calculate the loss of the student network, update the parameters through backpropagation, and train to obtain the final student network;

[0012] The low-resolution image is fed into the final student network to reconstruct the final high-resolution image.

[0013] Preferably, the teacher network is a dual-attention feature fusion network comprising a shallow feature extraction module, a deep feature extraction module, an upsampling module, and a reconstruction module; the shallow feature extraction module is used to extract shallow features of the image; the deep feature extraction module includes a dual-attention module, a residual-within-residual structure, and a feature fusion structure, used to extract deep features from the shallow features, and add the shallow features and deep features to obtain the final features; the upsampling module is used to obtain upsampled features based on the final features; the reconstruction module is used to obtain a super-resolution image based on the upsampled features; the dual-attention feature fusion network uses a pixel L1 loss function to train the network.

[0014] Preferably, the student network adopts a dual-attention feature fusion network architecture, and the number of residual blocks in each residual group in the student network is 30% of that in the teacher network, so as to achieve a lightweight network.

[0015] Preferably, knowledge is extracted from both inter-channel and intra-channel perspectives. Inter-channel information is used to extract attention information from feature maps, while intra-channel information is used to extract activation information from feature maps.

[0016] Preferably, the dual attention module of the deep feature extraction module is composed of spatial attention and channel attention connected in series;

[0017] Output feature Z of the spatial attention module s Represented as:

[0018]

[0019] Where X represents the input feature, Conv1(·) and Conv2(·) represent convolutional layers used to adjust the number of channels, ReLU(·) represents a modified linear unit used to add non-linearity to the attention function, and sigmoid(·) represents an activation function used to restrict the weight coefficients to the range of 0-1. This represents the element-wise multiplication operation;

[0020] Output feature Z of the channel attention module c Represented as:

[0021]

[0022] Where pool(·) represents a global average pooling layer, used to compress the spatial dimension of features;

[0023] The output feature Z of the dual attention module is represented as:

[0024] Z = F channel (F spatial (X))

[0025] Among them, F chann (·) indicates the channel attention module, F spatial (·) indicates the spatial attention module.

[0026] Preferably, the residual nested residual structure of the deep feature extraction module consists of multiple residual groups and global skip connections. Each residual group consists of multiple residual dual-attention blocks and local skip connections. Each residual dual-attention block is composed of ordinary residual blocks and dual-attention modules. The output feature Z of the residual dual-attention block is represented as:

[0027] Z=DA(Conv2(relu(Conv1(X))))+X

[0028] Where DA(·) represents a dual attention module, Conv1(·) and Conv2(·) represent convolutional layers, and relu(·) represents a modified linear unit.

[0029] Preferably, the feature fusion structure of the deep feature extraction module consists of the outputs of multiple residual groups and a dual attention module; the feature fusion structure concatenates the outputs of multiple residual groups in the channel direction to obtain fused features, and the dual attention module adaptively learns weights to readjust the fused features.

[0030] Preferably, the attention information of the teacher network feature map is represented as follows: assuming the intermediate feature is... Where C, H, and W represent the number of channels, the height of the spatial dimension, and the width of the spatial dimension, respectively. Let X represent the spatial feature of the i-th channel of the intermediate feature. The spatial attention calculation process for the intermediate feature is as follows: first, square the channel values ​​at each spatial location of the intermediate feature, then sum them, compressing them to a single channel dimension to obtain the two-dimensional spatial feature. Finally, the spatial attention is obtained by normalizing the two-dimensional spatial feature Q.

[0031] Preferably, the activation information of the teacher network feature graph is represented as follows: Indicate intermediate features The spatial characteristics of the i-th channel, and the negative value m of the i-th channel. i Represented by the expected value of all negative values ​​within this channel:

[0032] m i =E[X i |X i <0]

[0033] We get a negative value m i Then, use m i Fill the entire H×W space to obtain the same X i Negative tensors of the same size Then, spatial feature X i With negative tensor M i Perform a maximization operation to obtain the transformation of the channel before feature transfer; apply the above calculation process to each channel of the intermediate feature X to obtain a negative value Zhangdu with the same magnitude as X. Then, the intermediate feature X is maximized by the negative tensor M of the same size to obtain the intermediate feature transferred from the teacher network to the student network.

[0034] Preferably, the total loss function for student network training is expressed as the sum of real high-resolution image supervision loss, teacher network supervision loss, attention transfer loss, and activation information transfer loss.

[0035] Beneficial effects:

[0036] 1. The teacher network proposed in this invention combines spatial attention and channel attention mechanisms to adaptively adjust features, making the network pay more attention to spatial regions and channels rich in high-frequency information, thereby enhancing the network's representation ability; a feature fusion module based on a dual attention mechanism is used to retain low-level features and pass them to higher levels, and then all hierarchical features are fused together, and the dual attention mechanism is used to adaptively adjust features at different levels, making full use of the network's hierarchical features; ultimately, more details can be reconstructed.

[0037] 2. The knowledge representation method proposed in this invention considers the feature information flow mechanism of neural networks, extracting information from the teacher network from both inter-channel and intra-channel perspectives. Inter-channel extraction extracts attention information, while intra-channel extraction extracts activation information. Compared to other knowledge representation methods, the proposed method recovers more details and achieves greater performance improvement. Attached Figure Description

[0038] Figure 1 This is a diagram of the DAFFN framework.

[0039] Figure 2 This is a diagram of spatial attention structure.

[0040] Figure 3 This is a diagram of the channel attention structure.

[0041] Figure 4 This is a diagram of the residual double attention block structure.

[0042] Figure 5 This is the AAKD architecture diagram.

[0043] Figure 6 The results show quantitative comparisons of PSNR / SSIM between AAKD and other methods on four benchmark datasets.

[0044] Figure 7 The qualitative comparison results of AAKD with other methods are shown at four times magnification on image 92 of the Urban100 dataset. Detailed Implementation

[0045] The method described in this invention is a single-image super-resolution method based on multi-angle knowledge distillation. By inputting a low-resolution image into a trained student network, a high-resolution image can be reconstructed. The specific steps include:

[0046] (1) Preprocessing the dataset: First, the original image is randomly rotated (90°, 180°, 270°) and horizontally flipped to achieve data augmentation;

[0047] (2) Train the teacher network using the dataset and obtain the teacher super-resolution image (TSR);

[0048] (3) Input low-resolution images into the student network to obtain student super-resolution (SSR) images;

[0049] (4) Extract knowledge of the teacher network and student network from the perspectives of inter-channel and intra-channel respectively;

[0050] (5) Calculate the pixel L1 loss between the student super-resolution (SSR) image and the real high resolution (HR) image, the pixel L1 loss between the student super-resolution (SSR) image and the teacher super-resolution (TSR) image, and the L1 loss between the knowledge extracted from the teacher network and the student network. Accumulate all losses and train the final student network by updating the parameters through backpropagation.

[0051] (6) Input the low-resolution image into the final student network to reconstruct the final high-resolution image.

[0052] Furthermore, the teacher network in step (2) adopts a Dual Attention Feature Fusion Network (DAFFN), which consists of four parts: a shallow feature extraction module, a deep feature extraction module, an upsampling module, and a reconstruction module. Its structure diagram is shown below. Figure 1 As shown. Given a low-resolution image I LR First, extract I using the shallow feature extraction module. LR shallow features F s :

[0053] F s =H SF (I LR )

[0054] Where H SF This represents a shallow feature extraction module, consisting of a single convolutional layer. Then, F is extracted using a deep feature extraction module. s Deep features F in D :

[0055] F D =H DF (F s )

[0056] Where H DF This represents a deep feature extraction module, comprising a dual attention module, a residual nested residual structure, and a feature fusion structure. The residual nested residual structure consists of multiple residual groups and global skip connections. Each residual group consists of multiple residual dual attention blocks and local skip connections. Each residual dual attention block consists of ordinary residual blocks and a dual attention mechanism. Then, through the global skip connections of the residual nested residual structure, the shallow features F are fused. s and deep features F D Adding them together yields the final feature F extracted from the network. F :

[0057] F F =F s +F D

[0058] Next, the final feature F F The upsampling feature F is obtained through the upsampling module. UP :

[0059] F UP =H UP (F F )

[0060] Where H UP This represents the upsampling module, which consists of sub-pixel convolutional layers. Finally, the upsampled feature F... UP The super-resolution image I is obtained through the reconstruction module. SR :

[0061] I SR =H R (F UP )=H DAFFN (I LR )

[0062] Where H R The reconstruction module consists of a single convolutional layer, H. DAFFN This represents the DAFFN super-resolution algorithm. DAFFN uses a pixel-based L1 loss function to train the network. It is given a training dataset with N pairs of low-resolution and high-resolution images. The pixel L1 loss function can be expressed as:

[0063]

[0064] Where Θ represents the training parameters of DAFFN, and ||·||1 represents the L1 norm. The training objective of DAFFN is to minimize the pixel L1 loss function.

[0065] Furthermore, the dual attention (DA) module of the DAFFN deep feature extraction module is composed of spatial attention (SA) and channel attention (CA) cascaded together. This allows for adaptive adjustment of features between channels and different spatial regions, thereby enhancing the network's representational capabilities. SA focuses on the contribution of features from different spatial regions to high-frequency detail recovery. Spatial regions containing more high-frequency information such as edges and textures should receive more attention. Therefore, the key to SA is generating weight coefficients for different spatial regions. Its structure diagram is shown below. Figure 2 As shown. Assuming the input feature of the SA module is X, then the output feature of the SA module is Z. s It can be represented as:

[0066]

[0067] Where Conv1(·) and Conv2(·) represent convolutional layers used to adjust the number of channels, ReLU(·) represents a rectified linear unit used to add non-linearity to the attention layer, and sigmoid(·) represents an activation function used to restrict the weight coefficients to the range of 0-1. This represents an element-wise multiplication operation. CA focuses on the contribution of different channel characteristics to high-frequency detail recovery; more important channels should receive more attention. Therefore, the key to CA is generating weighting coefficients for different channels, as shown in the diagram below. Figure 3 As shown. Assuming the input feature of the CA module is X, then the output feature of the CA module is Z. c It can be represented as:

[0068]

[0069] Where pool(·) represents a global average pooling layer, used to compress the spatial dimension of features. Assuming the input feature of the DA module is X, the output feature Z of the DA module can be expressed as:

[0070] Z = F channel (F spatial (X))

[0071] Where F channel (·) indicates the CA module, F spatial (·) indicates the SA module.

[0072] Furthermore, the Residual-within-Residual (RIR) structure of the DAFFN deep feature extraction module consists of multiple Residual Groups (RGs) and Global Skip Connections (GSCs). Each Residual Group consists of multiple Residual Double Attention Blocks (RDABs) and Local Skip Connections (LSCs). Each RDAB consists of ordinary residual blocks and double attention modules. The RDAB structure diagram is shown below. Figure 4 As shown, assuming the input feature is X, then the output feature Z of RDAB can be expressed as:

[0073] Z=DA(Conv2(relu(Conv1(X))))+X

[0074] Where DA(·) represents the dual attention module. In the RIR structure, GSC allows the network to bypass the rich low-frequency information in the image at a coarse level, while LSC in RG further bypasses low-frequency information at a fine level, allowing the network to focus more on learning high-frequency information. This greatly simplifies the information flow between RGs and also helps the network avoid the gradient vanishing problem during training. The RIR structure allows training very deep networks to expand the receptive field, while the dual attention modules in the stacked RDAB layers help the network extract more important information, thereby enhancing the network's representational ability.

[0075] Furthermore, the feature fusion (FF) structure of the DAFFN deep feature extraction module consists of the outputs of multiple RGs and a DA module. First, the FF concatenates the outputs of multiple RGs along the channel direction to obtain fused features. Then, the fused features pass through the DA module, which adaptively learns weights to readjust the fused features. In this way, the FF can effectively preserve low-level features and transmit them to higher levels, and due to the DA module, the network pays more attention to hierarchical features rich in high-frequency information.

[0076] Furthermore, in step (3), the student network also adopts the DAFFN framework described above. Unlike the teacher network, the number of residual blocks in each residual group in the student network is 30% of that in the teacher network, in order to achieve a lightweight network.

[0077] Furthermore, in step (4), knowledge is extracted from both inter-channel and intra-channel perspectives. Feature information flow is the core of convolutional neural networks, and it operates at both inter-channel and intra-channel levels. Inspired by this, knowledge is extracted from the intermediate feature map of the teacher network from both inter-channel and intra-channel perspectives. Specifically, attention information of the intermediate feature map of the teacher network is extracted from the inter-channel perspective, and activation information of the intermediate feature map of the teacher network is extracted from the intra-channel perspective.

[0078] Attention information: Assuming the intermediate feature is Where C, H, and W represent the number of channels, the height of the spatial dimension, and the width of the spatial dimension, respectively. Let X represent the spatial feature of the i-th channel of the intermediate feature. The spatial attention calculation process for the intermediate feature is as follows: first, square the channel values ​​at each spatial location of the intermediate feature, then sum them to compress it to a single channel dimension, thus obtaining a two-dimensional spatial feature. Finally, the spatial attention is obtained by normalizing the two-dimensional spatial feature Q. The mathematical representation of this process is as follows:

[0079]

[0080]

[0081] Activation information: Activation information consists of activation response values ​​and activation states. Positive activation response values ​​are beneficial to the network and should be transferred to the student network, while negative activation response values ​​are detrimental and should not be transferred to the student network. To ensure the student network obtains the same activation state as the teacher network, the teacher network should generate a negative value for negative activation response values ​​for the student network to learn. Here, the negative value is the expected value of all negative values ​​in a single feature map. Assume... Indicate intermediate features If the spatial characteristics of the i-th channel are given, then the negative value m of the i-th channel is... i Expressed mathematically as the expected value of all negative values ​​within the channel, it can be represented as follows:

[0082] m i =E[X i |X i <0]

[0083] We get a negative value m i Then, use m i Fill the entire H×W space to obtain the same X i Negative tensors of the same size Then, spatial feature X i With negative tensor M i Performing a maximization operation yields the transformation of that channel before feature transfer. Applying this calculation process to each channel of the intermediate feature X results in a negative tensor of the same size as X. By maximizing the intermediate feature X with a negative tensor M of the same size, we can obtain the intermediate features transferred from the teacher network to the student network.

[0084] Y = max(M, X)

[0085] Furthermore, the loss function of the student network in step (5) consists of four parts, namely the pixel loss between the student SR (SSR) image and the HR image. Pixel loss between SSR images and Teacher SR (TSR) images Loss of attention Activation information transfer loss

[0086] Pixel loss between SSR and HR images This represents the supervision of the student network by the HR image. In SR tasks, the L1 loss function often achieves better performance and convergence compared to the L2 loss function. The L1 loss function is used to calculate...

[0087]

[0088] Pixel loss between SSR and TSR images This represents the supervision of the student network by the teacher network. Similarly, the L1 loss function is used to calculate...

[0089]

[0090] Suppose the intermediate features of the teacher network and the student network are represented by T and S, respectively. Attention shift loss. Spatial attention, representing the intermediate features of student-teacher network learning, is used to transfer information between channels. It is calculated as the difference in spatial attention between T and S, using the following formula:

[0091]

[0092] Activation information transfer loss This represents the activation information of intermediate features after the transformation of the student's online learning and the teacher's online learning, used to transfer information within the channel. The calculation formula is as follows:

[0093]

[0094] The overall loss function of the student network This can be expressed as the sum of HR supervision loss, teacher supervision loss, attention shift loss, and activation information shift loss:

[0095]

[0096] Where α, β, γ, and δ represent the weights of the four loss functions.

[0097] This embodiment uses the DIV2K dataset for training and four benchmark datasets for testing: Set5, Set14, Urban100, and BSD100. The entire AAKD method is trained in two phases: first, the teacher network is trained, and then the trained teacher network guides the student network training. The AAKD structure diagram is shown below. Figure 5 As shown.

[0098] Teacher Network Training: The teacher network employs the DAFFN architecture, trained on the PyTorch framework using an NVIDIA GTX 1080 Ti GPU for acceleration. In the experiments, data augmentation was achieved by randomly rotating the original images (90°, 180°, 270°) and horizontally flipping them. During training, all images were preprocessed by subtracting the average RGB values ​​from the DIV2K dataset. In each training batch, 16 low-resolution image patches of size 48×48 were randomly cropped from the low-resolution images and used as input to the network. The teacher DAFFN was trained using the Adam optimizer, with parameters β1 set to 0.9, β2 set to 0.999, and ε set to 10. -8 The learning rate is initialized to 10. -4 The total training epochs are 150, with the learning rate halved after 100 training epochs. In the teacher DAFFN, the number of residual groups G is set to 10, and the number of residual dual attention blocks B in each residual group is set to 18. The kernel size of the convolutional layers in the channel attention module is 1×1, while the kernel size of the remaining convolutional layers in the network is 3×3, with zero padding to ensure that the output and input feature sizes remain unchanged. The number of filters in convolutional layer 1 of the spatial attention module is set to 128, the number of filters in convolutional layer 2 is set to 1, the number of filters in convolutional layer 1 of the channel attention module is set to 4, and the number of filters in the sub-pixel convolutional layers of the upsampling module is set to 64*s. 2 Where 's' is the amplification factor, the number of filters in the convolutional layers of the reconstruction module is set to 3, and the number of filters in the remaining convolutional layers is set to 64. The activation functions used are ReLU and Sigmoid. The pixel L1 loss function is used to train the network. Finally, the RGB images generated by the model are converted to the YCbCr space, and the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) are quantitatively evaluated on the Y channel. Both PSNR and SSIM are taken as the mean of all image results in the dataset. Generally, higher PSNR and SSIM values ​​indicate better image quality and better model performance.

[0099] Student network training: The student network also adopts the DAFFN architecture, with 10 residual groups and 6 residual blocks. Other parameter settings are the same as those for the teacher network training. Attention distillation and activation information distillation are performed on the 2nd, 6th, and 10th feature maps of the teacher network and the corresponding student network.

[0100] Figure 6This paper presents quantitative comparisons of PSNR / SSIM between AAKD and other methods on four benchmark datasets under different magnification factors. Bicubic is a traditional bicubic upsampling interpolation method; NoKD is a student network without knowledge distillation; FSP transfers the relationships between different layers as knowledge to the student network; FitNet transfers the feature map of any intermediate layer as knowledge to the student network; AT transfers the attention map of the intermediate feature map as knowledge to the student network; MR transfers the activation information of the intermediate feature map as knowledge to the student network; FAKD transfers the spatial affinity matrix of the intermediate feature map as knowledge to the student network; AAKD, proposed in this invention, transfers the attention map and activation information of the intermediate feature map as knowledge to the student network; DAFFN represents the teacher network. Figure 6 It can be seen that among various knowledge distillation methods, AAKD achieves the best performance regardless of the magnification factor or the dataset. Furthermore, a comparison between NoKD and AAKD shows that knowledge distillation significantly improves the performance of student networks. Figure 7 The image shows qualitative results of AAKD and other methods magnified 4 times on image 92 of the Urban100 dataset. It can be seen that AAKD reconstructs the image better.

Claims

1. A single-image super-resolution method based on multi-angle knowledge distillation, characterized in that, Includes the following steps: Preprocess the dataset; The teacher network was trained using the dataset, and super-resolution images of the teachers were obtained. Inputting a low-resolution image into the student network yields a high-resolution image of the student. Knowledge of the teacher network and student network is extracted from two perspectives: between channels and within channels. Calculate the loss of the student network, update the parameters through backpropagation, and train to obtain the final student network; The low-resolution image is fed into the final student network to reconstruct the final high-resolution image; Knowledge is extracted comprehensively from both inter-channel and intra-channel perspectives. Inter-channel information is used to extract attention information from feature maps, while intra-channel information is used to extract activation information from feature maps. The attention information of the teacher network feature map is represented as follows: assuming the intermediate feature is... Where C, H, and W represent the number of channels, the height of the spatial dimension, and the width of the spatial dimension, respectively. This represents the spatial feature of the i-th channel of intermediate feature X; The spatial attention calculation process for intermediate features is as follows: First, square the channel values ​​at each spatial location of the intermediate feature, then sum them to compress it to a single channel dimension, thus obtaining a two-dimensional spatial feature. Finally, the spatial attention is obtained by normalizing the two-dimensional spatial features Q. ; The activation information of the teacher network feature graph is represented as: Indicate intermediate features The spatial characteristics of the i-th channel, and the negative value m of the i-th channel. i Represented by the expected value of all negative values ​​within this channel: ； We get a negative value m i Then, use m i Fill the entire H×W space to obtain the same X i Negative tensors of the same size Then, the spatial feature X i With negative tensor M i Perform a maximization operation to obtain the transformation of the channel before feature transfer; Applying the above calculation process to each channel of the intermediate feature X yields a negative tensor of the same size as X. Then, the intermediate feature X is maximized along with a negative tensor M of the same size to obtain the intermediate features transferred from the teacher network to the student network. ; The teacher network is a dual-attention feature fusion network comprising a shallow feature extraction module, a deep feature extraction module, an upsampling module, and a reconstruction module. The shallow feature extraction module extracts shallow features from the image. The deep feature extraction module includes a dual-attention module, a residual-within-residual structure, and a feature fusion structure, used to extract deep features from the shallow features and add the shallow and deep features to obtain the final features. The upsampling module obtains upsampled features based on the final features. The reconstruction module obtains a super-resolution image based on the upsampled features. The dual-attention feature fusion network uses a pixel L1 loss function for training. The dual attention module of the deep feature extraction module is composed of spatial attention and channel attention connected in series; Output feature Z of the spatial attention module s Represented as: ； Where X is the input feature. and This indicates a convolutional layer, used to adjust the number of channels. This represents a modified linear unit, used to add nonlinear capabilities to the attention mechanism. This represents the activation function, used to restrict the weight coefficients to the range of 0-1. This represents the element-wise multiplication operation; Output feature Z of the channel attention module c Represented as: ； in, This represents a global average pooling layer, used to compress the spatial dimension of features; The output feature Z of the dual attention module is represented as: ； in, This indicates the channel attention module. This represents a spatial attention module; The residual nested residual structure of the deep feature extraction module consists of multiple residual groups and global skip connections. Each residual group consists of multiple residual dual-attention blocks and local skip connections. Each residual dual-attention block is composed of ordinary residual blocks and dual-attention modules. The output feature Z of the residual dual-attention block is represented as: ； in This indicates a dual attention module. and Represents a convolutional layer This indicates a modified linear unit.

2. The single-image super-resolution method based on multi-angle knowledge distillation according to claim 1, characterized in that, The student network adopts a dual-attention feature fusion network architecture, and the number of residual blocks in each residual group in the student network is 30% of that in the teacher network, in order to achieve a lightweight network.

3. The single-image super-resolution method based on multi-angle knowledge distillation according to claim 1, characterized in that, The feature fusion structure of the deep feature extraction module consists of the outputs of multiple residual groups and a dual attention module. The feature fusion structure concatenates the outputs of multiple residual groups in the channel direction to obtain fused features, and the dual attention module adaptively learns weights to readjust the fused features.

4. The single-image super-resolution method based on multi-angle knowledge distillation according to claim 1, characterized in that, The total loss function for student network training is expressed as the sum of real high-resolution image supervision loss, teacher network supervision loss, attention transfer loss, and activation information transfer loss.

Images