A super-resolution image reconstruction method for low-resolution blurred scenes

Abstract

The present invention discloses an image super-resolution reconstruction method for low-resolution blurred scenes, and belongs to the field of image super-resolution reconstruction. The present invention proposes a cross-scale feature projection network, and applies it to the single image super-resolution reconstruction task. A cross-scale local feature projection module is proposed to extract and fuse feature information of multiple images of different scales to reduce the loss of context information in network transmission. A cross-scale global feature projection module is proposed to further learn deep feature information of different scales through different receptive fields. The modules are connected in a distillation manner to reduce the number of parameters while ensuring the performance of network reconstruction. After extensive experimental comparison, the algorithm proposed in the present invention can achieve good reconstruction effects when reconstructing low-resolution images of different zoom factors, and the reconstructed images are of high quality, which proves the robustness of the network.

Description

Technical Field

[0001] The invention relates to the field of image super-resolution reconstruction, and in particular to an image super-resolution reconstruction method for low-resolution blurred scenes. Background Art

[0002] Image super-resolution reconstruction algorithms are widely used in various fields, such as surveillance video processing, medical image processing, satellite image processing, etc., to improve image clarity, details and recognizability.

[0003] Therefore, image super-resolution reconstruction algorithms for blurred images play an important role in different application scenarios, bringing significant effects and application potential in improving image quality, enhancing detail restoration capabilities, and enhancing image recognition and analysis.

[0004] At present, the main super-resolution reconstruction methods include interpolation, reconstruction and learning. The interpolation method generally uses bicubic interpolation and geometric feature-based interpolation. Although it is simple, it is easy to produce image oscillation and artifacts. In contrast, reconstruction methods such as maximum a posteriori estimation and iterative back projection have improved the reconstruction effect, but the computational efficiency is low and is greatly affected by the magnification factor. As for learning methods, such as neighborhood embedding and sparse representation, HR image reconstruction is achieved by learning the mapping function of HR image and LR image sample pairs or using the internal similarity of images.

[0005] In recent years, with the rapid development of deep learning technology, convolutional neural networks (CNNs) have been widely used in super-resolution tasks, and a series of efficient image super-resolution networks have been proposed. SR methods based on deep learning have achieved significant performance improvements because we can use a large amount of data to train larger or deeper neural networks for image SR. For example, RCAN (Residual Channel Attention Networks) is a representative CNN-based image SR network with 15.59M parameters and more than 400 layers in depth. One of the most significant drawbacks of these large models is that they require high computational costs, which makes them very challenging to deploy.

[0006] In summary, the problems existing in the prior art are:

[0007] (1) End-to-end image super-resolution networks often refine image features by deepening feature extraction modules. However, this excessive stacking of modules will greatly increase the number of network parameters and computational cost, and lead to redundant calculations.

[0008] (2) Currently, most end-to-end networks only use direct connections to design the network, which greatly reduces the utilization rate of features. Although direct connections can show the dependencies between features, their mapping relationships are too simple and there is also information redundancy.

[0009] (3) In some cases, the details in the image may not be important for feature extraction and reconstruction. Existing algorithms find it difficult to remove some unnecessary detail information, and the robustness and generalization ability of the algorithms are low. Summary of the invention

[0010] The technical problem to be solved by the present invention is to provide an image super-resolution reconstruction method for low-resolution blurred scenes in view of the deficiencies of the above-mentioned prior art.

[0011] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0012] An image super-resolution reconstruction method for low-resolution blurred scenes, comprising the following steps: S1, designing an image super-resolution reconstruction model framework for low-resolution blurred scenes, wherein the image super-resolution reconstruction model framework for low-resolution blurred scenes comprises an image shallow feature extraction module, an image deep feature extraction module and an image reconstruction module;

[0013] S2, using an image shallow feature extraction module to extract shallow features of the image;

[0014] S3, extracting deep features of the image using an image deep feature extraction module; the image deep feature extraction module includes a cross-scale local feature projection module, a difference feature extraction module and a cross-scale global feature projection module, and the modules are connected in a knowledge distillation manner;

[0015] S4. Reconstruct the image to obtain a reconstructed super-resolution image.

[0016] As a further preferred embodiment of the present invention, the overall relationship between the image shallow feature extraction module, the image deep feature extraction module and the image reconstruction module is as follows:

[0017] F s =H conv3 (I LR ) (1)

[0018] F1=H FBB (F S ) (2)

[0019] F2=H FBB (F1) (3)

[0020] F3=H FBB(F2) (4)

[0021] I SR =H conv3 (H upsample (F3⊙(H conv1 (F s ))⊙(H conv1 (F1))⊙(H conv1 (F2)))) (5)

[0022] Among them, I LR Represents a low-resolution image, H conv3 (·) indicates a convolution with a kernel size of 3*3, H conv1 (·) indicates a convolution with a kernel size of 1*1, F s Represents a low-resolution image I LR The shallow features obtained after a 3×3 convolution; H FBB (·) represents the feature basic block module, F n＝1,2,3 Represents the shallow feature F s The deep features obtained by the FBB module; ⊙ is the feature concatenation operation, I SR is the final reconstructed super-resolution image.

[0023] As a further preferred embodiment of the present invention, the cross-scale local feature projection module downsamples the input features by four times and two times respectively to complete the extraction of multi-scale features, and selects an alternating downsampling and feature extraction method in the model structure.

[0024] As a further preferred embodiment of the present invention, the difference feature extraction module learns the differences between different features and performs feature fusion through residual connection, and the relationship between the difference feature extraction module and the cross-scale local feature projection module is as follows:

[0025] X scale＝1 =GELU(DFB(H conv1 (X))) (6)

[0026] X scale＝2 =GELU(DFB(H d (X))) (7)

[0027] X scale＝4 =GELU(DFB(H d (X))) (8)

[0028] X out =X scale＝1 |H u (X scale＝2 )|H u (X scale＝4 ) (9)

[0029] Among them, X is the given input feature, DFB is the difference feature extraction module, H d is the downsampling achieved by the maximum pooling convolution, X scale＝1,2,4 H is the output feature after DFB processing at different scales. u is the upsampling achieved by nearest neighbor interpolation, | is the feature concatenation method, X out is the final output of the cross-scale local feature projection module; GELU is the activation function.

[0030] As a further preferred embodiment of the present invention, the cross-scale global feature projection module designs a self-calibrated convolutional attention composed of a depth-separable convolution and an activation function, and the features selectively learned by the attention are then processed by convolution, including the following algorithm;

[0031] X'=H conv1 (X) (10)

[0032] W = Sigmoid(H conv1 (H dwconv7 (H conv1 (X')))) (11)

[0033]

[0034] X b =GELU(H dwconv3 (H conv1 (X'))) (13)

[0035] X out =X a +X b (14)

[0036] Among them, X' is the feature to be input into the two branches after being processed by the convolution layer, and H conv1 is a convolutional layer with a 1×1 convolution kernel, H dwconv7 is the depth-separable convolution of the 7×7 convolution kernel, Sigmoid is the activation function Sigmoid, W is the feature weight output by the Sigmoid function, H dwconv3 is a depth-separable convolution with a 3×3 convolution kernel, X a is the characteristic of the main output, X b is the feature of the branch output, GELU is the activation function, X out The feature is the final output feature after adding the features of the two branches of this module.

[0037] As a further preferred embodiment of the present invention, the image shallow feature extraction module is a convolution with a convolution kernel size of 3×3.

[0038] As a further preference of the present invention, the three feature basic block modules of the image deep feature extraction module, namely, the image deep feature extraction module includes a cross-scale local feature projection module, a difference feature extraction module and a cross-scale global feature projection module, each of which is composed of four feature fusion modules and one attention module.

[0039] As a further preference of the present invention, the image reconstruction module consists of a sub-pixel convolution layer and a convolution with a convolution kernel size of 3×3. The deep features of the image and the three-layer distillation features are fused through a feature splicing operation, and finally the reconstructed super-resolution image is obtained after the sub-pixel convolution layer and the 3×3 convolution.

[0040] The present invention has the following beneficial effects:

[0041] While keeping the model lightweight, the present invention uses a multi-scale sampling network to extract more accurate low-frequency features. First, an image super-resolution reconstruction model framework for low-resolution blurred scenes is constructed, including image shallow feature extraction, image deep feature extraction, and image reconstruction. In the deep feature extraction module, a cross-scale local feature projection module is proposed, which selects an alternating downsampling and feature extraction method in the model structure. This alternating downsampling method can learn features of different scales. Subsequently, a simple and effective cross-scale global feature projection module is proposed, which applies self-calibrated convolutional attention to input features to better perform feature aggregation, and then outputs nonlinear features through deep separable convolution processing of different branches to obtain better performance. Finally, the modules are connected in a distillation manner to reduce the number of parameters while ensuring the performance of network reconstruction. By shrinking the image, some unnecessary detail information can be removed, so that the model pays more attention to the overall structure and important features of the image, thereby improving the robustness and generalization ability of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flow chart of an image super-resolution reconstruction algorithm for low-resolution blurred scenes provided by the present invention;

[0043] Figure 2 It is a structural diagram of an image super-resolution reconstruction algorithm for low-resolution blurred scenes provided by the present invention;

[0044] Figure 3 is a structural diagram of a cross-scale local feature projection module provided by the present invention;

[0045] Figure 4 is a structural diagram of a cross-scale global feature projection module provided by the present invention;

[0046] Figure 5This is a visual comparison result of the image reconstructed by the algorithm proposed in the present invention and four times that of other algorithms on 86000 (B100 dataset);

[0047] Figure 6 This is a visual comparison result of the image reconstructed by the algorithm proposed in the present invention and other algorithms on 216081 (B100 dataset) 4 times;

[0048] Figure 7 This is a visual comparison result of the image reconstructed by the algorithm proposed in the present invention and other algorithms on img017 (Urban100 dataset), which is 4 times the size. DETAILED DESCRIPTION

[0049] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific preferred embodiments.

[0050] In the description of the present invention, it should be understood that the terms "left side", "right side", "upper part", "lower part" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. "First", "second" and the like do not indicate the importance of the components, and therefore cannot be understood as limiting the present invention. The specific dimensions used in this embodiment are only for illustrating the technical solution, and do not limit the scope of protection of the present invention.

[0051] like Figure 1-7 As shown, an image super-resolution reconstruction method for low-resolution blurred scenes includes the following steps: S1, designing an image super-resolution reconstruction model framework for low-resolution blurred scenes, the image super-resolution reconstruction model framework for low-resolution blurred scenes includes an image shallow feature extraction module, an image deep feature extraction module and an image reconstruction module; S2, using the image shallow feature extraction module to extract shallow features of the image; S3, using the image deep feature extraction module to extract deep features of the image; the image deep feature extraction module includes a cross-scale local feature projection module, a difference feature extraction module and a cross-scale global feature projection module, and the modules are connected by knowledge distillation; S4, reconstructing the image to obtain a reconstructed super-resolution image.

[0052] The algorithm framework of the present invention is as follows Figure 2 As shown in the figure, the algorithm consists of three modules: image shallow feature extraction, image deep feature extraction, and image reconstruction. The shallow feature extraction module consists of a convolution with a convolution kernel size of 3×3. LR First, a 3×3 convolution is performed to extract shallow features to obtain shallow features F SThe deep feature extraction module is composed of three feature basic block modules, each of which is composed of four feature fusion modules and one attention module. S The deep feature F3 is obtained through three feature basic modules. The image reconstruction module consists of a sub-pixel convolution layer and a convolution with a convolution kernel size of 3×3. The deep feature F3 is fused with the three-layer distillation feature through feature splicing operation, and finally the reconstructed super-resolution image I is obtained after the sub-pixel convolution layer and 3×3 convolution. SR .

[0053] The overall relationship between the image shallow feature extraction module, the image deep feature extraction module and the image reconstruction module is as follows:

[0054] F s =H conv3 (I LR ) (1)

[0055] F1=H FBB (F S ) (2)

[0056] F2=H FBB (F1) (3)

[0057] F3=H FBB (F2) (4)

[0058] I SR =H conv3 (H upsample (F3⊙(H conv1 (F s ))⊙(H conv1 (F1))⊙(H conv1 (F2)))) (5)

[0059] Among them, I LR Represents a low-resolution image, H conv3 (·) indicates a convolution with a kernel size of 3*3, H conv1 (·) indicates a convolution with a kernel size of 1*1, F s Represents a low-resolution image I LR The shallow features obtained after a 3×3 convolution; H FBB (·) represents the feature base module (FeatureBaseModule, FBB) module. Each feature base module here is composed of four feature fusion modules (Feature Fusion Block, FFB) and one enhanced spatial attention module (Enhanced SpatialAttention, ESA). n＝1,2,3 Represents the shallow feature F sThe deep features obtained by the FBB module; ⊙ is the feature concatenation operation, I SR is the final reconstructed super-resolution image.

[0060] The cross-scale local feature projection module extracts feature information of multiple images of different scales and fuses them to reduce the loss of context information in network transmission, and proposes a difference feature extraction block to learn the differences between different features and fuse features through residual connections.

[0061] In addition to using lightweight structures, the computational cost can be reduced by reducing the size of feature maps. At the same time, the reduced-size features can also improve the feature learning performance of the network by combining with higher-resolution features in many tasks. The cross-scale local feature projection module CSLPB proposed in the present invention extracts multi-scale features by downsampling the input features four times and twice, respectively, and selects alternating downsampling and feature extraction methods in the model structure. This alternating downsampling method can learn features of different scales. After the first downsampling, feature learning is performed, and some relatively rough features with richer upper and lower information can be learned. After that, downsampling by 2 times and feature learning are performed, and some relatively fine features can be learned on the coarser feature map. The present invention also proposes a Differentiated Feature Extraction Block (DFB), which learns the differences between different features and performs feature fusion through residual connections.

[0062] X scale＝1 =GELU(DFB(H conv1 (X))) (6)

[0063] X scale＝2 =GELU(DFB(H d (X))) (7)

[0064] X scale＝4 =GELU(DFB(H d (X))) (8)

[0065] X out =X scale＝1 |H u (X scale＝2 )|H u (X scale＝4 ) (9)

[0066] Among them, X is the given input feature, DFB is the difference feature extraction module, H d is the downsampling achieved by the maximum pooling convolution, X scale＝1,2,4 H is the output feature after DFB processing at different scales.u is the upsampling achieved by nearest neighbor interpolation, “|” is the feature concatenation method, X out It is the final output of the cross-scale local feature projection module.

[0067] A cross-scale global feature projection module is proposed to further learn deep feature information of different scales through different receptive fields. The CSGPB of the present invention designs a self-calibrated convolutional attention composed of deep separable convolution and activation function, and the features selectively learned by attention are further processed by convolution.

[0068] X'=H conv1 (X) (10)

[0069] W = Sigmoid(H conv1 (H dwconv7 (H conv1 (X')))) (11)

[0070]

[0071] X b =GELU(H dwconv3 (H conv1 (X'))) (13)

[0072] X out =X a +X b (14)

[0073] Among them, X' is the feature to be input into the two branches after being processed by the convolution layer, and H conv1 is a convolutional layer with a 1×1 convolution kernel, H dwconv7 is the depth-separable convolution of the 7×7 convolution kernel, Sigmoid is the activation function Sigmoid, W is the feature weight output by the Sigmoid function, H dwconv3 The self-calibrated convolution attention multiplies the generated feature weights with the original features, which helps the network model automatically learn the features that are more critical and relevant to the current task. a is the characteristic of the main output, X b is the feature of the branch output, GELU is the activation function, X out The feature is the final output feature after adding the features of the two branches of this module.

[0074] Finally, each module is connected by distillation to reduce the number of parameters while ensuring the performance of network reconstruction.

[0075] The following is a detailed description of the technical effects of the present invention in combination with performance tests and experimental analysis. Tabular data analysis shows that the model of the present invention is significantly better than other classic image super-resolution reconstruction models at a 4-fold scaling factor. The comparison results are shown in Table 1:

[0076] Table 1 Comparison of indicators under 4x scaling factor

[0077]

[0078] From Table 1, it can be found that under a scaling factor of 4, the PSNR and SSIM indicators of the model proposed in the present invention are significantly higher than those of other models while controlling the number of parameters.

[0079] In order to further demonstrate the superiority of the network model proposed in this paper, a 4-fold image reconstruction experiment was conducted on the Urban100 and B100 datasets with complex textures. The images reconstructed using this algorithm were compared with the results of other algorithms. The comparison results are shown in Figure 2. Figures 5 to 7 As shown. It can be clearly seen from these image results that compared with other models, the algorithm model proposed by the present invention can reconstruct images more realistically. Compared with the model of this patent, other models will have different degrees of blur and distortion after reconstructing the image, while the image reconstructed by the model of this patent shows better clarity. Figure 5 In the reconstruction of the window frames by other models, the images reconstructed by IMDN and LatticeNet are blurred to varying degrees. The images reconstructed by the patented model can clearly show the intersecting window frames in the middle of the window. Figure 6 The wheel is reconstructed in , and compared with the image reconstructed in this patent, which has complete line contours, other reconstructed images have partial line distortion. Figure 7 Different models are used to reconstruct the stripes on the zebra's body. It can be seen from the figure that the reconstruction effect of other models is poor, and the reconstructed stripes are even in the opposite direction of the HR image. The image reconstructed by the model proposed in the present invention not only has the same stripe direction, but also has clearer stripe lines. This shows that the algorithm model proposed in the present invention has obvious advantages over other algorithm models in image super-resolution reconstruction.

[0080] In order to verify the effectiveness of the cross-scale local and global feature projection modules CSLPB and CSGPB, this paper makes a comparison of ablation experimental data in Table 2. The implementation data shows that the innovations proposed in this paper have a positive effect on improving the model reconstruction performance. CSLPB and CSGPB improve the PSNR index by 0.03dB and 0.02dB respectively, while the algorithm in this paper improves the PSNR index by 0.06dB, which proves the effectiveness of the innovations.

[0081] Table 2 Effects of CSLPB and CSGPB on model performance

[0082]

[0083] In order to verify the impact of different numbers of FMBs on network performance, the number of FMBs in the model is set to 3, 4, and 5, and the results of reconstruction on the Urban100 dataset are compared. Table 3 shows the experimental comparison results of this part. As can be seen from the table, with the increase in the number of FMB modules, the number of parameters of the network model increases from 1.24M to 1.82M. However, when the number of modules increases from 4 to 5, the increase in the number of parameters leads to a decrease in the reconstruction index. Therefore, this paper chooses to fix the number of FMB modules to 4.

[0084] Table 3 Effect of different FMB numbers on model performance

[0085]

[0086]

[0087] This paper further analyzes the impact of the number of BMs in the distillation network on the reconstruction performance, sets the number of BMs to 2, 3, and 4, trains the model, and compares the reconstruction performance. The comparative experimental results are shown in Table 4, and the test data set is Urban100. When the number of basic module BM blocks increases to 4, the number of parameters increases by 0.4M, but the reconstruction index PSNR only increases by 0.04dB and SSIM increases by 0.0006. Compared with the model with fewer parameters, although the increase in the number of parameters brings about an increase in reconstruction time, the reconstruction index is greatly increased. After comprehensive consideration, this paper sets the number of BM modules to 3.

[0088] Table 4 Effect of different BM numbers on model performance

[0089]

[0090] The present invention is implemented in this way. When learning deep features of images, traditional neural networks only propagate features in the network through direct connections, while using projection can transform features so as to transmit features more smoothly. In order to ensure that the neural network has good robustness for images of different scales in practical applications, the present invention proposes a cross-scale feature projection network (CSPN) for low-resolution blur scenes and applies it to the single image super-resolution reconstruction task. First, a cross-scale local feature projection module (CSLPB) is proposed to extract and fuse feature information of multiple images of different scales to reduce the loss of contextual information in network transmission. Second, a cross-scale global feature projection module (CSGPB) is proposed to further learn deep feature information of different scales through different receptive fields. Third, each module is connected in a distillation manner to reduce the number of parameters while ensuring the performance of network reconstruction.

[0091] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. An image super-resolution reconstruction method for low-resolution blurred scenes, characterized by: The following steps are involved: S1. Designing an image super-resolution reconstruction model framework for low-resolution blurred scenes, wherein the image super-resolution reconstruction model framework for low-resolution blurred scenes comprises an image shallow feature extraction module, an image deep feature extraction module and an image reconstruction module; S2, using an image shallow feature extraction module to extract shallow features of the image; S3, extracting deep features of the image using an image deep feature extraction module; the image deep feature extraction module includes a cross-scale local feature projection module, a difference feature extraction module and a cross-scale global feature projection module, and the modules are connected in a knowledge distillation manner; S4, reconstructing the image to obtain a reconstructed super-resolution image; The overall relationship between the image shallow feature extraction module, the image deep feature extraction module and the image reconstruction module is as follows: F s =H conv3 (I LR ) (1) F1=H FBB (F S ) (2) F2=H FBB (F1) (3) <h2 style=";text-align:left;direction:ltr">F3=H<h2 style=";text-align:left;direction:ltr"> FBB <h2 style=";text-align:left;direction:ltr"> (F2) (4) I SR =H conv3 (H upsample (F3⊙(H conv1 (F s ))⊙(H conv1 (F1))⊙(H conv1 (F2)))) (5) Among them, I LR Represents a low-resolution image, H conv3 (·) indicates a convolution with a kernel size of 3*3, H conv1 (·) represents the convolution with a kernel size of 1*1, and Fs represents the low-resolution image I LR The shallow features obtained after a 3×3 convolution; H FBB (·) represents the feature basic block module, F n＝1,2,3 Represents the shallow feature F s The deep features obtained by the FBB module; ⊙ is the feature concatenation operation, I SR is the final reconstructed super-resolution image; The cross-scale local feature projection module performs four-fold and two-fold downsampling on the input features respectively to complete the extraction of multi-scale features, and selects the method of alternating downsampling and feature extraction in the model structure; The difference feature extraction module learns the differences between different features and fuses features through residual connections, and its relationship with the cross-scale local feature projection module is as follows: X scale＝1 =ENTER(DFB(H conv1 (X))) (6) X scale＝2 =ENTER(DFB(H d (X))) (7) X scale＝4 =ENTER(DFB(H d (X))) (8) X out =X scale＝1 |H u (X scale＝2 )|H u (X scale＝4 ) (9) Among them, X is the given input feature, DFB is the difference feature extraction module, H d is the downsampling achieved by the maximum pooling convolution, X scale＝1,2,4 H is the output feature after DFB processing at different scales. u is the upsampling achieved by nearest neighbor interpolation, | is the feature concatenation method, X out is the final output of the cross-scale local feature projection module; GELU is the activation function; The cross-scale global feature projection module designs a self-calibrated convolutional attention composed of a depth-separable convolution and an activation function. The features selectively learned by attention are then processed by convolution, including the following algorithms; X'=H conv1 (X) (10) W=Sigmoid(H conv1 (H dwconv7 (H conv1 (X')))) (11) X b =GEL(H dwconv3 (H conv1 (X'))) (13) X out =X a +X b (14) Among them, X' is the feature to be input into the two branches after being processed by the convolution layer, and H conv1 is a convolutional layer with a 1×1 convolution kernel, H dwconv7 is the depth-separable convolution of the 7×7 convolution kernel, Sigmoid is the activation function Sigmoid, W is the feature weight output by the Sigmoid function, H dwconv3 is a depth-separable convolution with a 3×3 convolution kernel, X a is the characteristic of the main output, X b is the feature of the branch output, GELU is the activation function, X out The feature is the final output feature after adding the features of the two branches of this module.

2. The image super-resolution reconstruction method for low-resolution blurred scenes according to claim 1, characterized in that: The image shallow feature extraction module is a convolution with a convolution kernel size of 3×3.

3. The image super-resolution reconstruction method for low-resolution blurred scenes according to claim 1, characterized in that: The three feature basic block modules of the image deep feature extraction module, namely, the image deep feature extraction module includes a cross-scale local feature projection module, a difference feature extraction module and a cross-scale global feature projection module, each of which is composed of four feature fusion modules and one attention module.

4. The image super-resolution reconstruction method for low-resolution blurred scenes according to claim 1, characterized in that: The image reconstruction module consists of a sub-pixel convolution layer and a convolution with a convolution kernel size of 3×3. The deep features of the image and the three-layer distillation features are fused through a feature splicing operation, and finally a reconstructed super-resolution image is obtained after a sub-pixel convolution layer and a 3×3 convolution.

Images