Image super-resolution reconstruction method based on feature adaptive attention

Abstract

The invention relates to an image super-resolution reconstruction method based on feature adaptive attention, a reconstructed image is obtained based on a constructed image super-resolution reconstruction model, and the image super-resolution reconstruction model comprises a feature extraction module, a feature adaptive attention module, a fusion module, a feature splicing module and a convolution layer. According to the method, the quality and precision of low-resolution image reconstruction are effectively improved; and the constructed image super-resolution reconstruction model has the characteristics of light weight and top-speed reasoning.

Description

Technical Field

[0001] The present application relates to the field of computer image processing technology, and in particular to an image super-resolution reconstruction method based on feature adaptive attention. Background Art

[0002] Image super-resolution is a cutting-edge technology that uses deep learning algorithms to improve image resolution. It is dedicated to reconstructing high-definition images from low-resolution images, making image details clearer and significantly improving quality. This technology has broad application prospects in many fields such as surveillance, medical imaging, satellite imaging, and the entertainment industry, showing great practical value.

[0003] At present, image processing algorithms in the field of computer vision have achieved remarkable achievements and excellent performance. However, in the field of image super-resolution, the generalization ability of existing models needs to be further improved and there are still certain limitations. In order to break through this bottleneck, researchers are constantly exploring innovative algorithms and models, striving to achieve stable and high-quality super-resolution effects in a wider range of application scenarios.

[0004] The current mainstream method in the field of image super-resolution is to enhance the network model's ability to reconstruct low-resolution image features by increasing the network depth and introducing a self-attention encoding mechanism, which is similar to convolution operations and non-local computations. This processing method makes it difficult to capture the features mapped to the latent space, and the features are very similar to the features of other objects, resulting in artifacts in the reconstructed results that are not consistent with the image logic.

[0005] Models trained with traditional interpolation algorithms cannot show good generalization capabilities on severely degraded images, and are often accompanied by unrecognizable feature information and artifacts. GAN-based image super-resolution faces two major difficulties: first, due to the lack of image pairs with various degradation information, it is impossible to obtain a dataset truly used for image super-resolution reconstruction; second, the performance of the super-resolution model depends to a large extent on the recognition ability of local information and texture features, but the method of extracting high-level features by simply increasing the depth of the dense layer cannot be directly applied to image data rich in texture information. The constraint of image super-resolution is still the inability to learn the mapping relationship from low-resolution images to high-resolution images, and the proposed ill-posed problem is still a huge challenge. Summary of the invention

[0006] In order to overcome at least one deficiency in the prior art, the present application provides an image super-resolution reconstruction method based on feature adaptive attention.

[0007] In a first aspect, a method for image super-resolution reconstruction based on feature adaptive attention is provided, comprising:

[0008] Constructing an image super-resolution reconstruction model; the image super-resolution reconstruction model includes a feature extraction module, a feature adaptive attention module, a fusion module, a feature splicing module and a convolution layer, the feature adaptive attention module includes a plurality of feature adaptive attention units, the fusion module includes a plurality of fusion units, and the feature adaptive attention units and the fusion units correspond to each other one by one;

[0009] The image super-resolution reconstruction model is trained based on the training data set to obtain a trained image super-resolution reconstruction model;

[0010] Inputting the image to be reconstructed into the trained image super-resolution reconstruction model to obtain a reconstructed image;

[0011] The feature extraction module extracts features from the input image to be reconstructed to obtain a feature map;

[0012] The feature map is input to the feature adaptive attention module for feature adaptive attention calculation. Each feature adaptive attention unit outputs a first calculation result and a second calculation result. The first calculation result is input to the next feature adaptive attention unit, and the second calculation result is input to the corresponding fusion unit. Each fusion unit outputs a fusion result, which is input to the next fusion unit.

[0013] The feature splicing module is used to splice the fusion result output by each fusion unit and the first calculation result output by the last feature adaptive attention module to obtain a splicing result;

[0014] The splicing result is input into the convolution layer to obtain the reconstructed image.

[0015] In one embodiment, the feature extraction unit includes a convolution layer, an activation function layer, a convolution layer and an activation function layer connected in sequence, and the feature extraction unit is used to map the image features to an implicit space and retain the features of the input image.

[0016] In one embodiment, the feature adaptive attention unit includes a first branch and a second branch, the first branch includes a group feature weighted fusion module VFM and a channel space scale attention mechanism module CSA connected in sequence, and the second branch includes a non-local self-attention module NLSA, a deconvolution layer, and an activation function layer; the input of the feature adaptive attention unit passes through the first branch and the second branch respectively, and the output of the first branch is subtracted from the output of the second branch to obtain a first subtraction result;

[0017] The output of the first subtraction result after passing through the convolution layer, the activation function layer and the convolution layer is added to the output of the first branch to obtain a first addition result;

[0018] The output of the first addition result after passing through the activation function layer and the strided convolution layer is subtracted from the input of the feature adaptive attention unit to obtain a second subtraction result;

[0019] The output of the second subtraction result after passing through the deconvolution layer and the activation function layer is added to the output of the first branch to obtain a first calculation result;

[0020] The first calculation result passes through the strided convolution layer, the activation function layer, the convolution layer, and the activation function layer to obtain the second calculation result.

[0021] In one embodiment, the group feature weighted fusion module VFM includes: a first average pooling layer, a second average pooling layer, a convolution layer, a first SoftMax layer, a third average pooling layer, a Sigmoid layer, a GroupNorm unit, a second SoftMax layer, and a fourth average pooling layer;

[0022] The input of the group feature weighted fusion module VFM passes through the first average pooling layer and the second average pooling layer respectively. The outputs of the first average pooling layer and the second average pooling layer are spliced ​​and then input into the Sigmoid layer. The output of the Sigmoid layer is multiplied by the input of the group feature weighted fusion module VFM to obtain the first multiplication result; the first multiplication result is input into the GroupNorm unit, and the output of the GroupNorm unit is input into the second SoftMax layer and the fourth average pooling layer respectively;

[0023] The input of the group feature weighted fusion module VFM passes through the convolution layer, and the output of the convolution layer is input to the first SoftMax layer and the third average pooling layer respectively;

[0024] The output of the first SoftMax layer is multiplied by the output of the fourth average pooling layer to obtain the second multiplication result;

[0025] The output of the third average pooling layer is multiplied by the output of the second SoftMax layer to obtain a third multiplication result;

[0026] The second multiplication result and the third multiplication result are added to obtain the output of the group feature weighted fusion module VFM.

[0027] In one embodiment, the channel space scale attention mechanism module CSA includes two branches with the same structure and a self-attention mechanism module; each branch includes a convolution layer, a batch normalization layer, an activation function layer, a convolution layer, and a batch normalization layer;

[0028] The input of the channel space scale attention mechanism module CSA is respectively input into two branches and self-attention mechanism modules with the same structure, and the three outputs are added to obtain the output of the channel space scale attention mechanism module CSA.

[0029] In one embodiment, the input of the first fusion unit among the multiple fusion units is the second calculation result output by the first feature adaptive attention unit, the feature map output by the feature extraction module, and the image to be reconstructed.

[0030] In a second aspect, a device for image super-resolution reconstruction based on feature adaptive attention is provided, comprising:

[0031] A model building module, used to build an image super-resolution reconstruction model; the image super-resolution reconstruction model includes a feature extraction module, a feature adaptive attention module, a fusion module, a feature splicing module and a convolution layer, the feature adaptive attention module includes a plurality of feature adaptive attention units, the fusion module includes a plurality of fusion units, and the feature adaptive attention units and the fusion units correspond to each other one by one;

[0032] A model training module is used to train the image super-resolution reconstruction model based on the training data set to obtain a trained image super-resolution reconstruction model;

[0033] A reconstruction module, used for inputting the image to be reconstructed into the trained image super-resolution reconstruction model to obtain a reconstructed image;

[0034] The feature extraction module extracts features from the input image to be reconstructed to obtain a feature map;

[0035] The feature map is input to the feature adaptive attention module for feature adaptive attention calculation. Each feature adaptive attention unit outputs a first calculation result and a second calculation result. The first calculation result is input to the next feature adaptive attention unit, and the second calculation result is input to the corresponding fusion unit. Each fusion unit outputs a fusion result, which is input to the next fusion unit.

[0036] The feature splicing module is used to splice the fusion result output by each fusion unit and the first calculation result output by the last feature adaptive attention module to obtain a splicing result;

[0037] The splicing result is input into the convolution layer to obtain the reconstructed image.

[0038] In a third aspect, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the above-mentioned image super-resolution reconstruction method based on feature adaptive attention is implemented.

[0039] In a fourth aspect, a computer program product includes a computer program / instructions, which, when executed by a processor, implement the above-mentioned image super-resolution reconstruction method based on feature adaptive attention.

[0040] Compared with the prior art, the present application has the following beneficial effects: the image super-resolution reconstruction method based on feature adaptive attention of the present application obtains a reconstructed image based on a constructed image super-resolution reconstruction model, and the image super-resolution reconstruction model includes a feature extraction module, a feature adaptive attention module, a fusion module, a feature splicing module and a convolutional layer. The method effectively improves the quality and accuracy of low-resolution image reconstruction; and the constructed image super-resolution reconstruction model has the characteristics of lightweight and extremely fast reasoning. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The present application may be better understood by referring to the following description given in conjunction with the accompanying drawings, which together with the following detailed description are included in this specification and form a part of this specification. In the drawings:

[0042] Figure 1 A flowchart of an image super-resolution reconstruction method based on feature adaptive attention according to an embodiment of the present application is shown;

[0043] Figure 2 The structural block diagram of the image super-resolution reconstruction model is shown;

[0044] Figure 3 shows the structural block diagram of the feature adaptive attention unit;

[0045] Figure 4 shows a structural block diagram of a group feature weighted fusion module;

[0046] Figure 5 The structural block diagram of the channel space scale attention mechanism module is shown;

[0047] Figure 6 The reconstructed image result obtained according to the method of the present application is shown;

[0048] Figure 7 A structural block diagram of an image super-resolution reconstruction device based on feature adaptive attention according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0049] The exemplary embodiments of the present application will be described below in conjunction with the accompanying drawings. For the sake of clarity and conciseness, not all features of the actual embodiments are described in the specification. However, it should be understood that many implementation-specific decisions can be made in the process of developing any such actual embodiments in order to achieve the specific goals of the developer, and these decisions may vary from embodiment to embodiment.

[0050] It is also necessary to explain here that, in order to avoid obscuring the present application due to unnecessary details, only the device structure closely related to the scheme according to the present application is shown in the drawings, while other details that are not closely related to the present application are omitted.

[0051] It should be understood that the present application is not limited to the described implementation forms due to the following description with reference to the accompanying drawings. In this article, where feasible, the embodiments can be combined with each other, features between different embodiments can be replaced or borrowed, and one or more features can be omitted in one embodiment.

[0052] The present invention provides an image super-resolution reconstruction method based on feature adaptive attention. Figure 1 A flowchart of an image super-resolution reconstruction method based on feature adaptive attention according to an embodiment of the present application is shown, see Figure 1 , methods include:

[0053] Step S1, constructing an image super-resolution reconstruction model VFSR; Figure 2 shows the structural block diagram of the image super-resolution reconstruction model, see Figure 2 The image super-resolution reconstruction model includes a feature extraction module, a feature adaptive attention module, a fusion module, a feature concatenation module Concatenate and a convolution layer Convolution. The feature adaptive attention module includes multiple feature adaptive attention units FFR, and the fusion module includes multiple fusion units Fusion. The feature adaptive attention units and the fusion units correspond one to one.

[0054] Step S2: training the image super-resolution reconstruction model based on the training data set to obtain a trained image super-resolution reconstruction model.

[0055] Here, the DIV2K public dataset is used for training, the image format and resolution of the dataset are unified, and then the dataset is randomly cropped and data enhanced to ensure sufficient data volume; the data images are processed using the traditional bicubic downsampling degradation model to obtain LR-HR image pairs, that is, low-resolution-high-resolution images, which constitute the training dataset; the model is trained based on the training dataset, and the parameters of the network are continuously adjusted when the image is generated after each iteration by operating the core module. Under the condition of sufficient data volume, the learning rate is continuously reduced to extract deep image features. The training is stopped after the final indicators PSNR (peak signal-to-noise ratio) and SSIM (structural similarity) stabilize to a certain range, and the parameter file with better effect is retained, that is, the trained model is obtained.

[0056] Step S3, inputting the image to be reconstructed into the trained image super-resolution reconstruction model to obtain a reconstructed image.

[0057] Here, the feature extraction module extracts features from the input image to be reconstructed to obtain a feature map.

[0058] The feature map is input to the feature adaptive attention module for feature adaptive attention calculation. Each feature adaptive attention unit outputs a first calculation result and a second calculation result. The first calculation result is input to the next feature adaptive attention unit, and the second calculation result is input to the corresponding fusion unit. Each fusion unit outputs a fusion result, which is input to the next fusion unit.

[0059] The feature splicing module is used to splice the fusion result output by each fusion unit and the first calculation result output by the last feature adaptive attention module to obtain a splicing result;

[0060] The splicing result is input into the convolution layer to obtain the reconstructed image.

[0061] Here, each feature adaptive attention unit outputs a feature deepened by the current unit, i.e., the first calculation result, and the upsampled feature for hierarchical fusion, i.e., the second calculation result. Each fusion unit receives two inputs, namely, the second calculation result output by the feature adaptive attention unit and the output of the previous fusion unit.

[0062] The image super-resolution reconstruction method based on feature adaptive attention of this embodiment obtains a reconstructed image based on a constructed image super-resolution reconstruction model. The image super-resolution reconstruction model includes a feature extraction module, a feature adaptive attention module, a fusion module, a feature stitching module and a convolution layer. The method effectively improves the quality and accuracy of low-resolution image reconstruction; and the constructed image super-resolution reconstruction model is lightweight and has the characteristics of extremely fast reasoning.

[0063] In one embodiment, see Figure 2 The feature extraction unit includes a convolution layer Convolution, an activation function layer PReLU, a convolution layer Convolution and an activation function layer PReLU connected in sequence. The feature extraction unit is used to map image features to an implicit space and retain the features of the input image.

[0064] In one embodiment, Figure 3 shows the structural block diagram of the feature adaptive attention unit, see Figure 3The feature adaptive attention unit includes a first branch and a second branch. The first branch includes a group feature weighted fusion module VFM and a channel space scale attention mechanism module CSA connected in sequence. The second branch includes a non-local self-attention module NLSA, a deconvolution layer Deconvolution, and an activation function layer PReLU. The input of the feature adaptive attention unit passes through the first branch and the second branch respectively, and the output of the first branch is subtracted from the output of the second branch to obtain a first subtraction result.

[0065] The output of the first subtraction result after passing through the convolution layer, the activation function layer and the convolution layer is added to the output of the first branch to obtain a first addition result;

[0066] The output of the first addition result after passing through the activation function layer and the strided convolution layer is subtracted from the input of the feature adaptive attention unit to obtain the second subtraction result;

[0067] The output of the second subtraction result after passing through the deconvolution layer and the activation function layer is added to the output of the first branch to obtain a first calculation result;

[0068] The first calculation result passes through the strided convolution layer, the activation function layer, the convolution layer, and the activation function layer to obtain the second calculation result.

[0069] In the field of image processing, the usual way to process features is to directly splice the feature maps output by multiple branches in the channel dimension to form a wider feature map. This method can retain the complete feature information from different branches, so it is often used in scenarios where multi-source heterogeneous features need to be fused. However, this will lead to a significant increase in the number of channels of the feature map, and the computational complexity and complexity of the model will also increase exponentially, greatly reducing efficiency. Considering that simply splicing features in the channel dimension cannot effectively fuse feature information in multiple dimensions, this will affect the final image reconstruction effect. Therefore, before fusing the output features of each branch, this embodiment first performs feature differentiation processing layer by layer, and then performs feature fusion.

[0070] In the above process, NLSA uses sparsity constraints to force the module to focus on relevant and information-rich areas while ignoring irrelevant and noisy features.

[0071] In one embodiment, Figure 4 The structural block diagram of the group feature weighted fusion module is shown in Figure 4 , the group feature weighted fusion module VFM includes: the first average pooling layer AugPool, the second average pooling layer AugPool, the convolution layer, the first SoftMax layer, the third average pooling layer AugPool, the Sigmoid layer, the GroupNorm unit, the second SoftMax layer, and the fourth average pooling layer AugPool;

[0072] The input of the group feature weighted fusion module VFM passes through the first average pooling layer and the second average pooling layer respectively. The outputs of the first average pooling layer and the second average pooling layer are spliced ​​and then input into the Sigmoid layer. The output of the Sigmoid layer is multiplied by the input of the group feature weighted fusion module VFM to obtain the first multiplication result; the first multiplication result is input into the GroupNorm unit, and the output of the GroupNorm unit is input into the second SoftMax layer and the fourth average pooling layer respectively;

[0073] The input of the group feature weighted fusion module VFM passes through the convolution layer, and the output of the convolution layer is input to the first SoftMax layer and the third average pooling layer respectively;

[0074] The output of the first SoftMax layer is multiplied by the output of the fourth average pooling layer to obtain the second multiplication result;

[0075] The output of the third average pooling layer is multiplied by the output of the second SoftMax layer to obtain a third multiplication result;

[0076] The second multiplication result and the third multiplication result are added to obtain the output of the group feature weighted fusion module VFM.

[0077] In this embodiment, the data input into the group feature weighted fusion module VFM is divided into two branches, one branch is used as a feature weight adaptive score to weighted reconstruct the overall features, that is, input to the first average pooling layer and the second average pooling layer, and the other branch uses the first feature weight adaptive score to fuse the features after feature dimension reduction and channel grouping, that is, input to the convolution layer. The GroupNorm (group normalization) in the first branch is less sensitive to the size of a batch, and GroupNorm is used to perform group feature normalization. In order to emphasize important features, reduce the weight of unimportant features and capture global contextual information, the features output by GroupNorm are divided into two branches, which are input to the second SoftMax layer and the fourth average pooling layer respectively.

[0078] In the above embodiment, unlike the traditional method, the grouping features of the image are processed in the first branch and self-attention weighted fusion is realized. The purpose of the second branch is to extract the spatial features of the image. To this end, the input features are first convolved and then branched.

[0079] In one embodiment, Figure 5 The structural block diagram of the channel space scale attention mechanism module is shown in Figure 5, the channel space scale attention mechanism module CSA, includes 2 branches with the same structure, the self-attention mechanism module; each branch includes a convolution layer, a batch normalization layer, an activation function layer, a convolution layer, and a batch normalization layer;

[0080] The input of the channel space scale attention mechanism module CSA is respectively input into two branches and self-attention mechanism modules with the same structure, and the three outputs are added to obtain the output of the channel space scale attention mechanism module CSA.

[0081] In one embodiment, L1 loss is used as the optimization target during model training:

[0082]

[0083] Among them, VF * is the optimized image super-resolution reconstruction model, VF is the image super-resolution reconstruction model, M is the number of image pairs in the training process, i is the label of the image pair, and I LR and I HR are the low-resolution image and high-resolution image in the image pair respectively; VF(I LR ) indicates that the input is I LR The output of the image super-resolution reconstruction model VF.

[0084] Figure 6 The reconstructed image results obtained according to the method of the present application are shown. At the same time, the method of the present application is compared with the prior art, and the results are shown in Table 1. The image finally reconstructed by the method of the present application has achieved good results, and has made great progress in improving the resolution and symmetry features. In addition, the network also has the characteristics of lightweight and extremely fast reasoning.

[0085] Table 1

[0086]

[0087] Using the same inventive concept as the image super-resolution reconstruction method based on feature adaptive attention, this embodiment also provides a corresponding image super-resolution reconstruction device based on feature adaptive attention. Figure 7 FIG. 1 shows a structural block diagram of an image super-resolution reconstruction device based on feature adaptive attention according to an embodiment of the present application, see Figure 7 , the device comprises:

[0088] A model construction module 71 is used to construct an image super-resolution reconstruction model; the image super-resolution reconstruction model includes a feature extraction module, a feature adaptive attention module, a fusion module, a feature splicing module and a convolution layer, the feature adaptive attention module includes a plurality of feature adaptive attention units, the fusion module includes a plurality of fusion units, and the feature adaptive attention units and the fusion units correspond to each other one by one;

[0089] A model training module 72 is used to train the image super-resolution reconstruction model based on the training data set to obtain a trained image super-resolution reconstruction model;

[0090] A reconstruction module 73, used for inputting the image to be reconstructed into the trained image super-resolution reconstruction model to obtain a reconstructed image;

[0091] The feature extraction module extracts features from the input image to be reconstructed to obtain a feature map;

[0092] The feature map is input to the feature adaptive attention module for feature adaptive attention calculation. Each feature adaptive attention unit outputs a first calculation result and a second calculation result. The first calculation result is input to the next feature adaptive attention unit, and the second calculation result is input to the corresponding fusion unit. Each fusion unit outputs a fusion result, which is input to the next fusion unit.

[0093] The feature splicing module is used to splice the fusion result output by each fusion unit and the first calculation result output by the last feature adaptive attention module to obtain a splicing result;

[0094] The splicing result is input into the convolution layer to obtain the reconstructed image.

[0095] The image super-resolution reconstruction device based on feature adaptive attention of this embodiment has the same inventive concept as the image super-resolution reconstruction method based on feature adaptive attention mentioned above. Therefore, the specific implementation method of the device can be seen in the embodiment part of the image super-resolution reconstruction method based on feature adaptive attention in the above text, and its technical effect corresponds to the technical effect of the above method, which will not be repeated here.

[0096] An embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned image super-resolution reconstruction method based on feature adaptive attention.

[0097] An embodiment of the present application provides a computer program product, including a computer program / instruction. When the computer program / instruction is executed by a processor, the above-mentioned image super-resolution reconstruction method based on feature adaptive attention is implemented.

[0098] The above are only various implementations of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for image super-resolution reconstruction based on feature adaptive attention, characterized in that: include: Constructing an image super-resolution reconstruction model; the image super-resolution reconstruction model includes a feature extraction module, a feature adaptive attention module, a fusion module, a feature splicing module and a convolution layer, the feature adaptive attention module includes a plurality of feature adaptive attention units, the fusion module includes a plurality of fusion units, and the feature adaptive attention units correspond to the fusion units one by one; Training the image super-resolution reconstruction model based on the training data set to obtain a trained image super-resolution reconstruction model; Inputting the image to be reconstructed into the trained image super-resolution reconstruction model to obtain a reconstructed image; The feature extraction module extracts features from the input image to be reconstructed to obtain a feature map; The feature map is input into the feature adaptive attention module for feature adaptive attention calculation, each feature adaptive attention unit outputs a first calculation result and a second calculation result, the first calculation result is input into the next feature adaptive attention unit, and the second calculation result is input into the corresponding fusion unit; Each fusion unit outputs a fusion result, and the fusion result is input to the next fusion unit; The feature splicing module is used to splice the fusion result output by each fusion unit and the first calculation result output by the last feature adaptive attention module to obtain a splicing result; The stitching result is input into the convolution layer to obtain a reconstructed image.

2. The method according to claim 1, characterized in that The feature extraction unit includes a convolution layer, an activation function layer, a convolution layer and an activation function layer connected in sequence, and the feature extraction unit is used to map the image features to an implicit space and retain the features of the input image.

3. The method according to claim 1, characterized in that The feature adaptive attention unit includes a first branch and a second branch, the first branch includes a group feature weighted fusion module VFM and a channel space scale attention mechanism module CSA connected in sequence, and the second branch includes a non-local self-attention module NLSA, a deconvolution layer, and an activation function layer; The input of the feature adaptive attention unit passes through the first branch and the second branch respectively, and the output of the first branch is subtracted from the output of the second branch to obtain a first subtraction result; The output of the first subtraction result after passing through the convolution layer, the activation function layer and the convolution layer is added to the output of the first branch to obtain a first addition result; Subtracting the output of the first addition result after passing through the activation function layer and the strided convolution layer from the input of the feature adaptive attention unit to obtain a second subtraction result; The output of the second subtraction result after passing through the deconvolution layer and the activation function layer is added to the output of the first branch to obtain the first calculation result; The first calculation result is obtained after passing through a strided convolution layer, an activation function layer, a convolution layer, and an activation function layer to obtain the second calculation result.

4. The method according to claim 3, characterized in that The group feature weighted fusion module VFM includes: a first average pooling layer, a second average pooling layer, a convolution layer, a first SoftMax layer, a third average pooling layer, a Sigmoid layer, a GroupNorm unit, a second SoftMax layer, and a fourth average pooling layer; The input of the group feature weighted fusion module VFM passes through the first average pooling layer and the second average pooling layer respectively. The outputs of the first average pooling layer and the second average pooling layer are spliced ​​and input into the Sigmoid layer. The output of the Sigmoid layer is multiplied by the input of the group feature weighted fusion module VFM to obtain a first multiplication result; the first multiplication result is input into the GroupNorm unit, and the output of the GroupNorm unit is respectively input into the second SoftMax layer and the fourth average pooling layer; The input of the group feature weighted fusion module VFM passes through the convolution layer, and the output of the convolution layer is respectively input to the first SoftMax layer and the third average pooling layer; The output of the first SoftMax layer is multiplied by the output of the fourth average pooling layer to obtain a second multiplication result; The output of the third average pooling layer is multiplied by the output of the second SoftMax layer to obtain a third multiplication result; The second multiplication result and the third multiplication result are added to obtain the output of the group feature weighted fusion module VFM.

5. The method according to claim 3, characterized in that The channel space scale attention mechanism module CSA includes two branches with the same structure and a self-attention mechanism module; each branch includes a convolution layer, a batch normalization layer, an activation function layer, a convolution layer, and a batch normalization layer; The input of the channel space scale attention mechanism module CSA is respectively input into two branches and self-attention mechanism modules with the same structure, and the three outputs are added to obtain the output of the channel space scale attention mechanism module CSA.

6. The method according to claim 1, characterized in that The input of the first fusion unit among the multiple fusion units is the second calculation result output by the first feature adaptive attention unit, the feature map output by the feature extraction module, and the image to be reconstructed.

7. An image super-resolution reconstruction device based on feature adaptive attention, characterized in that: include: A model construction module, used to construct an image super-resolution reconstruction model; the image super-resolution reconstruction model includes a feature extraction module, a feature adaptive attention module, a fusion module, a feature splicing module and a convolution layer, the feature adaptive attention module includes a plurality of feature adaptive attention units, the fusion module includes a plurality of fusion units, and the feature adaptive attention units correspond to the fusion units one by one; A model training module, used to train the image super-resolution reconstruction model based on a training data set to obtain a trained image super-resolution reconstruction model; A reconstruction module, used for inputting the image to be reconstructed into the trained image super-resolution reconstruction model to obtain a reconstructed image; The feature extraction module extracts features from the input image to be reconstructed to obtain a feature map; The feature map is input into the feature adaptive attention module for feature adaptive attention calculation, each feature adaptive attention unit outputs a first calculation result and a second calculation result, the first calculation result is input into the next feature adaptive attention unit, and the second calculation result is input into the corresponding fusion unit; Each fusion unit outputs a fusion result, and the fusion result is input to the next fusion unit; The feature splicing module is used to splice the fusion result output by each fusion unit and the first calculation result output by the last feature adaptive attention module to obtain a splicing result; The stitching result is input into the convolution layer to obtain a reconstructed image.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the image super-resolution reconstruction method based on feature adaptive attention as described in any one of claims 1-6.

9. A computer program product, characterized in that It includes a computer program / instruction, which, when executed by a processor, implements the image super-resolution reconstruction method based on feature adaptive attention as described in any one of claims 1 to 6.

Images