CN113362223B - Image Super-Resolution Reconstruction Method Based on Attention Mechanism and Two-Channel Network - Google Patents

Abstract

The invention belongs to the field of artificial intelligence, deep learning and image processing, and particularly relates to an attention mechanism and two-channel network-based image super-resolution reconstruction method, which comprises the following steps: acquiring an image to be detected in real time, and preprocessing the image to be detected; inputting the preprocessed image into a trained image super-resolution reconstruction model to obtain a high-definition reconstruction image; evaluating the reconstructed image by adopting the peak signal-to-noise ratio and the structural similarity, and marking the high-definition reconstructed image according to an evaluation result; the basis of the image super-resolution reconstruction model is a convolutional neural network; the invention uses a dual-channel network, one network uses an improved residual error structure to extract valuable high-frequency characteristics, namely high-grade characteristics, and the other network uses an improved VGG network, so that the sizes of input and output images are consistent, abundant low-frequency characteristics are extracted, and finally, the characteristics are fused, so that the reconstructed images are clearer.

Description

Image super-resolution reconstruction method based on attention mechanism and two-channel network

technical field

The invention belongs to the fields of artificial intelligence, deep learning and image processing, in particular to an image super-resolution reconstruction method based on an attention mechanism and a dual-channel network.

Background technique

Image super-resolution reconstruction technology uses a set of low-quality, low-resolution images (or motion sequences) to generate high-quality, high-resolution images. Image super-resolution reconstruction is applied in various computer vision tasks, including surveillance imaging, medical imaging, object recognition, etc. In real life, limited by the cost of image acquisition equipment, video image transmission bandwidth, or the technical bottleneck of the imaging modality itself, we do not always have the conditions to obtain large-scale high-definition images with sharp edges and no blocky blur. . In the context of this demand, super-resolution reconstruction techniques emerge as the times require, and the traditional image super-resolution (SR) reconstruction problem is defined as recovering a high-resolution (HR) image from its low-resolution (LR) observations. The use of image super-resolution reconstruction technology can improve the recognition ability and accuracy of images, and can achieve focused analysis of the target object, so that higher spatial resolution images of the region of interest can be obtained without directly using high spatial resolution with huge amounts of data. rate image configuration. Learning-based image super-resolution reconstruction technology is a popular direction in recent years. With the help of the convolutional neural network technology of end-to-end mapping in deep learning, low-resolution images are estimated by learning the mapping relationship between low-resolution and high-resolution images. High-frequency details that are lost in the mid-range to obtain high-quality images with sharp edges and rich texture details. In recent years, with the introduction of the attention mechanism, more and more models have begun to try to use the attention mechanism to improve the effect. In the image super-resolution reconstruction, the attention mechanism is embedded into the model to improve the accuracy of the results.

In traditional image reconstruction methods, CNN convolutional neural network is often used for calculation, and a lot of smooth information will be gradually lost in the calculation process. The main purpose of image super-resolution reconstruction is to obtain high-precision high-definition pictures, any loss of information will It affects the final reconstruction accuracy; and the calculation of the CNN convolutional neural network adopts the method of local receptive field. During the calculation, limited by the size of the convolution kernel, the obtained information is the local information of the picture, and the entire picture cannot be obtained. The global information of the image, and the pixels in the picture are related, and the long-distance dependent information is also an important dependent information for image reconstruction.

SUMMARY OF THE INVENTION

In order to solve the above problems in the prior art, the present invention proposes an image super-resolution reconstruction method based on an attention mechanism and a dual-channel network. The method includes: acquiring an image to be detected in real time, and preprocessing the image to be detected; The preprocessed image is input into the trained image super-resolution reconstruction model to obtain a high-definition reconstructed image; the peak signal-to-noise ratio and structural similarity are used to evaluate the reconstructed image, and the high-definition reconstructed image is marked according to the evaluation results; The resolution reconstruction model is based on a convolutional neural network;

The process of training an image super-resolution reconstruction model includes:

S1: Obtain the original high-definition picture data set, and use the bicubic interpolation degradation model to scale the pictures in the data set;

S2: Preprocess the scaled data set to obtain a training data set;

S3: Input each image data in the training dataset into the shallow feature channel and the deep feature channel in the image super-resolution reconstruction model for feature extraction;

S4: Use the first convolution layer to extract the initial features of the input image; input the initial features into the information cascade module to aggregate the hierarchical feature information of the convolution layer;

S5: Input the hierarchical feature information aggregated by the information cascade module into the improved residual module to obtain the correlation on the channel and the dependency information on the global space;

S6: Use non-local hole convolution to perform global feature extraction on dependency information to obtain the final deep feature map;

S7: use the second convolution layer to extract the initial features of the input image; input the initial features into the improved VGG network, extract the shallow features of the image, and obtain a shallow feature map;

S8: fuse the deep feature map and the shallow feature map, and upsample the fused feature map to obtain a high-definition reconstruction map;

S9: Use the loss function to constrain the difference between the high-definition reconstructed image and the original high-definition image, and continuously adjust the parameters of the model until the model converges, and the training of the model is completed.

Preferably, the bicubic interpolation degradation model is used to zoom the pictures in the data set by 2 times, 3 times, 4 times and 8 times.

Preferably, the formula of the bicubic interpolation degradation model is:

I ^LR =H _dn I ^HR +n

Preferably, the process of preprocessing the scaled data set includes performing enhancement processing on the image, including performing translation processing on the image and flipping the image in horizontal and vertical directions; dividing the enhanced data into different small image blocks, Assemble the segmented images to obtain a training data set.

Preferably, the information cascading module includes a feature aggregation structure stacked 10 times; the feature aggregation structure includes at least three layers of convolutional neural networks, a feature channel merging layer, a channel attention layer and a channel number transformation layer, and each layer of the convolutional neural network is connected in turn. , and the output branch of each layer of convolutional neural network except the last layer of convolutional neural network is connected to the feature channel merging layer, the feature channel merging layer, the channel attention layer and the channel number transformation layer are connected in turn to form an information cascade module ; The process of processing image data by this module includes: firstly using each layer of convolutional neural network to extract feature information of the input image in turn, and then merging the feature information extracted by each layer of convolution on the feature channel merging layer, using channel attention The mechanism distinguishes the importance of the merged information, and finally reduces the number of channels to the number of input channels, and repeats the above steps 10 times to obtain the hierarchical feature information of the aggregated convolutional layer.

Preferably, the improved residual module includes: a residual network structure, a channel attention mechanism layer and a spatial attention mechanism layer, and the residual network structure includes a convolutional neural network layer, a nonlinear activation layer and a convolutional neural network layer; the The process of the module processing image data includes: inputting the hierarchical feature information into the residual network structure to extract feature information, using the extracted feature information to use the channel attention mechanism to obtain the correlation on the channel, and then passing it down, using spatial attention. Mechanism to obtain dependencies on the global space.

Preferably, the non-local atrous convolution block includes: four parallel atrous convolutional layers with dilation parameters of 1, 2, 4, and 6 and three ordinary convolutional neural network layers; the process of processing image data by this module includes: Firstly, four kinds of atrous convolutions with different dilation parameters and two kinds of ordinary convolutional neural networks are used to extract feature information from the dependency information input by the improved residual network; For fusion, the feature information extracted by the ordinary convolutional neural network is fused according to the value of the pixel matrix; finally, the two kinds of fused feature information are added to obtain the global feature information

Preferably, the improved VGG network structure includes: 10 layers of ordinary convolution layers and 3 layers of pooling layers, and each pooling layer is embedded in the ordinary convolution layer to obtain the VGG network structure; the process of processing image data by this module includes: First use 2 layers of convolution and one layer of pooling to extract 64 channel feature information, then use 2 layers of convolution and one layer of pooling to extract 128 channel feature information, and then use 3 layers of convolution and one layer of pooling to extract 512 channel feature information channel feature information, and finally use 3 layers of convolution to extract the information of 512 channels and restore it to 64 channels; among them, the pooling layer uses padding to keep the feature scale unchanged.

Preferably, the loss function expression of the image super-resolution reconstruction model is:

Preferably, the formula for evaluating the reconstructed image using peak signal-to-noise ratio and structural similarity is:

Advantages of the present invention:

1. The present invention uses a dual-channel network, one network uses an improved residual structure to extract valuable high-frequency features-that is, high-level features, and one network uses an improved VGG (parameters of the vgg convolutional layer and pooling layer are used. fine-tuning to ensure that the input and output image scales are consistent, while discarding the last fully connected layer) to extract rich low-frequency features, and feature fusion at the end.

2. The present invention uses a dense connection method in a specific position of the model (two information cascade modules at the head and tail), and aggregates the information of each convolutional layer to achieve the purpose of making full use of the information of the convolutional layer, And at the end, the channel attention mechanism is used to calculate the channel weights of the combined information, instead of simply descending the channel.

3. The present invention uses a spatial attention mechanism, adding spatial attention after the existing channel attention mechanism, so that the extraction of global information is more sufficient, and the utilization of features is more comprehensive. At the same time, before upsampling, non-local hole convolution is used, and the previous result information is extracted with a global dependency, so that the output results are more closely related and the feature information is richer.

Description of drawings

Fig. 1 is the overall structure diagram of the image super-resolution reconstruction model of the present invention;

Fig. 2 is the information cascade structure diagram of the present invention;

3 is a residual structure diagram of the present invention;

Fig. 4 is the channel attention and spatial attention structure diagram of the present invention;

FIG. 5 is a non-local hole convolution diagram of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

An image super-resolution reconstruction method based on an attention mechanism and a dual-channel network, the method comprises: acquiring an image to be detected in real time, preprocessing the to-be-detected image; inputting the preprocessed image into the trained image super-resolution In the reconstruction model, a high-definition reconstruction image is obtained; the peak signal-to-noise ratio and structural similarity are used to evaluate the reconstructed image, and the high-definition reconstructed image is marked according to the evaluation results; the image super-resolution reconstruction model is based on a convolutional neural network.

An image super-resolution reconstruction model structure, as shown in Figure 1, the structure includes a deep feature channel, a shallow feature channel, an upsampling layer and a third convolution layer; the deep feature channel includes a first convolution layer, The information cascade module, the improved residual module and the non-local hole convolution block; the input image is processed by the first convolution layer through the information cascade module, the improved residual module and the non-local hole convolution block in turn. processing to obtain a deep feature map; the shallow feature channel includes a second convolution layer and an improved VGG network, and the input image is processed by the second convolution layer and processed by the improved VGG network to obtain a shallow feature map; The deep feature map and the shallow feature map are fused, and the upsampling layer is used to upsample the fused image, and the third convolutional layer is used to convolve the upsampled image to obtain a high-definition reconstructed image.

Optionally, the deep feature channel includes n information cascade modules and m improved residual modules, wherein all the information cascade modules are connected in series to obtain an information cascade module group; all the improved residual modules are connected in series, An improved set of residual modules.

Preferably, 2n information cascade modules are included in the deep feature channel, n information cascade modules are connected in series to obtain a first information cascade module group, and the remaining n information cascade modules are connected in series to obtain a second information cascade module group. an information cascading module group; the first information cascading module group and the second information cascading module group are respectively arranged at the input end and the output end of the improved residual module group.

The process of training an image super-resolution reconstruction model includes:

S1: Obtain the original high-definition picture data set, and use the bicubic interpolation degradation model to scale the pictures in the data set;

S2: Preprocess the scaled data set to obtain a training data set;

S3: Input each image data in the training dataset into the shallow feature channel and the deep feature channel in the image super-resolution reconstruction model for feature extraction;

S4: Use the first convolution layer to extract the initial features of the input image; input the initial features into the information cascade module to aggregate the hierarchical feature information of the convolution layer;

S5: Input the hierarchical feature information aggregated by the information cascade module into the improved residual module to obtain the correlation on the channel and the dependency information on the global space;

S6: Use non-local hole convolution to perform global feature extraction on dependency information to obtain the final deep feature map;

S7: use the second convolution layer to extract the initial features of the input image; input the initial features into the improved VGG network, extract the shallow features of the image, and obtain a shallow feature map;

S8: fuse the deep feature map and the shallow feature map, and upsample the fused feature map to obtain a high-definition reconstruction map;

S9: Use the loss function to constrain the difference between the high-definition reconstructed image and the original high-definition image, and continuously adjust the parameters of the model until the model converges, and the training of the model is completed.

The data set is the DIV2K data set, in which 800 high-definition (HR) pictures and their corresponding low-resolution (LR) pictures that have undergone degradation model (bicubic interpolation degradation) are used as the training set, and five pictures are used. as a validation set. Use Set5, Set14, Urban100, Manga109, BSD100 five datasets as test sets. The characteristics of these test datasets are that the texture information is very rich, and the degraded low-resolution images will lose most of the texture information, which is a very test of the image. Accuracy of super-resolution reconstructions. The evaluation indicators are traditional PSNR and SSIM, where PSNR represents peak signal-to-noise ratio and SSIM represents structural similarity.

All the data in the training set has undergone one forward propagation and one backpropagation in the neural network, which is called one round. Each round will update the parameters of the model, and the maximum number of rounds is set to 1000 rounds. We set the learning rate to be updated every 200 epochs, and during the 1000 epochs of training the model, the model and its parameters that performed best on the test dataset were saved.

In the original high-definition picture data set, the bicubic interpolation degradation model is used to scale the pictures in the data set by 2 times, 3 times, 4 times and 8 times. The formula for the degradation model is:

I ^LR =H _dn I ^HR +n

where I ^LR represents the low-resolution image, H _dn represents the degradation model, I ^HR represents the original high-resolution image, and n represents the additional noise.

The process of preprocessing the scaled data set includes enhancing the image, including translating the image, flipping the image horizontally and vertically; dividing the enhanced data into different small image blocks, and dividing the The images are collected to obtain the training data set.

As shown in Figure 2, the structure of the information cascade module includes: stacking the structure below 10 times - followed by a three-layer convolutional neural network, a feature channel merging layer, a channel attention layer, and a channel number transformation layer. The process of processing image data by this module includes: firstly extracting feature information of the input image by using each layer of convolutional neural network in turn, then combining the feature information extracted by each layer of convolution on the feature channel combining layer, using the channel attention mechanism The importance of the merged information is distinguished, and finally the number of channels is reduced to the number of input channels, and the above steps are repeated 10 times to obtain the hierarchical feature information of the aggregated convolutional layer.

Use the information cascade module to aggregate image information, fully retain the information of each convolutional layer, because the image has just entered the convolutional neural network, the low-frequency information is sufficient and abundant, but with the deepening of the network layer, more attention has been paid to it. For more abstract features, many edge texture information and smoothing information are gradually lost, so using the information cascade module at this time can well capture more low-frequency information and fuse it into the model.

F _IC = H _IC (I ^LR )

Among them, I ^LR represents the low-resolution input image, H _IC represents the convolution operation of the cascade module, and F _IC represents the result of the convolution calculation.

As shown in Figure 3, the structure of the improved residual module includes: residual network structure, channel attention mechanism layer and spatial attention mechanism layer. The residual network structure includes convolutional neural network layer, nonlinear activation layer and convolutional layer. Neural network layer; the process of processing image data by this module includes: inputting hierarchical feature information into the residual network structure to extract feature information, using the extracted feature information to use the channel attention mechanism to obtain the correlation on the channel, and then passing it down , using the spatial attention mechanism to obtain dependencies on the global space.

The output of the cascade module is used as the input of the improved residual module. After each ResNetBlock, the channel attention mechanism and the spatial attention mechanism are connected, and the correlation information on the channel and the dependency information on the global space are captured at the same time, and integrated into In the convolutional neural network, the feature information is enriched to provide stability for the training of the deep network.

F _RBC = H _RBC (F _IC )

F _CA = H _CA (F _RBC )

F _SA = H _SA (F _CA )

Among them, H _RBC represents the residual block structure convolution operation with input source connection, that is, the input information is fused with the output of the residual block. F _RBC represents the output feature information of the residual block, which can be expressed as [ _{f 1} , f ₂ , f ₃ . Then, the channel attention mechanism is used for each channel feature to obtain the product of the weight value of each channel and the original input data, that is, H _CA represents the convolution operation of channel attention, and F _CA represents the feature information after channel attention. Then, the spatial attention mechanism is used to calculate the global dependency information for the output feature information, which is fused with the original input data, that is, _HSA represents the convolution operation of the spatial attention mechanism, and _FSA represents the feature information after spatial attention.

As shown in Figure 4, the channel attention structure includes: the structure is composed of a global draw pooling, a 1x1 convolutional layer, a nonlinear activation layer, and a 1x1 convolutional layer in order. The process of processing image data by this module includes: first obtain the weight representation symbol of each channel through global average pooling, then use a 1x1 convolution layer to reduce the number of channels, then use a nonlinear activation layer to introduce nonlinear information, and then use a 1x1 convolution layer The layer transforms the number of channels back, and finally multiplies back the original input feature information to obtain the correlation on the feature channels. The spatial attention structure includes: the structure consists of a 1x1 convolutional layer, a softmax activation layer, a 1x1 convolutional layer, and a nonlinear activation layer. The process of processing image data by this module includes: firstly converting the input feature information CxHxW into a global feature map of HWx1x1 through a 1x1 convolutional layer, then using the softmax function to normalize the global feature map, and then multiplying it back to the original input information, Finally, the dependency information in the global space is obtained through a 1x1 convolutional layer and a non-linear activation layer.

As shown in Figure 5, the structure of the non-local atrous convolution block includes: the structure consists of four parallel atrous convolution layers with dilation parameters of 1, 2, 4, and 6, and a three-layer ordinary convolutional neural network Floor. The process of processing image data by this module includes: first, the feature information will use four kinds of dilated convolutions with different dilation parameters and two kinds of ordinary convolutional neural networks to extract the feature information respectively, and then on the one hand, the feature information obtained by the four kinds of dilated convolution will be used. Fusion is performed on the feature channel. On the other hand, the feature information extracted by the ordinary convolutional neural network is fused according to the value of the pixel matrix, and finally the two fused feature information is added to obtain the global feature information.

The improved VGG network structure includes: 10 layers of ordinary convolution layers and 3 layers of pooling layers, and each pooling layer is embedded into the ordinary convolution layer to obtain the VGG network structure; the process of processing image data by this module includes: first use 2 Layer convolution and one layer pooling to extract 64 channel feature information, then use 2 layers of convolution and one layer of pooling to extract 128 channel feature information, and then use 3 layers of convolution and one layer of pooling to extract 512 channel features Finally, 3 layers of convolution are used to extract the information of 512 channels and restore it to 64 channels; among them, the pooling layer uses padding to keep the feature scale unchanged.

Use non-local hole convolution to extract global features, and upsample the extracted feature information to expand the output result to the size we need. The atrous convolution can expand the receptive field without increasing the parameters by setting the expansion rate. Embedding it into the non-local convolution can significantly reduce the amount of calculation. At the same time, it can also obtain global information from different scales and extract features. more comprehensive.

F _NLHC = H _NLHC (F _SA )

Among them, H _NLHC represents the convolution operation of non-local hole convolution, and F _NLHC represents the feature information obtained after non-local hole convolution. After upsampling the final feature information, we output the corresponding high-definition reconstructed image, that is, the formula for the reconstructed image is:

F _Up = H _Up (F _NLHC )

where H _Up represents the up-sampled convolution operation, and F _Up represents the up-sampled output feature.

The loss function expression of the image super-resolution reconstruction model is:

和

分别表示第i张低分辨率图像和第i张对应的高分辨率图像，N表示数据集中图像的数量，HR表示高分辨率，LR表示低分辨率。Among them, θ represents the parameter quantity of the model, C _HR represents the super-resolution calculation equation,

and

represent the ith low-resolution image and the ith corresponding high-resolution image, respectively, N represents the number of images in the dataset, HR represents high resolution, and LR represents low resolution.

The expression of the super-resolution calculation equation is:

C _HR = F _UP (F _NLHC (F _SA (F _CA (F _RBC (F _IC (I ^LR )))))))

Among them, F _UP represents the output information after up-sampling, F _NLHC represents the information extracted by the non-local hole convolution, F _SA represents the information extracted by the spatial attention mechanism, F _CA represents the information extracted by the channel attention mechanism, and F _RBC represents the residual information. The information extracted by the difference block, F _IC represents the information output by the cascade module.

Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) were used as results evaluation metrics:

Among them, MSE represents the mean square error, MAX represents the maximum value of the pixel values, μ _X and μ _Y represent the average value of the pixels of image X and image Y, σ _X and σ _Y represent the standard value of the pixels of image X and image Y, _σXY represents the covariance of image X and image Y.

The above-mentioned embodiments further describe the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above-mentioned embodiments are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made to the present invention within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims (10)

1. an image super-resolution reconstruction method based on attention mechanism and dual-channel network, is characterized in that, comprises: obtains the image to be detected in real time, and carries out preprocessing to the image to be detected; In the image super-resolution reconstruction model, a high-definition reconstructed image is obtained; the peak signal-to-noise ratio and structural similarity are used to evaluate the reconstructed image, and the high-definition reconstructed image is marked according to the evaluation results; The process of training an image super-resolution reconstruction model includes: S1: Obtain the original high-definition picture data set, and use the bicubic interpolation degradation model to scale the pictures in the data set; S2: Preprocess the scaled data set to obtain a training data set; S3: Input each image data in the training dataset into the shallow feature channel and the deep feature channel in the image super-resolution reconstruction model for feature extraction; S4: Use the first convolution layer to extract the initial features of the input image; input the initial features into the information cascade module to aggregate the hierarchical feature information of the convolution layer; S5: Input the hierarchical feature information aggregated by the information cascade module into the improved residual module to obtain the correlation on the channel and the dependency information on the global space; S6: Use the non-local hole convolution block to extract global features from the dependency information to obtain the final deep feature map; S7: use the second convolution layer to extract the initial features of the input image; input the initial features into the improved VGG network, extract the shallow features of the image, and obtain a shallow feature map; S8: fuse the deep feature map and the shallow feature map, and upsample the fused feature map to obtain a high-definition reconstruction map; S9: Use the loss function to constrain the difference between the high-definition reconstructed image and the original high-definition image, and continuously adjust the parameters of the model until the model converges, and the training of the model is completed. 2. a kind of image super-resolution reconstruction method based on attention mechanism and dual-channel network according to claim 1, is characterized in that, the multiple that adopts bicubic interpolation degradation model to zoom the picture in the data set is 2 times, 3x, 4x and 8x. 3. a kind of image super-resolution reconstruction method based on attention mechanism and dual-channel network according to claim 1, is characterized in that, the formula of bicubic interpolation degradation model is: I ^LR =H _dn I ^HR +n where I ^LR represents the low-resolution image, H _dn represents the degradation model, I ^HR represents the original high-resolution image, and n represents the additional noise. 4. An image super-resolution reconstruction method based on an attention mechanism and a dual-channel network according to claim 1, wherein the process of preprocessing the scaled data set comprises performing enhancement processing on the image, comprising: Translate the image and flip it in horizontal and vertical directions; divide the enhanced data into different small image blocks, and collect the divided images to obtain a training data set. 5. An image super-resolution reconstruction method based on an attention mechanism and a dual-channel network according to claim 1, wherein the information cascade module comprises a feature aggregation structure stacked 10 times; the feature aggregation structure comprises at least three layers Convolutional neural network, feature channel merging layer, channel attention layer and channel number transformation layer, each layer of convolutional neural network is connected in turn, and the output branch of each layer of convolutional neural network except the last layer of convolutional neural network The feature channel merging layer is connected, and the feature channel merging layer, the channel attention layer and the channel number transformation layer are connected in turn to form an information cascade module; the process of processing image data by this module includes: first, use each layer of convolutional neural network to sequentially analyze the input image. Extract feature information, then combine the feature information extracted by each layer of convolution on the feature channel combining layer, use the channel attention mechanism to distinguish the importance of the combined information, and finally reduce the number of channels to the number of input channels, repeat The above steps are repeated 10 times to obtain the hierarchical feature information of the aggregated convolutional layer. 6. An image super-resolution reconstruction method based on an attention mechanism and a dual-channel network according to claim 1, wherein the improved residual module comprises: a residual network structure, a channel attention mechanism layer and a space Attention mechanism layer, residual network structure includes convolutional neural network layer, nonlinear activation layer and convolutional neural network layer; the process of processing image data by this module includes: inputting hierarchical feature information into the residual network structure to extract feature information , the extracted feature information uses the channel attention mechanism to obtain the correlation on the channel, and then passes it down, and uses the spatial attention mechanism to obtain the dependence on the global space. 7. An image super-resolution reconstruction method based on an attention mechanism and a dual-channel network according to claim 1, wherein the non-local hole convolution block comprises: four parallel dilation parameters are 1 and 2, respectively. , 4 and 6 atrous convolutional layers and three layers of ordinary convolutional neural network layers; the process of processing image data by this module includes: first, atrous convolution with four different dilation parameters and two ordinary convolutional neural networks are used to improve the Then, the feature information obtained by the four kinds of hole convolutions is fused on the feature channel, and the feature information extracted by the ordinary convolutional neural network is fused according to the value of the pixel matrix; finally The two kinds of fused feature information are added to obtain the global feature information. 8. An image super-resolution reconstruction method based on an attention mechanism and a dual-channel network according to claim 1, wherein the improved VGG network structure comprises: 10 layers of ordinary convolution layers and 3 layers of pooling layers , Embed each pooling layer into the ordinary convolutional layer to obtain the VGG network structure; the process of processing image data by this module includes: firstly using 2 layers of convolution and one layer of pooling to extract 64 channel feature information, and then using 2 layers Convolution and one-layer pooling extract 128 channel feature information, then use 3-layer convolution and one-layer pooling to extract 512 channel feature information, and finally use 3-layer convolution to extract 512-channel information and restore it to 64 channels; Among them, the pooling layer uses padding to keep the feature scale unchanged. 9. a kind of image super-resolution reconstruction method based on attention mechanism and dual-channel network according to claim 1, is characterized in that, the loss function expression of image super-resolution reconstruction model is:

Among them, θ represents the parameter quantity of the model, C _HR represents the super-resolution calculation equation,

and

represent the ith low-resolution image and the ith corresponding high-resolution image, respectively, N represents the number of images in the dataset, HR represents high resolution, and LR represents low resolution. 10. A kind of image super-resolution reconstruction method based on attention mechanism and dual-channel network according to claim 1, is characterized in that, adopts peak signal-to-noise ratio and structural similarity to evaluate the formula of reconstructed image:

where PSNR represents the peak signal-to-noise ratio, MSE represents the mean square error, MAX represents the maximum value among the pixel values, SSIM represents the structural similarity, μX and μY represent the mean of the pixels of image _X and image _Y , respectively, _σX and σ _Y represents the standard value of the pixels of image X and image Y, respectively, and σ _XY represents the covariance of image X and image Y.

Images