An infrared image super-resolution reconstruction method based on a dynamic attention mechanism

Abstract

The application discloses an infrared image super-resolution reconstruction method based on a dynamic attention mechanism and relates to the technical field of image processing methods. The method comprises the following steps: constructing a light-weight deep neural network model based on a dynamic attention mechanism; preparing a training data set of the neural network model; training the neural network model using the training data set; providing pairs of low-resolution infrared images and target images to the network model, setting hyperparameters to start training the network, and terminating the training until the maximum number of iterations is reached; saving the model; after the training is completed, saving a group of network weights corresponding to the highest peak signal-to-noise ratio index in the evaluation process as the final network model parameters for later testing. The method can improve the infrared image super-resolution performance, has a small number of parameters, and is fast in calculation.

Description

Technical Field

[0001] This invention relates to the field of image processing methods, and in particular to an infrared image super-resolution reconstruction method based on a dynamic attention mechanism. Background Technology

[0002] Thanks to the rapid development of microelectronics technology, visible light-based optical cameras (RGB cameras) have quickly gained widespread application in fields such as security monitoring, drones, military, and agriculture due to their fast imaging speed and good image quality. However, due to limitations in their imaging principles, the image quality of visible light-based RGB cameras deteriorates rapidly in scenarios with poor lighting quality (such as at night) and adverse weather conditions. In these situations, infrared cameras, with their strong ability to penetrate smoke and fog, can replace traditional RGB cameras for image capture. However, the spatial resolution of sensors in general civilian infrared cameras is low, unable to match the megapixel resolution of RGB cameras, resulting in low resolution infrared images. Therefore, improving the resolution of infrared images efficiently and economically using new technologies and methods has become an urgent problem to be solved. Compared to investing heavily in the research and development of high-resolution infrared sensors, improving infrared image resolution through algorithms has become the current mainstream development direction.

[0003] Single-image super-resolution aims to recover a high-resolution version of a single observed image frame. Over the past decade, with the increase in computer processing speed and the rapid development of dedicated graphics processing units (GPUs), single-image super-resolution technology based on deep neural networks has made significant progress. Leveraging the powerful nonlinear fitting capabilities of deep neural networks, image super-resolution algorithms based on this technology significantly outperform traditional algorithms based on interpolation and dictionary learning. Furthermore, with the advent of neural processing units (NPUs), deploying large-scale deep neural network models in edge devices such as mobile phones has become a reality. However, current infrared image super-resolution models based on deep neural networks have a large number of parameters, limiting their further application. In conclusion, developing lightweight infrared image super-resolution algorithms based on deep neural networks has clear practical significance and tangible socio-economic benefits. Summary of the Invention

[0004] The technical problem to be solved by this invention is how to provide an infrared image super-resolution reconstruction method based on dynamic attention mechanism that can improve the super-resolution performance of infrared images, has a small number of parameters, and a fast calculation speed.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: an infrared image super-resolution reconstruction method based on dynamic attention mechanism, characterized by comprising the following steps:

[0006] Construct a lightweight deep neural network model based on a dynamic attention mechanism;

[0007] Create the training dataset for the neural network model;

[0008] The neural network model is trained using the training dataset: a pair of low-resolution infrared images and target images are provided to the network model, hyperparameters are set to start training the network until the maximum number of iterations is reached, and then training is terminated.

[0009] Save the model: After training is completed, save the set of network weights corresponding to the highest peak signal-to-noise ratio during the evaluation process as the final network model parameters for later testing.

[0010] A further technical solution is that the lightweight deep neural network model includes a shallow feature extraction module, a deep feature extraction module, and a feature reconstruction module.

[0011] A further technical solution is that the shallow feature extraction module (SFE) includes a convolutional layer with a kernel size of 3×3; after the input low-resolution infrared image (LR) is processed by the shallow feature extraction module, its shallow features such as edges and textures are extracted. The above process can be described as follows:

[0012] I LFE =f 3×3 (I LR )

[0013] Among them, I LR f represents the input low-resolution infrared image. 3×3 (·) represents a convolution operation with a kernel size of 3×3, I LFE This represents the output of the Shallow Feature Extraction (SFE) module.

[0014] A further technical solution is that the deep feature extraction module includes 16 dynamic attention residual blocks, and the feature map after shallow feature extraction is continuously passed through the deep feature extraction module to extract deep features of the image.

[0015] Each dynamic attention residual block includes: a dynamic attention weight module, an attention module, and a non-attention module. A dynamic attention weight module is designed to generate two dynamic weights, which are used to weight the attention module and the non-attention module respectively, to efficiently suppress unwanted attention adjustments. Simultaneously, to further reduce the number of parameters, 1-D convolutions are partially used in the dynamic attention weight module instead of fully connected layers. The operations in the dynamic attention weight module can be described as follows:

[0016]

[0017] Where, xn-1 This represents the input to the nth dynamic attention module. and Representing the Softmax, fully connected, 1-D convolution, and global average pooling operations, respectively, ω att and ω n-att These are the weights of the attention module and the non-attention module, which are dynamically generated by each dynamic attention weight module based on the input features;

[0018] The attention module includes a pixel attention module and a channel attention module, which are used to highlight important features in the pixel and channel dimensions, respectively. The channel attention module multiplies the feature vectors after global average pooling and standard deviation pooling of the input features by a learnable parameter α and β, respectively, to highlight the different effects of global average pooling and standard deviation pooling. Simultaneously, the two pooling results are fused to further enhance the features. After fusing the above three features, a 1-D convolution is used instead of a fully connected layer to further extract inter-channel features. The operations in the attention module can be described as follows:

[0019]

[0020] Among them, f 1×1 (·) represents a convolution operation with a kernel size of 1×1, f CA (·) and f PA (·) represent the channel attention and pixel attention module operations, respectively. This is the output of the nth attention module;

[0021] The non-attention module includes a convolutional layer with a kernel size of 3×3. The operations in the non-attention module can be represented as follows:

[0022]

[0023] in, The output of the nth non-attention module is the same convolutional operation in both the attention module and the non-attention module, where the convolutional layer with a kernel size of 1×1 is the same convolutional operation.

[0024] The operation in the nth dynamic attention residual block can be represented as:

[0025]

[0026] The operation of the deep feature extraction module can be described as follows:

[0027]

[0028] in, I represents the operation within the nth dynamic attention module. DFEThis represents the output of the entire Deep Feature Extraction (DFE) module.

[0029] A further technical solution is as follows: the feature reconstruction module includes a nearest neighbor interpolation layer, three convolutional layers with a kernel size of 3×3, and a pixel attention layer. The operations of the feature reconstruction module corresponding to the ×2 and ×3 scale factors can be represented as follows:

[0030]

[0031] Among them, f Nearest (·) indicates the nearest neighbor interpolation operation. This represents the output of the feature reconstruction module corresponding to the ×2 and ×3 scale factors;

[0032] Since the ×4 scale factor super-resolution reconstruction requires the recovery of more high-frequency information, such as details, the feature reconstruction block design corresponding to this scale factor includes: two nearest neighbor interpolation layers, five convolutional layers with a kernel size of 3×3, and two pixel attention layers. The operation of the feature reconstruction block corresponding to the ×4 scale factor can be represented as follows:

[0033]

[0034] in, This represents the output of the feature reconstruction block corresponding to the ×4 scale factor;

[0035] The final output infrared image super-resolution result can be expressed as:

[0036] I SR =I FR +f Nearest (I LR ).

[0037] A further technical solution is to use existing high-resolution infrared images as target images when constructing the training dataset, and use bicubic interpolation algorithm to downsample and obtain low-resolution infrared images corresponding to scale factors of ×2, ×3 and ×4, respectively, as inputs to the network model.

[0038] A further technical solution involves using the Adam optimizer during network training, with an initial learning rate α = 5 × 10⁻⁶. -4 Batch size is 32; calculate the generated I SR and true high-resolution images I HR The L1 loss between the two is used to update the network weights through the backpropagation algorithm. The above steps are repeated until the set number of iterations is reached to complete the training.

[0039] A further technical solution is to save the set of network weights corresponding to the highest peak signal-to-noise ratio during the evaluation process as the final network model parameters for later testing after training is completed.

[0040] The beneficial effects of adopting the above technical solution are as follows: 1) The method of the present invention designs a dynamic attention weighting module, which dynamically generates weighted values ​​of the attention module and the non-attention module according to different input feature maps, so as to further enhance the feature representation. The dynamic attention weighting module proposed in the present invention can suppress unwanted attention adjustments and improve the super-resolution performance of the final infrared image. At the same time, both the dynamic attention weighting module and the non-attention module proposed in the present invention are based on the lightweight design concept, with a small number of parameters, which is convenient for deployment in embedded devices and edge devices.

[0041] 2) The channel attention module designed in this invention applies weights again to the features after global average pooling and standard deviation pooling, overcoming the defect of traditional channel attention module designs that ignore the effects of two different pooling methods. Furthermore, unlike the channel attention modules used in most current networks, this invention uses 1-D convolutions instead of fully connected layers for the fused channel attention features in order to reduce the number of parameters.

[0042] 3) The infrared image super-resolution reconstruction method based on a lightweight deep neural network with dynamic attention mechanism designed in this invention has a significant effect on infrared image super-resolution reconstruction while maintaining a small number of parameters. Due to comparison algorithms such as bicubic interpolation, it has the conditions to be deployed on edge devices and has good super-resolution reconstruction performance. Attached Figure Description

[0043] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0044] Figure 1 This is a flowchart of a preferred embodiment of the method of the present invention;

[0045] Figure 2 This is a network structure diagram of the method according to a preferred embodiment of the present invention;

[0046] Figure 3 This is a diagram of the dynamic attention residual block structure of the method described in a preferred embodiment of the present invention;

[0047] Figure 4 This is a pixel attention module structure diagram of the method described in a preferred embodiment of the present invention;

[0048] Figure 5 This is a structural diagram of the channel attention module of the method according to a preferred embodiment of the present invention;

[0049] Figure 6 This is a feature reconstruction block structure diagram corresponding to the ×2 and ×3 scale factors of the method described in a preferred embodiment of the present invention;

[0050] Figure 7 This is a feature reconstruction block structure diagram corresponding to the ×4 scale factor of the method described in a preferred embodiment of the present invention;

[0051] Figure 8 This is a qualitative comparison of the super-resolution reconstruction performance of infrared images at two resolutions (640×480 and 384×288) with a scale factor of 2, using the method described in a preferred embodiment of the present invention and the methods of bicubic interpolation, FSRCNN, and SR-LUT.

[0052] Figure 9 This is a qualitative comparison of the super-resolution reconstruction performance of infrared images at two resolutions (640×480 and 384×288) with a scale factor of 3, using the method described in a preferred embodiment of the present invention and the methods of bicubic interpolation, FSRCNN, and SR-LUT.

[0053] Figure 10 This is a qualitative comparison of the super-resolution reconstruction performance of infrared images at two resolutions (640×480 and 384×288) with a scale factor of 4, using the method described in a preferred embodiment of the present invention and the methods of bicubic interpolation, FSRCNN, and SR-LUT.

[0054] Figure 11 Comparison chart of parameter quantities - PSNR between the present invention and the comparative method. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and 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 effort are within the scope of protection of the present invention.

[0056] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0057] like Figure 1 As shown in the figure, this invention discloses an infrared image super-resolution reconstruction method based on a dynamic attention mechanism, comprising the following steps:

[0058] S1) Constructing a Deep Neural Network Model: This model mainly includes a Shallow Feature Extraction (SFE) module, a Deep Feature Extraction (DFE) module, and a Feature Reconstruction (FR) module. The original low-resolution infrared image input is interpolated by nearest neighbor and then added to the output of the feature reconstruction module to produce the final high-resolution infrared image output. To meet the requirements of lightweight design, the constructed Shallow Feature Extraction (SFE) module consists of only one convolutional layer with a kernel size of 3×3. The structure diagram of this module is shown below. Figure 2 As shown, the input low-resolution infrared image (LR) is processed by the shallow feature extraction (SFE) module, where features such as edges, texture, and color are extracted. This process can be described as follows:

[0059] I LFE =f 3×3 (I LR )

[0060] Among them, I LR f represents the input low-resolution infrared image. 3×3 (·) represents a convolution operation with a kernel size of 3×3, I LFE This represents the output of the Shallow Feature Extraction (SFE) module. The feature maps extracted from the shallow features are then continuously passed through the Deep Feature Extraction (DFE) module, which consists of 16 dynamic attention residual blocks, to extract deeper features from the image, such as semantic information. The overall structure of this module is shown in the diagram below. Figure 2 As shown, each dynamic attention residual module consists of three parts: a dynamic attention weight module, an attention module, and a non-attention module. Since the attention modules exist at different network depths, their contributions to the final super-resolution reconstruction performance vary. Currently, most deep neural network-based infrared image super-resolution methods simply stack attention modules with the same structure, which is insufficient to fully utilize the feature reshaping advantages of attention modules.

[0061] Therefore, this invention designs a dynamic attention weighting module to generate two dynamic weights to weight the attention module and the non-attention module respectively, in order to efficiently suppress unwanted attention adjustments. The structure diagram of the dynamic attention weighting module is shown below. Figure 3 As shown. Furthermore, to further reduce the number of parameters and achieve lightweight design, this invention partially uses 1-D convolutions instead of fully connected layers in the dynamic attention weight module. The operation in the dynamic attention weight module can be described as follows:

[0062]

[0063] Where, x n-1This represents the input to the nth dynamic attention weight module. and Representing the Softmax, fully connected, 1-D convolution, and global average pooling operations, respectively, ω att and ω n-att These represent the weights of the attention module and the non-attention module, dynamically generated by the dynamic attention weight module based on the input features. To fully extract the deep information of the feature map, this invention designs two branch modules within the dynamic attention residual block: an attention module and a non-attention module. The attention module consists of a pixel attention module and a channel attention module connected sequentially, used to highlight important features in the pixel and channel dimensions, respectively. Their structure diagrams are shown below. Figure 4 and Figure 5 As shown. Unlike typical channel attention modules, the channel attention module designed in this invention multiplies the feature vectors after global average pooling and standard deviation pooling by learnable parameters α and β, respectively, to highlight the different effects of global average pooling and standard deviation pooling. Simultaneously, this invention also fuses the two pooling results to further enhance the features. After fusing the above three features, to reduce parameters, this invention uses 1-D convolutions instead of fully connected layers to further extract inter-channel features. The operations in the attention module can be described as follows:

[0064]

[0065] Among them, f 1×1 (·) represents a convolution operation with a kernel size of 1×1, f CA (·) and f PA (·) represent the channel attention and pixel attention module operations, respectively. This is the output of the nth attention module. To reduce the number of parameters, the non-attention module designed in this invention consists of only one convolutional layer with a kernel size of 3×3, as shown in the diagram below. Figure 3 As shown. The operations in the non-attention module can be represented as:

[0066]

[0067] in, The output of the nth non-attention module is shown. Convolutional layers with a kernel size of 1×1 in both the attention and non-attention modules represent the same convolutional operation. In summary, the operation in the nth dynamic attention module can be represented as:

[0068]

[0069] In summary, the Deep Feature Extraction (DFE) operation proposed in this invention can be described as follows:

[0070]

[0071] in, I represents the operation within the nth dynamic attention module. DFE This represents the output of the entire Deep Feature Extraction (DFE) module. For scale factors of ×2 and ×3, the Feature Reconstruction (FR) module consists of a nearest neighbor interpolation block, three convolutional layers with kernel size of 3×3, and a pixel attention layer, as shown in the diagram below. Figure 6 As shown, the operations of the feature reconstruction modules corresponding to the ×2 and ×3 scale factors can be represented as follows:

[0072]

[0073] Among them, f Nearest (·) indicates the nearest neighbor interpolation operation. This represents the output of the feature reconstruction modules corresponding to the ×2 and ×3 scale factors. Since the ×4 scale factor super-resolution reconstruction requires the recovery of more high-frequency information, such as details, making reconstruction more difficult, the feature reconstruction block for this ×4 scale factor in this invention includes the following components: two nearest-neighbor interpolation layers, five convolutional layers with a kernel size of 3×3, and two pixel attention layers, as shown in the diagram below. Figure 7 As shown. The operation of the feature reconstruction block FR corresponding to the ×4 scale factor can be expressed as:

[0074]

[0075] in, This represents the output of the feature reconstruction block corresponding to the ×4 scale factor. The final output infrared image super-resolution result SR can be expressed as:

[0076] I SR =I FR +f Nearest (I LR )

[0077] S2) Creating a training dataset: For existing publicly available high-resolution infrared image datasets, use bicubic interpolation downsampling algorithms with scale factors of ×2, ×3, and ×4 to obtain the corresponding low-resolution infrared images to form the training set.

[0078] S3) Training the deep neural network: The Adam optimizer is used when training the network, with an initial learning rate α = 5 × 10⁻⁶. -4 The batch size is 32. Calculate the generated I... SR and true high-resolution images I HR The L1 loss between the two is used to update the network weights through the backpropagation algorithm. The above steps are repeated until the maximum number of iterations is reached to complete the training.

[0079] S4) Save the model: After training is completed, save the set of network weights corresponding to the highest peak signal-to-noise ratio during the evaluation process as the final network model parameters for testing.

[0080] This invention uses high-resolution infrared images with resolutions of 640×480 and 384×288 as test target images, and low-resolution infrared images corresponding to scale factors of ×2, ×3, and ×4 obtained by bicubic interpolation as network inputs. Qualitative comparisons of this invention with bicubic interpolation, FSRCNN, and SR-LUT methods on the above two resolution infrared image datasets are as follows: Figure 8 , Figure 9 and Figure 10 As shown. By Figure 8 , Figure 9 and Figure 10 It can be seen that the infrared image super-resolution reconstruction algorithm proposed in this invention can improve the resolution of infrared images, and can better reconstruct details such as edges and textures with fewer artifacts compared with the comparison method. This qualitatively proves the effectiveness of the proposed method in infrared image super-resolution reconstruction.

[0081] This invention uses high-resolution infrared images with resolutions of 640×480 and 384×288 as test target images, and low-resolution infrared images corresponding to scale factors of ×2, ×3, and ×4 obtained by a bicubic interpolation algorithm as network input. The high-resolution infrared image I generated by the network is calculated... SR Peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) index between the network and the target image were used as quantitative evaluation indicators. Table 1 shows the quantitative comparison results between the network proposed in this invention and the Bicubic, FSRCNN, and SR-LUT methods. The 640×480 and 384×288 resolution infrared images included 50 and 20 test images, respectively, and all experimental results were averaged from multiple tests. As can be seen from Table 1, the infrared image super-resolution reconstruction method proposed in this invention outperforms the Bicubic, FSRCNN, and SR-LUT methods in both PSNR and SSIM evaluation indicators, quantitatively demonstrating the effectiveness of this invention in infrared image super-resolution.

[0082] Table 1. Quantitative comparison results of infrared image super-resolution reconstruction of the present invention and comparative methods.

[0083]

[0084] The network proposed in this invention performs well on publicly available datasets compared to FSRCNN, SRCNN, DRRN, VDSR, MemNet, IMDN, SRMNDF, DRCN, and A. 2 NM, A 2 N and MANDNet-L F Algorithm parameter-PSNR comparison Figure 11 As shown. By Figure 11 It can be seen that the network proposed in this invention achieves a high PSNR while maintaining a relatively small number of parameters, thus achieving a good balance between parameter quantity and performance. Furthermore, it can be observed that the infrared image super-resolution network proposed in this invention has the potential for deployment in embedded devices and edge devices.

Claims

1. A method for super-resolution reconstruction of infrared images based on a dynamic attention mechanism, characterized in that... Includes the following steps: Construct a lightweight deep neural network model based on a dynamic attention mechanism; Create the training dataset for the neural network model; The neural network model is trained using the training dataset: a pair of low-resolution infrared images and target images are provided to the network model, hyperparameters are set to start training the network until the maximum number of iterations is reached, and then training is terminated. Save the model: After training is completed, save the set of network weights corresponding to the highest peak signal-to-noise ratio during the evaluation process as the final network model parameters for later testing; The lightweight deep neural network model includes a shallow feature extraction module, a deep feature extraction module, and a feature reconstruction module; The shallow feature extraction module (SFE) includes a convolutional kernel with a size of [missing information]. The convolutional layer; after the input low-resolution infrared image (LR) is processed by the shallow feature extraction module, its shallow features are extracted. The above process is described as follows: ； in, This represents the input low-resolution infrared image. Indicates the kernel size as Convolution operation, This represents the output of the Shallow Feature Extraction (SFE) module. The deep feature extraction module includes 16 dynamic attention residual blocks. The feature map after shallow feature extraction is continuously passed through the deep feature extraction module to extract deep features of the image. Each dynamic attention residual block includes: a dynamic attention weighting module, an attention module, and a non-attention module. A dynamic attention weighting module is designed to generate two dynamic weights, which are used to weight the attention module and the non-attention module respectively, to efficiently suppress unwanted attention adjustments. Simultaneously, to further reduce the number of parameters, 1-D convolutions are partially used in the dynamic attention weighting module instead of fully connected layers. The operations in the dynamic attention weighting module are described as follows: ； in, Indicates the first The input to a dynamic attention module, , , and These represent the Softmax, fully connected, 1-D convolution, and global average pooling operations, respectively. and These are the weights of the attention module and the non-attention module, which are dynamically generated by each dynamic attention weight module based on the input features; The attention module includes a pixel attention module and a channel attention module, which are used to highlight important features in the pixel and channel dimensions, respectively. The channel attention module multiplies the feature vectors after global average pooling and standard deviation pooling of the input features by a learnable parameter. and This is used to highlight the different roles of global average pooling and standard deviation pooling; simultaneously, the results of the two pooling methods are fused to further enhance the features; after fusing the above three features, 1-D convolutions are used instead of fully connected layers to further extract inter-channel features. The operations in the attention module are described as follows: ； in, Indicates the kernel size as Convolution operation, and These represent the channel attention and pixel attention module operations, respectively. For the first The output of each attention module; The non-attention module includes a convolutional kernel with a size of [missing information]. The operations in the non-attention module of the convolutional layer are represented as follows: ； in, For the first The outputs of the non-attention modules, where the kernel size in both the attention and non-attention modules is... The convolutional layers all involve the same convolution operation; No. The operations in a dynamic attention residual block are represented as follows: ； The operation of the deep feature extraction module is described as follows: ； in, Indicates the first Operations within a dynamic attention module This represents the output of the entire Deep Feature Extraction (DFE) module.

2. The infrared image super-resolution reconstruction method based on dynamic attention mechanism as described in claim 1, characterized in that: The feature reconstruction module includes a nearest neighbor interpolation layer and three convolutional kernels of size [missing information]. A convolutional layer and a pixel attention layer, and The operation of the feature reconstruction module corresponding to the scaling factor is represented as follows: ； in, This indicates the nearest neighbor interpolation operation. express and The output of the feature reconstruction module corresponding to the scaling factor; because Scale factor super-resolution reconstruction requires the recovery of more high-frequency information; therefore, the feature reconstruction block design corresponding to this scale factor includes: two nearest-neighbor interpolation layers and five convolutional kernels of size [missing information]. The convolutional layer and two pixel attention layers, The operation representation of the feature reconstruction block corresponding to the scaling factor is as follows: ； in, express The output of the feature reconstruction block corresponding to the scaling factor; The final output infrared image super-resolution result is represented as follows: 。 3. The infrared image super-resolution reconstruction method based on dynamic attention mechanism as described in claim 1, characterized in that: When constructing the training dataset, existing high-resolution infrared images are used as target images, and downsampling is performed using a bicubic interpolation algorithm to obtain the following results: , and The low-resolution infrared image corresponding to the scale factor is used as the input to the network model.

4. The infrared image super-resolution reconstruction method based on dynamic attention mechanism as described in claim 1, characterized in that: The Adam optimizer is used when training the network, with an initial learning rate of... Batch size is 32; Calculated and generated and true high-resolution images Between The loss is used to update the network weights through backpropagation, and the above steps are repeated until the set number of iterations is reached to complete the training.

5. The infrared image super-resolution reconstruction method based on dynamic attention mechanism as described in claim 1, characterized in that: After training is completed, the set of network weights corresponding to the highest peak signal-to-noise ratio during the evaluation process will be saved as the final network model parameters for later testing.

Images