Image denoising method based on convolutional neural network

Abstract

The invention discloses an image denoising method based on a convolutional neural network. The method comprises the following steps that S10, an image containing noise passes through a multi-scale cavity convolution module; s20, through a feature extraction stage in which a 3CBlockn module and a Conv module are combined, the input of each layer is a feature map formed by combining the output of the previous layer in a jump connection mode, and the feature maps of different layers are spliced together by using Concat operation; and S30, taking the output of the EndConv module as a noise image, and subtracting the noise image from the original image to obtain a denoised image. The denoising method provided by the invention has wide applicability and remarkable practical value in practical application, can effectively solve the image noise problem encountered in electric power facility monitoring, and provides important technical support for safe operation of an electric power system.

Description

Technical Field

[0001] The invention belongs to the field of image processing, and in particular relates to an image denoising method based on a convolutional neural network. Background Art

[0002] In the power transmission equipment detection system, multispectral images are often affected by noise to varying degrees due to sensor noise and environmental interference, resulting in a decrease in image quality, which in turn affects the accuracy of detection. Image denoising technology is a key means to solve this problem. In recent years, convolutional neural networks (CNNs), as a powerful deep learning model, have shown excellent performance in the field of image processing.

[0003] Image denoising is an important step in image processing, which aims to restore a clear image as close to the original image as possible from an image disturbed by noise. Traditional image denoising methods mainly include various techniques based on the spatial domain and frequency domain. These methods usually use the local characteristics and frequency characteristics of the image to distinguish between noise and useful information. Common image noises include Poisson noise, Gaussian noise, salt and pepper noise, etc. Generally, Gaussian Poisson noise is more consistent with real image noise. Common image denoising methods include mean filtering, median filtering, Gaussian filtering, etc.

[0004] Mean filtering: Taking the target pixel as the center, the average value of all pixels in its local area is used as the pixel value after filtering.

[0005] The mean filter algorithm is simple, has a small amount of calculation, is easy to implement and apply, and is suitable for real-time processing applications. It can effectively smooth high-frequency noise and reduce random noise in the image. However, while removing noise, the mean filter will blur the details and edges in the image, resulting in a decrease in image quality. In addition, it only considers information in the local area and cannot effectively use the global characteristics of the image to remove noise.

[0006] Gaussian filtering: Use the Gaussian function to calculate the weighted coefficients of the pixel values ​​at each point in the filter kernel, and calculate the weighted average as the pixel value at the center of the filter kernel. The Gaussian function generally uses formula (1).

[0007]

[0008] Where:

[0009] σ is the standard deviation of the Gaussian function

[0010] Δx, Δy are the coordinate differences between this point and the center of the filter kernel

[0011] g(Δx,Δy) represents the Gaussian function value at that point

[0012] Taking a 3×3 filter kernel as an example, its (Δx, Δy) distribution is (-1,-1) to (1,1). The Gaussian function calculated according to formula (1) and the pixel coordinates of the Gaussian filter kernel needs to be normalized according to formula (2) to ensure that the sum of the weighted coefficients is 1.

[0013]

[0014] Median filtering: Median filtering is a nonlinear filtering technology commonly used for image denoising. It mainly replaces the central pixel value with the median of the pixel values ​​in the window to achieve the effect of removing noise.

[0015] The median filter has an excellent effect on removing impulse noise (such as salt and pepper noise) because the median can effectively ignore extreme outliers. Compared with the mean filter, the median filter better preserves the edges and details of the image while removing noise, and does not blur the image details like the mean filter. Although the median filter can protect the edges well, it may still lose some details when processing images with rich details. Summary of the invention

[0016] In view of this, the present invention proposes an image denoising method based on a convolutional neural network, comprising the following steps:

[0017] S10, the noisy image passes through a multi-scale dilated convolution module;

[0018] S20, after the feature extraction stage composed of 3CBlock_n module and Conv module, the input of each layer is the feature map merged with the output of the previous layer by using skip connection, and the feature maps of different layers are spliced ​​together by Concat operation;

[0019] S30, using the output of the EndConv module as the noise image, and subtracting the noise image from the original image to obtain a denoised image.

[0020] Preferably, the multi-scale dilated convolution module includes: a standard convolution module and three dilated convolution modules with dilation rates of 2, 3, and 4, respectively, and the four convolution outputs are merged together.

[0021] Preferably, the multi-scale hole convolution module includes 4 Conv modules in parallel, each Conv module input is a 3-channel noise image, the output is an 8-channel feature map, a total of 32 channels, and each convolution kernel size is 3×3, the step size is 1, the hole rates are 1, 2, 3 and 4, and the filling is 1, 2, 3 and 4; the larger the hole rate, the larger the convolution receptive field R, and the formula is:

[0022] R=(k-1)*d*s+1 (3)

[0023] In the formula, k is the convolution kernel size, d is the convolution dilation rate, and s is the step size;

[0024] The Conv module uses convolution followed by batch normalization and then SiLU function activation. The formula is:

[0025] F out =SiLU(bn(conv(F in ))) (4)

[0026] The convolution kernel size of the multi-scale dilated convolution module is k×k=3×3, the step size is s=1, and the dilation rate is d={1, 2, 3, 4}. According to formula (3), the receptive field of the multi-scale dilated convolution module is {3, 5, 7, 9}.

[0027] Preferably, the 3CBlock_n module in S20 includes a Conv module with a convolution kernel k=3 and a step size s=1, a C2f_n module and an attention mechanism module.

[0028] Preferably, the C2f_n module is a feature extraction module in YOLOv8, which consists of two 1x1 convolutional layers and several Bottleneck modules, wherein the first 1x1 convolutional layer reduces the number of input channels and divides them into two parts, and several Bottleneck modules extract features from one part, and finally all features are spliced ​​and converted into the number of output channels through the second 1x1 convolutional layer.

[0029] Preferably, the attention mechanism module includes a channel attention module and a spatial attention module connected in series.

[0030] Preferably, the channel attention module first performs global average pooling on each channel, and the output size is (batch, channels, 1, 1), that is, a global average value of each channel, batch is the number of batches, and channels is the number of channels; then a 1x1 convolution layer is used to perform a linear transformation on the pooled output; then the sigmoid activation function is used to limit the output to between 0 and 1, representing the weight of each channel; the channel attention module passes the pooled features through the convolution layer and the activation function, and then multiplies them element by element with the input features to generate a weighted feature map.

[0031] Preferably, the spatial attention module calculates the average value and the maximum value of the input feature map in the channel dimension respectively, generates two single-channel feature maps, and then splices the two single-channel feature maps in the channel dimension to generate a feature map with two channels; then uses a 3×3 or 7×7 convolution layer to perform a convolution operation on the spliced ​​feature map to obtain a single-channel feature map; then uses a sigmoid activation function to limit the output to between 0 and 1 to represent the weight of each position of the feature map; the spatial attention module passes the concatenated feature of the average feature and the maximum feature through the convolution layer and the activation function, and then multiplies it element by element with the input feature to generate a weighted feature map.

[0032] Compared with existing technologies, convolutional neural networks (CNNs) have attracted widespread attention due to their outstanding performance in the field of image processing. Through multi-layer convolution and pooling operations, CNNs can effectively extract image features and achieve efficient and accurate results in various image processing tasks. In the field of image denoising, CNNs automatically learn complex nonlinear denoising models through an end-to-end learning framework, which has significant advantages over traditional linear or nonlinear filtering methods.

[0033] Combining the performance of various denoising methods under various noise conditions and their denoising capabilities in real scenes, the neural network denoising method proposed in the present invention performs well and can be well qualified for image denoising tasks in power facility scenes such as power plants. This method not only shows high stability and robustness under different noise levels, but also can significantly remove image noise and ensure good image details when dealing with the real noise environment of power plants. Therefore, the denoising method proposed in the present invention has wide applicability and significant practical value in practical applications, can effectively solve the image noise problem encountered in power facility monitoring, and provide important technical support for the safe operation of power systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to make the purpose, technical solution and beneficial effects of the present invention clearer, the present invention provides the following drawings for illustration:

[0035] Figure 1 A network structure diagram of an image denoising method based on a convolutional neural network according to an embodiment of the present invention;

[0036] Figure 2 Schematic diagram of a multi-scale dilated convolution module of an image denoising method based on a convolutional neural network according to an embodiment of the present invention;

[0037] Figure 3 Schematic diagram of C2f_n module in the image denoising method S20 based on convolutional neural network according to an embodiment of the present invention;

[0038] Figure 4Schematic diagram of the attention mechanism module in the image denoising method S20 based on convolutional neural network according to an embodiment of the present invention;

[0039] Figure 5 Schematic diagram of a channel attention module in the image denoising method S20 based on a convolutional neural network according to an embodiment of the present invention;

[0040] Figure 6 Schematic diagram of a spatial attention module in S20 of an image denoising method based on a convolutional neural network according to an embodiment of the present invention;

[0041] Figure 7 It is a PSNR index diagram of the image denoising method based on convolutional neural network according to an embodiment of the present invention and other methods in the prior art;

[0042] Figure 8 This is an SSIM index diagram of the image denoising method based on convolutional neural network according to an embodiment of the present invention and other methods in the prior art. DETAILED DESCRIPTION

[0043] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0044] See also Figure 1 The image denoising method based on convolutional neural network of the present invention comprises the following steps:

[0045] S10, the noisy image passes through a multi-scale dilated convolution module StartConv;

[0046] S20, after the feature extraction stage composed of 3CBlock_n module and Conv module, the input of each layer is a feature map merged with the output of the previous layer using the skip connection method, and the Concat operation is used ( Figure 1 C in the middle represents concatenating feature maps of different layers together;

[0047] S30, using the output of the EndConv module as the noise image, and subtracting the noise image from the original image to obtain a denoised image.

[0048] The "Concat" operation is used multiple times to achieve jump connections by splicing feature maps of different layers together. This connection method allows the network to directly transfer information between layers of different depths, integrating the features of shallow and deep layers, so that the model can capture both low-level and high-level features. Through this strategy, the network can alleviate the gradient vanishing problem and ensure that the information of the previous layer can be effectively transferred to the next layer, thereby enhancing the feature expression ability of the model and improving the overall performance. Specifically, each "Concat" operation splices the outputs of the previous layers together, so that subsequent convolution operations can be processed based on richer input features, thereby improving the denoising effect.

[0049] The multi-scale dilated convolution module includes: a standard convolution module and three dilated convolution modules with dilation rates of 2, 3, and 4, respectively, and the four convolution outputs are merged together.

[0050] See also Figure 2 The multi-scale dilated convolution module includes 4 Conv modules in parallel. Each Conv module inputs a 3-channel noise image and outputs an 8-channel feature map, with a total of 32 channels. The size of each convolution kernel is 3×3, the step size is 1, the dilation rates are 1, 2, 3 and 4, and the filling rates are 1, 2, 3 and 4. The larger the dilation rate, the larger the convolution receptive field R. The formula is:

[0051] R=(k-1)*d*s+1 (3)

[0052] In the formula, k is the convolution kernel size, d is the convolution dilation rate, and s is the step size;

[0053] The Conv module uses convolution conv and then batch normalization bn and then SiLU function activation. The formula is:

[0054] F out =SiLU(bn(conv(F in ))) (4)

[0055] The convolution kernel size of the multi-scale dilated convolution module is k×k=3×3, the step size is s=1, and the dilation rate is d={1, 2, 3, 4}. According to formula (3), the receptive field of the multi-scale dilated convolution module is {3, 5, 7, 9}. Generally, the larger the receptive field of the convolution, the more global features the model tends to capture and some local details may be ignored; the smaller the receptive field, the more attention the model pays to local details. The StartConv module contains both a large receptive field {7, 9} and a small receptive field {3, 5}, so that the module provides the ability of multi-scale feature extraction. This design not only enhances the feature representation ability of the model, but also better balances the capture of details and global information, improving the performance and robustness of the model in image denoising tasks.

[0056] The 3CBlock_n module in S20 includes a Conv module with convolution kernel k=3 and step size s=1, a C2f_n module, and an attention mechanism module. Before the 3CBlock_n module, some also use a Conv module with k=1 and s=1, which is mainly used to adjust the number of channels of the feature map after Concat and integrate channel information.

[0057] See also Figure 3 , C2f_n module is the feature extraction module in YOLOv8, which consists of two 1x1 convolutional layers and several Bottleneck modules. The first 1x1 convolutional layer reduces the number of input channels and divides them into two parts. Several Bottleneck modules extract features from one part, and finally all features are concatenated and converted into the number of output channels through the second 1x1 convolutional layer. Its function is to reduce the amount of calculation and parameters, improve the feature extraction capability and gradient flow efficiency, thereby enhancing the performance and training efficiency of the model.

[0058] See also Figure 4 The attention mechanism module consists of a channel attention module and a spatial attention module connected in series, which enhances the convolutional neural network's attention to important channels and spatial positions in the feature map, thereby improving the model's sensitivity to key information and overall performance. The channel attention module emphasizes discriminative channels, while the spatial attention module focuses on important spatial positions. The two work together to significantly improve the feature extraction effect.

[0059] See also Figure 5 , the channel attention module first performs global average pooling on each channel, and the output size is (batch, channels, 1, 1), that is, a global average of each channel, batch is the number of batches, channels is the number of channels; then a 1x1 convolution layer is used to linearly transform the pooled output; then the sigmoid activation function is used to limit the output between 0 and 1, representing the weight of each channel; the channel attention module passes the pooled features through the convolution layer and activation function, and then multiplies them element by element with the input features to generate a weighted feature map. The formula is:

[0060] F out =F in ·M c (F in )=F in ·σ(fc(AP(F in )) (5)

[0061] In the formula, F in is the input feature map, F out is the output feature map, M cis the feature map channel weight, σ is the sigmoid activation function, fc is the 1×1 convolution layer processing, and AP is the global average pooling processing of the channel.

[0062] The channel attention module captures the importance of each channel and weights each channel of the input feature map, highlighting important channels and suppressing unimportant channels. In this way, the model can pay more attention to discriminative features.

[0063] See also Figure 6 , the spatial attention module calculates the average and maximum values ​​of the input feature map in the channel dimension respectively, generates two single-channel feature maps, and then splices the two single-channel feature maps in the channel dimension to generate a feature map with two channels; then uses a 3×3 or 7×7 convolution layer to perform a convolution operation on the spliced ​​feature map to obtain a single-channel feature map; then uses a sigmoid activation function to limit the output between 0 and 1 to represent the weight of each position in the feature map; the spatial attention module passes the concatenated features of the average feature and the maximum feature through the convolution layer and activation function, and then multiplies them element by element with the input feature to generate a weighted feature map. The formula is:

[0064] F out =F in ·M s (F in )

[0065] =F in ·σ(conv(cat([F Max ,F Mean ])) (6)

[0066] In the formula, F in is the input feature map, F out is the output feature map, M s is the feature map channel weight, σ is the sigmoid activation function, conv is the 3×3 (or 7×7) convolution layer processing, cat is the concatenation (Concat) operation of the feature map in the channel dimension, F Max and F Mean is the feature map F in Maximum and average values ​​over the channel dimension.

[0067] The spatial attention module captures the importance of the feature map in the spatial dimension and weights each spatial position of the input feature map, highlighting important positions and suppressing unimportant positions. This can help the model focus more on the spatial distribution of the feature map.

[0068] In the 3CBlock_n module, the combination of the Conv module and the C2f_n module can extract and fuse features of different scales. The Conv module extracts preliminary spatial features, and the C2f_n module performs deep feature extraction and fusion through multiple Bottleneck layers. By adding the CBAM module at the end, important features are further enhanced and irrelevant information is suppressed, so that the model can focus more on key parts and improve the overall feature expression ability. This combination can not only improve the feature extraction ability of the model, but also enhance the robustness and accuracy of the model through the attention mechanism, making the denoising effect better.

[0069] Although subjective evaluation can provide intuitive visual comparison, its results are affected by subjective feelings, differences between evaluators and the evaluation environment, and have certain subjectivity and limitations. In order to overcome the limitations of subjective evaluation, this paper uses the DIV2K dataset as the benchmark for objective evaluation, and uses classic evaluation indicators such as peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) for objective quantitative evaluation. The DIV2K dataset provides high-quality reference images that can effectively measure the performance of different algorithms in noise suppression.

[0070] Due to the lack of real data sets under objective quantitative evaluation, the present invention selects 100 DIV2K high-definition validation set images as experimental data, cuts these 100 images into 640×640 size, and adds Poisson noise, Gaussian Poisson noise (Gaussian standard deviation is 15, referred to as Gaussian noise in this invention), Gaussian noise with standard deviation of 20 (referred to as Gaussian 20), Gaussian noise with standard deviation of 30 (Gaussian 30), Gaussian noise with standard deviation of (Gaussian 40) (Gaussian 40) and Gaussian noise with standard deviation of 50 (Gaussian 50). Compare the denoising ability of each convolutional neural network under different noise levels, and use the average structural similarity (SSIM) and average peak signal-to-noise ratio (PSNR) to comprehensively evaluate the denoising ability of each model.

[0071] Peak signal-to-noise ratio (PSNR) is a commonly used evaluation metric in image denoising, which evaluates image quality by quantifying the difference between the denoised image and the original image. A higher PSNR value usually indicates higher image fidelity and less noise, so when comparing the effects of different denoising algorithms, PSNR provides an intuitive and standardized benchmark.

[0072] The calculation steps of PSNR are as follows:

[0073] (1) Calculate Mean Squared Error (MSE)

[0074] MSE measures the average difference in pixel values ​​between two images. Assuming the original image is I and the denoised image is K, and their sizes are m×n, the MSE calculation formula is:

[0075]

[0076] Where I(i,j) and K(i,j) represent the pixel value of the image at position (i,j).

[0077] (2) Calculate the peak signal-to-noise ratio (PSNR)

[0078] PSNR uses the MSE value to calculate the signal-to-noise ratio. The formula is as follows:

[0079]

[0080] In the formula, MAX I is the maximum pixel value in image I.

[0081] The unit of PSNR is decibel (dB). The higher the value, the more similar the two images are, that is, the less noise there is and the better the image quality. Generally speaking, when the PSNR value is above 30dB, the denoising effect is considered good; the higher the value, the better the denoising effect.

[0082] Structural similarity (SSIM) is an indicator used to evaluate the structural similarity between images. It can effectively reflect the similarity of images in terms of brightness, contrast and texture. Compared with PSNR, SSIM can better capture the differences in image quality perceived by the human visual system.

[0083] The calculation steps of SSIM are as follows:

[0084] (1) Brightness comparison

[0085] Brightness comparison is used to measure the average brightness difference of images. The formula is as follows:

[0086]

[0087] In the formula, μ x and μ y They represent the average brightness of image x and image y respectively, and C1 is a constant used to avoid the denominator being zero.

[0088] (2) Contrast comparison

[0089] Contrast similarity is used to measure the contrast difference of images. The formula is as follows:

[0090]

[0091] In the formula, σ x and σy They represent the standard deviation of image x and image y respectively, and C2 is a constant used to avoid the denominator being zero.

[0092] (3) Texture comparison

[0093] Texture comparison is used to measure the similarity of texture information of images. The formula is as follows:

[0094]

[0095] In the formula, σ xy Represents the covariance of images x and y, and C3 is a constant used to avoid the denominator being zero.

[0096] (4) Comprehensive calculation of SSIM index

[0097] The comparison results of brightness, contrast and texture are combined to obtain the final SSIM index. The formula is as follows:

[0098] SSIM(x,y)=[L(x,y)] α ·[C(x,y)] β ·[S(x,y)] γ (12)

[0099] Where α, β and γ are adjustment parameters, which are usually set to 1.

[0100] The SSIM value range is between -1 and 1, where 1 means that the two images are exactly the same and -1 means that the pixel values ​​of the two images are completely opposite. Unlike PSNR, SSIM pays more attention to the structural information of the image, so it is more in line with human visual perception in evaluating image quality.

[0101] By using the two indicators of PSNR and SSIM, the present invention can comprehensively and objectively evaluate the denoising performance of different convolutional neural networks under various noise conditions. Combined with the high-quality reference images in the DIV2K dataset, these indicators provide a reliable quantitative basis for the effect of image denoising algorithms.

[0102] Calculate the PSNR and SSIM indicators of 100 high-definition images in the DIV2K dataset after various noise pollution and the original images, and take the average value, see Figure 7 , Figure 8 .

[0103] The DnCNN model performs very well at lower noise levels (such as Poisson noise, Gaussian Poisson noise, Gaussian noise 20 and 30), with the highest or near-highest SSIM and PSNR values. However, the performance of DNCNN drops significantly at high noise levels (such as Gaussian noise 40 and 50).

[0104] The FFDNet model has good denoising ability only under its specified noise level. For example, the FFDNet30 model achieves the best denoising effect only when the Gaussian noise has a standard deviation of 30, and the FFDNet50 model achieves relatively good denoising effect only when the Gaussian noise has a standard deviation of 50. However, this operation is unrealistic in real scenarios because we cannot accurately judge the noise level.

[0105] The DPSR, IMDN and IRCNN models have poor denoising capabilities under various noise levels.

[0106] The model of the present invention performs best under high noise levels (such as Gaussian noise 40 and 50), with the highest SSIM and PSNR values, showing its superiority in dealing with severe noise. Although the method of the present invention does not always perform best under low noise levels (such as Poisson noise, Gaussian Poisson noise, Gaussian noise 20 and 30), its performance is relatively stable and there is no significant performance degradation. The performance of the method of the present invention is relatively consistent under various noise levels, and it can provide higher denoising effects and consistency regardless of low or high noise conditions. This stability enables it to provide more reliable denoising effects when facing real scenes with unknown noise levels and types.

[0107] In addition to the above embodiments, the present invention may also have other implementation modes. Any technical solution formed by equivalent replacement or equivalent transformation is within the protection scope required by the present invention.

[0108] The present invention is described in detail above, but the specific implementation of the present invention is not limited thereto. Those skilled in the art may make various modifications or alterations without departing from the spirit and scope of the claims of the present application.

Claims

1. An image denoising method based on convolutional neural network, characterized in that: The following steps are involved: S10, the noisy image passes through a multi-scale dilated convolution module; S20, after the feature extraction stage composed of 3CBlock_n module and Conv module, the input of each layer is the feature map merged with the output of the previous layer by using skip connection, and the feature maps of different layers are spliced ​​together by Concat operation; S30, using the output of the EndConv module as the noise image, and subtracting the noise image from the original image to obtain a denoised image.

2. The image denoising method based on convolutional neural network according to claim 1, characterized in that: The multi-scale dilated convolution module includes: a standard convolution module and three dilated convolution modules with dilated rates of 2, 3, and 4, respectively, and the four convolution outputs are merged together.

3. The image denoising method based on convolutional neural network according to claim 1, characterized in that: The multi-scale dilated convolution module includes 4 Conv modules in parallel. Each Conv module inputs a 3-channel noise image and outputs an 8-channel feature map, with a total of 32 channels. The size of each convolution kernel is 3×3, the step size is 1, the dilation rates are 1, 2, 3 and 4, and the filling rates are 1, 2, 3 and 4. The larger the dilation rate, the larger the convolution receptive field R. The formula is: R=(k-1)*d*s+1 (3) In the formula, k is the convolution kernel size, d is the convolution dilation rate, and s is the step size; The Conv module uses convolution followed by batch normalization and then SiLU function activation. The formula is: F out =SiLU(bn(conv(F in ))) (4) The convolution kernel size of the multi-scale dilated convolution module is k×k=3×3, the step size is s=1, and the dilation rate is d={1, 2, 3, 4}. According to formula (3), the receptive field of the multi-scale dilated convolution module is {3, 5, 7, 9}.

4. The image denoising method based on convolutional neural network according to claim 1, characterized in that: The 3CBlock_n module in S20 includes a Conv module with a convolution kernel k=3 and a step size s=1, a C2f_n module and an attention mechanism module.

5. The image denoising method based on convolutional neural network according to claim 4, characterized in that: The C2f_n module is a feature extraction module in YOLOv8, which consists of two 1x1 convolutional layers and several Bottleneck modules. The first 1x1 convolutional layer reduces the number of input channels and divides them into two parts. Several Bottleneck modules extract features from one part, and finally all features are concatenated and converted into the number of output channels through the second 1x1 convolutional layer.

6. The image denoising method based on convolutional neural network according to claim 4, characterized in that: The attention mechanism module includes a channel attention module and a spatial attention module connected in series.

7. The image denoising method based on convolutional neural network according to claim 5, characterized in that: The channel attention module first performs global average pooling on each channel, and the output size is (batch, channels, 1, 1), that is, a global average of each channel, batch is the number of batches, and channels is the number of channels; then a 1x1 convolution layer is used to perform a linear transformation on the pooled output; then the sigmoid activation function is used to limit the output between 0 and 1, representing the weight of each channel; the channel attention module passes the pooled features through the convolution layer and activation function, and then multiplies them element by element with the input features to generate a weighted feature map.

8. The image denoising method based on convolutional neural network according to claim 5, characterized in that: The spatial attention module calculates the average value and the maximum value of the input feature map in the channel dimension respectively, generates two single-channel feature maps, and then splices the two single-channel feature maps in the channel dimension to generate a feature map with two channels; then a 3×3 or 7×7 convolution layer is used to perform a convolution operation on the spliced ​​feature map to obtain a single-channel feature map; then a sigmoid activation function is used to limit the output to between 0 and 1 to represent the weight of each position of the feature map; the spatial attention module passes the concatenated feature of the average feature and the maximum feature through the convolution layer and the activation function, and then multiplies it element by element with the input feature to generate a weighted feature map.

Images