An image denoising method based on convolutional neural network

Abstract

The application discloses a kind of based on convolutional neural network's image denoising method, comprising the following steps: S10, containing noise image passes through a multiscale hole convolution module;S20, after being extracted by the feature of 3CBlock_n module and Conv module combination stage, using the way of skip connection, each layer input is the feature map that is merged with the output of previous layer, using Concat operation to splice the feature map of different layers together;S30, with the output of EndConv module as noise image, the image after denoising is obtained by subtracting noise image from original image.The denoising method proposed in the application has wide applicability and significant practical value in practical application, and can effectively solve the image noise problem encountered in power facility monitoring, providing important technical support for safe operation of power system.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of image processing, and particularly relates to an image denoising method based on a convolutional neural network. BACKGROUND

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

[0003] Image denoising is an important step in image processing, aiming to recover a clear image as close as possible to the original image from a noisy image. Traditional image denoising methods mainly include various techniques based on spatial and frequency domains, which usually utilize local and frequency characteristics of images to distinguish noise and useful information. Common image noise includes Poisson noise, Gaussian noise, and salt and pepper noise, etc., and Gaussian Poisson noise is more consistent with real image noise. Common image denoising methods include mean filtering, median filtering, and Gaussian filtering, etc.

[0004] Mean filtering: taking the target pixel point as the center, the average value of all pixel points in its local domain is taken as the filtered pixel value of the point.

[0005] Mean filtering algorithm is simple, has small computational complexity, and is easy to implement and apply, suitable for real-time processing applications. It can effectively smooth high-frequency noise and reduce random noise in images. However, it blurs the details and edges in the image while removing noise, resulting in a decline in image quality. Moreover, it only considers local region information and cannot effectively utilize global characteristics of the image for denoising.

[0006] Gaussian filtering: the weighted coefficients of pixel values of each point in the filter kernel are calculated using a Gaussian function, and the weighted average is taken as the pixel value of the center point of the filter kernel. The Gaussian function generally uses formula (1).

[0007]

[0008] In the formula:

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

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

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

[0012] Taking a 3x3 filter kernel as an example, its (Δx, Δy) distribution is (-1, -1)~(1, 1). The Gaussian function calculated according to formula (1) and the pixel point coordinates of the Gaussian filter kernel also needs to be normalized according to formula (2) to ensure that the sum of the weighting coefficients is 1.

[0013]

[0014] Median filtering: Median filtering is a nonlinear filtering technique commonly used for image denoising, which mainly replaces the center pixel value with the median value of the pixel values in the window to achieve the effect of removing noise.

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

[0016] Therefore, the present application proposes an image denoising method based on a convolutional neural network, comprising the following steps:

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

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

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

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

[0021] Preferably, the multi-scale dilated convolution module comprises four Conv modules in parallel, each Conv module inputs a 3-channel noise image, and outputs an 8-channel feature map, a total of 32 channels, and each convolution kernel size is 3x3, the step is 1, the dilated rate is 1, 2, 3 and 4, and the padding is 1, 2, 3 and 4; the larger the dilated rate, the larger the convolution receptive field R, and its formula is:

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

[0023] where k is the size of the convolution kernel, d is the convolution dilation rate, and s is the step size.

[0024] The Conv module is activated by the SiLU function after batch normalization processing, and the formula is:

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

[0026] The size of the convolution kernel of the multi-scale dilated convolution module is k x k = 3 x 3, the step size s = 1, and the dilation rate 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 the 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 is composed of two 1x1 convolution layers and a plurality of Bottleneck modules. The first 1x1 convolution layer reduces the input channel number and divides it into two parts. The plurality of Bottleneck modules extract features from one part. Finally, the second 1x1 convolution layer splices and converts all features into an output channel number.

[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, where batch is the batch size and channels is the number of channels. Then, a 1x1 convolution layer is used to linearly transform the pooled output. Then, the output is limited to between 0 and 1 by a sigmoid activation function, representing the weight of each channel. The channel attention module multiplies the weighted feature map generated by the pooled features, the convolution layer, and the activation function with the input features element by element.

[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, 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 convolution operation on the spliced feature map to obtain a single-channel feature map, and then limits the output to 0 to 1 through a sigmoid activation function to represent the weight of each position of the feature map. The spatial attention module splices the average feature and the maximum feature, and then performs convolution layer and activation function on the spliced feature, and then element-wise multiplies the input feature to generate a weighted feature map.

[0032] Compared with the prior art, the convolutional neural network CNN is widely concerned due to its excellent performance in the field of image processing. The CNN can effectively extract the features of the image through multiple layers of convolution and pooling operation, and realize efficient and accurate results in various image processing tasks. In the field of image denoising, the CNN automatically learns a complex nonlinear denoising model through an end-to-end learning framework, which has a significant advantage over traditional linear or nonlinear filtering methods.

[0033] In summary, the performance of various denoising methods under various noise conditions and their denoising ability in real scenes, the neural network denoising method proposed in the present application performs well and can well perform the image denoising task in the power plant and other power facility scenes. The method not only has high stability and robustness under different noise levels, but also can significantly remove image noise and ensure good image details when processing real noise environment of power plants. Therefore, the denoising method proposed in the present application has wide applicability and significant practical value in practical application, and can effectively solve the image noise problem encountered in power facility monitoring, and provide important technical support for the safe operation of power system. BRIEF DESCRIPTION OF DRAWINGS

[0034] In order to make the purpose, technical scheme and beneficial effects of the present application more clear, the present application provides the following drawings for illustration:

[0035] Figure 1 The network structure diagram of the image denoising method based on the convolutional neural network of the embodiment of the present application is shown in the figure;

[0036] Figure 2 The multi-scale dilated convolution module schematic diagram of the image denoising method based on the convolutional neural network of the embodiment of the present application is shown in the figure;

[0037] Figure 3 The C2f_n module schematic diagram in S20 of the image denoising method based on the convolutional neural network of the embodiment of the present application is shown in the figure;

[0038] Figure 4Fig. 1 is a schematic diagram of an attention mechanism module in the image denoising method based on a convolutional neural network according to an embodiment of the present application;

[0039] Figure 5 Fig. 2 is a schematic diagram of a channel attention module in the image denoising method based on a convolutional neural network according to an embodiment of the present application;

[0040] Figure 6 Fig. 3 is a schematic diagram of a spatial attention module in the image denoising method based on a convolutional neural network according to an embodiment of the present application;

[0041] Figure 7 Fig. 4 is a PSNR index graph of the image denoising method based on a convolutional neural network according to an embodiment of the present application and other methods in the prior art;

[0042] Figure 8 Fig. 5 is a SSIM index graph of the image denoising method based on a convolutional neural network according to an embodiment of the present application and other methods in the prior art. DETAILED DESCRIPTION

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

[0044] Referring to Figure 1 , the image denoising method based on a convolutional neural network according to the present application includes the following steps:

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

[0046] S20, a feature extraction stage composed of a 3CBlock_n module and a Conv module is adopted, and a skip connection is used, wherein the input of each layer is a feature map obtained by merging the output of the previous layer, and different layers of feature maps are spliced together using a Concat operation (C represents) in the middle. Figure 1

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

[0048] ​The "Concat" operation is used multiple times to concatenate the feature maps of different layers, realizing a skip connection. This connection allows the network to directly pass information between different depth layers, fusing 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, ensuring that the information of the previous layers can be effectively passed to the later layers, thereby enhancing the feature expression ability of the model and improving the overall performance. Specifically, each "Concat" operation concatenates the outputs of the previous layers, allowing the subsequent convolution operation to be based on richer input features, thereby improving the denoising effect.

[0049] The multi-scale dilated convolution module includes one standard convolution module and three dilated convolution modules with dilated rates of 2, 3, and 4, and the outputs of the four convolutions are combined together.

[0050] Referring to Figure 2 , the multi-scale dilated convolution module includes four Conv modules in parallel, each with a 3-channel noise image as input and an 8-channel feature map as output, a total of 32 channels, and each convolution kernel size is 3x3, the step is 1, the dilated rate is 1, 2, 3, and 4, and the padding is also 1, 2, 3, and 4; the larger the dilated rate, the larger the convolution receptive field R, and its formula is:

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

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

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

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

[0055] The convolution kernel size k x k of the multi-scale dilated convolution module is 3 x 3, the step s is 1, and the dilated rate d is {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 convolution, the more global features the model tends to capture, which may ignore some local details; the smaller the receptive field, the more the model focuses on local details. The StartConv module contains both large receptive fields {7, 9} and small receptive fields {3, 5}, enabling the module to provide multi-scale feature extraction capabilities. 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 the image denoising task.

[0056] 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. Before the 3CBlock_n module, some also use a Conv module with k=1 and s=1, mainly to adjust the channel number of the concatenated feature map and integrate channel information.

[0057] Referring to Figure 3 , the C2f_n module is a feature extraction module in YOLOv8, which is composed of two 1x1 convolution layers and several Bottleneck modules. The first 1x1 convolution layer reduces the input channel number and divides it into two parts. The several Bottleneck modules extract features from one part. Finally, the second 1x1 convolution layer splices and converts all features into the output channel number. Its role is to reduce the amount of calculation and parameters, improve the feature extraction ability and gradient flow efficiency, and thus enhance the performance and training efficiency of the model.

[0058] Referring to Figure 4 , the attention mechanism module includes a channel attention module and a spatial attention module connected in series, which enhances the attention of the convolutional neural network to important channels and spatial positions in the feature map, thereby improving the sensitivity of the model 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 feature extraction.

[0059] Referring to Figure 5 , the channel attention module first performs global average pooling on each channel, outputting a size of (batch, channels, 1, 1), i.e., a global average value for each channel, with batch being the batch size and channels being the number of channels. Then, a 1x1 convolution layer is used to perform 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 multiplies the weighted feature map generated by the pooled features after the convolution layer and the activation function with the input features element by element. 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, and M cF in is the input feature map, F out is the output feature map, M s is the channel weight of the feature map, σ is the sigmoid activation function, fc is the 1x1 convolution layer processing, and AP is the global average pooling processing on the channel.

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

[0063] Referring to 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 concatenates the two single-channel feature maps in the channel dimension to generate a feature map with two channels. Then, a 3x3 or 7x7 convolution layer is used to perform convolution operation on the concatenated 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 average feature and the maximum feature after concatenation are passed through the convolution layer and the activation function, and then multiplied 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 channel weight of the feature map, σ is the sigmoid activation function, conv is the 3x3 (or 7x7) convolution layer processing, cat is the concatenation (Concat) operation of the feature map in the channel dimension, F Max and F Mean are the maximum and average values of the feature map F in in the channel dimension.

[0067] The spatial attention module weights each spatial position of the input feature map by capturing the importance of the feature map in the spatial dimension, highlights important positions, and suppresses 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 extracts and fuses deep features through multiple Bottleneck layers. By adding a CBAM module at the end, important features are further enhanced, irrelevant information is suppressed, and the model can focus more on key parts, improving the overall feature expression capability. This combination not only improves the feature extraction capability of the model, but also enhances the robustness and precision 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, evaluator differences and evaluation environment, and there is certain subjectivity and limitation. In order to overcome the limitations of subjective evaluation, the present application uses the DIV2K data set as the benchmark for objective evaluation, and uses classical evaluation indicators such as peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) for objective quantitative evaluation. The DIV2K data set provides high-quality reference images, which 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 application selects 100 DIV2K high-definition verification set images as experimental data, crops the 100 images to 640x640 size, and respectively adds Poisson noise, Gaussian Poisson noise (Gaussian standard deviation is 15, which is referred to as high-peak noise in the present application), Gaussian noise with a standard deviation of 20 (referred to as Gaussian 20), Gaussian noise with a standard deviation of 30 (Gaussian 30), Gaussian noise with a standard deviation of 40 (Gaussian 40), and Gaussian noise with a 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 the 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 index in image denoising, which quantifies the difference between the denoised image and the original image to evaluate image quality. Higher PSNR values generally indicate higher image fidelity and less noise, so PSNR provides an intuitive and standardized benchmark when comparing the effects of different denoising algorithms.

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

[0073] (1) Calculate the 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, both of size m x n, the MSE calculation formula is:

[0075]

[0076] where I(i,j) and K(i,j) represent the pixel values at position (i,j) in the images.

[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] where MAX I is the maximum pixel value in the image I.

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

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

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

[0084] (1) Brightness comparison

[0085] The brightness comparison is used to measure the average brightness difference of the image, the formula is as follows:

[0086]

[0087] where μ x and μ y represent the average brightness of images x and y respectively, C1 is a constant to avoid the case of zero denominator.

[0088] (2) Contrast comparison

[0089] The contrast similarity is used to measure the contrast difference of the image, the formula is as follows:

[0090]

[0091] where σ x and σy respectively, C2 is a constant to avoid the case that the denominator is zero.

[0092] (3) Texture comparison

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

[0094]

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

[0096] (4) Comprehensive calculation of SSIM index

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

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

[0099] In the formula, α, β and γ are adjustment parameters, usually all taking 1.

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

[0101] By using the two indexes of PSNR and SSIM, the 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 indexes provide reliable quantitative basis for the effect of image denoising algorithm.

[0102] The PSNR and SSIM indexes between the 100 high-definition images in the DIV2K dataset after being polluted by various noises and the original images are calculated, and the average value is taken, as shown in Figure 7 、 Figure 8 .

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

[0104] The FFDNet model has good denoising ability only at its specified noise level, for example, the FFDNet30 model achieves the best denoising effect only in the case of Gaussian noise with a standard deviation of 30, and the FFDNet50 model achieves relatively good denoising effect only in the case of Gaussian noise with a standard deviation of 50, however, this operation is not realistic for real scenes, because we cannot accurately judge the noise level.

[0105] The denoising ability of the DPSR, IMDN and IRCNN models is poor at multiple noise levels.

[0106] The model of the present application performs best at high noise levels (such as Gaussian noise 40 and 50), with the highest SSIM and PSNR values, showing its superiority in handling severe noise. Although the method of the present application does not always perform best at 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 decline. The method of the present application performs consistently at various noise levels, whether it is low noise or high noise condition, and can provide high denoising effect and consistency. This stability enables it to provide more reliable denoising effect when facing real scenes with unknown noise level and type.

[0107] In addition to the above embodiments, the present application can have other implementation manners. Any technical solutions formed by equivalent substitution or equivalent transformation are within the scope of protection required by the present application.

[0108] The above has been a detailed description of the present application, but the specific implementation form of the present application is not limited thereto. Those skilled in the art can make various modifications or modifications without departing from the spirit and scope of the claims of the present application.

Claims

1. A method for image denoising based on a convolutional neural network, characterized in that, The method comprises the following steps: S10, the image containing noise is input into a multi-scale hollow convolution module; S20, a feature extraction stage combined by a 3CBlock_n module and a Conv module is adopted, and each layer input is a feature map combined with the output of the previous layer, and the feature maps of different layers are spliced together by using a Concat operation; S30, the output of the EndConv module is taken as the noise image, and the denoised image is obtained by subtracting the noise image from the original image; The multi-scale hollow convolution module comprises a standard convolution module and three hollow convolution modules with hollow rates of 2, 3 and 4 respectively, and the outputs of the four convolution modules are combined together; The multi-scale hollow convolution module comprises four Conv modules in parallel, each Conv module inputs a 3-channel noise image, and outputs an 8-channel feature map, a total of 32 channels, and each convolution kernel has a size of 3x3, a step of 1, a hollow rate of 1, 2, 3 and 4, and a padding of 1, 2, 3 and 4; the greater the hollow rate, the greater the convolution receptive field R, and the formula is: （3）； In the formula, k is the size of the convolution kernel, d is the convolution hollow rate, and s is the step; The Conv module adopts a SiLU function activation after batch normalization processing after convolution, and the formula is: （4）； The size of the convolution kernel of the multi-scale hollow convolution module is kxk=3x3, the step is s=1, the hollow rate is d={1, 2, 3, 4}, and according to the formula (3), the receptive field of the multi-scale hollow convolution module is {3, 5, 7, 9}.

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

3. The method of claim 2, wherein the CNN is a U-Net. The C2f_n module is a feature extraction module in YOLOv8, which comprises two 1x1 convolution layers and a plurality of Bottleneck modules, wherein the first 1x1 convolution layer reduces the input channel number and divides it into two parts, the plurality of Bottleneck modules perform feature extraction on one part, and finally all features are spliced and converted into the output channel number through the second 1x1 convolution layer.

4. The image denoising method based on a convolutional neural network according to claim 2, characterized in that, The attention mechanism module comprises a channel attention module and a spatial attention module connected in series.

5. The method of claim 4, wherein the CNN-based image denoising method is a method of training a CNN-based image denoising model. The channel attention module firstly performs global average pooling on each channel, and outputs a global average value of each channel with a size of (batch, channels, 1, 1), wherein batch is the batch number and channels is the channel number; then a 1x1 convolution layer is used to linearly transform 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 combines the pooled features through the convolution layer and the activation function, and then element-wise multiplies the input features to generate a weighted feature map.

6. The method of claim 4, wherein the CNN-based image denoising method is a method of training a CNN-based image denoising model. The spatial attention module respectively calculates the average value and the maximum value of the input feature map in the channel dimension, generates two single-channel feature maps, splices the two single-channel feature maps in the channel dimension, and generates a feature map with two channels; subsequently, a 3*3 or 7*7 convolution layer is used to perform convolution operation on the spliced feature map, and a single-channel feature map is obtained; 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 splices the average feature and the maximum feature, and then performs convolution layer and activation function on the spliced feature, and then elementwisely multiplies the input feature to generate a weighted feature map.

Images