Deep learning denoising model and method based on feature fusion and attention mechanism

Abstract

The application belongs to the field of image denoising, and relates to a deep learning denoising model and method based on feature fusion and an attention mechanism, comprising a first feature fusion module, which is used for extracting first common features and first attention features from an input noise image, and fusing the extracted first common features and first attention features to obtain first fused features; a VGGB module; a second feature fusion module; a dilated convolution module, which is used for optimizing the spatio-temporal prediction of the second fused features to obtain optimized receptive field features; an attention mechanism module, which is used for extracting image local information in the optimized receptive field features through the attention mechanism to obtain an attention feature map; and a residual module. The first feature fusion module, the VGGB module, the second feature fusion module, the dilated convolution module, the attention mechanism module and the residual module are combined to form a denoising model, so that the problem of low image quality caused by the loss of image local information during image denoising and the poor image denoising effect is solved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of image denoising, and particularly discloses a deep learning denoising model and method based on feature fusion and an attention mechanism. BACKGROUND

[0002] During shooting, transmission and storage, images are easily polluted by noise due to imaging equipment and external noise interference, thereby affecting the visual quality of the images. Image denoising is widely used in actual production and life and is a main way to improve image quality. Therefore, image denoising is an important research topic in the field of computer vision.

[0003] In recent years, deep learning has been rapidly developed and widely applied in the field of image denoising. Deep learning can adapt to different types and complex noises by learning a large number of noisy images, and achieves better image denoising performance due to its learnable characteristics. Therefore, researchers have designed many neural network structures for image denoising. For example, Jain et al. designed a convolutional neural network denoising model for the first time, proving that a convolutional neural network can be used for image denoising; Burger et al. proposed a general multilayer perceptron MLP for image denoising, proving that a multilayer perceptron can be comparable to the then advanced image denoising method; Chen et al. proposed a trainable nonlinear diffusion model TNRD, which improves the calculation speed due to parameter training and structural simplicity; Zhang et al. first introduced the residual idea into the denoising network and proposed a denoising convolutional neural network DnCNN, which realizes high-quality image denoising and is one of the most excellent denoising algorithms at present. Although the above deep learning denoising algorithms achieve high image denoising effects, the convolution parameter learning process pays more attention to the overall information and ignores the local information of the image, resulting in the loss of image structural features and affecting the image denoising effect. The local information of the image is a main factor reflecting the quality of the image and is an important task in the field of image denoising. Therefore, an image denoising model capable of retaining more local information of the image is urgently needed. SUMMARY

[0004] The application aims to provide a deep learning denoising model and method based on feature fusion and an attention mechanism, and solve the problem of low image denoising effect caused by the loss of local information of the image, thereby improving the quality of the image.

[0005] The specific scheme of the application is as follows:

[0006] The deep learning denoising model based on feature fusion and an attention mechanism comprises:

[0007] The first feature fusion module is configured to extract first common features and first attention features from the input noise image, and fuse the extracted first common features and first attention features to obtain first fusion features.

[0008] The VGGB module is configured to extract feature information from the first fusion features layer by layer to obtain reinforced fusion features.

[0009] The second feature fusion module is configured to extract second common features and second attention features from the reinforced fusion features, and fuse the extracted second common features and second attention features to obtain second fusion features.

[0010] The dilated convolution module is configured to optimize the second fusion features to obtain optimized receptive field features.

[0011] The attention mechanism module is configured to extract image local information in the optimized receptive field features by using an attention mechanism to obtain an attention feature map.

[0012] The residual module is configured to fit the attention feature map to obtain a noise component map, and subtract the noise component map from the noise image by using a skip connection to obtain a clean image.

[0013] Preferably, the first feature fusion module comprises:

[0014] The first convolution unit and the first attention unit, the first common features extracted by the first convolution unit are used as input data of the first attention unit, and the first attention features output by the first attention unit are fused with the first common features extracted by the first convolution unit to obtain the first fusion features.

[0015] Preferably, the first convolution unit comprises:

[0016] The first convolution layer, the BN layer and the first ReLU layer are connected in sequence.

[0017] Preferably, the VGGB module comprises: a plurality of fifth convolution layers, a pooling layer and a third ReLU layer connected in sequence.

[0018] Preferably, the second feature fusion module comprises:

[0019] The second convolution unit and the second attention unit, the second common features extracted by the second convolution unit are used as input data of the second attention unit, and the second attention features output by the second attention unit are fused with the second common features extracted by the second convolution unit to obtain the second fusion features.

[0020] The second convolution unit comprises a second convolution layer and a second ReLU layer connected in sequence.

[0021] Preferably, the dilated convolution module comprises:

[0022] three DCB units, the DCB unit comprising a third convolutional layer, a fourth convolutional layer and a dilated convolutional layer.

[0023] Preferably, the attention mechanism is a spatial attention mechanism, and the spatial attention mechanism comprises:

[0024] The maximum pooling is obtained by selecting the maximum value of each sub-region, and the average pooling is obtained by calculating the average value of each sub-region;

[0025] Based on the maximum pooling and the average pooling, the channel information of the optimized receptive field feature is mapped to obtain maximum pooling features and average pooling features, and the maximum pooling features and the average pooling features are spliced and fused to generate fused pooling features.

[0026] The fused pooling features are sequentially subjected to standard 3*3 convolution processing, batch normalization processing and rectified linear unit activation processing to obtain spatial information weights, and the spatial information weights are multiplied by the optimized receptive field features to obtain an attention feature map.

[0027] The present application also relates to a deep learning denoising method based on feature fusion and attention mechanism, comprising the following steps:

[0028] S1, obtaining a training set and a test set, and preprocessing the training set;

[0029] S2, training a denoising model according to the preprocessed training set to obtain a trained denoising model;

[0030] S3, testing and verifying the trained denoising model based on the test set to obtain a verified denoising model;

[0031] S4, inputting the obtained noise image into the verified denoising model to obtain a clean image.

[0032] Preferably, the preprocessing comprises:

[0033] S11, standardization processing, uniformly cropping the original images in the training set to obtain original images of the same size, and performing flip processing on the original images of the same size to obtain standardized images;

[0034] S12, normalization processing, compressing the amplitude range of the standardized images to a preset interval.

[0035] Preferably, S2 comprises:

[0036] According to the preprocessed training set, a training number is obtained, and the training is divided into different stage training based on the training number, and different learning rates are used in different stage training.

[0037] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0038] 1、The application preserves more image information through the first feature fusion module and the second feature fusion module, strengthens image information through the VGGB module, expands the receptive field through the dilated convolution module, maps more context information, extracts image local information through the attention mechanism module, and obtains a clean image through the residual module, so that higher peak signal-to-noise ratio is pursued while rich image local information is preserved, the image denoising effect is improved, the image quality is improved, and the image visual effect is improved. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 A block diagram of the deep learning denoising model based on feature fusion and attention mechanism in the embodiment of the application.

[0040] Figure 2 A flowchart of the deep learning denoising method based on feature fusion and attention mechanism in the embodiment of the application. DETAILED DESCRIPTION

[0041] To make the objectives, technical solutions and advantages of the embodiments of the application clearer, the technical solutions in the embodiments of the application will be described below in detail with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are some but not all of the embodiments of the application. The components of the embodiments of the application described and shown in the drawings herein can be arranged and designed in various different configurations.

[0042] The deep learning denoising model based on feature fusion and attention mechanism, as shown in Figure 1 Fig. 1, comprises a first feature fusion module, a VGGB module, a second feature fusion module, a dilated convolution module, an attention mechanism module and a residual module connected in sequence.

[0043] The denoising model of the application is a backbone network mainly based on part of the network structure of VGG16, and VGG16 is a deep convolutional neural network.

[0044] The first feature fusion module FFB1 is used for extracting first ordinary features and first attention features from the input noise image, and fusing the extracted first ordinary features and first attention features to obtain first fusion features, so as to preserve more image information.

[0045] The first feature fusion module comprises a first convolution unit and a first attention unit. The first convolution unit comprises a first convolution layer Conv, a BN layer and a first ReLU layer, wherein the first convolution layer Conv performs local feature extraction on input data (i.e., a noisy image) through a sliding filter (i.e., a convolution kernel); the BN layer, as a batch normalization layer, performs standardization processing on the output data of the first convolution layer to solve problems such as small gradient and gradient explosion in the training process, and improve the model training efficiency and generalization ability; and the first ReLU layer, as an activation function, performs nonlinear mapping on the output data of the BN layer. The first ordinary feature extracted by the first convolution unit is taken as input data of the first attention unit, the first attention feature output by the first attention unit is fused with the first ordinary feature extracted by the first convolution unit to obtain first fusion features.

[0046] The first feature fusion module can obtain rich semantic feature information and position information, can retain more detailed features, and can reduce the phenomenon that the influence of shallow features is weakened due to the increase in network depth; ensure that shallow information (such as edges, textures and other details) is not excessively weakened at an early stage, highlight the significant areas of noise, so that the shallow details can be retained and strengthened, and local information is avoided to be lost.

[0047] The VGGB module is configured to extract feature information from the first fusion features layer by layer from shallow to deep to obtain enhanced fusion features.

[0048] The VGGB module comprises a plurality of fifth convolution layers, a pooling layer and a third ReLU layer connected in sequence; the number of the fifth convolution layers can be 3, the fifth convolution layers can adopt 3x3 small convolution kernels, each fifth convolution layer extracts higher-level semantic features step by step, and retains certain edge, texture and structure information; the third ReLU layer is used to accelerate convergence, suppress gradient disappearance, enhance nonlinear expression ability, strengthen the stability of shallow features and deep features, and ensure that robust features can still be extracted in the presence of noise; and the pooling layer is used to reduce the dimension, realize spatial down-sampling, gradually compress the size of the feature map, increase the receptive field, and improve the ability of the model to capture context information.

[0049] The second feature fusion module FFB2 is configured to extract second ordinary features and second attention features from the enhanced fusion features, and fuse the extracted second ordinary features and second attention features to obtain second fusion features.

[0050] The second feature fusion module comprises a second convolution unit and a second attention unit. The second convolution unit comprises a second convolution layer Conv and a second ReLU layer, wherein the second convolution layer Conv performs local feature extraction through sliding filters (i.e., convolution kernels), and the ReLU layer is an activation function that performs nonlinear mapping on the output data of the second convolution layer; the second ordinary features extracted by the second convolution unit are taken as input data of the second attention unit, and the second convolution unit and the second attention unit are spliced and then subjected to a convolution algorithm to fuse the second attention features output by the second attention unit and the second ordinary features extracted by the second convolution unit to obtain second fusion features.

[0051] The convolution algorithm can make the fusion size of the second attention features and the second ordinary features consistent, avoid the mismatching problem caused by the simple addition of the second attention features and the second ordinary features, and make the fusion effect better. The convolution algorithm is a special integral operation in mathematical analysis, which is realized through the integral of the product of the flipped, shifted and overlapped parts of two signals.

[0052] The reinforced fusion features are high semantic features, and the high semantic features usually focus on global information and ignore detailed information. The deep features of the reinforced fusion features and the attention features are fused again through the second feature fusion module, which can make up for the insufficient sensitivity of the deep network to details, ensure the consideration of both the detailed information and the global information, realize the integrity of feature expression at a higher semantic level, and consider both the global context and the local details, thereby significantly improving the detail recovery capability and the overall visual quality of image denoising.

[0053] The dilated convolution module DCBs is used for optimizing the spatio-temporal prediction of the second fusion features to obtain optimized receptive field features, realizes the expansion of the receptive field without increasing the convolution layer, and maps more context information to improve the denoising capability of the model.

[0054] In the field of image denoising, the pixel points in the noise neighborhood are very important for denoising, and the traditional convolution denoising network usually uses a stacked convolution manner to increase the receptive field and increase the connection between the context information between pixels to improve the denoising effect of the image. However, the increase in the number of convolution layers will cause problems such as gradient disappearance or gradient explosion, resulting in poor denoising effect. Therefore, the dilated convolution module is adopted to expand the receptive field without increasing the number of convolution layers to increase the connection of the image context information and improve the image denoising effect.

[0055] The expansion convolution module comprises three DCB units, each of which comprises a third convolution layer, a fourth convolution layer and an expansion convolution layer connected in sequence, wherein the DCB is a core component for strengthening the spatio-temporal prediction learning in the spatio-temporal prediction MoDeRNN model, which improves the prediction quality and long-term correlation by extracting and updating the detailed features of the input and context state, and improves the prediction accuracy; the expansion convolution layer refers to increasing the receptive field by inserting a hole in the convolution kernel, i.e. increasing the interval between the pixel points in the convolution kernel, so that each convolution kernel point processes a larger range of input data, thereby expanding the receptive field without increasing the convolution layer, and mapping more context information.

[0056] An attention mechanism module AMB is used to obtain an attention feature map by refining local image information in the optimized receptive field feature through an attention mechanism.

[0057] Since image denoising mainly focuses on part of the input noise image, and there are also some complex background areas that distract people's attention, directly using convolution features to predict saliency may lead to suboptimal results, because non-salient regions will distract attention, therefore, the application adopts an attention mechanism to assist convolutional neural network training, and the attention mechanism can pay more attention to detailed areas, which helps to strengthen detailed features for image denoising detail retention.

[0058] The attention mechanism is a spatial attention mechanism, comprising:

[0059] The maximum pooling is obtained by selecting the maximum value of each sub-region as the output, and the average pooling is obtained by calculating the average value of each sub-region as the output; based on the maximum pooling and the average pooling, the channel information of the optimized receptive field feature is mapped to obtain maximum pooling features and average pooling features, the maximum pooling features and the average pooling features are spliced and fused to generate fused pooling features; the fused pooling features are sequentially subjected to standard 3*3 convolution processing, batch normalization processing and rectified linear unit activation processing to obtain spatial information weights, and the spatial information weights are multiplied by the optimized receptive field feature to obtain a local spatial information expression feature map, i.e. an attention feature map.

[0060] The calculation formula of the fused pooling feature is:

[0061] ,

[0062] Among them, is the fused pooling feature, F is the optimized receptive field feature, is the maximum pooling, is the average pooling, and cat is splicing and fusion.

[0063] The calculation formula of the spatial information weight is:

[0064] ,

[0065] wherein, is a spatial information weight, is a fusion pooling feature, is a standard 3*3 convolution processing, is a batch normalization processing, is a rectified linear unit activation processing.

[0066] The calculation formula of the attention feature map is:

[0067] ,

[0068] wherein, is a spatial information weight, is a fusion pooling feature.

[0069] Pooling refers to an operation in deep learning for reducing the dimension of a feature map, reducing the amount of calculation and the number of parameters.

[0070] Through the spatial attention mechanism, spatial features are paid attention to and the information of the effective position part is improved, and the spatial feature expression is strengthened, so that the noise distribution sensitive area can be highlighted, that is, through the spatial attention mechanism, key spatial information features can be obtained, noise position spatial information features can be enhanced, noise points can be removed in a targeted manner, local expression of the image can be improved, and image smoothing can be reduced.

[0071] The residual module RB is used for fitting the attention feature map to obtain a noise component map, and a clean image is obtained by using a noise image to subtract the noise component map through a jump connection, so that the image denoising process is completed.

[0072] More image information is retained through the first feature fusion module and the second feature fusion module, image information is strengthened through the VGGB module, the receptive field is expanded through the dilated convolution module, more context information is mapped, image local information is refined through the attention mechanism module, and a clean image is obtained through the residual module, so that higher peak signal-to-noise ratio is pursued while rich image local information is retained, the image denoising effect is improved, the image quality is improved, and the image visual effect is improved.

[0073] The application also relates to a deep learning denoising method based on feature fusion and an attention mechanism, as shown in Figure 2 , comprising:

[0074] S1, obtaining a training set and a test set, and preprocessing the training set;

[0075] A large number of original images are acquired, the original images include noise-containing samples and noise-free samples, the original images contain multiple styles, such as people, landscapes, animals, buildings and the like, and the problem that the denoising model training result is poor due to single image type is avoided; the original images are divided into a training set and a test set, and the ratio of the training set to the test set can be 8:2.

[0076] The training set is preprocessed, and the preprocessing includes:

[0077] S11, standardization processing. The original images in the training set are uniformly cropped to obtain original images of the same size, and the size of the cropped original images can be 40*40; the original images of the same size are flipped to obtain standardized images, and the flipping includes up-down flipping, 90° rotation, 180° rotation, 270° rotation, 90° rotation followed by up-down flipping, 180° rotation followed by up-down flipping, and 270° rotation followed by up-down flipping.

[0078] The size of the original image is generally large, which can be 180*180, if the large-size original image is directly used for training of the denoising model, the image detail features will be lost, the denoising result will be poor, and the like, therefore, the size of the original image in the training set is reduced by cropping, and the small-size image can facilitate extraction of the detail features and improve the training efficiency of the denoising model.

[0079] S12, normalization processing. The amplitude range of the standardized image is compressed to a preset interval, and the preset interval can be [0, 1] interval or [-1, 1] interval.

[0080] The test set includes original images of different sizes and original images of different styles, which can be 256*256 gray images and 512*512 gray images. The test of image denoising through images of different sizes and different styles can verify the applicability and generalizability of the denoising model.

[0081] S2, the denoising model is trained according to the preprocessed training set to obtain a trained denoising model;

[0082] The standardized images in the preprocessed training set are sequentially input into the denoising model, the denoising model is trained through the learning ability of the denoising model, and the trained denoising model is obtained.

[0083] Since the fixed learning rate of the denoising model is not conducive to gradient descent in the later training, the learning rate gradually decreases with the increase of the training number, which is more suitable for the training process, therefore, the learning rate of the denoising model of the present application adopts a floating learning rate, that is, the training number is obtained based on the preprocessed training set, the training is divided into different stage training based on the training number, and different learning rates are used in different stage training.

[0084] For example, if the training times are 70, the training is divided into first stage training, second stage training and third stage training, the first value learning rate is used in the first stage training, the first stage training can be the first training to the 30th training, the first value learning rate can be 0.001; the second value learning rate is used in the second stage training, the second stage training can be the 31st training to the 50th training, the second value learning rate can be 0.0001; the third value learning rate is used in the third stage training, the third stage training can be the 51st training to the 70th training, and the third value learning rate can be 0.00001.

[0085] The gradient descent optimizer used in the training of the denoising model is an Adam optimizer.

[0086] S3, based on the test set, the trained denoising model is tested and verified to obtain a denoising model that passes the verification.

[0087] The test verification includes test comparison of each module in the denoising model and test comparison of the denoising model and the existing model.

[0088] In order to show the role of each module in the denoising model of the application, an ablation experiment is used to compare the denoising results, which include peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). The peak signal-to-noise ratio (PSNR) is an important indicator for measuring image quality, which is used to evaluate the denoising result. Its core is to calculate the mean square error (MSE) between the original image and the distorted image, and then to quantify the evaluation through the ratio of the maximum power of the signal to the power of the noise. The higher the peak signal-to-noise ratio, the better the denoising result. The structural similarity index (SSIM) is an evaluation index for measuring the structural similarity between the processed image and the original image. Its value range is 0-1, and the larger the value, the higher the structural similarity, i.e. the better the denoising effect.

[0089] The test comparison of each module in the denoising model includes:

[0090] The test set is input into the trained denoising model, the denoising model without the first feature fusion module and the second feature fusion module, the denoising model without the attention mechanism module, the denoising model without the dilated convolution module and the denoising model without the residual module, and the first denoising result, the second denoising result, the third denoising result, the fourth denoising result and the fifth denoising result are output respectively, as shown in Table 1.

[0091] Table 1 Denoising results of ablation experiment

[0092]

[0093] The first denoising result is compared with the second denoising result, the third denoising result, the fourth denoising result and the fifth denoising result respectively, and it can be known from Table 1 that the peak signal-to-noise ratio (PSNR) of the second denoising result decreases by 0.09 dB, and the structural similarity index (SSIM) decreases by 0.0202, thereby proving that the first feature fusion module and the second feature fusion module can retain more image details and improve the image denoising effect and image details; the peak signal-to-noise ratio (PSNR) of the third denoising result decreases by 0.04 dB, and the structural similarity index (SSIM) decreases by 0.0272, thereby proving that the attention mechanism module can improve the denoising effect and enhance the image details; the peak signal-to-noise ratio (PSNR) of the fourth denoising result decreases by 0.05 dB, and the structural similarity index (SSIM) decreases by 0.0273, thereby proving that the dilated convolution module can increase the receptive field, improve the denoising effect and increase the detail information; the peak signal-to-noise ratio (PSNR) of the fifth denoising result decreases by 0.18 dB, and the structural similarity index (SSIM) decreases by 0.0293, thereby proving that the image denoising can obtain a better denoising effect by fitting the noise characteristics through the residual module.

[0094] The test comparison of the denoising model and the existing model includes:

[0095] The test set with different noise levels is respectively input into the trained denoising model, BM3D, EPLL, WNNM, MLP and TNRD, and the peak signal-to-noise ratio (PSNR) under different noise levels is obtained, as shown in Table 2.

[0096] Table 2 Peak signal-to-noise ratio of each model under different noise levels

[0097]

[0098] It can be known from Table 2 that the peak signal-to-noise ratio (PSNR) of the existing models BM3D, EPLL, WNNM, MLP and TNRD under different noise levels is lower than the peak signal-to-noise ratio (PSNR) of the denoising model under the corresponding noise level, thereby proving that the denoising effect of the denoising model of the present application is better, and the retained image details are more significant.

[0099] S4, input the obtained noise image into the verified denoising model to obtain a clean image.

[0100] The obtained noise image is input into the verified denoising model, and the clean image is obtained by sequentially passing through the first feature fusion module, the VGGB module, the second feature fusion module, the dilated convolution module, the attention mechanism module and the residual module in the denoising model, and the calculation formula of the clean image is:

[0101]

[0102] wherein, is the clean image,​ is a noise image, is a first feature fusion processing, is a strengthened fusion feature processing, is a second feature fusion processing, is an expanded convolution processing, is an attention mechanism processing, is a residual processing, is a convolution processing.

[0103] The above only is the preferred embodiment of the present application, and is not used to limit the present application, for the person skilled in the art, the present application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A deep learning denoising model based on feature fusion and attention mechanism, characterized in that, The method comprises the following steps: The first feature fusion module is used for extracting first common features and first attention features from the input noise image, and fusing the extracted first common features and first attention features to obtain first fusion features; The VGGB module is used for extracting feature information from the first fusion features layer by layer to obtain reinforced fusion features; The second feature fusion module is used for extracting second common features and second attention features from the reinforced fusion features, and fusing the extracted second common features and second attention features to obtain second fusion features; The dilated convolution module is used for optimizing the spatio-temporal prediction of the second fusion features to obtain optimized receptive field features; The attention mechanism module is used for extracting image local information in the optimized receptive field features through the attention mechanism to obtain an attention feature map; The residual module is used for fitting the attention feature map to obtain a noise component map, and subtracting the noise component map from the noise image through the jump connection to obtain a clean image. 2.The deep learning denoising model based on feature fusion and attention mechanism of claim 1, characterized in that, The first feature fusion module comprises: The first convolution unit extracts first common features as input data of the first attention unit, and the first fusion features are obtained by fusing the first attention features output by the first attention unit and the first common features extracted by the first convolution unit. 3.The deep learning denoising model based on feature fusion and attention mechanism of claim 2, characterized in that, The first convolution unit comprises a first convolution layer, a BN layer and a first ReLU layer connected in sequence. 4.The deep learning denoising model based on feature fusion and attention mechanism of claim 1, characterized in that, The VGGB module comprises a plurality of fifth convolution layers, a pooling layer and a third ReLU layer connected in sequence. 5.The deep learning denoising model based on feature fusion and attention mechanism of claim 1, characterized in that, The second feature fusion module comprises: The second convolution unit extracts second common features as input data of the second attention unit, and the second fusion features are obtained by fusing the second attention features output by the second attention unit and the second common features extracted by the second convolution unit through convolution algorithm after the second convolution unit and the second attention unit are spliced; The second convolution unit comprises a second convolution layer and a second ReLU layer connected in sequence. 6.The deep learning denoising model based on feature fusion and attention mechanism of claim 1, characterized in that, The dilated convolution module comprises three DCB units, and each DCB unit comprises a third convolution layer, a fourth convolution layer and a dilated convolution layer.

7. The deep learning denoising model based on feature fusion and attention mechanism according to claim 1, characterized in that, The attention mechanism is a spatial attention mechanism, and the spatial attention mechanism comprises: The maximum pooling is obtained by selecting the maximum value of each sub-region, and the average pooling is obtained by calculating the average value of each sub-region; Based on the maximum pooling and the average pooling, the channel information of the optimized receptive field features is mapped to obtain maximum pooling features and average pooling features, and the maximum pooling features and the average pooling features are spliced and fused to generate fused pooling features; The spatial information weight is obtained by sequentially performing standard 3*3 convolution processing, batch normalization processing and rectified linear unit activation processing on the fused pooling features, and the attention feature map is obtained by multiplying the spatial information weight by the optimized receptive field features.

8. A deep learning denoising method based on feature fusion and attention mechanism, characterized in that, The deep learning denoising model based on feature fusion and attention mechanism comprises the following steps: S1, obtaining a training set and a test set, and preprocessing the training set; S2, training the denoising model according to the preprocessed training set to obtain a trained denoising model; S3, test and verify the trained denoising model based on the test set to obtain a passed denoising model; S4, input the obtained noise image into the passed denoising model to obtain a clean image. 9.The deep learning denoising method based on feature fusion and attention mechanism according to claim 8, characterized in that, The preprocessing comprises: S11, standardization processing, uniformly cropping the original images in the training set to obtain original images of the same size, and performing flip processing on the original images of the same size to obtain standardized images; S12, normalization processing, compressing the amplitude range of the standardized images to a preset interval.

10. The deep learning denoising method based on feature fusion and attention mechanism according to claim 8, characterized in that S2 Comprise: According to the training set after preprocessing, obtain the training times, based on the training times, divide the training into different stage training, in different stage training, adopt different learning rate.

Images