Image denoising method based on negative image and double-branch attention residual network

Abstract

The invention discloses an image denoising method based on a negative image and a double-branch attention residual network, which comprises the following steps of: firstly, constructing a data set through paired clean and noise images; then, an image denoising model is constructed, the model comprises two branch networks, the original noise image and the negative image of the original noise image serve as input, the upper branch network comprises five attention residual error multi-scale blocks, and the lower branch network comprises five attention residual error expansion blocks; then training the model; the negative image which is not deeply researched at present is introduced to serve as input of the lower branch network, the negative image can reverse pixel values and highlight features which are difficult to extract in an original noise image, more image details are reserved by the model while the complexity of the model is not increased, and the method has the advantages of being high in accuracy and high in robustness. Modules in the upper and lower branch networks can perform effective complementary feature extraction on two inputs and perform screening through an attention mechanism, so that the method becomes an effective choice of an actual image denoising task.

Description

Technical Field

[0001] The present invention relates to the technical field of image denoising, and more particularly to an image denoising method based on negative images and a dual-branch attention residual network. Background Art

[0002] Image denoising is a very classic and popular low-level visual task, preparing for tasks such as object detection and image segmentation. The goal of image denoising is to reduce the noise introduced by internal or external sources during image acquisition, transmission and reception, and then obtain a clean image. In the past, some traditional denoising methods such as NLM smoothed the area by averaging similar image blocks, and BM3D used non-local self-similarity and sparse representation to improve denoising performance. Although these methods have good flexibility and performance, they are difficult to optimize and therefore difficult to be widely used in practice. In recent years, convolutional neural networks have been widely used due to their high performance and high efficiency. The denoising method using convolutional neural networks uses paired clean images and noisy images for end-to-end training and learning, and finally uses the trained network model to denoise and reconstruct the image, all of which have achieved good results. However, there are still some problems, such as the performance degradation and increased computational cost caused by deeper networks, and the existing methods do not pay attention to the angle of the negative image, which results in many image details affected by noise cannot be effectively restored, and many models are limited to extracting local features while ignoring global features. Summary of the invention

[0003] In view of the shortcomings of the prior art, the present invention provides an image denoising method based on negative image and dual-branch attention residual network. The method introduces the negative image of the original noisy image as the input of the lower branch network, and respectively extracts effective complementary features of the original noisy image and the negative image through the attention residual multi-scale block of the upper branch network and the attention residual expansion block of the lower branch network, and enriches the reuse of features through multiple long jump connections, so as to effectively remove the additive Gaussian white noise in the image and improve the visual effect of the image.

[0004] The present invention is achieved through the following technical solutions:

[0005] An image denoising method based on negative image and dual-branch attention residual network, comprising the following steps:

[0006] Step 1: Preprocess the public dataset, including image cropping, image enhancement, and noise addition. The preprocessing operation is used to obtain a noise image that corresponds to the clean image one by one, and then construct a training dataset for the image denoising model.

[0007] Step 2: Construct an image denoising model. The model adopts a dual-branch convolutional neural network structure, taking the original noisy image and its negative image as input respectively. The upper branch network includes five attention residual multi-scale blocks and up- and down-sampling operations, and the lower branch network includes five attention residual expansion blocks.

[0008] Step 3: Train the model based on the constructed training data set.

[0009] Step 4: Use the test data set to test the trained model and calculate the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of the denoised image.

[0010] Furthermore, the public dataset is selected as Flick2K, and 2650 high-resolution images are used for model training. After carefully calculating the receptive field size of the model, each image is randomly cropped into 128×128 image blocks and converted into grayscale images. At the same time, each image is flipped and rotated at angles of 90 degrees, 180 degrees, and 270 degrees to enhance the generalization ability of the model. Finally, additive Gaussian white noise in the noise range of [0, 50] is randomly added to the clean image to obtain pairs of clean images and noisy images to form a training dataset.

[0011] Furthermore, the model is designed as a dual-branch network, with a widened network replacing a deepened network, taking the original noisy image and its negative image as input respectively. The upper branch network consists of 5 attention residual multi-scale blocks and up- and down-sampling operations to extract rich local features from the original noisy image. The lower branch network consists of 5 attention residual expansion blocks to extract rich global complementary features from the negative image. Both blocks contain dual attention blocks to screen features in a targeted manner. The network contains multiple long jump connections and residual connections for feature reuse.

[0012] For the input original noise image, the negative image is first obtained by subtracting the original noise image pixel value from 255, and it is used as the input of the next branch network. Before each input enters the first attention residual multi-scale block or attention residual expansion block, a convolution layer with a convolution kernel size of 3×3 and an output channel number of 128 is used to perform shallow feature extraction on the input image, mapping the input from a simple low-dimensional space to a more abstract high-dimensional space. After that, the shallow feature map is input into the first attention residual multi-scale block or attention residual expansion block to start further feature extraction, and the number of channels of the two modules is kept at 128. In this process, the attention residual multi-scale block or attention residual expansion block extracts the complementary features of the original noise image and the negative image respectively, enhancing the network model's performance in restoring image details. Afterwards, after 5 modules of feature extraction, the network model is trained. The output of feature extraction is input into a convolution layer with a convolution kernel size of 3×3 and an output channel number of 1 for further feature extraction. Subsequently, the features extracted by the dual-branch network are concatenated. After that, the obtained deep feature map is passed through a convolution layer with a convolution kernel size of 3×3 and an output channel number of 1 to generate a residual image, which is residually connected with the input image to obtain the final output denoised image. In addition, in order to make full use of the extracted features, the features extracted by the first 3×3 convolution layer and the last 3×3 convolution layer of the upper branch network, the first attention residual multi-scale block and the fifth attention residual multi-scale block, the second attention residual multi-scale block and the fourth attention residual multi-scale block are added through jump connections to further improve network performance and training efficiency. The same is true for the lower branch network.

[0013] Furthermore, the negative image is obtained by subtracting the original noise pixel value from 255. The negative image can invert the pixel value of the original noise image, making bright pixels darker and dark pixels brighter. Therefore, the negative image provides an input with a completely opposite brightness distribution. The low-brightness features in the grayscale image may be masked by high-brightness noise. The negative image can reverse this relationship and highlight the hidden features. The negative image is a complement to the original noise image. The features that cannot be extracted from the original noise image due to the influence of noise are easier to extract from the negative image, which can effectively improve the performance of the model in retaining image details.

[0014] Furthermore, the attention residual multi-scale block consists of a multi-scale pixel attention block, a residual convolution block, and a dual attention block. The multi-scale pixel attention block consists of three convolution kernels with sizes of 1×1, 3×3, and 5×5, respectively. It extracts feature information of different scales in parallel, then multiplies the results pixel by pixel, and then adds them to the module input pixel by pixel and outputs them to the subsequent residual convolution block. There are three groups of residual convolution blocks, each of which consists of a 3×3 convolution kernel, a ReLU activation function, and a 3×3 convolution kernel. Multiple outputs are added to other inputs through residual connections to enhance feature reuse, so that the network model can extract rich local features. The dual attention block consists of a channel attention mechanism and a spatial attention mechanism. Spatial information is integrated on top of the channel attention mechanism, and maximum pooling is introduced to further refine the features as the pre-input of the spatial attention part. Finally, the initial input of the module and the output of the dual attention block are added pixel by pixel through residual connections to obtain the final output.

[0015] Furthermore, the attention residual expansion block is composed of a residual expansion block and a dual attention block. There are 4 groups of residual expansion blocks, each of which consists of an expansion convolution, a ReLU activation function, and an expansion convolution. In the combination of the 4 groups of expansion convolutions, the expansion coefficients of the 8 expansion convolutions are 1, 2, 3, 4, 3, 2, 1, and 1, respectively, to extract global information with a larger receptive field. At the same time, the combination of the 4 groups of expansion convolutions uses multiple residual connections to reuse features. Then, the extracted information is input into the dual attention block, and the output is obtained through channel and spatial level attention. Finally, the initial input of the module and the output of the dual attention block are added pixel by pixel through the residual connection to obtain the final output.

[0016] Furthermore, during the training process, the preprocessed noisy image is passed into the model. After being processed into a negative image, it passes through 5 attention residual multi-scale blocks and attention residual expansion blocks in sequence, and the Concat feature splicing operation is performed at the end of the dual-branch network. Finally, a 3×3 convolution kernel is used for the final processing to obtain a denoised image. The loss is calculated based on the output denoised image and clean image, and then the loss is back-propagated to update the neural network weight parameters.

[0017] Furthermore, when training the image denoising model, the mean square error loss function was used, and the Adam optimizer was used. The batch size was set to 8, the initial value of the learning rate was configured to 1e-4, and it was halved every 100,000 iterations. Finally, the model parameters were saved after 785,000 iterations of model training.

[0018] Compared with the prior art, the present invention has the following advantages:

[0019] 1. This paper proposes a dual-branch network to replace the deep network, which can save computational costs and obtain more stable and effective denoising performance.

[0020] 2. In the blind denoising network, the present invention introduces negative images to change the distribution of light and dark contrast pixels, making certain details and features more prominent. This enhances network performance without increasing any model complexity.

[0021] 3. The attention residual multi-scale block and attention residual expansion block proposed in this paper can fully extract and filter the complementary feature information of the original noisy image and the negative image by combining the dual attention mechanism, which greatly improves the performance of image denoising. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solution in the embodiment of the present invention, the following will briefly introduce the drawings required for use in the description of the embodiment.

[0023] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0024] Figure 2 is a diagram of the image denoising model architecture of an embodiment of the present invention;

[0025] Figure 3 is a structural diagram of the attention residual multi-scale block of an embodiment of the present invention;

[0026] Figure 4 is a structural diagram of a dual attention block according to an embodiment of the present invention;

[0027] Figure 5 is a structural diagram of an attention residual expansion block in an embodiment of the present invention;

[0028] Figure 6 This is a comparison diagram of the visual effects of the images reconstructed by the embodiment of the present invention and other mainstream models when the noise level is 15;

[0029] Figure 7 This is a comparison of the visual effects of the image reconstructed by the embodiment of the present invention with other mainstream models when the noise level is 50. DETAILED DESCRIPTION

[0030] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only provided for the convenience of explanation. The embodiments described below are only part of the present invention and should not be construed as being limited to all examples of the above-mentioned methods.

[0031] like Figure 1-7 As shown in Figure 1, an image denoising method based on negative image and dual-branch attention residual network has the following specific steps:

[0032] Step 1: First, prepare the training data set. Preprocess the selected public data set Flick2K, and use 2650 high-resolution images for model training. After carefully calculating the receptive field size of the model, randomly crop each image into 128×128 image blocks and convert them into grayscale images. After that, perform data augmentation operations such as flipping and rotating each image by 90 degrees, 180 degrees, and 270 degrees to enhance the generalization ability of the model. Finally, add additive Gaussian white noise in the noise range of [0,50] to the clean image to obtain pairs of clean images and noisy images to form the training data set of the blind denoising model.

[0033] Step 2: Construct an image denoising model based on negative images and a two-branch attention residual network, such as Figure 2 As shown in the figure, the model is designed as a dual-branch network, with a widened network replacing a deepened network, taking the original noisy image and its negative image as input respectively. The upper branch network consists of 5 attention residual multi-scale blocks and up- and down-sampling operations to extract rich local features from the original noisy image. The lower branch network consists of 5 attention residual expansion blocks to extract rich global complementary features from the negative image. Both blocks contain dual attention blocks to screen features in a targeted manner. The network contains multiple long jump connections and residual connections for feature reuse.

[0034] For the input original noise image, the negative image is first obtained by subtracting the original noise image pixel value from 255, and it is used as the input of the next branch network. Before each input enters the first attention residual multi-scale block or attention residual expansion block, a convolution layer with a convolution kernel size of 3×3 and an output channel number of 128 is used to perform shallow feature extraction on the input image, mapping the input from a simple low-dimensional space to a more abstract high-dimensional space. After that, the shallow feature map is input into the first attention residual multi-scale block or attention residual expansion block to start further feature extraction, and the number of channels of the two modules is kept at 128. In this process, the attention residual multi-scale block or attention residual expansion block extracts the complementary features of the original noise image and the negative image respectively, enhancing the network model's performance in restoring image details. Afterwards, after 5 modules of feature extraction, the network model is trained. The output of feature extraction is input into a convolution layer with a convolution kernel size of 3×3 and an output channel number of 1 for further feature extraction. Subsequently, the features extracted by the dual-branch network are concatenated. After that, the obtained deep feature map is passed through a convolution layer with a convolution kernel size of 3×3 and an output channel number of 1 to generate a residual image, which is residually connected with the input image to obtain the final output denoised image. In addition, in order to make full use of the extracted features, the features extracted by the first 3×3 convolution layer and the last 3×3 convolution layer of the upper branch network, the first attention residual multi-scale block and the fifth attention residual multi-scale block, the second attention residual multi-scale block and the fourth attention residual multi-scale block are added through jump connections to further improve network performance and training efficiency. The same is true for the lower branch network.

[0035] At present, research on image denoising pays little attention to the perspective of negative images, which also leads to the fact that in many grayscale images, due to the addition of additive Gaussian white noise, the low-brightness detail features in the grayscale image are covered by noise and are treated as noise by the model, resulting in the loss of image details. In order to solve this problem, negative images are introduced into the model as input. The negative image is obtained by subtracting the original noise pixel value from 255. The negative image can invert the pixel value of the original noise image, making the bright pixels darker and the dark pixels brighter. Therefore, the negative image provides an input with a completely opposite brightness distribution. The low-brightness features in the grayscale image may be masked by the high-brightness noise. The negative image can reverse this relationship and highlight the hidden features. The negative image is a supplement to the original noise image. The features that cannot be extracted from the original noise image due to the influence of noise are easier to extract from the negative image, which can effectively improve the performance of the model in retaining image details.

[0036] Most of the current models are limited to local features and ignore global features. To solve this problem, an attention residual multi-scale block applied to the upper branch network and an attention residual dilation block applied to the lower branch network are proposed to effectively extract the complementary local and global features of the original noisy image and the negative image, respectively, to enhance the model performance and improve the model's ability to extract and retain details.

[0037] like Figure 3 As shown in , the attention residual multi-scale block consists of a multi-scale pixel attention block, a residual convolution block, and a dual attention block. The multi-scale pixel attention block consists of three convolution kernels with sizes of 1×1, 3×3, and 5×5, respectively. It extracts feature information of different scales in parallel, then multiplies the results pixel by pixel, and then adds them to the module input pixel by pixel and outputs them to the subsequent residual convolution block. There are three groups of residual convolution blocks, each of which consists of a 3×3 convolution kernel, a ReLU activation function, and a 3×3 convolution kernel. And through residual connections, multiple outputs are added to other inputs to enhance feature reuse, so that the network model can extract rich local features. The dual attention block consists of a channel attention mechanism and a spatial attention mechanism, such as Figure 4 As shown in the figure, spatial information is integrated on top of the channel attention mechanism, and maximum pooling is introduced to further refine the features as the pre-input of the spatial attention part. Finally, the initial input of the module and the output of the dual attention block are added pixel by pixel through the residual connection to obtain the final output.

[0038] like Figure 5 As shown in the figure, the attention residual expansion block consists of a residual expansion block and a dual attention block; there are 4 groups of residual expansion blocks, each of which consists of an expansion convolution, a ReLU activation function, and an expansion convolution. In the 4 groups of expansion convolution combinations, the expansion coefficients of the 8 expansion convolutions are 1, 2, 3, 4, 3, 2, 1, and 1, respectively, to extract global information with a larger receptive field. At the same time, the combination of the 4 groups of expansion convolutions uses multiple residual connections to reuse features. Then, the extracted information is input into the dual attention block, and the output is obtained through channel and spatial level attention. Finally, the initial input of the module and the output of the dual attention block are added pixel by pixel through the residual connection to obtain the final output.

[0039] Step 3: Use the constructed training data set to train the image denoising model. In the embodiment of the present invention, the mean square error loss function is used, and the Adam optimizer is used. The batch size is set to 8, the initial value of the learning rate is configured to 1e-4, and it is halved every 100,000 iterations. The model parameters are saved after 785,000 iterations. The experimental environment is based on the Pytorch framework, runs on Ubuntu 20.04, and uses Nvidia GeForce 4090 for training. During the training process, after the noisy image is processed into a negative image, it passes through 5 attention residual multi-scale blocks and attention residual expansion blocks in sequence together with the original noisy image, and is merged at the end of the dual-branch network, and finally processed by a convolution kernel of size 3 to obtain a denoised image. The loss is calculated based on the noise image, and then the neural network weight parameters are updated using the loss back propagation.

[0040] Step 4: Calculate the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) of the reconstructed image using the trained model and current mainstream image denoising models and compare them to verify the performance of the model. These include traditional methods and deep learning methods, specifically BM3D, DnCNN, FFDNet, ADNet, DudeNet, RDDCNN, and MWDCNN. PSNR is a measure of the difference between the reconstructed image and the clean image. The unit is decibel (dB). The higher the PSNR value, the closer the reconstructed image is to the clean image, and the better the denoising effect. SSIM pays more attention to the texture features of the image. The closer the SSIM value is to 1, the better the texture features of the reconstructed image are restored. The test data sets select BSD68 and Set12 data sets. The BSD68 data set is derived from 68 images of buildings, plants, people, etc. in the BSD data set. The Set12 data set consists of 7 256×256 and 5 512×512 grayscale images. These two data sets are widely used in evaluating image denoising effects. Table 1 shows the comparison of PSNR and SSIM results on the BSD68 data set for different models at different noise levels. You can see It can be seen that the model proposed in the present invention has achieved excellent performance in both PSNR and SSIM indicators. When the noise level is 15, the model of the present invention is significantly better than BM3D by 0.71dB. When the noise level is 50, the SSSIM of the model of the present invention is improved by 0.023 compared with RDDCNN. This shows the superiority of the model proposed in the present invention. Table 2 shows the comparison of single images and average PSNR and SSIM results on the Set12 dataset of different models when the noise level is 50. It can be seen that the model of the present invention not only achieves the best average PSNR and SSIM values, but also has obvious advantages in the measurement of single images, which further confirms the excellent performance of the model proposed in the present invention.

[0041] Table 1: Comparison of average PSNR and SSIM results of BSD68 dataset under different noise levels for different models

[0042]

[0043] Table 2: Comparison of average and single image PSNR and SSIM results of different models on the Set12 dataset at noise level 50

[0044]

[0045] In order to verify the effectiveness of the model proposed in this invention, we conducted ablation experiments on the Set12 dataset with noise levels of 15, 20, 25, and 30 respectively using four different models. Table 3 shows the average PSNR and SSIM of these models. The results show that the model proposed in this invention outperforms the other three models. Therefore, the ablation experiment shows the effectiveness of the proposed model.

[0046] Table 3: Comparison of average PSNR and SSIM results of four models in ablation experiments

[0047]

[0048] Step 5: Compare the visual effects of the reconstructed images of each model to verify the model performance from a subjective level. Figure 6 As shown in Figure 1, the "Lena" image in the Setl2 dataset is selected and the experiment is carried out under the condition of noise level 15. It can be observed that the model proposed in the present invention can restore more detailed texture information that other models cannot restore. In addition, the PSNR index of the proposed model shows significant advantages. Figure 7 As shown in the figure, the "test003" image in the BSD68 dataset is selected, and the experiment is carried out under the condition of noise level 50, and the image is focused on the roof of the image. The reason for choosing to focus here is that the image details in this part are relatively complex and rich, which is more suitable for comparing the visual effects of the reconstructed images of various models. It can be observed that BM3D is the most blurred, while DnCNN and FFDNet have achieved good denoising effects, but there are many artifacts. Although ADNet and RDDCNN have a better overall appearance, a lot of details and edge information are lost. In contrast, the model proposed in the present invention has achieved the best results in terms of overall visual quality and retention of details and edges. In addition, the proposed model also achieves the highest PSNR index. In short, excellent results have been achieved both subjectively and objectively.

[0049] The above contents are further detailed descriptions of the present invention in combination with specific preferred implementation modes, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For technicians in the technical field to which the present invention belongs, several equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and the performance or use is the same, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. An image denoising method based on negative image and dual-branch attention residual network, characterized in that: The following steps are involved: Step 1: Preprocess the public dataset, where the preprocessing operations include image cropping, image enhancement, and noise addition; Through preprocessing operations, a noisy image that corresponds to a clean image is obtained, and then a training data set for the image denoising model is constructed; Step 2: Construct an image denoising model. The model adopts a dual-branch convolutional neural network structure, which takes the original noisy image and its negative image as input respectively. The upper branch network includes five attention residual multi-scale blocks and up- and down-sampling operations, and the lower branch network includes five attention residual expansion blocks. Step 3: Train the model based on the constructed training data set; Step 4: Use the test data set to test the trained model and calculate the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of the denoised image.

2. The image denoising method based on negative image and dual-branch attention residual network according to claim 1, characterized in that: The public dataset is selected as Flick2K, and 2650 high-resolution images are used for model training; after carefully calculating the receptive field size of the model, each image is randomly cropped into a 128×128 image block and converted into a grayscale image; at the same time, each image is flipped and rotated at angles of 90 degrees, 180 degrees, and 270 degrees to enhance the generalization ability of the model; finally, additive Gaussian white noise in the noise range of [0, 50] is randomly added to the clean image to obtain pairs of clean images and noisy images to form a training dataset.

3. The image denoising method based on negative image and dual-branch attention residual network according to claim 1, characterized in that: The model is designed as a dual-branch network, with a widened network replacing a deepened network, taking the original noisy image and its negative image as input respectively; the upper branch network consists of 5 attention residual multi-scale blocks and up- and down-sampling operations to extract rich local features from the original noisy image; the lower branch network consists of 5 attention residual expansion blocks to extract rich global complementary features from the negative image; both blocks contain dual attention blocks to screen features in a targeted manner; the network contains multiple long jump connections and residual connections for feature reuse; For the input original noise image, the pixel value of the original noise image is first subtracted from 255 to obtain a negative image, which is used as the input of the next branch network; before each input enters the first attention residual multi-scale block or attention residual expansion block, a convolution layer with a convolution kernel size of 3×3 and an output channel number of 128 is used to perform shallow feature extraction on the input image, mapping the input from a simple low-dimensional space to a more abstract high-dimensional space; after that, the shallow feature map is input into the first attention residual multi-scale block or attention residual expansion block to start further feature extraction. The number of channels of the two modules is kept at 128. In this process, the attention residual multi-scale block or attention residual expansion block extracts complementary features of the original noise image and the negative image respectively, enhancing the network model's performance in restoring image details; after that, the output of the feature extraction of the five modules is input into a convolution layer with a convolution kernel size of 3×3 and an output channel number of 1 for further feature extraction; subsequently, the features extracted by the dual-branch network are subjected to the Concat feature splicing operation; Afterwards, the obtained deep feature map is passed through a convolution layer with a convolution kernel size of 3×3 and an output channel number of 1 to generate a residual image, which is residually connected with the input image to obtain the final output denoised image; in addition, in order to make full use of the extracted features, the features extracted by the first 3×3 convolution layer and the last 3×3 convolution layer of the upper branch network, the first attention residual multi-scale block and the fifth attention residual multi-scale block, the second attention residual multi-scale block and the fourth attention residual multi-scale block of the upper branch network are added through jump connections to further improve network performance and training efficiency. The same is true for the lower branch network.

4. The image denoising method based on negative image and dual-branch attention residual network according to claim 3, characterized in that: The negative image is obtained by subtracting the original noise pixel value from 255. The negative image can invert the pixel value of the original noise image, making bright pixels darker and dark pixels brighter. Therefore, the negative image provides an input with a completely opposite brightness distribution. The low-brightness features in the grayscale image may be masked by high-brightness noise. The negative image can reverse this relationship and highlight the hidden features. The negative image is a complement to the original noisy image. Features that cannot be extracted from the original noisy image due to the influence of noise are easier to extract from the negative image, which can effectively improve the model's performance in retaining image details.

5. The image denoising method based on negative image and dual-branch attention residual network according to claim 3, characterized in that: The attention residual multi-scale block consists of a multi-scale pixel attention block, a residual convolution block, and a dual attention block; the multi-scale pixel attention block consists of three convolution kernels with sizes of 1×1, 3×3, and 5×5, respectively, which extracts feature information of different scales in parallel, then multiplies the results pixel by pixel, and then adds them to the module input pixel by pixel and outputs them to the subsequent residual convolution block; there are three groups of residual convolution blocks, each of which consists of a 3×3 convolution kernel, a ReLU activation function, and a 3×3 convolution kernel, and multiple outputs are added to other inputs through residual connections to enhance feature reuse, so that the network model can extract rich local features; the dual attention block consists of a channel attention mechanism and a spatial attention mechanism, which integrates spatial information on top of the channel attention mechanism, and introduces maximum pooling to further refine the features as the pre-input of the spatial attention part; Finally, the initial input of the module and the output of the dual attention block are added pixel by pixel through the residual connection to obtain the final output.

6. The image denoising method based on negative image and dual-branch attention residual network according to claim 3, characterized in that: The attention residual expansion block consists of a residual expansion block and a dual attention block. There are 4 groups of residual expansion blocks, each of which consists of an expansion convolution, a ReLU activation function, and an expansion convolution. In the combination of the 4 expansion convolutions, the expansion coefficients of the 8 expansion convolutions are 1, 2, 3, 4, 3, 2, 1, and 1, respectively, to extract global information with a larger receptive field. At the same time, the combination of the 4 expansion convolutions uses multiple residual connections to reuse features. Then, the extracted information is input into the dual attention block, and the output is obtained through channel and spatial level attention; Finally, the initial input of the module and the output of the dual attention block are added pixel by pixel through the residual connection to obtain the final output.

7. The image denoising method based on negative image and dual-branch attention residual network according to claim 1, characterized in that: During the training process, the preprocessed noisy image is passed into the model. After being processed into a negative image, it passes through 5 attention residual multi-scale blocks and attention residual expansion blocks in sequence, and the Concat feature splicing operation is performed at the end of the dual-branch network. Finally, it is processed through a convolution kernel of size 3×3 to obtain a denoised image; the loss is calculated based on the output denoised image and clean image, and then the loss backpropagation is used to update the neural network weight parameters.

8. The image denoising method based on negative image and dual-branch attention residual network according to claim 1, characterized in that: When training the image denoising model, the mean square error loss function was adopted, and the Adam optimizer was used. The batch size was set to 8, the initial value of the learning rate was configured to 1e-4, and it was halved every 100,000 iterations. Finally, the model parameters were saved after 785,000 iterations of model training.

Images