Image Denoising Method Based on Feature Fusion and Complementary Learning

Abstract

This invention discloses an image denoising method based on feature fusion and complementary learning, primarily addressing the problems of unstable performance and poor quality of generated denoised images in existing image denoising methods. The implementation scheme involves constructing two predictors, each including a multi-level encoder and a multi-level decoder; a feature interaction module including a conditional weight generator and a convolutional network; and a fusion module including convolutional layers. The output of each decoder stage of the two predictors is weighted and coupled through the feature interaction module and then connected to the input of the next stage decoder. The last stage decoder of the first predictor outputs a first denoised image, while the last stage decoder of the second predictor outputs noise. This noise is then subtracted from the original noisy image to obtain a second denoised image. The first and second denoised images are then concatenated in parallel along the channel dimension and input into the fusion module to obtain the final denoised image. This invention can stably and effectively remove image noise while also restoring image texture details relatively well, and can be used for target recognition, image conversion, and resource exploration.

Description

Technical Field

[0001] This invention belongs to the field of digital image processing technology, specifically relating to an image denoising method that can be used for target recognition, image conversion, and resource exploration. Background Technology

[0002] Due to environmental and human factors, as well as sensor limitations, imaging devices inevitably introduce noise during the acquisition, manual editing, and subsequent transmission of signals collected from the imaged object. Noisy images adversely affect many image processing applications. Therefore, image denoising is a crucial step in digital image processing, aiming to reduce or eliminate noise in images and improve their quality and usability. In the field of image denoising, deep learning-based image denoising methods have developed rapidly and have attracted widespread attention from researchers.

[0003] In their paper "A Remote Sensing Hyperspectral Image Noise Removal Method Based on Multipriors Guidance," Wu, Yinhu, and others proposed a multipriors-guided method for denoising remote sensing hyperspectral images. This method extracts image features from a variance-stabilized hyperspectral image using a linear self-attention module, and then reconstructs a noise-free image from the extracted features. While this method can stably remove noise from images, the quality of the resulting denoised image needs improvement.

[0004] Huang, Yongqiang, and others proposed a method for denoising optical coherence tomography images based on disentangled representation in their paper "Noise-Powered Disentangled Representation for Unsupervised Speckle Reduction of Optical Coherence Tomography Images." This method first extracts noise information from the noisy image using a noise encoder, and then subtracts the extracted noise information from the noisy image to obtain the denoised image. While this noise prediction-based method can produce relatively high-quality denoised images, its performance is not robust enough, especially in challenging scenarios with high noise levels where its denoising performance significantly degrades.

[0005] Therefore, how to design an image denoising method that can accurately remove noise while preserving image details well has become an urgent problem to be solved in this field. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of the existing technology by proposing an image denoising method based on feature fusion complementary learning convolutional neural networks, so as to improve the quality of denoised images and ensure the stability of denoising performance in high noise intensity scenes.

[0007] The technical idea of ​​this invention is to combine the advantages of deep learning image denoising methods based on noise prediction and denoised content prediction, and integrate the fusion of prediction results and the interaction of intermediate feature maps into a unified framework to achieve effective fusion of the two methods and effectively improve the quality of denoised images.

[0008] Based on the above ideas, the implementation scheme of the present invention includes the following:

[0009] 1. An image denoising method based on feature fusion complementary learning, characterized by comprising the following steps:

[0010] 1) Obtain noisy and denoised image sets in pairs from the public dataset, perform normalization preprocessing on them to change the pixel value range of the images from [0, 255] to [-1, 1]; and divide the preprocessed image sets into training and testing sets in an 8:2 ratio.

[0011] 2) Construct a feature interaction complementary learning denoising network:

[0012] 2a) Construct two predictors, each consisting of an N-level encoder and an M-level decoder, to predict the denoised image and noise, respectively;

[0013] 2b) Establish a feature interaction module including a conditional weight generator and a convolutional network to weightedly couple the outputs of each stage decoder of the two predictors;

[0014] 2c) Establish a fusion module including convolutional layers to fuse the denoised images predicted by the output of the last stage decoder of the two predictors to obtain the fused denoised image;

[0015] 2d) The outputs of the first to M-1th level decoders of the two predictors are weighted and coupled by the feature interaction module and then connected to the input of the next level decoder. The Mth level decoders of the two predictors output the first denoised image and noise respectively. The noise is subtracted from the original noise image to obtain the second denoised image. The first denoised image and the second denoised image are then connected in parallel according to the channel dimension and connected to the input of the fusion module to form a feature interaction complementary learning denoising network.

[0016] 3) The loss function L of the feature interaction complementary learning denoising network is designed as follows:

[0017] L = L fin +0.5L n +0.5L rsi

[0018] Where Lfin is the loss function of the denoised image output by the fusion module, Lrsi is the loss function of the first denoised image, and Ln is the loss function of the noise output by the predictor.

[0019] 4) Input the training set images into the feature interaction complementary learning denoising network, calculate the gradient of the loss function using the backpropagation method, and optimize the parameters of the two predictors, the feature interaction module, and the fusion module respectively with the goal of predicting the denoised image and predicting the noise, until the loss function tends to stabilize and the trained feature interaction complementary learning denoising network is obtained, which includes the trained denoised image predictor and noise predictor.

[0020] 5) Input the noisy images from the test set into the trained feature interaction complementarity learning denoising network to obtain denoised images.

[0021] 2. An image denoising system based on feature fusion complementary learning, characterized in that it comprises: a denoised image predictor, a noise predictor, a feature interaction module, and a fusion module;

[0022] The denoised image predictor includes an N-level encoder and an M-level decoder. The N-level encoder is used to extract features of the image content in the noisy image and input the features into the first-level decoder. The outputs of the first to M-1 levels of the M-level decoder are weighted and coupled through a feature interaction module and then input into the next level decoder. The M-level decoder outputs the preliminary denoised image.

[0023] The noise predictor includes an N-level encoder and an M-level decoder. The N-level encoder is used to extract the noise features in the noisy image and input them into the first-level decoder. The outputs of the first to M-1 levels of the M-level decoder are weighted and coupled through a feature interaction module and then input into the next level decoder. The M-level decoder outputs the image noise, and the image noise output by the predictor is subtracted from the original noisy image to obtain a preliminary denoised image.

[0024] The feature interaction module includes a weight generator and a convolutional network layer, which are used to weighted couple the output features of each decoder in the two predictors;

[0025] The fusion module includes a convolutional network layer for fusing the preliminary denoised images output from the last stage decoder of the two predictors to obtain the fused final denoised image.

[0026] Compared with the prior art, the present invention has the following advantages:

[0027] First, this invention effectively combines the advantages of deep learning image denoising methods based on noise prediction and denoised image prediction, and constructs an image denoising network based on feature fusion complementary learning, thereby improving the quality of denoised images.

[0028] Secondly, by integrating the feature interactions between predictors and the predictor output results into a single system, this invention achieves a full combination of the two methods, which not only effectively improves the denoising performance of images but also ensures the stability of denoising performance in high noise intensity scenes. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating the implementation of the present invention;

[0030] Figure 2 This is a schematic diagram of the feature interaction complementary learning denoising network structure constructed in this invention;

[0031] Figure 3 for Figure 2 The predictor structure diagram in the image;

[0032] Figure 4 for Figure 2 Structure diagram of the feature interaction module in the middle;

[0033] Figure 5 A structural block diagram of the noise reduction system for the invention;

[0034] Figure 6 The figure shows a comparison of the denoising effects of this invention and five existing denoising methods in high-noise scenarios. Detailed Implementation

[0035] The embodiments and effects of the present invention will be further described in detail below with reference to the accompanying drawings.

[0036] Reference Figure 1 The implementation steps of the present invention include the following:

[0037] Step 1: Acquisition and preprocessing of the dataset.

[0038] Download publicly available image denoising datasets from the public network and perform normalization preprocessing on them to change the pixel value range of the image from [0, 255] to [-1, 1].

[0039] The preprocessed image dataset was divided into a training set and a test set in an 8:2 ratio.

[0040] Step 2: Construct an image denoising network based on feature fusion complementary learning.

[0041] 2.1) Two predictors, each consisting of an N-level encoder and an M-level decoder, are established to predict the denoised image and noise, respectively:

[0042] like Figure 3As shown, each predictor includes a convolutional layer Conv with a kernel size of 4×4 and a stride of 2, a deconvolutional layer DeConv with a kernel size of 4×4 and a stride of 2, a batch normalization layer BN, a LeakReLU activation function, and a Tanh activation function, wherein:

[0043] The convolutional layer and the batch normalization layer are connected to form the first-level encoder. The convolutional layer, the batch normalization layer, and the LeakRelu activation function are connected in sequence to form other levels of encoders. Each level of encoder is then connected in sequence to form an N-level encoder.

[0044] The deconvolutional layer, the batch normalization layer, and the Tanh activation function are connected in sequence to form the last stage decoder. The deconvolutional layer, the batch normalization layer, and the LeakRelu activation function are connected in sequence to form other stages of decoders. The decoders at each stage are connected in sequence to form the M-stage decoder.

[0045] Connect the Nth level encoder to the first level decoder, and connect the output of each level in the encoder to the corresponding level in the decoder to obtain a predictor.

[0046] 2.2) Construct a feature interaction module to weightedly couple the outputs of each stage decoder of the two predictors:

[0047] like Figure 4 As shown, the implementation of this step includes the following:

[0048] 2.2.1) Establish a conditional weight generator (CWG) consisting of a cascaded average pooling layer, a linear layer, and a sigmoid activation function to generate the weight parameters of the feature map;

[0049] 2.2.2) Construct a convolutional layer Conv with a kernel size of 1×1 and a stride of 1. 1×1 This is used to weight each channel of the feature map;

[0050] 2.2.3) Based on the weight generator CWG and the convolutional layer Conv 1×1 The following interactive relationships constitute the feature interaction module:

[0051] The initial feature PF1 of the first predictor, the PF1 processed by CWG, and the Conv... 1×1 The processed initial features PF2 of the second predictor are multiplied together to obtain the intermediate features MPF1 of the first predictor:

[0052] MPF1 = PF1 × CWG(PF1) × Conv 1×1 (PF2)

[0053] The initial feature PF2 of the second predictor, the PF2 processed by CWG, and the feature processed by Conv... 1×1The initial features PF1 of the processed first predictor are multiplied together to obtain the intermediate features MPF2 of the second predictor:

[0054] MPF2 = PF2 × CWG(PF2) × Conv 1×1 (PF1);

[0055] The coupling feature CPF1 of the first predictor is obtained by subtracting MPF1 from the initial feature PF1 of the first predictor and then adding MPF2:

[0056] CPF1 = PF1 - MPF1 + MPF2;

[0057] The coupling feature CPF2 of the second predictor is obtained by subtracting MPF2 from the initial feature PF2 of the second predictor and then adding MPF1:

[0058] CPF2 = PF2 - MPF2 + MPF1;

[0059] 2.3) The fusion module is established, which consists of a convolutional layer with a kernel size of 1×1 and a stride of 1. It is used to fuse the denoised images predicted by the output of the last stage decoder of the two predictors to obtain the fused denoised image.

[0060] 2.4) The outputs of the first to M-1th level decoders of the two predictors are weighted and coupled by the feature interaction module and then connected to the input of the next level decoder. The Mth level decoders of the two predictors output the first denoised image and noise respectively. The noise is subtracted from the original noise image to obtain the second denoised image. The first denoised image and the second denoised image are then connected in parallel according to the channel dimension and connected to the input of the fusion module to form an image denoising network based on feature fusion complementary learning.

[0061] Step 3: Construct the loss function for the image denoising network based on feature fusion complementary learning.

[0062] 3.1) Set the loss function L for the denoised image output by the fusion module. fin :

[0063] L fin =L char +0.15L SSIM +0.2L edge

[0064] in For the pixel loss between the predicted result and the standard result, Pre f and gt i These represent the fused denoised image and the standard noise-free image in the training set, respectively, with ε being a constant of 10. -3 ;

[0065] L SSIM =1-SSIM(pref ,gt i The sum of the predicted and standard results is the SSIM loss, where SSIM is the structural similarity between the predicted and standard results.

[0066] The edge loss between the predicted result and the standard result is Δ, where Δ is the Laplace operator.

[0067] 3.2) Set the first denoised image loss function L rsi :

[0068] L rsi =L' char +0.15L' SSIM +0.2L' edge

[0069] in For the pixel loss between the predicted result and the standard result, pre r and gt i These represent the first denoised image and the standard noise-free image in the training set, respectively.

[0070] L' SSIM =1-SSIM(pre r ,gt i The sum of the predicted and standard results is the SSIM loss, where SSIM is the structural similarity between the predicted and standard results.

[0071] The edge loss between the predicted result and the standard result;

[0072] 3.3) Set the loss function L for the predictor output noise. n :

[0073] L n =L char-n +0.5L asymm

[0074] in For noise and standard result pixel loss, pre n The noise represented by gt is the predicted noise. n For standard noise, gt n It is obtained by subtracting the standard noise-free image from the noisy image in the training set;

[0075] For the asymmetric loss of noise estimation, For all pre n -gt n <0 pre n -gt n The factorial of α, where α is a constant set to 0.3;

[0076] 3.4) Using the loss functions from steps 3.1) to 3.3), the loss function L of the constructed feature interaction complementarity learning denoising network is:

[0077] L = L fin +0.5L n +0.5L rsi .

[0078] Step 4: Train the image denoising network based on feature fusion complementary learning.

[0079] 4.1) Set the feature interaction complementarity learning denoising network to training mode;

[0080] 4.2) Input the noisy image in the training set into the network. The noisy image is forward propagated through the feature interaction complementary learning denoising network. The first predictor is trained on the denoised image predictor and calculates and outputs the first denoised image. The second predictor is trained on the noise predictor and calculates and outputs the noise. The noise output by the second predictor is subtracted from the original noisy image to obtain the second denoised image. The fusion module is trained on the fusion of the first denoised image and the second denoised image and calculates and outputs the fused denoised image.

[0081] 4.3) Calculate the loss function L using the first denoised image and the fused denoised image, respectively, against the standard noise-free image. rsi L fin The loss function L is calculated using the noise output by the second predictor and the standard noise. n ;

[0082] 4.4) Calculate the loss function L = L using the backpropagation method. fin +0.5L n +0.5L rsi The gradient;

[0083] 4.5) Use the Adam optimizer to update the model parameters based on the changes in the gradient of the loss function;

[0084] 4.6) Repeat steps 4.2) to 4.5) until the loss function stabilizes, and you will get the trained image denoising network.

[0085] Step 5: Use the trained denoising network to denoise the image.

[0086] 5.1) Set the feature interaction complementarity learning denoising network to test mode;

[0087] 5.2) Input the noisy images in the test set into the trained image denoising network, and convert them into the first denoised image through its denoised image predictor; and convert the noisy images in the test set into noise through the noise predictor, and then subtract the noise from the original noisy image to obtain the second denoised image;

[0088] 5.3) The first denoised image and the second denoised image are concatenated in parallel along the channel dimension, and then transformed into the final denoised image through the fusion module.

[0089] Reference Figure 5 This example demonstrates an image denoising system based on feature fusion and complementary learning, comprising: a denoised image predictor, a noise predictor, a feature interaction module, and a fusion module, wherein:

[0090] The denoised image predictor includes an N-level encoder and an M-level decoder. The N-level encoder is used to extract features of the image content in the noisy image and input the features into the first-level decoder. The outputs of the first to M-1 levels of the M-level decoder are weighted and coupled through a feature interaction module and then input into the next level decoder. The M-level decoder outputs the preliminary denoised image.

[0091] The noise predictor includes an N-level encoder and an M-level decoder. The N-level encoder is used to extract the noise features in the noisy image and input them into the first-level decoder. The outputs of the first to M-1 levels of the M-level decoder are weighted and coupled through a feature interaction module and then input into the next level decoder. The M-level decoder outputs the image noise, and the image noise output by the predictor is subtracted from the original noisy image to obtain a preliminary denoised image.

[0092] The feature interaction module includes a weight generator and a convolutional network layer, which are used to weighted couple the output features of each decoder in the two predictors;

[0093] The fusion module includes a convolutional network layer for fusing the preliminary denoised images output from the last stage decoder of the two predictors to obtain the fused final denoised image.

[0094] The effects of the present invention will be further explained below with reference to experiments.

[0095] 1. Experimental conditions:

[0096] The experiments of this invention were conducted in a hardware environment with an Intel 7800X CPU, 32GB of memory, and an RTX 3090Ti graphics card, and a Windows 10 operating system and Python 3.8.16 software environment.

[0097] 2. Experiment Content:

[0098] In this experiment, Gaussian white noise with σ ranging from 0 to 60 was added to the NWPU-RESISC45 remote sensing dataset to evaluate the performance stability of the proposed method under different noise intensities. The denoising performance of the proposed method was evaluated using the PolyU and SIDD real denoising datasets. On each dataset, the detection performance of the proposed method was compared with that of five existing mainstream methods—BM3D, DnCNN, DUMRN, CBDNet, and DGCL—based on two metrics: Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM).

[0099] Experiment 1: Gaussian white noise with σ values ​​of 5, 10, 20, 30, 40, and 50 was added to the NWPU-RESISC45 remote sensing dataset. Image denoising was performed using five mainstream methods and the method of this invention. The PSNR and SSIM of the denoised images obtained by each method and the standard denoised image are shown in Table 1.

[0100] Table 1. Comparison of NWPU-RESISC45 dataset metrics between the present invention and five existing mainstream methods.

[0101]

[0102] The higher the Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) in the table, the closer the denoised image is to the standard image in the dataset, and the better the image denoising effect.

[0103] As shown in Table 1, the method proposed in this invention outperforms other mainstream methods in both low-noise scenarios with small σ values ​​and high-noise scenarios with large σ values.

[0104] The comparison of the denoising effects of this invention and five existing denoising methods in a high-noise scene with σ=50 is shown in the figure below. Figure 6 As shown, 6(a) is a noisy image, 6(b) is a denoised image, 6(c) is a denoising effect image of the BM3D method, 6(d) is a denoising effect image of the DnCNN method, 6(e) is a denoising effect image of the DUMNP method, 6(f) is a denoising effect image of the CBDNet method, 6(g) is a denoising effect image of the DGCL method, and 6(h) is a denoising effect image of the method of the present invention.

[0105] from Figure 6 It can be seen that other comparative methods exhibit varying degrees of performance degradation in high-noise scenarios and cannot effectively complete the denoising task. In contrast, the method of this invention can not only stably remove noise but also restore image details as much as possible.

[0106] Experiment 2: Image denoising was performed on the test set of the PolyU dataset using five mainstream methods and the method of this invention. The PSNR and SSIM of the denoised images obtained by each method and the standard denoised image are shown in Table 2.

[0107] Table 2 Comparison of PolyU dataset metrics between this invention and five existing mainstream methods

[0108]

[0109]

[0110] As shown in Table 2, the denoised image generated by this invention has significantly higher performance than other comparative methods, and the image denoising effect is the best.

[0111] Experiment 3: Image denoising was performed on the test set of the SIDD dataset using five mainstream methods and the method of this invention. The PSNR and SSIM of the denoised images obtained by each method and the standard denoised image are shown in Table 3.

[0112] Table 3 Comparison of SIDD dataset metrics between this invention and five existing mainstream methods

[0113]

[0114] As shown in Table 3, the denoised image generated by this invention has significantly higher performance than other comparative methods, and the image denoising effect is the best.

[0115] The sources of the above five mainstream methods are as follows:

[0116] BM3D: [Reference: Dabov K, Foi A, Katkovnik V, et al.Image Denoising bySparse 3-DTransform-Domain Collaborative Filtering[J].IEEE Transactions onImage Processing,2007,16(8):p.2080-2095.].

[0117] DnCNN: [Reference: Zhang K, Zuo W, Chen Y, et al. Beyond a GaussianDenoiser: Residual Learning of Deep CNN for Image Denoising[J]. IEEE Transactions on Image Processing, 2016, 26(7): 3142-3155.].

[0118] CBDNet: [Reference: Guo S, Yan Z, Zhang K, et al. "Toward ConvolutionalBlind Denoising of Real Photographs," In Proceedings of the 2019IEEE / CVFConference on Computer Vision and Pattern Recognition (CVPR)., pp.1712–1722, Jun.2019.].

[0119] DUMRN: [Reference: Xu J, Yuan M, Yan DM, et al. Deep unfolding multi-scaleregularizer network for image denoising[J]. Computational Visual Media, 2023, 9(2): 335-350.].

[0120] DGCL: [Reference: Zhao S.; Lin S.; Cheng

[0121] The above experimental results demonstrate the correctness and effectiveness of the method proposed in this invention.

[0122] The above description is only a preferred embodiment of the present invention and is not intended to limit the technical scope of the present invention. Any modifications, equivalent substitutions, improvements, etc., made without departing from the principle of the present invention should be included within the protection scope of the present invention.

[0123] It should be noted that the step numbers in the specification and claims of this invention are only for the purpose of clearly describing the embodiments of this invention and facilitating understanding, and their order is not limited.

Claims

1. An image denoising method based on feature fusion and complementary learning, characterized in that, Includes the following steps: 1) Obtain noisy and denoised image sets in pairs from the public dataset, perform normalization preprocessing on them to change the pixel value range of the images from [0, 255] to [-1, 1]; and divide the preprocessed image sets into training and testing sets in an 8:2 ratio. 2) Construct a feature interaction complementary learning denoising network: 2a) Construct two predictors, each consisting of an N-level encoder and an M-level decoder, to predict the denoised image and noise, respectively; 2b) Establish a feature interaction module including a conditional weight generator and a convolutional network to weightedly couple the outputs of each stage decoder of the two predictors; the implementation steps include the following: 2b1) Concatenate the average pooling layer, the linear layer, and the sigmoid activation function to form a conditional weight generator (CWG); 2b2) Construct a convolutional layer Conv with a kernel size of 1×1 and a stride of 1. 1×1 ; 2b3) The initial feature PF1 of the first predictor, the PF1 after CWG processing, and the PF1 after Conv 1×1 The initial feature PF2 of the processed second predictor is multiplied to obtain the intermediate feature MPF1 of the first predictor. The initial feature PF2 of the second predictor is multiplied by the PF2 processed by CWG and the feature MPF1 processed by Conv. 1×1 The initial features PF1 of the processed first predictor are multiplied together to obtain the intermediate features MPF2 of the second predictor: 2b4) Subtracting MPF1 from the initial feature PF1 of the first predictor and adding MPF2 gives the first predictor coupling feature CPF1. Subtracting MPF2 from the initial feature PF2 of the second predictor and adding MPF1 gives the second predictor coupling feature CPF2, thus achieving weighted coupling of the two predictor features. 2c) Establish a fusion module including convolutional layers to fuse the denoised images predicted by the output of the last stage decoder of the two predictors to obtain the fused denoised image; 2d) The outputs of the first to M-1th level decoders of the two predictors are weighted and coupled by the feature interaction module and then connected to the input of the next level decoder. The Mth level decoders of the two predictors output the first denoised image and noise respectively. The noise is subtracted from the original noise image to obtain the second denoised image. The first denoised image and the second denoised image are then connected in parallel according to the channel dimension and connected to the input of the fusion module to form a feature interaction complementary learning denoising network. 3) The loss function L of the feature interaction complementary learning denoising network is designed as follows: ； Among them, L fin L is the loss function for the fusion module to output the denoised image. rsi It is the first denoising image loss function, L n It is the loss function of the predictor output noise; 4) Input the training set images into the feature interaction complementary learning denoising network, calculate the gradient of the loss function using the backpropagation method, and optimize the parameters of the two predictors, the feature interaction module, and the fusion module respectively with the goal of predicting the denoised image and predicting the noise, until the loss function tends to stabilize and the trained feature interaction complementary learning denoising network is obtained, which includes the trained denoised image predictor and noise predictor. 5) Input the noisy images from the test set into the trained feature interaction complementarity learning denoising network to obtain denoised images.

2. The method according to claim 1, characterized in that, Step 2a) involves establishing two predictors, each comprising a multi-level encoder and a multi-level decoder. The implementation steps include the following: 2a1) The encoder is composed of a convolutional layer with a kernel size of 4×4 and a stride of 2, a batch normalization layer, and a LeakRelu activation function; 2a2) The decoder is composed of a deconvolutional layer with a kernel size of 4×4 and a stride of 2, a batch normalization layer, LeakRelu and Tanh activation function; 2a3) Connect the convolutional layer and the batch normalization layer to form the first-level encoder, and connect the convolutional layer, the batch normalization layer, and LeakReLU; Activation functions are sequentially connected to form other levels of encoders; The deconvolutional layer, batch normalization layer, and Tanh activation function are connected in sequence to form the last stage decoder. The deconvolutional layer, batch normalization layer, and LeakRelu activation function are connected in sequence to form other stages decoders. The output of each stage in the encoder is connected to the corresponding stage in the decoder to obtain a predictor. 2a4) Repeat step 2a3) to obtain another predictor.

3. The method according to claim 1, characterized in that, Step 2c) forms the convolutional layer of the fusion module, with a kernel size of 1×1 and a stride of 1.

4. The method according to claim 1, characterized in that, In step 3), the fusion module outputs the loss function L of the denoised image. fin and the first denoised image loss function L rsi They are represented as follows: ； ； in For the pixel loss between the predicted result and the standard result, Pre f and gt i These represent the fused denoised image and the standard noise-free image in the training set, respectively, with ε being a constant of 10. −3 ; The SSIM loss is calculated between the predicted results and the standard results, where SSIM represents the structural similarity between the predicted results and the standard results. The edge loss between the predicted result and the standard result is Δ, where Δ is the Laplace operator. For the pixel loss between the predicted result and the standard result, pre r and gt i These represent the first denoised image and the standard noise-free image in the training set, respectively. The SSIM loss is calculated between the predicted results and the standard results, where SSIM represents the structural similarity between the predicted results and the standard results. The edge loss is calculated by comparing the predicted result with the standard result.

5. The method according to claim 1, characterized in that, The loss function L of the predictor output noise in step 3) n It is expressed as follows: ； in For noise and standard result pixel loss, pre n The noise represented by gt is the predicted noise. n For standard noise, gt n It is obtained by subtracting the standard noise-free image from the noisy image in the training set, where ε is a constant of 10. −3 ; For the asymmetric loss of noise estimation, For all pre n -gt n <0 o'clockgt n -pre n The product of the products, with α being a constant set to 0.

3.

6. The method according to claim 1, characterized in that, In step 4), the gradient of the loss function is calculated using backpropagation, and the parameters of the two predictors, the feature interaction module, and the fusion module are optimized using optimizers with the goal of predicting the denoised image and predicting the noise, respectively. The implementation steps include the following: 4a) Set the feature interaction complementarity learning denoising network to training mode; 4b) The noisy images in the training set are forward propagated through a feature interaction complementary learning denoising network. The first predictor is trained on the denoised image predictor and outputs the first denoised image. The second predictor is trained on the noise predictor and outputs the noise. The noise output by the second predictor is subtracted from the original noisy image to obtain the second denoised image. The fusion module is trained on the fusion of the first denoised image and the second denoised image and outputs the fused denoised image. 4c) Calculate the loss function L using the first denoised image and the fused denoised image, respectively, against the standard noise-free image. rsi L fin The loss function L is calculated using the noise output by the second predictor and the standard noise. n ; 4d) Calculate the loss function using the backpropagation method. The gradient; 4e) Use the Adam optimizer to update the model parameters based on the changes in the gradient of the loss function; 4f) Repeat process 4b) to 4e) until the loss function reaches a stable state.

7. The method according to claim 1, characterized in that, Step 5) involves inputting the noisy images from the test set into the trained feature interaction complementarity learning denoising network to obtain denoised images. The steps include the following: 5a) Set the feature interaction complementarity learning denoising network to test mode; 5b) The denoising image predictor converts noisy images in the test set into first denoised images; 5c) The noise predictor converts the noisy images in the test set into noise, and then subtracts the noise from the original noisy image to obtain the second denoised image; 5d) The first denoised image and the second denoised image are concatenated in parallel along the channel dimension, and then transformed into the final denoised image through the fusion module.

8. An image denoising system for implementing the method of claim 1, characterized in that, include: Denoising image predictor, noise predictor, feature interaction module, and fusion module; The denoised image predictor includes an N-level encoder and an M-level decoder. The N-level encoder is used to extract features of the image content in the noisy image and input the features into the first-level decoder. The outputs of the first to M-1 levels of the M-level decoder are weighted and coupled through a feature interaction module and then input into the next level decoder. The M-level decoder outputs the preliminary denoised image. The noise predictor includes an N-level encoder and an M-level decoder. The N-level encoder is used to extract the noise features in the noisy image and input them into the first-level decoder. The outputs of the first to M-1 levels of the M-level decoder are weighted and coupled through a feature interaction module and then input into the next level decoder. The M-level decoder outputs the image noise, and the image noise output by the predictor is subtracted from the original noisy image to obtain a preliminary denoised image. The feature interaction module includes a weight generator and a convolutional network layer, which are used to weighted couple the output features of each decoder in the two predictors; The fusion module includes a convolutional network layer for fusing the preliminary denoised images output from the last stage decoder of the two predictors to obtain the fused final denoised image.

Images