Image denoising method and system based on adversarial frequency mixing technology

Abstract

The present application discloses an image denoising method and system based on adversarial frequency mixing technology, which belongs to the field of image processing. The method includes: obtaining a training data set of noisy images and noise-free images, inputting the noisy images into a denoising network, predicting an initial denoised image, converting the noisy images and the initial denoised images into the frequency domain using Fourier transform and extracting low-frequency bands, medium-frequency bands and high-frequency bands in layers, generating mixed masks for each frequency band, obtaining and reorganizing mixed frequency bands, obtaining a mixed noise image through inverse Fourier transform, inputting the mixed noise image into a denoising network, and obtaining a final denoised image; determining a first loss function based on the initial denoised image, the final denoised image and the noise-free image, training and optimizing the denoising network; inputting the image to be denoised into the optimized denoising network, and outputting a target image. The present application can effectively solve the problem of insufficient generalization ability of existing denoising networks when facing unknown noise distributions.

Description

Technical Field

[0001] The present application belongs to the field of image processing technology, and specifically relates to an image denoising method and system based on adversarial frequency mixing technology. Background Art

[0002] Image denoising methods mainly improve image clarity and contrast by effectively eliminating noise in images, while retaining key features such as image edges and textures to enhance image recognition capabilities. This technology is not only widely used in medical imaging diagnosis, security monitoring, and unmanned driving, improving diagnostic accuracy, monitoring safety, and driving safety, but also plays an important role in aerospace, image restoration, and art processing, and has promoted technological progress and application development in related fields. In the process of image recognition and analysis, the image to be processed is usually denoised through a trained image denoising model. Among them, the denoising convolutional neural network is one of the most well-known models, which has laid the foundation for further innovation in the field of image processing technology. Later, with the widespread application of the Transformer architecture, relying on its outstanding performance in identifying complex dependencies, it effectively reduced the impact of noise on the image.

[0003] Traditional denoising networks artificially add noise to images to train and evaluate the denoising networks. The denoising networks trained in this way are often only applicable to specific types of noise distributions, such as Gaussian noise, Poisson noise, etc., while image noise in the real world is often affected by many factors of image signal processing, such as the difference of color filter arrays, gamma correction, etc. Therefore, even the most advanced denoising networks will have insufficient generalization capabilities when faced with unseen real-world noise distributions. Summary of the invention

[0004] The purpose of the embodiments of the present application is to provide an image denoising method and system based on adversarial frequency mixing technology, which can effectively solve the problem of insufficient generalization ability of existing image denoising networks when facing unknown real-world noise distribution.

[0005] In order to solve the above technical problems, this application is implemented as follows:

[0006] In a first aspect, an embodiment of the present application provides an image denoising method based on adversarial frequency mixing technology, the method comprising:

[0007] Obtain a training data set, the training data set including a noisy image and a corresponding noise-free image;

[0008] Input the noisy image into the denoising network and predict the initial denoised image;

[0009] According to the Fourier transform principle, the noise image and the initial denoised image are converted to the frequency domain and hierarchically extracted to obtain a first low-frequency band, a first intermediate-frequency band and a first high-frequency band of the noise image, and a second low-frequency band, a second intermediate-frequency band and a second high-frequency band of the initial denoised image, respectively;

[0010] Generate a low-frequency mixing mask, a medium-frequency mixing mask, and a high-frequency mixing mask respectively through a mask generation network, wherein the low-frequency mixing mask corresponds to the first low-frequency band and the second low-frequency band, the medium-frequency mixing mask corresponds to the first medium-frequency band and the second medium-frequency band, and the high-frequency mixing mask corresponds to the first high-frequency band and the second high-frequency band;

[0011] Based on the low-frequency mixing mask, the middle-frequency mixing mask, and the high-frequency mixing mask, a mixed low-frequency band, a mixed middle-frequency band, and a mixed high-frequency band are respectively obtained through a frequency mixing module;

[0012] The mixed low-frequency band, the mixed middle-frequency band and the mixed high-frequency band are recombined, and a mixed noise image is obtained through inverse Fourier transformation;

[0013] Input the mixed noise image into the denoising network to obtain the final denoised image;

[0014] Determining a first loss function based on the initial denoised image, the final denoised image, and the noise-free image;

[0015] The denoising network is trained by the first loss function to obtain an optimized denoising network;

[0016] Get the image to be denoised and input it into the optimized denoising network to output the target image.

[0017] Furthermore, the first loss function is:

[0018] ,

[0019] in, represents the first loss function for optimizing the denoising network, represents the peak signal-to-noise ratio, represents the initial denoised image, represents the final denoised image, Indicates balance and The hyperparameters of represents a noise-free image.

[0020] Further, in the step of respectively generating a low-frequency mixed mask, a medium-frequency mixed mask and a high-frequency mixed mask through a mask generation network, it specifically includes:

[0021] Generate a mixing mask according to an element-wise mask generation network;

[0022] The element-by-element mask generation network uses pixel segmentation techniques to generate a corresponding hybrid mask for each frequency domain element of the noisy image and the initial denoised image;

[0023] Based on the U-Net network, the mixed mask of each frequency domain element is encoded and decoded to generate a low-frequency mixed mask, a medium-frequency mixed mask, and a high-frequency mixed mask.

[0024] Further, in the step of respectively generating a low-frequency mixed mask, a medium-frequency mixed mask and a high-frequency mixed mask through a mask generation network, it specifically includes:

[0025] generating a mixed mask according to a band mask generation network, wherein the band mask generation network includes at least one convolutional layer and a fully connected layer;

[0026] Extract feature information of input data through convolutional layer;

[0027] Based on the feature information, the fully connected layer generates a set of mixed masks through nonlinear transformation, where each mask in the mixed mask corresponds to an element in a frequency band area;

[0028] According to different frequency ranges, a low-frequency mixing mask, a middle-frequency mixing mask and a high-frequency mixing mask are output.

[0029] Furthermore, the image denoising method based on the adversarial frequency mixing technology also includes:

[0030] Building an AFM module, the AFM module is used to generate a mixed low-frequency band, a mixed mid-frequency band, and a mixed high-frequency band;

[0031] Wherein, the AFM module at least includes a mask generation network and a frequency mixing module;

[0032] The mask generation network includes at least one deep learning model for automatically generating a low-frequency mixed mask, a medium-frequency mixed mask, and a high-frequency mixed mask according to the characteristics of the input image;

[0033] The frequency mixing module mixes the frequency components of the image according to the low-frequency mixing mask, the middle-frequency mixing mask and the high-frequency mixing mask to obtain a mixed low-frequency band, a mixed middle-frequency band and a mixed high-frequency band.

[0034] Furthermore, the image denoising method based on the adversarial frequency mixing technology also includes:

[0035] A complementary mixed mask is defined, and the complementary value of the complementary mixed mask and the mixed mask of the corresponding frequency is 1;

[0036] Based on the low-frequency complementary mixing mask, the middle-frequency complementary mixing mask and the high-frequency complementary mixing mask, a complementary mixed low-frequency band, a complementary mixed middle-frequency band and a complementary mixed high-frequency band are obtained respectively;

[0037] Recombining the complementary mixed low-frequency band, the complementary mixed middle-frequency band and the complementary mixed high-frequency band, and obtaining a complementary mixed noise image through inverse Fourier transformation;

[0038] Input the complementary mixed noise image into the denoising network to obtain the final complementary denoised image;

[0039] Determining a second loss function based on the final complementary denoised image, the final denoised image and the noise-free image;

[0040] The mask generation network is adversarially trained through the second loss function to update the AFM module.

[0041] Further, the step of determining the second loss function based on the final complementary denoised image, the final denoised image and the noise-free image specifically includes:

[0042] The second loss function is:

[0043] ,

[0044] in, Represents the second loss function updated for the AFM module, represents the peak signal-to-noise ratio, represents the final denoised image, represents the final complementary denoised image, Indicates balance and The hyperparameters of represents a noise-free image.

[0045] In a second aspect, an embodiment of the present application provides an image denoising network system based on adversarial frequency mixing technology, the system comprising:

[0046] Construct a data set module to obtain a training data set, where the training data set includes a noisy image and a corresponding noise-free image;

[0047] An image processing module is used to input the noisy image into the denoising network and predict the initial denoised image;

[0048] A frequency band layering module, for converting the noise image and the initial denoised image into the frequency domain and performing layered extraction according to the Fourier transform principle, so as to obtain a first low-frequency band, a first intermediate-frequency band and a first high-frequency band of the noise image, and a second low-frequency band, a second intermediate-frequency band and a second high-frequency band of the initial denoised image respectively;

[0049] A mask generation module, used to generate a low-frequency mixing mask, a medium-frequency mixing mask and a high-frequency mixing mask respectively through a mask generation network, wherein the low-frequency mixing mask corresponds to the first low-frequency band and the second low-frequency band, the medium-frequency mixing mask corresponds to the first medium-frequency band and the second medium-frequency band, and the high-frequency mixing mask corresponds to the first high-frequency band and the second high-frequency band;

[0050] A frequency band mixing module, for obtaining a mixed low frequency band, a mixed middle frequency band and a mixed high frequency band respectively through a frequency mixing module based on a low frequency mixing mask, a middle frequency mixing mask and a high frequency mixing mask;

[0051] An image conversion module is used to recombine the mixed low-frequency band, the mixed middle-frequency band and the mixed high-frequency band, and obtain a mixed noise image through inverse Fourier transformation;

[0052] An image output module is used to input the mixed noise image into the denoising network to obtain the final denoised image;

[0053] A function construction module for determining a first loss function based on an initial denoised image, a final denoised image and a noise-free image;

[0054] A training module, used to train the denoising network through a first loss function to obtain an optimized denoising network;

[0055] The final output module is used to obtain the image to be denoised and input it into the optimized denoising network to output the target image.

[0056] In a third aspect, an embodiment of the present application provides an electronic device, which includes a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the steps of the method described in the first aspect.

[0057] In a fourth aspect, an embodiment of the present application provides a readable storage medium, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the method described in the first aspect are implemented.

[0058] In the embodiment of the present application, a training data set is first obtained. The noisy images and noise-free images in the training data set provide a reliable basis for the training and accurate evaluation of the image denoising network; the noisy image is input into the denoising network to predict the initial denoised image; then, according to the Fourier transform principle, the noisy image and the initial denoised image are converted to the frequency domain and hierarchically extracted, and each frequency band is extracted by hierarchical Fourier transform to improve the denoising effect and the ability to retain image details; a low-frequency mixed mask, a medium-frequency mixed mask and a high-frequency mixed mask are respectively generated by a mask generation network, and a mask generation network is introduced to generate a mixed mask for each frequency band, which can more finely control the noise of each frequency band; each frequency band is subjected to a mixed mask according to the frequency band. The frequency bands are mixed and recombined, and a mixed noise image is obtained by inverse Fourier transform, so that the generated noise image is more consistent with the noise distribution in the actual scene; the mixed noise image is input into the denoising network to obtain the final denoised image, and the first loss function is determined based on the initial denoised image, the final denoised image and the noise-free image. The denoising network is trained according to the first loss function and the optimized denoising network is obtained. In the training process, a training module of adversarial hybrid technology is introduced to generate a new noise image similar to the real-world noise distribution in the frequency domain, and enhance the robustness of any denoising network to unknown noise types; finally, any image to be denoised is input into the optimized denoising network to provide users with a high-quality target image. The present application can effectively solve the problem of insufficient generalization ability of existing image denoising networks when facing unknown real-world noise distributions. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A flowchart of an implementation of an image denoising method based on adversarial frequency mixing technology provided by an embodiment of the present invention is shown;

[0060] Figure 2 A method for obtaining a mixed noise image and a complementary mixed noise image in an image denoising method based on adversarial frequency mixing technology provided by an embodiment of the present invention is shown;

[0061] Figure 3 A schematic diagram of the structure of an image denoising system based on adversarial frequency mixing technology provided by an embodiment of the present invention is shown;

[0062] Figure 4 A schematic structural diagram of an electronic device applying an image denoising method based on adversarial frequency mixing technology provided by an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0063] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0064] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described here. In addition, the "and / or" in the specification and claims represents at least one of the connected objects, and the character " / " generally represents that the objects associated with each other are in an "or" relationship.

[0065] The following, in conjunction with the accompanying drawings, describes in detail the image denoising method and system based on adversarial frequency mixing technology provided by the embodiments of the present application through specific embodiments and their application scenarios.

[0066] It is worth noting that the image denoising method based on adversarial frequency mixing technology disclosed in this application has demonstrated its advantages and application potential in multiple fields. In the field of medical imaging, it can clearly present the details of lesions, providing doctors with more accurate and reliable diagnostic basis, thereby greatly improving the accuracy and efficiency of medical diagnosis; in monitoring and security, in the face of complex and changeable monitoring environments and low light conditions, it can effectively improve the clarity and detail resolution of monitoring images, making blurred images sharp and vivid.

[0067] The image denoising method based on adversarial frequency mixing technology disclosed in this application not only solves the problem of insufficient generalization ability of existing image denoising networks when facing unknown noise distribution, but also realizes accurate layered extraction and mixing of different frequency bands in the image through unique frequency domain processing and adversarial mechanism, thereby greatly improving the clarity and detail restoration of the denoised image. At the same time, this method has strong adaptability and generalization ability, can be widely adapted to various types of image data, and has opened up a new path for the development and application of image denoising technology.

[0068] See also Figure 1 to Figure 2 , which is an implementation of an image denoising method based on adversarial frequency mixing technology provided by an embodiment of the first aspect of the present invention, the method comprises steps S101 to S110, wherein:

[0069] S101: Acquire a training data set, where the training data set includes a noisy image and a corresponding noise-free image.

[0070] In this embodiment, the training dataset is the Smartphone Image Denoising Dataset (SIDD dataset). The SIDD dataset is a training dataset specifically used for smartphone image denoising tasks. It contains 30,000 noisy images from 10 scenes taken by five representative smartphone cameras under different photo conditions. At the same time, the SIDD dataset also provides real noise-free images corresponding to the noisy images.

[0071] It is worth mentioning that the SIDD dataset provides a high-quality, large-scale denoising training resource, which contains images from different lighting conditions and different scenes, making the trained image denoising network more generalizable. In addition, the SIDD dataset contains noisy images and corresponding noise-free images, providing a reliable basis for the training and accurate evaluation of image denoising networks.

[0072] S102: Input the noisy image into a denoising network to predict an initial denoised image.

[0073] In some embodiments of the present application, the formula for denoising a noisy image is:

[0074] ,

[0075] in, represents the denoising network, represents a noisy image, represents the initial denoised image.

[0076] In this embodiment, the denoising network adopts a denoising convolutional neural network (Denoising Convolutional Neural Network, hereinafter referred to as DnCNN). DnCNN is a deep learning model specifically used for image and video denoising. It combines the latest advances in deep learning technologies such as convolutional neural networks and provides advanced performance in image denoising. The main structure of DnCNN includes convolutional layers, batch normalization and activation functions. By adding batch normalization between convolution and activation functions, DnCNN helps prevent overfitting and improve generalization capabilities.

[0077] It is worth noting that DnCNN uses the principle of residual learning. Its input is a noisy image. It predicts the noise by training a residual map and subtracts the predicted noise from the noisy image to obtain the initial denoised image. This method makes it easier to optimize the denoising network. Batch normalization is a technique used to accelerate the training of deep neural networks and improve their stability. In DnCNN, batch normalization is used to accelerate the training process and improve denoising performance.

[0078] S103: According to the Fourier transform principle, the noise image and the initial denoised image are converted to the frequency domain and hierarchically extracted to obtain the first low-frequency band, the first intermediate-frequency band and the first high-frequency band of the noise image, and the second low-frequency band, the second intermediate-frequency band and the second high-frequency band of the initial denoised image.

[0079] In some embodiments of the present application, the noisy image and the initial denoised image are converted to the frequency domain using a fast Fourier transform.

[0080] In some embodiments of the present application, a hierarchical Fourier transform is used to perform hierarchical extraction on the frequency domains of the noise image and the initial noise image to obtain a first low-frequency band, a first mid-frequency band and a first high-frequency band of the noise image, and a second low-frequency band, a second mid-frequency band and a second high-frequency band of the initial denoised image, respectively.

[0081] In this embodiment, the formula for performing hierarchical extraction on the frequency domain of the noise image and the initial noise image is:

[0082] ，

[0083] ,

[0084] in, represents a noisy image, represents the frequency domain of the noisy image, Respectively represent the low frequency band, medium frequency band, and high frequency band of the noise image. represents the initial denoised image, represents the frequency domain of the initial denoised image, Represent the low-frequency band, mid-frequency band and high-frequency band of the initial denoised image, respectively. Indicates hierarchical extraction in the frequency domain.

[0085] S104: Generate a low-frequency mixing mask, a medium-frequency mixing mask and a high-frequency mixing mask respectively through a mask generation network, wherein the low-frequency mixing mask corresponds to the first low-frequency band and the second low-frequency band, the medium-frequency mixing mask corresponds to the first medium-frequency band and the second medium-frequency band, and the high-frequency mixing mask corresponds to the first high-frequency band and the second high-frequency band.

[0086] In this embodiment, the low-frequency mixed mask is generated by the mask generation network. , IF Mixing Mask and high frequency mixed mask .

[0087] It is worth noting that the mixing mask is used to control the mixing ratio of each frequency band of the noisy image and the initial denoised image. In this way, new images with different noise characteristics can be generated, which are both challenging for the denoising network and can enhance its robustness to changes in real-world noise distribution.

[0088] In some embodiments of the present application, a mixed mask is generated according to an element-wise mask generation network (hereinafter referred to as AFM-E). AFM-E uses pixel cutting technology to generate a corresponding mixed mask for each frequency domain of the noisy image and the initial denoised image. Based on the U-Net network, the mixed mask of each frequency domain element is encoded and decoded to generate a low-frequency mixed mask, a medium-frequency mixed mask and a high-frequency mixed mask.

[0089] It is worth noting that AFM-E can perform fine operations on each frequency domain element of the image and accurately control the mixing ratio of different frequency components. During the processing, AFM-E can flexibly adjust the information fusion method of low-frequency bands, medium-frequency bands and high-frequency bands by generating element-by-element mixing masks for each frequency band, so as to retain the detailed features of the image to the greatest extent while suppressing noise, and enhance the adaptability of the entire denoising network to complex noise environments and different image types.

[0090] In some embodiments of the present application, a mixed mask is generated by a band-wise mask generation network (Band-wise Mask Generation Network, hereinafter referred to as AFM-B). AFM-B includes at least one convolutional layer and a fully connected layer. The convolutional layer extracts feature information of the input data. According to the feature information of the input data, the fully connected layer generates a set of mixed masks through nonlinear transformation. Each mask in the mixed mask corresponds to an element in a frequency band area. According to different frequency ranges, a low-frequency mixed mask, a medium-frequency mixed mask and a high-frequency mixed mask are output.

[0091] It is worth noting that AFM-B can accurately and efficiently generate frequency masks based on specific requirements and algorithms, effectively filter out target frequency band signals, and reduce the impact of interference signals. The frequency band masks generated by AFM-B can effectively enhance or suppress signals in specific frequency bands, thereby achieving more precise control of signals.

[0092] S105: Based on the low-frequency mixing mask, the middle-frequency mixing mask, and the high-frequency mixing mask, respectively obtain a mixed low-frequency band, a mixed middle-frequency band, and a mixed high-frequency band through a frequency mixing module.

[0093] In this embodiment, the frequency band mixing formula is:

[0094] ,

[0095] ,

[0096] ,

[0097] in, represents element-wise multiplication, , , Respectively represent low-frequency mixing mask, medium-frequency mixing mask and high-frequency mixing mask, represents a noisy image, Respectively represent the first low frequency band, the first mid frequency band, and the first high frequency band, represents the initial denoised image, Respectively represent the second low frequency band, the second middle frequency band, and the second high frequency band, , , They represent mixed low-frequency band, mixed mid-frequency band and mixed high-frequency band respectively.

[0098] It is worth mentioning that the frequency mixing module is used to mix the frequency bands of the noise image and the initial noise image, which can accurately control the fusion ratio of different frequency bands, so that the mixed noise image retains the basic characteristics of the initial noise image while adding new frequency components in the noise image, thereby improving the fineness and realism of the noise simulation.

[0099] S106: recombining the mixed low-frequency band, the mixed middle-frequency band and the mixed high-frequency band, and performing inverse Fourier transform to obtain a mixed noise image.

[0100] In this embodiment, the formula for recombining the mixed low frequency band, the mixed middle frequency band and the mixed high frequency band is:

[0101] ,

[0102] in, represents the frequency domain of the mixed noise image, , , They represent mixed low-frequency band, mixed mid-frequency band and mixed high-frequency band respectively.

[0103] In this embodiment, the formula for obtaining the mixed noise image through inverse Fourier transform is:

[0104] ,

[0105] in, represents a mixed noise image, Indicates inverse Fourier transform.

[0106] S107: Input the mixed image into the denoising network to obtain the final denoised image.

[0107] In this embodiment, by recombining the mixed low-frequency band, the mixed mid-frequency band and the mixed high-frequency band, and obtaining the mixed noise image through inverse Fourier transform, the information fusion of the image in different frequency bands can be accurately controlled, so that the generated noise image retains the smooth structure and overall contour of the image in the low-frequency band, enhances the details and texture features of the image in the mid-frequency band, and introduces fine noise patterns and edge information in the high-frequency band. The mixed noise image can better simulate the complexity and diversity of real-world images, and the noise it carries is closer to the noise distribution in the real world, so as to obtain new training samples.

[0108] In this embodiment, the final denoised image is compared and analyzed with the noise-free image. During the comparison process, direct feedback on the effect of the denoising network can be obtained, and the subsequent training of the denoising network can be ensured to be more refined based on the feedback.

[0109] S108: Determine a first loss function based on the initial denoised image, the final denoised image and the noise-free image.

[0110] In some embodiments of the present application, the first loss function is determined by a peak signal-to-noise ratio between an initial denoised image and a noise-free image and a peak signal-to-noise ratio between a final denoised image and a noise-free image.

[0111] It is worth mentioning that Peak Signal-to-Noise Ratio (PSNR) is a widely used image quality assessment indicator that measures the similarity between the reconstructed image and the original image. PSNR is usually used to evaluate the quality of tasks such as image compression and image restoration. The higher the PSNR value, the smaller the difference between the reconstructed image and the original image, that is, the better the image quality. The calculation formula for the peak signal-to-noise ratio is:

[0112] ,

[0113] ,

[0114] Where MSE stands for root mean square error, is the maximum possible pixel value, is the original image at position

[0115] The pixel value at is the reconstructed image at position The pixel value at and are the height and width of the image respectively.

[0116] S109: Train the denoising network using the first loss function to obtain an optimized denoising network.

[0117] In this embodiment, the first loss function is:

[0118] ,

[0119] in, represents the first loss function for optimizing the denoising network, represents the peak signal-to-noise ratio, represents the initial denoised image, represents the final denoised image, Indicates balance and The hyperparameters of represents a noise-free image.

[0120] It is worth mentioning that choosing a suitable loss function can enhance the stability and robustness of the denoising network. The loss function provides a quantitative evaluation criterion for the denoising network. By calculating the difference between the denoised image and the true noise-free image, the performance of the network can be accurately measured.

[0121] By way of example and not limitation, The value is determined according to the actual situation. In the embodiment of the present application, According to the training data test, when When the value is 0.8, the training effect of the denoising network is the best.

[0122] S110: Obtain the image to be denoised and input it into the optimized denoising network, and output the target image.

[0123] It is worth mentioning that an AFM module is built to generate a mixed low-frequency band, a mixed mid-frequency band and a mixed high-frequency band.

[0124] In some embodiments of the present application, the AFM module includes at least a mask generation network and a frequency mixing module.

[0125] Among them, the mask generation network includes at least one deep learning model, which is used to automatically generate a low-frequency mixing mask, a medium-frequency mixing mask and a high-frequency mixing mask according to the characteristics of the input image; the frequency mixing module mixes the frequency components of the image according to the low-frequency mixing mask, the medium-frequency mixing mask and the high-frequency mixing mask to obtain a mixed low-frequency band, a mixed medium-frequency band and a mixed high-frequency band.

[0126] In some examples of the present application, the AFM module is updated by fixing the denoising network.

[0127] In some examples of the present application, a complementary mixing mask is defined, and the complementary value of the complementary mixing mask and the mixing mask of the corresponding frequency is 1.

[0128] In this embodiment, the low-frequency complementary mixing mask, the medium-frequency complementary mixing mask and the high-frequency mixing mask are respectively , , .

[0129] In some examples of the present application, based on the low frequency complementary mixing mask, the middle frequency complementary mixing mask and the high frequency complementary mixing mask, a complementary mixed low frequency band, a complementary mixed middle frequency band and a complementary mixed high frequency band are obtained respectively.

[0130] In this embodiment, the mixing formula for the complementary frequency bands is:

[0131] ,

[0132] ,

[0133] ,

[0134] in, represents element-wise multiplication, , , Respectively represent low-frequency mixing mask, medium-frequency mixing mask and high-frequency mixing mask, represents a noisy image, Respectively represent the first low frequency band, the first mid frequency band, and the first high frequency band, represents the initial denoised image, Respectively represent the second low frequency band, the second middle frequency band, and the second high frequency band, , , They represent complementary hybrid low frequency band, complementary hybrid middle frequency band and complementary hybrid high frequency band respectively.

[0135] In some embodiments of the present application, the complementary mixed low frequency band, the complementary mixed middle frequency band and the complementary mixed high frequency band are recombined and subjected to inverse Fourier transformation to obtain a complementary mixed noise image.

[0136] In this embodiment, the formula for recombining the complementary hybrid low frequency band, the complementary hybrid middle frequency band and the complementary hybrid high frequency band is:

[0137] ,

[0138] in, Represents the frequency domain of the complementary mixed noise image.

[0139] In this embodiment, the formula for obtaining the complementary mixed noise image through inverse Fourier transform is:

[0140] ,

[0141] in, represents a complementary mixed noise image, Indicates inverse Fourier transform.

[0142] In some embodiments of the present application, a complementary mixed noise image is input into a denoising network to obtain a final complementary denoised image.

[0143] In some embodiments of the present application, a second loss function is determined based on the final complementary denoised image, the final denoised image and the noise-free image.

[0144] In this embodiment, the second loss function is:

[0145] ,

[0146] in, Represents the second loss function updated for the AFM module, represents the peak signal-to-noise ratio, represents the final denoised image, represents the final complementary denoised image, Indicates balance and The hyperparameters of represents a noise-free image.

[0147] By way of example and not limitation, The value is determined according to the actual situation. In the embodiment of the present application, According to the training data test, when When the value is 0.3, the training effect of the denoising network is the best.

[0148] In some embodiments of the present application, the mask generation network is adversarially trained using a second loss function to update the AFM module.

[0149] In this embodiment, The difference between the peak signal-to-noise ratio of the final denoised image and the noise-free image is minimized to generate an adversarial mixed noise image that is difficult to be processed by the denoising network; the peak signal-to-noise ratio between the final complementary denoised image and the noise-free image is maximized to obtain a final complementary denoised image that is complementary to the mixed noise image, and the final complementary denoised image is used as a relatively easy input to the denoising network to generate. By adversarially training the mask generation network, the mask generated by the mask generation network is closer to the real world when controlling the ratio of the initial denoised image to the mixed image.

[0150] In the embodiment of the present application, a training data set is first obtained. The noisy images and noise-free images in the training data set provide a reliable basis for the training and accurate evaluation of the image denoising network; the noisy image is input into the denoising network to predict the initial denoised image; then, according to the Fourier transform principle, the noisy image and the initial denoised image are converted to the frequency domain and hierarchically extracted, and each frequency band is extracted by hierarchical Fourier transform to improve the denoising effect and the ability to retain image details; a low-frequency mixed mask, a medium-frequency mixed mask and a high-frequency mixed mask are respectively generated by a mask generation network, and a mask generation network is introduced to generate a mixed mask for each frequency band, which can more finely control the noise of each frequency band; each frequency band is subjected to a mixed mask according to the frequency band. The frequency bands are mixed and recombined, and a mixed noise image is obtained by inverse Fourier transform, so that the generated noise image is more consistent with the noise distribution in the actual scene; the mixed noise image is input into the denoising network to obtain the final denoised image, and the first loss function is determined based on the initial denoised image, the final denoised image and the noise-free image. The denoising network is trained according to the first loss function and the optimized denoising network is obtained. In the training process, a training module of adversarial hybrid technology is introduced to generate a new noise image similar to the real-world noise distribution in the frequency domain, and enhance the robustness of any denoising network to unknown noise types; finally, any image to be denoised is input into the optimized denoising network to provide users with a high-quality target image. The present application can effectively solve the problem of insufficient generalization ability of existing image denoising networks when facing unknown real-world noise distributions.

[0151] See also Figure 3 , shown is a structural schematic diagram of an image denoising system based on adversarial frequency mixing technology provided in an embodiment of the second aspect of the present invention.

[0152] It should be noted that the image denoising method based on adversarial frequency mixing technology provided in the embodiment of the present application can be executed by an image denoising system based on adversarial frequency mixing technology, or a control module in the system for executing and loading the image denoising method based on adversarial frequency mixing technology. In the embodiment of the present application, the image denoising method based on adversarial frequency mixing technology provided in the embodiment of the present application is explained by taking the image denoising system based on adversarial frequency mixing technology executing the image denoising method based on adversarial frequency mixing technology as an example.

[0153] The image denoising system based on adversarial frequency mixing technology in the embodiment of the present application can be a device, or a component, integrated circuit, or chip in a terminal. The device can be a mobile electronic device or a non-mobile electronic device. Exemplarily, the mobile electronic device can be a mobile phone, a tablet computer, a laptop computer, a PDA, an in-vehicle electronic device, a wearable device, an ultra-mobile personal computer (UMPC), a netbook, or a personal digital assistant (PDA), etc. The non-mobile electronic device can be a server, a personal computer (PC), a television (TV), a teller machine or a self-service machine, etc., which is not specifically limited in the embodiment of the present application.

[0154] The image denoising system based on adversarial frequency mixing technology in the embodiment of the present application may be a device having an operating system. The operating system may be an Android operating system, an iOS operating system, or other possible operating systems, which are not specifically limited in the embodiment of the present application.

[0155] The image denoising system based on adversarial frequency mixing technology provided in the embodiment of the present application can achieve Figure 1 to Figure 2 The various processes implemented in the method embodiment.

[0156] In some embodiments of the present application, the above-mentioned image denoising system based on adversarial frequency mixing technology includes:

[0157] The dataset building module 201 is used to obtain a training dataset, where the training dataset includes a noisy image and a corresponding noise-free image.

[0158] In this embodiment, the main function of the dataset building module 201 is to obtain a training dataset for denoising network training, which consists of pairs of noisy images and corresponding noise-free images. By providing the correspondence between the noisy images and the noise-free images, the denoising network can learn how to restore a clear noise-free image from the noisy images. In addition, different types of noise can also be used, such as Gaussian noise, salt and pepper noise, Poisson noise, etc., to increase the diversity and authenticity of the dataset.

[0159] The image processing module 202 is used to input the noisy image into the denoising network to predict an initial denoised image.

[0160] In this embodiment, the image processing module 202 is responsible for inputting the noisy image into the denoising network. When the noisy image is input into the denoising network, the denoising network will use its internal complex structure and parameters to identify and remove the noise in the image. After being processed by the denoising network, the image processing module 202 can obtain an initial denoised image. The initial denoised image is the result of preliminary denoising, which greatly reduces the noise in the image and provides a basis for subsequent further processing and optimization.

[0161] The frequency band layering module 203 is used to convert the noise image and the initial denoised image into the frequency domain and perform layered extraction according to the Fourier transform principle to obtain the first low frequency band, the first intermediate frequency band and the first high frequency band of the noise image, and the second low frequency band, the second intermediate frequency band and the second high frequency band of the initial denoised image.

[0162] In this embodiment, the frequency band layering module 203 processes the noise image and the initial denoised image according to the Fourier transform principle. First, the noise image and the initial denoised image are converted to the frequency domain, in which the information of the image is represented by components of different frequencies. Through this conversion, the characteristics of different frequency components in the image can be analyzed more clearly. Then, a layered extraction operation is performed. This frequency band layering method helps to process different frequency components in the image more specifically.

[0163] The mask generation module 204 is used to generate a low-frequency mixing mask, a medium-frequency mixing mask and a high-frequency mixing mask respectively through a mask generation network, wherein the low-frequency mixing mask corresponds to the first low-frequency band and the second low-frequency band, the medium-frequency mixing mask corresponds to the first medium-frequency band and the second medium-frequency band, and the high-frequency mixing mask corresponds to the first high-frequency band and the second high-frequency band.

[0164] In this embodiment, the mask generation module 204 generates low-frequency mixed masks, medium-frequency mixed masks, and high-frequency mixed masks respectively through a mask generation network. These mixed masks are used to screen and fuse the various frequency bands of the noisy image and the initial denoised image, so as to retain the image details as much as possible while removing the noise. By generating these mixed masks of different frequencies, the mask generation module 204 provides important guidance and basis for subsequent image fusion and optimization, which helps to more accurately perform image denoising and detail restoration.

[0165] The frequency band mixing module 205 is used to obtain a mixed low frequency band, a mixed middle frequency band and a mixed high frequency band through a frequency mixing module based on the low frequency mixing mask, the middle frequency mixing mask and the high frequency mixing mask.

[0166] In this embodiment, the frequency band mixing module 205 determines how to mix the corresponding frequency bands of the noise image and the initial denoised image according to the mixing mask. If the value of the mixing mask at a certain position is biased towards the noise image part, then mixing the low frequency band at this position will retain more relevant information of the noise image; conversely, if the mixing mask is biased towards the initial denoised image part, the mixing frequency band will adopt more relevant information of the initial denoised image. The frequency band mixing module 205 can perform fine mixing processing on different frequency bands in a targeted manner, fully combining the advantages of the noise image and the initial denoised image, and laying the foundation for finally generating a high-quality denoised image.

[0167] The image conversion module 206 is used to reassemble the mixed low-frequency band, the mixed middle-frequency band and the mixed high-frequency band, and obtain a mixed noise image through inverse Fourier transformation.

[0168] In this embodiment, the main task of the image conversion module 206 is to recombine the mixed low-frequency band, mixed middle-frequency band and mixed high-frequency band obtained by the frequency band mixing module, and integrate the frequency band information of these three different frequency ranges according to specific rules and algorithms. Then, the recombined image is converted from the frequency domain back to the spatial domain through inverse Fourier transform, so that the recombined mixed noise image can be intuitively understood and observed.

[0169] The image output module 207 is used to input the mixed noise image into the denoising network to obtain the final denoised image.

[0170] In this embodiment, the image output module 207 receives the mixed noise image output from the image conversion module 206, and the denoising network further analyzes and processes the input mixed noise image. During the processing, the denoising network can identify the noise part in the mixed noise image and try to remove the noise while retaining the details and structural information of the image as much as possible.

[0171] The function construction module 208 is used to determine a first loss function based on the initial denoised image, the final denoised image and the noise-free image.

[0172] In this embodiment, the initial denoised image is the result obtained by preliminary processing of the denoising network, which reduces the noise to a certain extent, and the final denoised image is the output after the entire denoising process, and it is expected to be as close to the noise-free image as possible. The noise-free image is used as a standard reference, representing the true state of the image without noise interference. The function construction module 208 constructs a first loss function, which can be used to guide the optimization of the denoising network and the entire denoising system, so that the denoised image outputted in the end is closer to the noise-free image.

[0173] The training module 209 is used to train the denoising network through the first loss function to obtain an optimized denoising network.

[0174] In this embodiment, the training module 209 uses the first loss function determined by the function construction module 208 in the image denoising system to train the denoising network to obtain an optimized denoising network. During the training process, the training module 209 calculates the difference between the final denoised image and the noise-free image according to the first loss function. If the loss function value is large, it means that the denoising effect is not good. At this time, the training module will adjust the parameters of the denoising network according to a specific optimization algorithm, so that the loss function value gradually decreases until the loss function value reaches a relatively ideal state or meets the preset stop condition.

[0175] The final output module 210 is used to obtain the image to be denoised and input it into the optimized denoising network to output the target image.

[0176] In this embodiment, after the denoising network is optimized by the training module 209 to obtain an optimized denoising network, the final output module 210 begins to function. At this time, the image to be denoised is input into the optimized denoising network to output a clean, noise-free target image required by the user. The final output module 210 ensures that the denoising network can apply the optimized denoising capability to the actual image to be denoised, and provide the user with a high-quality denoised image output.

[0177] In the embodiment of the present application, a training data set is first obtained. The noisy images and noise-free images in the training data set provide a reliable basis for the training and accurate evaluation of the image denoising network; the noisy image is input into the denoising network to predict the initial denoised image; then, according to the Fourier transform principle, the noisy image and the initial denoised image are converted to the frequency domain and hierarchically extracted, and each frequency band is extracted by hierarchical Fourier transform to improve the denoising effect and the ability to retain image details; a low-frequency mixed mask, a medium-frequency mixed mask and a high-frequency mixed mask are respectively generated by a mask generation network, and a mask generation network is introduced to generate a mixed mask for each frequency band, which can more finely control the noise of each frequency band; each frequency band is subjected to a mixed mask according to the frequency band. The frequency bands are mixed and recombined, and a mixed noise image is obtained by inverse Fourier transform, so that the generated noise image is more consistent with the noise distribution in the actual scene; the mixed noise image is input into the denoising network to obtain the final denoised image, and the first loss function is determined based on the initial denoised image, the final denoised image and the noise-free image. The denoising network is trained according to the first loss function and the optimized denoising network is obtained. In the training process, a training module of adversarial hybrid technology is introduced to generate a new noise image similar to the real-world noise distribution in the frequency domain, and enhance the robustness of any denoising network to unknown noise types; finally, any image to be denoised is input into the optimized denoising network to provide users with a high-quality target image. The present application can effectively solve the problem of insufficient generalization ability of existing image denoising networks when facing unknown real-world noise distributions.

[0178] Optionally, an embodiment of the third aspect of the present application provides an electronic device, including a processor 301, a memory 302, and a computer program 303 stored in the memory and executable on the processor. When the computer program is executed by the processor 301, each process of the above-mentioned image denoising method embodiment based on adversarial frequency mixing technology is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0179] An embodiment of the fourth aspect of the present application provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, each process of the above-mentioned image denoising method embodiment based on adversarial frequency mixing technology is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0180] The processor is a processor in the electronic device described in the above embodiment. The readable storage medium includes a computer readable storage medium, such as a computer read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0181] It should be noted that, in this article, the terms "comprise", "include" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise one..." do not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0182] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0183] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. An image denoising method based on adversarial frequency mixing technology, characterized in that: The method comprises: Acquire a training data set, wherein the training data set includes a noisy image and a corresponding noise-free image; Inputting the noisy image into a denoising network to predict an initial denoised image; According to the Fourier transform principle, the noise image and the initial denoised image are converted into the frequency domain and hierarchically extracted to obtain a first low-frequency band, a first intermediate-frequency band and a first high-frequency band of the noise image, and a second low-frequency band, a second intermediate-frequency band and a second high-frequency band of the initial denoised image respectively; Generate a low-frequency mixing mask, a medium-frequency mixing mask, and a high-frequency mixing mask respectively through a mask generation network, wherein the low-frequency mixing mask corresponds to the first low-frequency band and the second low-frequency band, the medium-frequency mixing mask corresponds to the first medium-frequency band and the second medium-frequency band, and the high-frequency mixing mask corresponds to the first high-frequency band and the second high-frequency band; Generate a mixed mask according to an element-by-element mask generation network; the element-by-element mask generation network uses pixel segmentation technology to generate a corresponding mixed mask for each frequency domain element of the noisy image and the initial denoised image; based on a U-Net network, encode and decode the mixed mask of each frequency domain element to generate the low-frequency mixed mask, the intermediate-frequency mixed mask and the high-frequency mixed mask; Generate a mixed mask according to a frequency band mask generation network, the frequency band mask generation network includes at least one convolution layer and a fully connected layer; extract feature information of input data through the convolution layer; based on the feature information, the fully connected layer generates a set of mixed masks through nonlinear transformation, wherein each mask in the mixed mask corresponds to an element in a frequency band area; output the low-frequency mixed mask, the medium-frequency mixed mask and the high-frequency mixed mask according to different frequency ranges; Based on the low-frequency mixing mask, the intermediate-frequency mixing mask, and the high-frequency mixing mask, respectively obtaining a mixed low-frequency band, a mixed intermediate-frequency band, and a mixed high-frequency band through a frequency mixing module; An AFM module is constructed, wherein the AFM module is used to generate the mixed low-frequency band, the mixed medium-frequency band and the mixed high-frequency band; wherein the AFM module at least includes the mask generation network and the frequency mixing module; the mask generation network includes at least one deep learning model, which is used to automatically generate the low-frequency mixing mask, the medium-frequency mixing mask and the high-frequency mixing mask according to the characteristics of the input image; the frequency mixing module mixes the frequency components of the image according to the low-frequency mixing mask, the medium-frequency mixing mask and the high-frequency mixing mask to obtain the mixed low-frequency band, the mixed medium-frequency band and the mixed high-frequency band; Recombining the mixed low-frequency band, the mixed mid-frequency band and the mixed high-frequency band, and obtaining a mixed noise image through inverse Fourier transformation; Inputting the mixed noise image into the denoising network to obtain a final denoised image; Based on the initial denoised image, the final denoised image and the noise-free image, a first loss function is determined, where the first loss function is: in, represents the first loss function for optimizing the denoising network, represents the peak signal-to-noise ratio, represents the initial denoised image, represents the final denoised image, Indicates balance and The hyperparameters of represents a noise-free image; Training the denoising network by using the first loss function to obtain an optimized denoising network; The image to be denoised is obtained and input into the optimized denoising network, and the target image is output.

2. The image denoising method based on adversarial frequency mixing technology according to claim 1, characterized in that: The image denoising method further comprises: A complementary mixing mask is defined, wherein the complementary value of the complementary mixing mask and the mixing mask of the corresponding frequency is 1; Based on the low-frequency complementary mixing mask, the middle-frequency complementary mixing mask and the high-frequency complementary mixing mask, a complementary mixed low-frequency band, a complementary mixed middle-frequency band and a complementary mixed high-frequency band are obtained respectively; Recombining the complementary mixed low frequency band, the complementary mixed middle frequency band and the complementary mixed high frequency band, and obtaining a complementary mixed noise image through inverse Fourier transformation; Inputting the complementary mixed noise image into a denoising network to obtain a final complementary denoised image; Determining a second loss function based on the final complementary denoised image, the final denoised image and the noise-free image; The mask generation network is adversarially trained using the second loss function to update the AFM module.

3. The image denoising method based on adversarial frequency mixing technology according to claim 2, characterized in that: In the step of determining a second loss function based on the final complementary denoised image, the final denoised image and the noise-free image, the second loss function is: in, Represents the second loss function updated for the AFM module, represents the peak signal-to-noise ratio, represents the final denoised image, represents the final complementary denoised image, Indicates balance and The hyperparameters of represents a noise-free image.

4. An image denoising system based on adversarial frequency mixing technology, characterized in that: The system comprises: Constructing a data set module to obtain a training data set, wherein the training data set includes a noisy image and a corresponding noise-free image; An image processing module, used for inputting the noise image into a denoising network to predict an initial denoised image; A frequency band layering module, for converting the noise image and the initial denoised image into a frequency domain and performing layered extraction according to the Fourier transform principle, so as to respectively obtain a first low frequency band, a first intermediate frequency band and a first high frequency band of the noise image, and a second low frequency band, a second intermediate frequency band and a second high frequency band of the initial denoised image; A mask generation module is used to generate a low-frequency mixed mask, a medium-frequency mixed mask and a high-frequency mixed mask respectively through a mask generation network, wherein the low-frequency mixed mask corresponds to the first low-frequency band and the second low-frequency band, the medium-frequency mixed mask corresponds to the first medium-frequency band and the second medium-frequency band, and the high-frequency mixed mask corresponds to the first high-frequency band and the second high-frequency band; a mixed mask is generated according to an element-by-element mask generation network; the element-by-element mask generation network adopts pixel segmentation technology to generate a corresponding mixed mask for each frequency domain element of the noisy image and the initial denoised image; based on a U-Net network, The mixed mask of each frequency domain element is encoded and decoded to generate the low-frequency mixed mask, the intermediate-frequency mixed mask and the high-frequency mixed mask; the mixed mask is generated according to the frequency band mask generation network, and the frequency band mask generation network includes at least one convolution layer and a fully connected layer; the feature information of the input data is extracted through the convolution layer; based on the feature information, the fully connected layer generates a set of mixed masks through nonlinear transformation, wherein each mask in the mixed mask corresponds to an element in a frequency band area; according to different frequency ranges, the low-frequency mixed mask, the intermediate-frequency mixed mask and the high-frequency mixed mask are output; A frequency band mixing module is used to obtain a mixed low frequency band, a mixed medium frequency band and a mixed high frequency band respectively through a frequency mixing module based on the low frequency mixing mask, the medium frequency mixing mask and the high frequency mixing mask; an AFM module is built, and the AFM module is used to generate the mixed low frequency band, the mixed medium frequency band and the mixed high frequency band; wherein the AFM module at least includes the mask generation network and the frequency mixing module; the mask generation network includes at least one deep learning model, which is used to automatically generate the low frequency mixing mask, the medium frequency mixing mask and the high frequency mixing mask according to the characteristics of the input image; the frequency mixing module mixes the frequency components of the image according to the low frequency mixing mask, the medium frequency mixing mask and the high frequency mixing mask to obtain the mixed low frequency band, the mixed medium frequency band and the mixed high frequency band; An image conversion module, used for recombining the mixed low-frequency band, the mixed middle-frequency band and the mixed high-frequency band, and obtaining a mixed noise image through inverse Fourier transformation; An image output module, used for inputting the mixed noise image into the denoising network to obtain a final denoised image; A function construction module is used to determine a first loss function based on the initial denoised image, the final denoised image and the noise-free image, wherein the first loss function is: in, represents the first loss function for optimizing the denoising network, represents the peak signal-to-noise ratio, represents the initial denoised image, represents the final denoised image, Indicates balance and The hyperparameters of represents a noise-free image; A training module, used for training the denoising network through the first loss function to obtain an optimized denoising network; The final output module is used to obtain the image to be denoised and input it into the optimized denoising network to output the target image.

5. An electronic device, characterized in that: It includes a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein when the program or instruction is executed by the processor, the steps of the image denoising method based on adversarial frequency mixing technology as described in any one of claims 1 to 3 are implemented.

6. A readable storage medium, characterized in that: The readable storage medium stores a program or instruction, and when the program or instruction is executed by the processor, the steps of the image denoising method based on adversarial frequency mixing technology as described in any one of claims 1 to 3 are implemented.

Images