Cascade neural network-based adjustable image denoising method and system

Abstract

The invention provides an adjustable and controllable image denoising method and system based on a cascade neural network, and particularly relates to the technical field of image processing, and the scheme comprises the steps: collecting a plurality of to-be-processed images; and based on preset noise reduction intensity, performing noise reduction processing on each to-be-processed image by using the trained cascade noise reduction network to obtain a noise reduction image corresponding to each to-be-processed image. According to the scheme, the noise reduction intensity of the to-be-processed image can be dynamically adjusted so as to adapt to the images with different noise levels and characteristics, and the personalized requirement of the cascaded noise reduction network for noise reduction processing of the image can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to an adjustable image denoising method and system based on a cascade neural network. Background Art

[0002] Image denoising technology has a wide range of applications in the fields of image processing and computer vision, especially in the field of medical imaging. Images are often affected by various noises during acquisition, including Gaussian noise, Poisson noise, salt and pepper noise, etc. These noises will cause image quality to deteriorate, thus affecting subsequent image analysis, interpretation and use. Therefore, how to effectively remove noise and improve image quality has become an important research direction in the field of image processing.

[0003] Traditional image denoising methods can be roughly divided into two categories: filter-based methods and transform-domain-based methods. Filter-based methods such as mean filtering, median filtering, and Wiener filtering remove noise through local smoothing, but often lead to the loss of image edges and details. Transform-domain-based methods, such as wavelet transform and Fourier transform, attempt to separate noise and useful signals in the frequency domain, but also face the trade-off between denoising and detail protection.

[0004] In recent years, denoising methods based on deep learning, such as the Denoising Convolutional Neural Network (DnCNN) and the very deep Residual Encoder-Decoder Networks (RED-Net), have used denoising neural network models to learn the mapping relationship between large-scale noisy images and clear images. Although there have been some improvements in detail preservation, most of the existing denoising neural network models are based on fixed network architectures and hyperparameters, lacking flexible adjustment for different image scenes or noise types. For example, in some low-noise scenes, overly strong denoising processing will lead to excessive smoothing of details, while in high-noise scenes, insufficient denoising strength may retain too much noise. It can be seen that the image denoising methods in the prior art cannot adjust the denoising strength according to the specific scene requirements, resulting in the inability to adapt to diverse image denoising tasks. Summary of the invention

[0005] In view of the above-mentioned deficiencies in the prior art, the object of the present invention is to provide an adjustable image denoising method and system based on a cascade neural network, aiming to solve the problem that the image denoising method in the prior art cannot adjust the denoising intensity according to the specific scene requirements, resulting in the inability to adapt to a variety of image denoising tasks.

[0006] In order to achieve the above object, the first aspect of the present invention provides an adjustable image denoising method based on a cascade neural network, comprising:

[0007] Collecting a number of images to be processed;

[0008] Based on the preset denoising strength, each of the images to be processed is denoised using a trained cascade denoising network to obtain a denoised image corresponding to each of the images to be processed.

[0009] Optionally, based on the preset denoising strength, using a trained cascade denoising network to perform denoising on each of the to-be-processed images respectively, and outputting a denoised image corresponding to each of the to-be-processed images, comprises:

[0010] Based on a preset noise reduction strength, determining a noise reduction level corresponding to each of the to-be-processed images;

[0011] Each of the images to be processed and the corresponding noise reduction level are respectively input into the cascade noise reduction network for denoising processing to obtain a noise reduction image corresponding to each of the images to be processed.

[0012] Optionally, the cascade denoising network is composed of a plurality of deep neural networks and data consistency modules stacked alternately, and each of the images to be processed and the corresponding denoising level are respectively input into the cascade denoising network for denoising, and a denoised image corresponding to each of the images to be processed is output, including:

[0013] Based on the noise reduction level corresponding to each of the images to be processed, using the deep neural network to perform noise reduction processing on the corresponding images to be processed respectively, so as to obtain an initial noise reduction image corresponding to each of the images to be processed;

[0014] The data consistency module is used to adaptively calibrate the initial denoised image to obtain an initial denoised image corresponding to each of the images to be processed, and the initial denoised image is input into the next set of deep neural networks and data consistency modules for iterative optimization to obtain a denoised image corresponding to each of the images to be processed.

[0015] Optionally, it is characterized in that the training process of the cascade denoising network includes:

[0016] Collect several pairs of noise-free images and noisy images to construct a training set;

[0017] Based on the deep neural network, construct the initial cascade denoising network;

[0018] Based on the noise-free images in the training set, estimating the noise intensity of the corresponding noisy images;

[0019] Based on the noise intensity, the initial cascade denoising network is iteratively trained using the training set until the denoising loss meets a preset loss threshold, thereby obtaining a trained cascade denoising network.

[0020] Optionally, estimating the noise intensity of the corresponding noisy image based on the noise-free image in the training set includes:

[0021] Collecting pixel values ​​at each pixel position of each pair of noise-free images and the corresponding noisy images;

[0022] Obtaining a variance of the noise-free image based on variances of all the pixel values ​​of the noise-free image;

[0023] Obtaining a residual image based on the deviation of the noise-free image and the corresponding noisy image at each pixel position;

[0024] Obtaining a variance of the residual image based on variances of all pixel values ​​of the residual image;

[0025] Based on the variance of the noise-free image and the variance of the residual image, the noise intensity of the corresponding noisy image is estimated.

[0026] Optionally, the iterative training of the initial cascade denoising network using the training set based on the noise intensity until the denoising loss meets a preset loss threshold to obtain a trained cascade denoising network includes:

[0027] Based on the noise intensity, performing denoising processing on the initial cascade denoising network using the training set to obtain a current denoised image of each noisy image in the training set;

[0028] Based on the noise-free images in the training set, estimating the noise intensity of the corresponding current denoised image;

[0029] The above steps of denoising and estimating noise intensity are repeated until the denoising loss meets a preset loss threshold, thereby obtaining a trained cascade denoising network.

[0030] Optionally, the calculation process of the noise reduction loss includes:

[0031] The difference between the current denoised image and the corresponding noise-free image is calculated using a preset image difference measurement loss function to obtain a denoising loss.

[0032] A second aspect of the present invention provides an adjustable image denoising system based on a cascade neural network, the system comprising:

[0033] A data acquisition module, used for acquiring a number of images to be processed;

[0034] The denoising processing module is used to perform denoising processing on the image to be processed based on a preset denoising intensity using a trained cascade denoising network to obtain a denoised image corresponding to each image to be processed.

[0035] A third aspect of the present invention provides an intelligent terminal, which includes a memory, a processor, and an adjustable image denoising program based on a cascade neural network stored in the memory and executable on the processor. When the adjustable image denoising program based on a cascade neural network is executed by the processor, the adjustable image denoising program based on a cascade neural network implements any step of the above-mentioned adjustable image denoising method based on a cascade neural network.

[0036] A fourth aspect of the present invention provides a computer-readable storage medium, on which is stored an adjustable image denoising program based on a cascade neural network. When the adjustable image denoising program based on a cascade neural network is executed by a processor, the adjustable image denoising program based on a cascade neural network implements any step of the above-mentioned adjustable image denoising method based on a cascade neural network.

[0037] Compared with the prior art, the beneficial effects of this solution are as follows:

[0038] This method proposes a new adjustable cascade denoising network framework, which allows users to dynamically adjust the denoising strength according to specific needs to adapt to the denoising needs of different images, and divides the image denoising process into multiple stages, each stage focusing on processing different levels of noise. This hierarchical processing method can ensure that the denoised image is more visually natural and consistent while retaining important information and details in the image, thereby improving the denoising effect. At the same time, the data consistency module included in the cascade denoising network uses an additional deep neural network to adaptively calibrate the output image during the denoising process, thereby preventing the denoised image content from deviating too much from the noisy image, thereby ensuring the stability of image denoising. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0040] Figure 1 A brief flow chart of the adjustable image denoising method based on cascade neural network of the present invention;

[0041] Figure 2 It is a flowchart of the adjustable image denoising process based on cascade neural network of the present invention;

[0042] Figure 3It is a flow chart of cascade denoising network training of the present invention;

[0043] Figure 4 It is a structural diagram of the cascade denoising network of the present invention;

[0044] Figure 5 It is a schematic diagram of the module of the adjustable image denoising system based on cascade neural network of the present invention;

[0045] Figure 6 It is a schematic diagram of the structure of the intelligent terminal of the present invention. DETAILED DESCRIPTION

[0046] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present invention. However, it should be clear to those skilled in the art that the present invention may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present invention.

[0047] It should be understood that when used in this specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof.

[0048] It should also be understood that the terms used in the present specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0049] It should be further understood that the term "and / or" used in the present description and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0050] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0052] In view of the problem that the image denoising methods in the prior art cannot adjust the denoising intensity according to the requirements of specific scenarios, resulting in the inability to adapt to diverse image denoising tasks, the present invention proposes a concept of adjustable image denoising based on a cascade neural network, mainly proposing a new adjustable cascade denoising network framework, which allows users to dynamically adjust the denoising intensity according to specific needs, and divides the image denoising process into multiple stages, each stage focusing on processing different levels of noise. This layered processing method can effectively reduce the loss of details caused by one-time denoising and improve the denoising effect. The flexible control mechanism can effectively respond to the needs of different medical scenarios and improve the usability and adaptability of the algorithm in practical applications. At the same time, the data consistency module contained in the cascade denoising network framework uses an additional deep neural network to adaptively calibrate the output image during the denoising process, thereby preventing the denoised image content from deviating too much from the noisy image, thereby ensuring the stability of image denoising.

[0053] The embodiment of the present invention provides an adjustable image denoising method based on a cascade neural network, which is deployed on electronic devices such as computers and servers and is applied to the scene of image denoising, and is aimed at the situation where the intensity of image denoising can be flexibly adjusted. The types of the above images are not limited, including but not limited to CT images, magnetic resonance images, ultrasound images, X-ray images and other medical images. Specifically, Figure 1 and Figure 2 As shown, the steps of the method in this embodiment include:

[0054] Step S100: collecting a number of images to be processed;

[0055] Specifically, a number of images to be processed containing noise are collected. These images to be processed may be images taken in the same scene containing the same or different levels of noise, or may be images taken in different scenes containing different types of noise. This embodiment does not impose any specific limitation on this.

[0056] Step S200: Based on the preset denoising strength, each of the to-be-processed images is denoised using a trained cascade denoising network to obtain a denoised image corresponding to each of the to-be-processed images.

[0057] Specifically, in order to achieve flexible adjustment of the denoising intensity of the image to be processed, the present embodiment sets the same or different denoising intensity for the image to be processed according to the specific denoising requirements of the actual application scenario, and then sequentially inputs the images to be processed with the same denoising intensity into the trained cascade denoising network, and uses the trained cascade denoising network to perform denoising on each image to be processed, and outputs a denoised image at the denoising intensity corresponding to each image to be processed; similarly, the above denoising steps are performed on the images to be processed corresponding to each denoising intensity, respectively, using the trained cascade denoising network, until a denoised image corresponding to each of the images to be processed is obtained.

[0058] Furthermore, if the denoising effect of the denoised image does not meet expectations, the denoising intensity is re-set for the denoised image, and the denoised image is regarded as an image to be processed, and the same steps as the above denoising process are performed until a denoised image with the expected denoising effect is obtained.

[0059] This embodiment can dynamically adjust the denoising intensity of the processed image according to the specific denoising requirements of the actual application scenario, and the cascade denoising network designed in this embodiment can perform denoising intensity processing on the processed image according to the preset denoising intensity, thereby realizing flexible denoising requirements for the processed image to meet the image clarity requirements in different scenarios, so that the usability and adaptability of the model in practical applications are significantly improved.

[0060] In a preferred embodiment, the step S200 uses a trained cascade denoising network to perform denoising on each of the images to be processed based on a preset denoising strength, and outputs a denoised image corresponding to each of the images to be processed, including:

[0061] Step S210: determining a noise reduction level corresponding to each of the to-be-processed images based on a preset noise reduction strength;

[0062] Step S220: inputting each of the images to be processed and the corresponding noise reduction levels into the cascade noise reduction network for denoising, to obtain a noise reduction image corresponding to each of the images to be processed.

[0063] Specifically, firstly, based on factors such as the noise level of the image and the complexity of the image content, several different noise reduction levels are set, and each noise reduction level corresponds to a different noise reduction strength; the specific division form of the noise reduction level can be by dividing the numerical range corresponding to the continuous noise reduction strength into multiple intervals, and each interval is represented as a noise reduction level, or several discrete noise reduction strengths can be set, and each noise reduction strength is used to represent a noise reduction level; the specific expression form of the noise reduction level can be represented by level one, level two, ..., etc., or by low level, medium level, high level, high level, etc. For each image to be processed, the noise reduction demand is comprehensively evaluated according to the noise level of the picture, the image content, the shooting environment conditions, the circuit structure and the performance of the electronic components, and the preset noise reduction strength standard, and the appropriate noise reduction level is selected for each image to be processed accordingly. Then, the image to be processed and its corresponding denoising level are input into a cascade denoising network. The cascade denoising network is composed of a plurality of denoising network units in cascade. Each denoising network unit denoises the image according to different noise characteristics or image details to obtain a corresponding denoised image and transmits it to the next denoising network unit for iterative denoising. Each denoising network unit is a deep neural network, which is trained according to the corresponding noise type or image feature to identify and remove different types of noise in the image while retaining important information and details in the image. Finally, after all the denoising network units are processed, the denoised image corresponding to each image to be processed is obtained. The denoised image that meets the expectations is significantly reduced in noise level while retaining the main content and details of the image.

[0064] The method of this embodiment is not only flexible and efficient, but also can adapt to images with different noise levels and characteristics, which is conducive to improving the usability and adaptability of applying the cascade denoising network to image denoising processing.

[0065] In a preferred embodiment, the cascade denoising network is composed of a plurality of deep neural networks and data consistency modules stacked alternately. In step S220, each of the images to be processed and the corresponding denoising level are respectively input into the cascade denoising network for denoising, and a denoised image corresponding to each of the images to be processed is output, including:

[0066] Step S221: Based on the noise reduction level corresponding to each of the images to be processed, the deep neural network is used to perform noise reduction processing on the corresponding images to be processed respectively, so as to obtain an initial noise reduction image corresponding to each of the images to be processed;

[0067] Step S222: Use the data consistency module to adaptively calibrate the initial denoised image to obtain an initial denoised image corresponding to each of the images to be processed, and input the initial denoised image into the next set of deep neural networks and data consistency modules for iterative optimization to obtain a denoised image corresponding to each of the images to be processed.

[0068] Specifically, the cascade denoising network in this embodiment is composed of several deep neural networks and data consistency modules stacked alternately, and each denoising network unit includes a group of deep neural networks and data consistency modules connected in series. The specific process of denoising processing is described below for one or a batch of images to be processed with the same denoising level, including: first, denoising processing is performed through the first denoising network unit, and based on the denoising level corresponding to each image to be processed, the corresponding images to be processed are denoised using a deep neural network respectively, so as to learn the mapping of extracting clean images from noisy images for specific noise types or image features, thereby obtaining an initial denoised image corresponding to each image to be processed; and using the data consistency module, an additional deep neural network is used to perform pixel-level adaptive calibration on the initial denoised image to obtain an initial denoised image corresponding to each image to be processed, so as to prevent the denoised image content from deviating too much from the noisy image, and ensure that the denoised image maintains consistency and visual quality with the original image in statistics or structure, thereby ensuring the stability of image denoising. Then, the initial denoised image is input into the next set of deep neural networks and data consistency modules for iterative optimization to obtain the denoised image corresponding to each image to be processed. After multiple rounds of denoising and data consistency calibration, the denoising process is gradually refined and regulated, which can significantly improve the denoising effect of the image to be processed and effectively reduce the detail loss caused by one-time denoising, thereby improving the quality of the denoised image.

[0069] It is easy to understand that for the images to be processed with different noise reduction levels, the same specific noise reduction process described above is used to perform noise reduction with different noise reduction intensities. The only difference is that the parameter values ​​set for the cascade noise reduction network are different, so as to dynamically adjust the noise reduction intensity to meet the noise reduction needs of different images, and ensure that the noise-reduced image is more natural and consistent visually while retaining important information and details in the image. At the same time, this flexible control mechanism can realize the dynamic adjustment of the noise reduction intensity, effectively meet the needs of different image noise reduction scenarios, and help improve the usability and adaptability of the algorithm in practical applications.

[0070] like Figure 3 and Figure 4 As shown, in a preferred embodiment, the training process of the cascade denoising network in step S200 includes:

[0071] Step M100: collect several pairs of noise-free images and noisy images to construct a training set;

[0072] Specifically, a sufficient amount of paired training image data is collected, each pair of training data includes a noise-free image and a noisy image, and a training set is constructed. The noise-free image can be obtained by filtering the noisy image, denoising it using a graph-based regularizer or a learning-based method; the noisy image can be obtained by artificially adding noise (such as Gaussian noise, Poisson noise, salt and pepper noise, etc.) to the noise-free image, or obtained from an actual scene. Furthermore, the training set can also be preprocessed, such as removing useless information, enhancing data diversity, resizing, normalizing, etc., to ensure the consistency of the input data.

[0073] Step M200: constructing an initial cascade noise reduction network based on a deep neural network;

[0074] Specifically, based on the deep neural network, multiple cascaded denoising network units are designed to form an initial cascade denoising network. Each unit contains a deep neural network (such as a convolutional neural network CNN, a recurrent neural network RNN ​​or its variants LSTM, GRU, etc.) and a data consistency module. Each denoising network unit is responsible for processing noise in different frequency bands or time periods and gradually reducing the impact of noise.

[0075] Step M300: estimating the noise intensity of the corresponding noisy image based on the noise-free image in the training set;

[0076] Specifically, the noise intensity is estimated by analyzing the noisy image itself, for example, by using the statistical characteristics of the image (such as mean, variance, histogram, etc.); or, by using image block-based analysis, the noise intensity is estimated by comparing the differences between image blocks. When the noise intensity value is large, the cascade denoising network model will increase the denoising intensity to more actively suppress the noise; when the noise intensity value is low, the cascade denoising network model will weaken the denoising intensity to avoid excessive processing of image details. This embodiment allows the user to dynamically adjust the denoising intensity according to the noise characteristics and needs of the specific image, so that the model has the flexibility of denoising, can adapt to different types of images, and provide more personalized processing.

[0077] If there are enough pairs of noise-free and noisy images, a supervised learning method can be used to train a noise intensity estimation model, such as linear regression, decision tree, or deep learning model. During the training process, the noise-free image and the corresponding noisy image are used as input, and the noise intensity is trained as output. If there are only a limited number of noise-free and noisy image pairs, semi-supervised or unsupervised methods can be used to estimate the noise intensity based on prior knowledge of the image, statistical characteristics of the image, or similarity measures of the image.

[0078] Step M400: Based on the noise intensity, the initial cascade denoising network is iteratively trained using the training set until the denoising loss meets a preset loss threshold, thereby obtaining a trained cascade denoising network.

[0079] Specifically, this embodiment is based on the noise-free image in the training set and its corresponding noisy image, as well as the noise intensity estimation value of the noisy image. For each denoising network unit, an unsupervised learning method, such as autoencoder, clustering and other methods, is used for pre-training to enable it to have basic denoising capabilities; on the basis of pre-training, a supervised learning method is used to calculate the loss by comparing the denoised output and the clean signal, and the network parameters are optimized to jointly train the entire cascade denoising network; further, according to the needs of the actual application scenario, data enhancement, regularization and other technologies are used to improve the generalization ability of the network, and the network is fine-tuned to further improve the denoising effect. During the entire training process, a preset denoising loss function is used to evaluate the denoising effect, and the network parameters (such as learning rate, batch size, number of iterations, etc.) are optimized through the back propagation algorithm until the denoising loss drops below the preset loss threshold or reaches the preset number of iterations, and the weights and parameters of the trained cascade denoising network are saved. The denoising loss function is used to measure the difference between the denoised image and the noise-free image. Any image difference measurement loss function can be selected, such as the mean square error loss function or the SSIM loss function, etc. This embodiment does not impose any specific restrictions on this. Furthermore, in order to prevent overfitting, a regularization term can be added to the loss function, such as L1 regularization, L2 regularization, etc.

[0080] A new adjustable cascade denoising network framework constructed in this embodiment allows users to dynamically adjust the denoising strength according to specific needs, and divides the image denoising process into multiple stages, each stage focusing on processing different levels of noise. This layered processing method can effectively reduce the detail loss caused by one-time denoising and improve the denoising effect.

[0081] In a preferred embodiment, the step M300 of estimating the noise intensity of the corresponding noisy image based on the noise-free image in the training set includes:

[0082] The pixel value of each pixel position of each pair of noise-free images and the corresponding noisy image is collected; the variance of the noise-free image is obtained based on the variance of all pixel values ​​of the noise-free image; the residual image is obtained based on the deviation of the noise-free image and the corresponding noisy image at each pixel position; the variance of the residual image is obtained based on the variance of all pixel values ​​of the residual image; the noise intensity of the corresponding noisy image is estimated based on the variance of the noise-free image and the variance of the residual image.

[0083] For example, the noise level estimation module is used to obtain the noise intensity of an image by comparing the difference between a noisy image and a noise-free image. The calculation process is as follows:

[0084]

[0085] Where l is the noise intensity, P signal is the variance of the noise-free image, P noise is the variance of the residual image between the noisy image and the noise-free image. It is a statistic that measures the degree of discreteness of all pixel positions. The residual image is used to represent the difference between the noisy image and the noise-free image at each pixel position. The variance of the residual image can be obtained by calculating the mean of the residual image, traversing each pixel position, calculating the square of the difference between the residual and the mean, and then adding all square values ​​and dividing by the product of the number of rows and columns of the image. It should be noted that before calculating the residual image, it is necessary to ensure that the noisy image and the noise-free image are accurately aligned to avoid calculation errors caused by image misalignment; when calculating the variance of the residual image, it is necessary to consider the influence of the noise model and make appropriate adjustments according to the actual situation to ensure the accuracy and effectiveness of the calculation results.

[0086] like Figure 3 and Figure 4 As shown, in a preferred embodiment, in step M400, based on the noise intensity, the initial cascade denoising network is iteratively trained using the training set until the denoising loss meets a preset loss threshold, thereby obtaining a trained cascade denoising network, including:

[0087] Step M410: Based on the noise intensity, the initial cascade denoising network is subjected to denoising processing using the training set to obtain a current denoised image of each noisy image in the training set;

[0088] Step M420: estimating the noise intensity of the corresponding current denoised image based on the noise-free image in the training set;

[0089] Step M430: Repeat the above steps of denoising and estimating noise intensity until the denoising loss meets a preset loss threshold, thereby obtaining a trained cascade denoising network.

[0090] Specifically, the noise intensity and the noisy image are input into the cascade denoising network, and the cascade denoising network performs denoising processing to output the denoised image. The structure of the cascade denoising network is as follows: Figure 3 As shown in Figure 1, it contains n cascade blocks, each of which is composed of a stack of deep neural networks (DNN) and data consistency modules (DC). The calculation of the i-th cascade block is as follows. First, the output of the DNN is calculated:

[0091] z i =f θ (x i-1 ,l) (2)

[0092] Among them, x i-1 is the output of the i-1th cascade block, l is the noise intensity, f θ represents a deep neural network with parameters θ, z i is the image after being processed by the current DNN.

[0093] Then, calculate the output of DC:

[0094]

[0095] Among them, x i is the output of DC, x0 is the noisy image, Indicates that the parameter is The input is a deep neural network, z i is the image processed by the current DNN. It can be regarded as a data consistency weight, which is passed through a deep neural network Learned. Deep Neural Network It refers to a network structure that can map x0 and l into learnable weight coefficients. This embodiment does not limit the specific network structure of the deep neural network.

[0096] Finally, after iterative denoising of n cascade blocks, a denoised image with a high signal-to-noise ratio can be obtained. It is worth noting that the cascade structure of the present embodiment allows the denoising process to be divided into multiple stages, each of which focuses on different levels of noise. This step-by-step processing can more effectively remove various types of noise and reduce the loss of details caused by one-time processing. At the same time, the data consistency module proposed in this embodiment uses an additional deep neural network to adaptively calibrate the output image during the denoising process, thereby preventing the denoised image content from deviating too much from the noisy image, thereby ensuring the stability of image denoising.

[0097] In a preferred embodiment, the calculation process of the noise reduction loss in step M430 includes:

[0098] The difference between the current denoised image and the corresponding noise-free image is calculated using a preset image difference measurement loss function to obtain a denoising loss.

[0099] Specifically, the noise-free image and the denoised image are input into the loss calculation module to obtain the training loss. The loss calculation module can be expressed as:

[0100] l=L(y,x) (4)

[0101] Where l is the loss value, x represents the denoised image obtained after denoising, y represents the corresponding noise-free image, and the loss function L(·) can be any image difference metric loss function, such as the mean square error loss function or the SSIM loss function.

[0102] like Figure 5 As shown, corresponding to the above-mentioned adjustable image denoising method based on cascade neural network, an embodiment of the present invention further provides an adjustable image denoising system based on cascade neural network, and the above-mentioned adjustable image denoising system based on cascade neural network includes:

[0103] A data acquisition module 510 is used to acquire a number of images to be processed;

[0104] The denoising processing module 520 is used to perform denoising processing on the image to be processed based on a preset denoising strength by using a trained cascade denoising network to obtain a denoised image corresponding to each image to be processed.

[0105] Specifically, in this embodiment, the specific functions of the above-mentioned adjustable image denoising system based on cascade neural network can also refer to the corresponding description in the above-mentioned adjustable image denoising method based on cascade neural network, which will not be repeated here.

[0106] Based on the above embodiments, the present invention further provides an intelligent terminal, whose principle block diagram can be shown as follows: Figure 6 As shown. The above-mentioned intelligent terminal includes a processor, a memory, a network interface and a display screen connected through a system bus. Among them, the processor of the intelligent terminal is used to provide computing and control capabilities. The memory of the intelligent terminal includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and an adjustable image denoising program based on a cascade neural network. The internal memory provides an environment for the operation of the operating system and the adjustable image denoising program based on a cascade neural network in the non-volatile storage medium. The network interface of the intelligent terminal is used to communicate with an external terminal through a network connection. When the adjustable image denoising program based on a cascade neural network is executed by the processor, the steps of any one of the above-mentioned adjustable image denoising methods based on a cascade neural network are implemented. The display screen of the intelligent terminal can be a liquid crystal display screen or an electronic ink display screen.

[0107] Those skilled in the art will understand that Figure 6 The principle block diagram shown in the figure is only a block diagram of a partial structure related to the scheme of the present invention, and does not constitute a limitation on the smart terminal to which the scheme of the present invention is applied. The specific smart terminal may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0108] In one embodiment, a smart terminal is provided, which includes a memory, a processor, and an adjustable image denoising program based on a cascade neural network stored in the memory and executable on the processor. When the adjustable image denoising program based on a cascade neural network is executed by the processor, the steps of any one of the adjustable image denoising methods based on a cascade neural network provided in an embodiment of the present invention are implemented.

[0109] An embodiment of the present invention further provides a computer-readable storage medium, on which is stored an adjustable image denoising program based on a cascade neural network. When the adjustable image denoising program based on a cascade neural network is executed by a processor, the steps of any one of the adjustable image denoising methods based on a cascade neural network provided in an embodiment of the present invention are implemented.

[0110] It should be understood that the serial numbers of the steps in the above embodiments do not imply a sequence of execution, and the execution sequence of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0111] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the above-mentioned device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of the present invention. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0112] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0113] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0114] In the embodiments provided by the present invention, it should be understood that the disclosed apparatus / terminal equipment and method can be implemented in other ways. For example, the apparatus / terminal equipment embodiments described above are only illustrative, for example, the division of the above modules or units is only a logical function division, and in actual implementation, other division methods can be used, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0115] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, it should be understood by those skilled in the art that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features therein may be replaced by equivalents. However, these modifications or replacements do not deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A controllable image denoising method based on cascade neural network, characterized in that: The following steps are involved: Collecting a number of images to be processed; Based on the preset denoising strength, each of the images to be processed is denoised using a trained cascade denoising network to obtain a denoised image corresponding to each of the images to be processed.

2. The adjustable image denoising method based on cascade neural network according to claim 1 is characterized in that: The method of performing denoising on each of the images to be processed based on the preset denoising strength using the trained cascade denoising network and outputting a denoised image corresponding to each of the images to be processed includes: Based on a preset noise reduction strength, determining a noise reduction level corresponding to each of the to-be-processed images; Each of the images to be processed and the corresponding noise reduction level are respectively input into the cascade noise reduction network for denoising processing to obtain a noise reduction image corresponding to each of the images to be processed.

3. The adjustable image denoising method based on cascade neural network according to claim 2 is characterized in that: The cascade denoising network is composed of a plurality of deep neural networks and data consistency modules stacked alternately, and each of the images to be processed and the corresponding denoising level are respectively input into the cascade denoising network for denoising, and a denoised image corresponding to each of the images to be processed is output, including: Based on the noise reduction level corresponding to each of the images to be processed, using the deep neural network to perform noise reduction processing on the corresponding images to be processed respectively, so as to obtain an initial noise reduction image corresponding to each of the images to be processed; The data consistency module is used to adaptively calibrate the initial denoised image to obtain an initial denoised image corresponding to each of the images to be processed, and the initial denoised image is input into the next set of deep neural networks and data consistency modules for iterative optimization to obtain a denoised image corresponding to each of the images to be processed.

4. The adjustable image denoising method based on cascade neural network according to any one of claims 1 to 3, characterized in that: The training process of the cascade denoising network includes: Collect several pairs of noise-free images and noisy images to construct a training set; Based on the deep neural network, construct the initial cascade denoising network; Based on the noise-free images in the training set, estimating the noise intensity of the corresponding noisy images; Based on the noise intensity, the initial cascade denoising network is iteratively trained using the training set until the denoising loss meets a preset loss threshold, thereby obtaining a trained cascade denoising network.

5. The adjustable image denoising method based on cascade neural network according to claim 4 is characterized in that: The estimating the noise intensity of the corresponding noisy image based on the noise-free image in the training set includes: Collecting pixel values ​​at each pixel position of each pair of noise-free images and the corresponding noisy images; Obtaining a variance of the noise-free image based on variances of all the pixel values ​​of the noise-free image; Obtaining a residual image based on the deviation of the noise-free image and the corresponding noisy image at each pixel position; Obtaining a variance of the residual image based on variances of all pixel values ​​of the residual image; Based on the variance of the noise-free image and the variance of the residual image, the noise intensity of the corresponding noisy image is estimated.

6. The adjustable image denoising method based on cascade neural network according to claim 4, characterized in that: The method of iteratively training the initial cascade denoising network based on the noise intensity using the training set until the denoising loss meets a preset loss threshold to obtain a trained cascade denoising network includes: Based on the noise intensity, performing denoising processing on the initial cascade denoising network using the training set to obtain a current denoised image of each noisy image in the training set; Based on the noise-free images in the training set, estimating the noise intensity of the corresponding current denoised image; The above steps of denoising and estimating noise intensity are repeated until the denoising loss meets a preset loss threshold, thereby obtaining a trained cascade denoising network.

7. The adjustable image denoising method based on cascade neural network according to claim 6, characterized in that: The calculation process of the noise reduction loss includes: The difference between the current denoised image and the corresponding noise-free image is calculated using a preset image difference measurement loss function to obtain a denoising loss.

8. A controllable image denoising system based on cascade neural network, characterized in that: The system comprises: A data acquisition module, used for acquiring a number of images to be processed; The denoising processing module is used to perform denoising processing on the image to be processed based on a preset denoising intensity using a trained cascade denoising network to obtain a denoised image corresponding to each image to be processed.

9. An intelligent terminal, characterized in that: The intelligent terminal includes a memory, a processor, and an adjustable image denoising program based on a cascade neural network stored in the memory and executable on the processor. When the adjustable image denoising program based on a cascade neural network is executed by the processor, the steps of the adjustable image denoising method based on a cascade neural network as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores an adjustable image denoising program based on a cascade neural network. When the adjustable image denoising program based on a cascade neural network is executed by a processor, the steps of the adjustable image denoising method based on a cascade neural network as described in any one of claims 1 to 7 are implemented.

Images