Ground magnetic resonance multi-type noise denoising network construction system and denoising method

Abstract

The application belongs to the field of nuclear magnetic resonance sounding signal noise suppression methods, and is a ground magnetic resonance multi-type noise denoising network construction system and denoising method, which comprises a magnetic resonance signal construction module, a plurality of groups of magnetic resonance effective signals and environmental noise are simulated, and a noisy data set affected by three types of noise is constructed; a denoising neural network building module, the magnetic resonance signal generated by the magnetic resonance signal construction module is trained and optimized based on three types of noisy data sets and noise data sets, three denoising networks for different types of noise are obtained, a classification and discrimination model adopts a support vector machine method to judge the noise type according to the characteristics of different noises, and outputs the judgment result to the corresponding denoising network model in the denoising neural network building module. The application can reduce the scale of the label data amount, shorten the training time of the model, and remove the noise in a targeted manner, so that the dependence of the network model on data diversity can be reduced.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of noise suppression methods for magnetic resonance sounding (MRS) signals, and in particular to a ground magnetic resonance multi-type noise denoising network and a denoising method. BACKGROUND

[0002] The magnetic resonance groundwater detection technology can directly and quantitatively interpret hydrological information such as the position of the aquifer, the size of the water content, the medium porosity and the conductivity, and has been widely used in the fields of groundwater resource detection and water source geological disaster prediction in recent years.

[0003] Due to the very weak magnetic resonance signal, which is in the order of nanovolts, the magnetic resonance signal is often disturbed by random noise, power frequency noise, spike noise and other interference in the environment during actual detection. The noise interference seriously affects the quality of the magnetic resonance signal and further affects the subsequent interpretation of the hydrological information.

[0004] With the rise of machine learning in geophysical methods, some deep learning methods have also been applied in the field of magnetic resonance noise reduction. Patent CN109870729A discloses a "deep neural network magnetic resonance signal denoising method based on discrete cosine transform", which uses discrete cosine transform to process the data, and then establishes a mapping between the noisy signal and the effective signal based on the deep neural network to achieve the purpose of noise reduction. Patent CN111860273A discloses a "magnetic resonance groundwater detection noise suppression method based on a convolutional neural network", which drives a convolutional neural network to establish a mapping between the noisy signal and the noise by using a large amount of labeled data, continuously optimizes the network, realizes end-to-end magnetic resonance denoising, and further introduces residual learning to enable the network to better learn the signal features and recover the effective signal from the noisy signal.

[0005] The above-mentioned magnetic resonance denoising method based on neural network, when constructing the data set, the noisy signal includes interference from multiple types of noise. In order to improve the denoising performance of the model, a large amount of training data set is required, which results in a long model training time. In addition, the applicability of the network model depends on the diversity of the data set. SUMMARY

[0006] The technical problem to be solved by the present application is to provide a construction system of a ground magnetic resonance multi-type noise denoising network and a denoising method, which solves the problem of dependence on data diversity of the existing method.

[0007] The present application is implemented as follows,

[0008] A construction system of a ground magnetic resonance multi-type noise denoising network, comprising:

[0009] The magnetic resonance signal construction module simulates a plurality of groups of magnetic resonance effective signals and environmental noises including spike noises, power frequency noises and random noises according to the principle of ground magnetic resonance groundwater detection, and constructs noise-containing data sets affected by the three types of noises respectively;

[0010] The denoising neural network building module trains and optimizes networks respectively based on the three types of noise-containing data sets and noise data sets generated by the magnetic resonance signal construction module, and obtains three denoising neural networks for different types of noises, including a model Net s (θ s ) for removing spike noises, a model Net h (θ h ) for reducing power harmonic noise, and a model Net r (θ r ) for eliminating random noise.

[0011] The classification and discrimination model building module constructs a classification and discrimination model using a support vector machine method to judge the type of noise according to the characteristics of different noises, and outputs the judgment result to the corresponding denoising network model in the denoising neural network building module.

[0012] Further, the data set is constructed by the magnetic resonance signal construction module, specifically including:

[0013] a1, constructing a magnetic resonance signal, the time domain expression is as follows

[0014]

[0015] wherein, E0∈[10,10]00, f L ∈[2200,2400]Hz, T2 * ∈[0.1,1]s, t=(0:N-1) / f s , f s =25kHz, N=6250, the effective signal data x1 is obtained according to the above formula, the signal parameters are changed, and n effective signals X={x1,x2,...,x n} are obtained.

[0016] a2, according to the characteristics of different types of noises, noises including spike noises s, power frequency noises h and random noises r are constructed by using the randn function and the triangular function in MATLAB, the noise parameters are changed, and there are n groups of each type of noise data set, including a spike noise data set S={s1,s2,...,s n}, a power frequency noise data set H={h1,h2,...,h n} and a random noise data set R={r1,r2,...,r n}.

[0017] a3, superimpose the noise from the constructed noise dataset with the valid signal from the signal dataset to obtain three noisy datasets, including N. s ={n s1 ,n s2 ,...,n sn},N h ={n h1 ,n h2 ,...,n hn} and N r ={n r1 ,n r2 ,...,n rn}

[0018] Furthermore, three denoising networks targeting different types of noise are constructed using a denoising neural network construction module, including:

[0019] A denoising network is built using convolution and upsampling operations, and skip layers are introduced to connect the convolution and upsampling operations. The convolution operation is described by the following formula:

[0020]

[0021] Where * represents the convolution operation, k is the convolution kernel, b is the bias, f is the activation function, and x represents the input and output features during the convolution operation. Upsampling is described by the following formula:

[0022]

[0023] Where upsample(·) is the upsampling function;

[0024] Data set {N s ,S},{N h ,H}, and {N r R and R are used to train the denoising network, and the objective function of the network is as follows:

[0025]

[0026]

[0027]

[0028] By continuously updating the parameters θ of the network model s θ h and θ r Optimize the objective function to obtain the Net model with spike noise removed. s (θ s Net, a model for reducing power frequency harmonic noise h (θ h), and a model Net for eliminating random noise r (θ r ).

[0029] Further, the classification and discrimination module comprises a classification model construction module and a discrimination model, the classification model construction module constructs a noise classification model based on a support vector machine, and specifically comprises:

[0030] c1, decomposing the multi-type noise classification into a plurality of application support vector machines to solve the two-classification problem;

[0031] c2, constructing a data set N s ,N h ,N r , and X, recombining N s and {N h ,N r ,X} as a training sample set of the first two-classification problem for distinguishing between severe spike noise and other conditions, and the corresponding labels are +1 and -1 respectively; N h and {N s ,N r ,X} as a training sample set of the second two-classification problem for distinguishing between severe power frequency noise and other conditions, and the corresponding labels are +1 and -1 respectively; N r and {N s ,N h ,X} as a training sample set of the third two-classification problem for distinguishing between severe random noise and other conditions, and the corresponding labels are +1 and -1 respectively;

[0032] c3, the support vector machine finds a hyperplane in a high-dimensional space to separate the training samples, where w is a weight vector and b is a threshold, is an objective function, and the problem of finding an optimal hyperplane is converted into solving the following problem:

[0033]

[0034]

[0035] where C is a penalty parameter for balancing weights, and ξ is a relaxation variable, and the final decision function obtained by calculation is

[0036]

[0037] where α is an optimal Lagrange multiplier, x is a training sample, y is a label taking ±1, and K(x i ,x j ) is a kernel function of the model, and a radial basis function is selected as the kernel function of the model, which is defined as follows:

[0038]

[0039] wherein σ is bandwidth, controlling the radial range of action;

[0040] c4, three hyperplanes are found according to step c3 using the three training samples in step c2 respectively and three support vector machine-based classification models V1, V2 and V3 capable of distinguishing noise types are obtained.

[0041] Further, the discrimination module performs denoising, and the process includes:

[0042] the noisy data set is input into the classification model V1, and the discrimination module performs judgment according to the following formula:

[0043]

[0044] If the above formula is true, it indicates that the signal is seriously affected by spike noise, and the noisy signal is input into the model Net s (θ s ) for processing, otherwise the data is input into the classification model V2, and the discrimination module performs judgment again:

[0045]

[0046] If the above formula is true, it indicates that the signal is seriously affected by power frequency noise, and the noisy signal is input into the model Net h (θ h ) for processing, otherwise the data is input into the classification model V3, and the discrimination module performs judgment again:

[0047]

[0048] If the above formula is true, it indicates that the signal is seriously affected by random noise, and the noisy signal is input into the model Net r (θ r ) for processing, otherwise it indicates that the noise is weak, and the effective signal features are recognized.

[0049] A construction method of a ground magnetic resonance multi-type noise denoising network, comprising:

[0050] a. According to the principle of ground magnetic resonance groundwater detection, simulate multiple groups of magnetic resonance effective signals and environmental noise, including spike noise, power frequency noise and random noise, and construct noisy data sets affected by three types of noise respectively;

[0051] b. Construct three denoising neural networks, and train and optimize the networks based on three noisy datasets and three noisy datasets respectively, to obtain three denoising networks for different types of noise, including a model Net for removing spike noise. s (θ s Net, a model for reducing power frequency harmonic noise h (θ h ), and the Net model for eliminating random noise. r (θ r );

[0052] c. Using the support vector machine method, a classification and discrimination model is established to determine the type of noise based on the characteristics of different noises. The classification and discrimination model is connected to the denoising network. After the noisy signal is judged, the result of the noise type judgment is input into the corresponding denoising network.

[0053] Furthermore, including:

[0054] The specific steps for constructing the dataset in step a are as follows:

[0055] a1, construct the magnetic resonance signal, the time-domain expression is as follows:

[0056]

[0057] Where E0 is the initial amplitude, f L For Larmor frequency, It is the relaxation time. It is the initial phase, t = (0:N-1) / f s f s Let N be the sampling rate and N be the data length. Using the above formula, we obtain the effective signal data x. By changing the signal parameters, we obtain n effective signals X = {x1, x2, ..., xn}. n};

[0058] a2. Based on the characteristics of different types of noise, construct noise datasets using the randn function and trigonometric functions in MATLAB, including spike noise s, power frequency noise h, and random noise r. Change the noise parameters. There are n datasets for each type of noise, including the spike noise dataset S = {s1, s2, ..., s...}. n}, Power frequency noise dataset H = {h1, h2, ..., h n} and a random noise dataset R = {r1, r2, ..., r n};

[0059] a3, superimpose the noise from the constructed noise dataset with the valid signal from the signal dataset to obtain three noisy datasets, including N. s ={n s1 ,ns2 ..., n sn}, N h = {n h1 , n h2 ..., n hn} and N r = {n r1 , n r2 ..., n rn}.

[0060] Further, step b specifically comprises:

[0061] A de-noising network is built by using convolution, up-sampling and other operations, and a skip layer connection is introduced, and the convolution operation is described by the following formula:

[0062]

[0063] Wherein, * is the convolution operation, k is the convolution kernel, b is the bias, f is the activation function, and x is the input and output features in the convolution operation process, and the up-sampling is described by the following formula:

[0064]

[0065] Wherein, upsample(·) is the up-sampling function;

[0066] The data sets {N s , S}, {N h , H} and {N r , R} are used to train the de-noising network, and the objective function of the network is as follows:

[0067]

[0068]

[0069]

[0070] By continuously updating the parameters θ s , θ h and θ r of the network model, the objective function is optimized to obtain the model Net s (θ s ) for removing spike noise, the model Net h (θ h ) for reducing power harmonic noise, and the model Net r (θ r ) for eliminating random noise.

[0071] Further, the construction in step c specifically comprises:

[0072] c1, decompose the multi-type noise classification into multiple application support vector machines to solve the binary classification problem;

[0073] c2, construct the data set N s , N h , N r , and X, N s , and {N h , N r , X} as the training sample set of the first binary classification problem, used to distinguish between spike noise and other cases, corresponding labels are +1 and -1 respectively; N h , and {N s , N r , X} as the training sample set of the second binary classification problem, used to distinguish between power frequency noise and other cases, corresponding labels are +1 and -1 respectively; N r , and {N s , N h , X} as the training sample set of the third binary classification problem, used to distinguish between random noise and other cases, corresponding labels are +1 and -1 respectively;

[0074] c3, support vector machine finds a hyperplane in high-dimensional space Separate the training samples, where w is the weight vector and b is the threshold, is the objective function, and the problem of finding the optimal hyperplane is converted to solving the following problem:

[0075]

[0076]

[0077] Where C is the penalty parameter, used to balance the weight, and ξ is the relaxation variable. The final decision function is obtained by calculation

[0078]

[0079] Where α is the optimal Lagrange multiplier, x is the training sample, y is the label taking ±1, K(x i , x j ) is the kernel function of the model, and the radial basis function is selected as the kernel function of the model, defined as follows:

[0080]

[0081] Where σ is the bandwidth, which controls the radial range.

[0082] c4, respectively, three training samples in step c2 are used to find three hyperplanes according to step c3 and Three support vector machine-based classification models V1, V2 and V3 capable of distinguishing noise types are obtained.

[0083] A method for denoising by using a denoising network constructed by the construction system of the ground magnetic resonance multi-type noise denoising network, comprising:

[0084] The noisy data set is input into the support vector machine classification model V1, and the discrimination module judges according to the following formula:

[0085]

[0086] If the above formula is true, it means that the signal is seriously affected by spike noise, and the noisy signal is input into the model Net s (θ s ) for processing, otherwise the data is input into the classification model V2, and the discrimination module judges again:

[0087]

[0088] If the above formula is true, it means that the signal is seriously affected by spike noise, and the noisy signal is input into the model Net h (θ h ) for processing, otherwise the data is input into the classification model V2, and the discrimination module judges again:

[0089]

[0090] If the above formula is true, it means that the signal is seriously affected by spike noise, and the noisy signal is input into the model Net r (θ r ) for processing, otherwise it means that the noise is weak, and the effective signal features are recognized.

[0091] Compared with the prior art, the present application has the beneficial effects that:

[0092] The present application applies a convolutional neural network to construct multiple denoising models capable of removing three different noises, i.e., spike noise, power harmonic noise and random noise, and uses a support vector machine method to classify data containing different types of noise and input them into corresponding denoising models for processing. BRIEF DESCRIPTION OF DRAWINGS

[0093] Figure 1A flowchart of network construction for denoising multiple types of ground magnetic resonance noise is provided for the embodiments of the present application.

[0094] Figure 2 A structural block diagram of the convolutional neural denoising network for different noise types (a) a model for removing spike noise, (b) a model for reducing power harmonic noise, and (c) a model for eliminating random noise.

[0095] Figure 3 A principle block diagram of the denoising network for multiple types of ground magnetic resonance noise. DETAILED DESCRIPTION

[0096] In order to make the purpose, technical solutions and advantages of the present application clearer and more apparent, the present application will be further described in detail below in combination with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0097] Referring to Figure 1 , the construction system of the ground magnetic resonance multiple type noise denoising network model comprises:

[0098] A magnetic resonance signal construction module simulates multiple sets of magnetic resonance effective signals and environmental noise, including spike noise, power frequency noise and random noise, according to the ground magnetic resonance groundwater detection principle, and constructs a noisy data set affected by the three types of noise respectively;

[0099] A denoising neural network construction module trains and optimizes the network based on the three noisy data sets and noise data sets generated by the magnetic resonance signal construction module, respectively, to obtain three denoising networks for different types of noise, including a model Net s (θ s ) for removing spike noise, a model Net h (θ h ) for reducing power harmonic noise, and a model Net r (θ r ) for eliminating random noise.

[0100] A classification and discrimination model uses a support vector machine method to determine the noise type according to the characteristics of different noise, and outputs the determination result to the corresponding denoising network model in the denoising neural network construction module, so as to ensure that the denoising network for different types of noise can play the maximum denoising role.

[0101] Specific steps for the magnetic resonance signal construction module to construct the data set are as follows:

[0102] a1, construct a magnetic resonance signal, the time domain expression is as follows

[0103]

[0104] wherein E0∈[10,10]00, f L ∈[2200,2400]Hz, t=(0:N-1) / f s , f s =25kHz, N=6250, the effective signal data x1 is obtained according to the above formula, the signal parameters are changed, and n effective signals X={x1,x2,...,x n} are obtained.

[0105] a2, according to the characteristics of different types of noise, the noise is constructed by using the randn function and the trigonometric function in MATLAB, including the spike noise s, the power frequency noise h and the random noise r, the noise parameters are changed, and each type of noise data set has n groups, including the spike noise data set S={s1,s2,...,s n}, the power frequency noise data set H={h1,h2,...,h n} and the random noise data set R={r1,r2,...,r n}.

[0106] a3, the noise in the constructed noise data set is superimposed with the effective signal in the signal data set respectively, and three noise-containing data sets are obtained, including N s ={n s1 ,n s2 ,...,n sn}, N h ={n h1 ,n h2 ,...,n hn} and N r ={n r1 ,n r2 ,...,n rn}.

[0107] The specific steps of building and determining the denoising network model by the denoising neural network building module are as follows:

[0108] Firstly, the denoising network is built by using convolution, upsampling and other operations, and skip connection is introduced to solve the model degradation problem caused by the deepening of network layers, and the convolution operation can be described by the following formula:

[0109]

[0110] wherein * is the convolution operation, k is the convolution kernel, b is the bias, f is the activation function, and x is the input and output feature in the convolution operation process, and the upsampling can be described by the following formula:

[0111]

[0112] where upsampl(·) is an up-sampling function;

[0113] Secondly, the data sets {N s ,S},{N h ,H}, and {N r ,R} are used to train the denoising network, and the objective function of the network is as follows:

[0114]

[0115]

[0116]

[0117] By continuously updating the parameters θ s , θ h and θ r of the network model, the objective function is optimized, so as to obtain the model Net s (θ s ) for removing spike noise, the model Net h (θ h ) for reducing power harmonic noise, and the model Net r (θ r ) for eliminating random noise.

[0118] Referring to Figure 2 , the models of the three kinds of noises all include a convolution layer and an up-sampling layer, and the noisy signal is output after being input from the convolution layer and passing through the up-sampling layer, and the convolution layer and the up-sampling layer are connected through a skip layer.

[0119] Referring to Figure 3 , the classification and discrimination module includes a classification model construction module and a discrimination model, and the classification model construction module constructs a noise classification model based on a support vector machine, and the specific steps are as follows:

[0120] c1, the classification of multiple types of noise is decomposed into multiple application support vector machines to solve the problem of two classification. The types of discrimination include being seriously affected by spike noise, being seriously affected by power frequency noise, being seriously affected by random noise, and being able to identify effective signals with less noise.

[0121] c2, the constructed data sets N s , N h , N r , and X are recombined, N s and {N h , N r , X} are used as the training sample set of the first two-classification problem, and are used to distinguish between serious spike noise and other conditions, and the corresponding labels are +1 and -1, respectively; N h and {N s , Nr , as the training sample set of the second binary classification problem, used to distinguish the serious power frequency noise and other conditions, and the corresponding labels are +1 and -1 respectively; N r and{N s ,N h , as the training sample set of the third binary classification problem, used to distinguish the serious random noise and other conditions, and the corresponding labels are +1 and -1 respectively;

[0122] c3, the support vector machine is based on statistical theory, trying to find a hyperplane in high-dimensional space separating the training samples, wherein w is a weight vector, and b is a threshold, is the objective function. The problem of solving the optimal hyperplane can be converted into solving the following problem:

[0123]

[0124]

[0125] wherein C is a penalty parameter, used to balance the weight, and ξ is a relaxation variable. The final decision function obtained by calculation is

[0126]

[0127] wherein α is the optimal Lagrange multiplier, x is the training sample, y is the label taking ±1, K(x i ,x j ) is the kernel function of the model, and the radial basis function is selected as the kernel function of the model in the application, and is defined as follows:

[0128]

[0129] wherein σ is the bandwidth, controlling the radial range of action.

[0130] c4, three hyperplanes are found according to step c3 by using the three training samples in step c2 respectively and three support vector machine-based classification models V1, V2 and V3 capable of distinguishing noise types are obtained, wherein w1, w2 and w3 are the weight vectors of the classification models V1, V2 and V3 respectively, b1, b2 and b3 are the corresponding thresholds, and are the objective functions of the three models. Referring to Figure 3 , the noise classification model is established by the support vector machine, and the denoising is carried out by the discrimination module, and the process includes:

[0131] the noise data set Input into the support vector machine classification model V1, the discrimination module judges according to the following formula:

[0132] Wherein, is the optimal Lagrange multiplier of the classification model V1, is the training sample, is the label corresponding to the training sample, which takes ±1, representing the presence or absence of spike noise, and K(·) is the kernel function of the model, and the radial basis function is selected as the kernel function of the model. If the above formula is true, it means that the signal is severely affected by spike noise, and the noisy signal is input into the model Net s (θ s ) for processing, otherwise the data is input into the classification model V2, and the discrimination module judges again:

[0133]

[0134] Wherein, is the optimal Lagrange multiplier of the classification model V2, is the training sample , and the label corresponding to the training sample takes ±1, representing the presence or absence of power frequency noise. If the above formula is true, the signal is severely affected by power frequency noise, and the noisy signal is input into the model Net h (θ h ) for processing, otherwise the data is input into the classification model V3, and the discrimination module judges again:

[0135] Wherein, is the optimal Lagrange multiplier of the classification model V3, is the training sample , and the label corresponding to the training sample takes ±1, representing the presence or absence of random noise. If the above formula is true, the signal is severely affected by random noise, and the noisy signal is input into the model Net r (θ r ) for processing, otherwise it means that the noise is weak, and the effective signal feature can be recognized. The above steps can improve the denoising performance through multiple cycles.

[0136] A ground magnetic resonance multi-type noise denoising network Construction method , comprising:

[0137] a. According to the principle of ground magnetic resonance groundwater detection, simulate multiple groups of magnetic resonance effective signals and environmental noise, including spike noise, power frequency noise and random noise, and construct a noisy data set affected by three types of noise respectively;

[0138] b. Construct three denoising neural networks, and train and optimize the networks based on three noisy datasets and three noisy datasets respectively, to obtain three denoising networks for different types of noise, including a model Net for removing spike noise. s (θ s Net, a model for reducing power frequency harmonic noise h (θ h ), and the Net model for eliminating random noise. r (θ r );

[0139] c. Using the support vector machine method, a classification and discrimination model is established to determine the type of noise based on the characteristics of different noises. The classification and discrimination model judges the noisy signal and inputs it into the corresponding denoising network according to the judgment result of the noise type, so as to ensure that the denoising network plays the maximum denoising role for different types of noise.

[0140] The specific steps for constructing the dataset in step a are as follows:

[0141] a1, construct the magnetic resonance signal, the time-domain expression is as follows:

[0142]

[0143] Where E0 is the initial amplitude, f L For Larmor frequency, It is the relaxation time. It is the initial phase, t = (0:N-1) / f s f s Let N be the sampling rate and N be the data length. Using the above formula, we obtain the effective signal data x. By changing the signal parameters, we obtain n effective signals X = {x1, x2, ..., xn}. n};

[0144] a2. Based on the characteristics of different types of noise, construct noise datasets using the randn function and trigonometric functions in MATLAB, including spike noise s, power frequency noise h, and random noise r. Change the noise parameters. There are n datasets for each type of noise, including the spike noise dataset S = {s1, s2, ..., s...}. n}, Power frequency noise dataset H = {h1, h2, ..., h n} and a random noise dataset R = {r1, r2, ..., r n};

[0145] a3, superimpose the noise from the constructed noise dataset with the valid signal from the signal dataset to obtain three noisy datasets, including N. s ={n s1 ,n s2 ,...,nsn},N h ={n h1 ,n h2 ,...,n hn} and N r ={n r1 ,n r2 ,...,n rn}

[0146] The specific steps for building and determining the denoising network model in step b are as follows:

[0147] First, a denoising network is built using operations such as convolution and upsampling, and skip connections are introduced to address the model degradation problem caused by increasing network depth. The convolution operation can be described by the following formula:

[0148]

[0149] Where * represents the convolution operation, k is the convolution kernel, b is the bias, f is the activation function, and x represents the input and output features during the convolution operation. Upsampling can be described by the following formula:

[0150]

[0151] Where upsample(·) is the upsampling function;

[0152] Secondly, the dataset {N s ,S},{N h ,H}, and {N r R and R are used to train the denoising network, and the objective function of the network is as follows:

[0153]

[0154]

[0155]

[0156] By continuously updating the parameters θ of the network model s θ h and θ r The objective function is optimized to obtain the Net model with spike noise removed. s (θ s Net, a model for reducing power frequency harmonic noise h (θ h ), and the Net model for eliminating random noise. r (θ r ).

[0157] The specific steps for building a noise classification model based on support vector machines in step c are as follows:

[0158] c1, the multi-type noise classification is decomposed into multiple application support vector machines to solve the binary classification problem. The classified categories include being seriously affected by spike noise, being seriously affected by power frequency noise, being seriously affected by random noise, and being less affected by noise to identify effective signals;

[0159] c2, the constructed data set N s , N h , N r , and X are recombined, N s and {N h , N r , X} are used as a training sample set of the first binary classification problem, for distinguishing between serious spike noise and other conditions, and the corresponding labels are +1 and -1 respectively; N h and {N s , N r , X} are used as a training sample set of the second binary classification problem, for distinguishing between serious power frequency noise and other conditions, and the corresponding labels are +1 and -1 respectively; N r and {N s , N h , X} are used as a training sample set of the third binary classification problem, for distinguishing between serious random noise and other conditions, and the corresponding labels are +1 and -1 respectively;

[0160] c3, the support vector machine is based on statistical theory, and tries to find a hyperplane in a high-dimensional space that separates the training samples, where w is a weight vector, and b is a threshold, is an objective function. The problem of solving the optimal hyperplane can be converted into solving the following problem:

[0161]

[0162]

[0163] where C is a penalty parameter for balancing the weight, and ξ is a relaxation variable. Through calculation, the final decision function is

[0164]

[0165] where α is an optimal Lagrange multiplier, x is a training sample, y is a label taking ±1, K(x i , x j ) is a kernel function of the model, and the present application selects a radial basis function as the kernel function of the model, and is defined as follows:

[0166]

[0167] where σ is a bandwidth, and controls the radial range.

[0168] c4, three hyperplanes are found according to the steps of c3 using the three training samples in c2 respectively and Three support vector machine-based classification models V1, V2 and V3 capable of distinguishing noise types are obtained.

[0169] A method for denoising by using a ground magnetic resonance multi-type noise denoising network, the method comprising:

[0170] inputting a noisy data set into the support vector machine classification model V1, and the discrimination module judging according to the following formula:

[0171]

[0172] If the above formula is true, it means that the signal is seriously affected by spike noise, and the noisy signal is input into the model Net s (θ s ) for processing, otherwise the data is input into the classification model V2, and the discrimination module judges again:

[0173]

[0174] If the above formula is true, it means that the signal is seriously affected by power frequency noise, and the noisy signal is input into the model Net h (θ h ) for processing, otherwise the data is input into the classification model V3, and the discrimination module judges again:

[0175]

[0176] If the above formula is true, it means that the signal is seriously affected by random noise, and the noisy signal is input into the model Net r (θ r ) for processing, otherwise it means that the noise is weak and the effective signal features can be recognized.

[0177] The above only describes the preferred embodiments of the present application and is not used to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A system for constructing a multi-type noise denoising network for ground magnetic resonance imaging, characterized in that, include: The magnetic resonance signal construction module simulates multiple sets of effective magnetic resonance signals and environmental noise, including spike noise, power frequency noise and random noise, based on the principle of ground magnetic resonance groundwater detection, and constructs noisy datasets affected by the three types of noise respectively. The denoising neural network construction module trains and optimizes the network based on the magnetic resonance signals generated by the magnetic resonance signal construction module using three noisy datasets and noisy datasets, resulting in three denoising neural networks for different types of noise, including a model Net for removing spike noise. s (θ s Net, a model for reducing power frequency harmonic noise h (θ h ), and the Net model for eliminating random noise. r (θ r ); The classification and discrimination model building module constructs a classification and discrimination model that uses the support vector machine method to determine the noise type based on the characteristics of different noises, and outputs the judgment result to the corresponding denoising network model in the denoising neural network building module. The dataset is constructed using the magnetic resonance signal construction module, specifically including: a1, construct the magnetic resonance signal, the time-domain expression is as follows: ， in, , , , , , =25kHz, N=6250, obtain the effective signal data x1 according to the above formula, change the signal parameters to obtain n effective signals X={x1,x2, ..., x n }; a2. Based on the characteristics of different types of noise, construct noise datasets using the randn function and trigonometric functions in MATLAB, including spike noise s, power frequency noise h, and random noise r. Change the noise parameters. There are n datasets for each type of noise, including the spike noise dataset S={s1, s2, ..., s...} n }, Power frequency noise dataset H={h1, h2, ..., h n } and a random noise dataset R={r1,r2, ..., r n }; a3, superimpose the noise from the constructed noise dataset with the valid signal from the signal dataset to obtain three noisy datasets, including N. s ={n s1 , n s2 , ..., n sn }, N h ={n h1 , n h2 , ..., n hn } and N r ={n r1 , n r2 , ...,n rn }; Three denoising networks targeting different types of noise are built using a denoising neural network construction module, including: A denoising network is built using convolution and upsampling operations, and skip layers are introduced to connect convolution and upsampling. The convolution operation is described by the following formula: ， Where * represents the convolution operation, k is the convolution kernel, b is the bias, f is the activation function, and x represents the input and output features during the convolution operation. Upsampling is described by the following formula: ， in, It is an upsampling function; Data set {N s , S},{N h , H}, and {N r R and R are used to train the denoising network, and the objective function of the network is as follows: ， ， ， By continuously updating the parameters of the network model , and Optimize the objective function to obtain the Net model with peak noise removed. s ( Net, a model for reducing power frequency harmonic noise h ( ), and the Net model for eliminating random noise. r ( ).

2. The system for constructing a multi-type noise denoising network for ground magnetic resonance imaging according to claim 1, characterized in that, The classification and discrimination module includes a classification model building module and a discrimination model. The classification model building module constructs a noise classification model based on support vector machines, specifically including: c1 decomposes the classification of multiple types of noise into multiple application support vector machines to solve the binary classification problem; c2, will construct the dataset N s , N h , N r Recombined with X, N s and {N h , N r X} serves as the training sample set for the first type of binary classification problem, used to distinguish between severe spike noise and other cases, with corresponding labels of +1 and -1, respectively; N h and {N s , N r X} serves as the training sample set for the second type of binary classification problem, used to distinguish between severe power frequency noise and other situations, with corresponding labels of +1 and -1, respectively; N r and {N s , N h , X} serves as the training sample set for the third type of binary classification problem, used to distinguish between severe random noise and other cases, with corresponding labels of +1 and -1, respectively; c3, Support Vector Machines find hyperplanes in high-dimensional space. The training samples are separated, where w is the weight vector and b is the threshold. Given the objective function, the problem of finding the optimal hyperplane is transformed into solving the following problem: ， ， Where C is the penalty parameter, used to balance the weights. Let be slack variables. The final decision function is obtained through calculation. ， in, The optimal Lagrange multiplier is given by x, where x is the training sample and y is the label. , For the kernel function of the model, the radial basis function is chosen and defined as follows: ， in It refers to bandwidth, which controls the radial range of action; c4, using the three training samples from step c2 respectively, find three hyperplanes according to step c3. , and This yields three support vector machine-based classification models V1, V2, and V3, which are capable of distinguishing noise types.

3. The system for constructing a multi-type noise denoising network for ground magnetic resonance imaging according to claim 2, characterized in that, The denoising neural network performs denoising, and the process includes: Noisy dataset The input is fed into classification model V1, and the discrimination module makes judgments according to the following formula: ， If the above formula holds true, it indicates that the signal is severely affected by spike noise. Therefore, the noisy signal should be input into the Net model to remove spike noise. s (θ s If the data is not processed internally, it will be input into the classification model V2, and the discrimination module will then make a judgment. ， If the above equation holds true, then the signal suffers from severe power frequency noise. The noisy signal is then input into a model Net to reduce power frequency harmonic noise. h (θ h If the data is not processed internally, it will be input into the classification model V3, and the discrimination module will then make a judgment. ， If the above equation holds true, then the signal suffers from severe random noise. The noisy signal is then input into a model Net to eliminate random noise. r (θ r If the noise is not processed within the internal processing, it indicates that the noise is weak and the characteristics of the effective signal are identified.

4. A method for constructing a multi-type noise denoising network for ground magnetic resonance imaging, characterized in that, include: a. Based on the principle of ground magnetic resonance groundwater detection, simulate multiple sets of effective magnetic resonance signals and environmental noise, including spike noise, power frequency noise and random noise, and construct noisy datasets affected by the three types of noise respectively. b. Construct three denoising neural networks, and train and optimize the networks based on three noisy datasets and three noisy datasets respectively, to obtain three denoising networks for different types of noise, including a model Net for removing spike noise. s (θ s Net, a model for reducing power frequency harmonic noise h (θ h ), and the Net model for eliminating random noise. r (θ r ); c. Using the support vector machine method, a classification and discrimination model is established to determine the type of noise based on the characteristics of different noises. The classification and discrimination model is connected to the denoising network. After the noisy signal is judged, the result of the noise type judgment is input into the corresponding denoising network. The specific steps for constructing the dataset in step a are as follows: a1, construct the magnetic resonance signal, the time-domain expression is as follows: ， in, The initial amplitude, For Larmor frequency, It is the relaxation time. It is the initial phase. , Let N be the sampling rate and N be the data length. Using the above formula, we obtain the effective signal data x. By changing the signal parameters, we obtain n effective signals X = {x1, x2, ..., x...}. n }; a2. Based on the characteristics of different types of noise, construct noise datasets using the randn function and trigonometric functions in MATLAB, including spike noise s, power frequency noise h, and random noise r. Change the noise parameters. There are n datasets for each type of noise, including the spike noise dataset S={s1, s2, ..., s...} n }, Power frequency noise dataset H={h1, h2, ..., h n } and a random noise dataset R={r1,r2, ..., r n }; a3, superimpose the noise from the constructed noise dataset with the valid signal from the signal dataset to obtain three noisy datasets, including N. s ={n s1 , n s2 , ..., n sn }, N h ={n h1 , n h2 , ..., n hn } and N r ={n r1 , n r2 , ...,n rn }; Step b specifically includes: A denoising network is built using convolution and upsampling operations, and skip connections are introduced. The convolution operation is described by the following formula: ， Where * represents the convolution operation, k is the convolution kernel, b is the bias, f is the activation function, and x represents the input and output features during the convolution operation. Upsampling is described by the following formula: ， in, It is an upsampling function; Data set {N s , S},{N h , H}, and {N r R and R are used to train the denoising network, and the objective function of the network is as follows: ， ， ， By continuously updating the parameters of the network model , and Optimize the objective function to obtain the Net model with peak noise removed. s ( Net, a model for reducing power frequency harmonic noise h ( ), and the Net model for eliminating random noise. r ( ).

5. The method for constructing a multi-type noise denoising network for ground magnetic resonance imaging according to claim 4, characterized in that, Step c involves building a noise classification model based on support vector machines, specifically including: c1 decomposes the classification of multiple types of noise into multiple application support vector machines to solve the binary classification problem; c2, will construct the dataset N s , N h , N r Recombined with X, N s and {N h , N r X} serves as the training sample set for the first type of binary classification problem, used to distinguish between severe spike noise and other cases, with corresponding labels of +1 and -1, respectively; N h and {N s , N r X} serves as the training sample set for the second type of binary classification problem, used to distinguish between severe power frequency noise and other situations, with corresponding labels of +1 and -1, respectively; N r and {N s , N h , X} serves as the training sample set for the third type of binary classification problem, used to distinguish between severe random noise and other cases, with corresponding labels of +1 and -1, respectively; c3, Support Vector Machines find hyperplanes in high-dimensional space. The training samples are separated, where w is the weight vector and b is the threshold. Given the objective function, the problem of finding the optimal hyperplane is transformed into solving the following problem: ， ， Where C is the penalty parameter, used to balance the weights. Let be slack variables. The final decision function is obtained through calculation. ， in, The optimal Lagrange multiplier is given by x, where x is the training sample and y is the label. , For the kernel function of the model, the radial basis function is chosen and defined as follows: ， in It refers to bandwidth, which controls the radial range of action; c4, using the three training samples from step c2 respectively, find three hyperplanes according to step c3. , and This yields three support vector machine-based classification models V1, V2, and V3, which are capable of distinguishing noise types.

6. The method for constructing a multi-type noise denoising network for ground magnetic resonance imaging according to claim 5, characterized in that, The denoising neural network performs denoising, and the process includes: Noisy dataset The input is fed into the Support Vector Machine classification model V1, and the discrimination module makes judgments according to the following formula: ， If the above formula holds true, it indicates that the signal is severely affected by spike noise. Therefore, the noisy signal should be input into the Net model to remove spike noise. s (θ s If the data is not processed internally, it will be input into the classification model V2, and the discrimination module will then make a judgment. ， If the above equation holds true, then the signal suffers from severe power frequency noise. The noisy signal is then input into a model Net to reduce power frequency harmonic noise. h (θ h If the data is not processed internally, it will be input into the classification model V3, and the discrimination module will then make a judgment. ， If the above equation holds true, then the signal suffers from severe random noise. The noisy signal is then input into a model Net to eliminate random noise. r (θ r If the noise is not processed within the internal processing, it indicates that the noise is weak and the characteristics of the effective signal are identified.

Images