CN103632347B - A kind of Magnetic Resonance Image Denoising based on wavelet shrinkage - Google Patents

Abstract

The invention discloses a kind of Magnetic Resonance Image Denoising based on wavelet shrinkage, the method in conjunction with the threshold shrink of ratio atrophy and Low threshold, to reach denoising and to retain the object of image detail as much as possible.Technical solution of the present invention chooses the mould square image background regions be made up of noise, obeys this condition of non-central card side distribution, estimate, overcome former algorithm and depend on the known restriction of noise the noise variance of mould image by it.By merging the accuracy that improve high frequency coefficient, thus improve the denoising effect of image: for Soft thresholding denoising, the situation of too much details is lost in insufficient and ratio atrophy method denoising, regarded as two kinds of limiting cases, thus merged by weights, obtain high frequency coefficient more accurately, achieve the denoising effect being better than above two kinds of methods.

Description

Magnetic resonance image denoising method based on wavelet shrinkage

Technical Field

The invention relates to the technical field of image processing, in particular to a magnetic resonance image denoising method based on wavelet shrinkage.

Background

The application background of the invention is that: magnetic Resonance Imaging (MRI) is a new imaging method in the field of medical imaging, and uses radio frequency pulses to make hydrogen nuclei in a magnetic field resonate to generate signals, and the signals are processed and imaged. MRI has the advantages of multiple imaging parameters, high contrast, free orientation tomography, no bone artifact interference, no ionizing radiation damage to human body, etc., and is one of the most widely used advanced techniques in clinical examination (see document [ 1 ] DuaG, varun rajd. mrideneinosing using wavetoatoms scattering kage [ J ]. globals journal of research in engineering,12(4-F), 2012.). However, limited imaging times often require one to make a tradeoff between resolution and signal-to-noise ratio [ 2 ] Pizuricaa, winkAM, Vansteenkspace. Aeeviewowave wavelentdenosingingm III and less rassouring imaging [ J ]. Current medical imaging reviews,2 (2): 247-260,2006 ]. High resolution images tend to contain strong noise (see document [ 3 ] wang hao. proces magnetic resonance imaging. chinjmagn resonance imaging, vol.3, No.3,2012.). Noise can obscure image details and affect clinical diagnosis (see [ 4 ] LiLingyuan, Zhang Yanhuaa. adaptive MRIdenosingbasedonlifting wall. computer engineering and applications,43(35):83-85,2007.). Therefore, the method has extremely important significance for denoising the MR image.

The related art analysis is as follows:

scheme one

Scheme name: magnetic resonance image denoising method based on threshold atrophy (see [ 5 ] DonohoDL, Johnston IM. Adaptinggtounknown noise and noise wave height set library [ J ]. journal of American Statisc, 12(90):1200-

The algorithm idea is as follows: the noise is discontinuous in a time domain, and an actual signal is usually continuous, so that the noise is represented as high-frequency information, and the amplitude of a high-frequency coefficient of the noise is usually small; the actual signal reflects the high-frequency part of the image detail information and is mainly concentrated at the place with larger amplitude of the high-frequency coefficient. Therefore, the threshold shrinking carries out shrinking zero setting on the high-frequency coefficient with smaller amplitude, namely removing the noise. The algorithm has the following defects: the accurate determination of the threshold is difficult, and the improper threshold can cause the problems of insufficient image denoising or killing image details and the like.

Scheme two

Scheme name: magnetic resonance image denoising method based on scale shrinkage (see [ 8 ]) KivancMihcakM, KozintsevI, Ramchandran K, et al, Low-complex imaging noise creating based on magnetic resonance imaging and denoising methods of wavelets and coefficients [ J ]. SignalProcessingLetters, IEEE,1999,6(12): 300-)

The algorithm idea is as follows: on the premise that the high-frequency coefficient of the image obeys Gaussian distribution, the degree of the high-frequency coefficient polluted by noise is estimated by a minimum mean square error estimation method, and an atrophy coefficient is determined, so that the high-frequency coefficient is shrunk, and the purpose of denoising is achieved.

The algorithm has the following defects: the algorithm requires a known noise variance, which is not readily available in practice; and the image after denoising is seriously blurred, and more detail information is lost in the image.

Disclosure of Invention

The invention aims to: the invention provides a magnetic resonance image denoising method based on wavelet shrinkage, which combines proportional shrinkage and low-threshold shrinkage to achieve the purposes of denoising and retaining image details as much as possible.

The technical scheme of the invention is as follows: a magnetic resonance image denoising method based on wavelet shrinkage comprises the following steps:

step 1.1, threshold atrophy of images

In the selection of the threshold function, the following formula is adopted:

$T_{ji}=\frac{1}{2^j-1}*{\mathrm{\λ}}_{ji}*M_{ji}$

where j represents the number of wavelet layers in a particular row, i represents the direction,for contrast of image, sigma represents the scale of high frequency coefficientThe standard deviation, mu represents the mean value of the high-frequency coefficient, and M is the absolute median of the high-frequency coefficient; the threshold function is derived from the propagation characteristics of the noise and the actual signal; the energy of the actual signal is increased along with the increase of decomposition levels, and the noise is just opposite, so that the higher the level is, the smaller the threshold value is, and the high-frequency coefficient of the actual signal is fully reserved; the larger the contrast is, the larger the energy difference between the noise energy and the actual signal is, so that a larger threshold is selected to remove the noise as much as possible; the addition of M takes the statistical characteristics of the high-frequency coefficient into consideration;

the corresponding threshold processing function adopts a soft threshold, and the specific formula is as follows:

${\hat{w}}_{j,k}=\left\{\begin{array}{cc}\mathrm{sgn}\left(w_{j,k}\right)\left(\right|w_{j,k}|-T),& \left|w_{j,k}\right|\&GreaterEqual;T\\ 0,& \left|w_{j,k}\right|<T\end{array}\right.$

where T is a threshold value, w_j,kIs the observed high-frequency coefficient of the signal,is an estimate of the high frequency coefficient;

the method comprises the following specific steps:

1) performing two-dimensional discrete orthogonal wavelet decomposition on the image;

2) calculating a threshold value of the high-frequency coefficient in each layer and each direction;

3) carrying out threshold atrophy on the high-frequency coefficient by using a soft threshold;

step 1.2, the image is subjected to scale atrophy

Obeying normal distribution N (0, sigma) at high frequency coefficient of signal²) Noise obeys normal distributionUnder the assumption that the actual signal is estimated according to the minimum mean square error:

$\hat{X}\left(k\right)=\frac{{\mathrm{\&sigma;}}^2\left(k\right)}{{\mathrm{\&sigma;}}^2\left(k\right)+{\mathrm{\&sigma;}}_n^2}Y\left(k\right)$

wherein sigma²(k) Is the variance of the high frequency coefficients without noise interference,is the variance of the noise, and Y (k) is the high-frequency coefficient of the noise-containing image; sigma²(k) Is unknown, under the assumption that the variance of each point is uniformly changed, it can be estimated through the neighborhood, and the estimation formula is as follows:

${\mathrm{\&sigma;}}^2\left(k\right)=\max (0,\frac{1}{H}\underset{k\&Element;\mathrm{\&Omega;}\left(k\right)}{\&Sigma;}{Y}^2(k)-{\mathrm{\&sigma;}}_n^2)$

where Ω (k) is the neighborhood window, M is the number of high frequency coefficients in the neighborhood, Y (k) is the high frequency coefficients in the neighborhood,is the noise variance;

the variance of noise is required for solving the shrinkage coefficient, and in practical application, the variance of noise is unknown, so that the variance of noise needs to be estimated; in the estimation of the noise variance, the invention adopts the following estimation method:

${\mathrm{\&sigma;}}_n^2=M_n^2/2$

whereinThe mean value of the pure background part of the representative mode square image is obtained mainly according to the characteristic that the noise of the representative mode square image obeys non-central chi-square distribution;

the module square image can be obtained by squaring the module image; extracting a pure background part, namely performing edge detection on the image by adopting a canny operator, and then scanning and communicating the detected edge, wherein the part outside the obtained communicated region is the pure background region of the image;

the method comprises the following specific steps:

1) performing two-dimensional discrete orthogonal wavelet decomposition on the image;

2) determining high-frequency shrinking coefficients of all points, and shrinking the high-frequency coefficients;

step 1.3 fusion of threshold atrophy and proportional atrophy

From the single denoising effect, the image noise obtained based on the soft threshold atrophy method is filtered, but the image noise is not thorough enough, which means that the high-frequency coefficient is not sufficiently shrunk in the shrinking process; the noise of the de-noised image obtained by shrinkage of the scale is filtered thoroughly, but the problems of image blurring and excessive detail loss exist, which means that the high-frequency coefficient is excessively shrunk in the shrinking process. Therefore, the two can be endowed with different weights for fusion, thereby obtaining an ideal denoising effect;

in the determination of the weights of the two methods, because the proportional shrinkage method is premised on that the variance of the high-frequency coefficient of the actual signal is uniformly changed, gradient information is calculated on the estimated variance by using a sobel operator so as to judge the accuracy of the high-frequency coefficient after shrinkage; in places with larger gradients, the scale shrinkage precondition is not satisfied, so the reliability of the coefficient is not high, and the weight of the high-frequency coefficient estimated by the coefficient is low, while in places with uniform changes, the reliability is high and the weight is high;

the method comprises the following specific steps:

1) on the basis of the first two steps, recording a high-frequency coefficient matrix processed by a threshold collapsing method as A, and recording a high-frequency coefficient matrix processed by a proportional collapsing method as B;

2) calculating gradient information of the high-frequency coefficient variance of each point by using a sobel operator, normalizing the gradient information, wherein if the gradient value is greater than a threshold value T (empirical value), a corresponding position 0 is arranged in a matrix C (the matrix with the same size as A and used for storing the scale shrinkage credibility condition), and if not, the corresponding position 1 is arranged;

3) calculating the fused high-frequency coefficient:

R＝α*(B*C)+β*(A*C)+γ*(B*D)+*(A*D)

where α + β ═ 1, γ + ═ 1, D ═ 1-C, α, γ are empirical values, and the above x represents the dot product of the corresponding elements of the matrix;

4) and performing inverse transformation on the image coefficient to obtain a denoised image.

Further, based on a number of experiments, T is 0.1, α is 0.75, γ is 0.45, and the neighborhood window is 3 × 3.

Further, step 1.4 is included after step 1.3: the denoising process comprises the following steps:

STEP 1: a magnetic resonance image containing noise is input.

STEP 2: and carrying out two-dimensional discrete orthogonal wavelet transform on the image containing the noise.

STEP 3: two sets of high-frequency coefficients are obtained by performing threshold shrinking treatment and proportional shrinking treatment on the high-frequency coefficients after STEP 2.

STEP 4: determining a fusion weight, and fusing the two high-frequency coefficients after STEP3 to obtain a new high-frequency coefficient.

STEP 5: and performing two-dimensional discrete orthogonal wavelet inverse transformation to obtain a denoised image.

STEP 6: and outputting the denoised magnetic resonance image.

The technical scheme of the invention has the advantages and positive effects that:

(1) the image noise variance is reasonably estimated, and the limit that the prior algorithm depends on the known noise variance is overcome: the technical scheme of the invention selects a background area of the model square image formed by noise, and estimates the noise variance of the model image by the condition that the background area obeys non-central chi-square distribution.

(2) The accuracy of the high-frequency coefficient is improved through fusion, so that the denoising effect of the image is improved: the method is used for solving the problems that the soft threshold method is insufficient in denoising and the proportional shrinkage method is excessive in denoising loss details, and is regarded as two limit conditions, so that weight is fused to obtain a more accurate high-frequency coefficient, and a denoising effect superior to the two methods is achieved.

Drawings

Fig. 1 shows a magnetic resonance image denoising process.

Fig. 2 is a diagram illustrating the denoising effect at 3% noise, wherein (a) a magnetic resonance image with 3% noise added; (b) denoising the image by a proportional shrinkage method; (c) denoising the image by a threshold shrinkage method; (d) and improving the denoised image of the algorithm.

Fig. 3 is a diagram illustrating the denoising effect at 5% noise, wherein (a) a magnetic resonance image with 5% noise added; (b) denoising the image by a proportional shrinkage method; (c) denoising the image by a threshold shrinkage method; (d) and improving the denoised image of the algorithm.

Detailed Description

The invention is further described with reference to the following figures and detailed description.

The invention relates to a magnetic resonance image denoising method based on wavelet shrinkage, which comprises the following specific implementation steps:

step 1.1, threshold atrophy of images

In the selection of the threshold function, the following formula is adopted:

$T_{ji}=\frac{1}{2^j-1}*{\mathrm{\&lambda;}}_{ji}*M_{ji}$

where j represents the number of wavelet layers in a particular row, i represents the direction,for the contrast of the image, σ represents the standard deviation of the high frequency coefficients, μ represents the mean of the high frequency coefficients, and M is the absolute median of the high frequency coefficients. The threshold function is derived from the propagation characteristics of the noise and the actual signal. The energy of the actual signal increases with the increase of the decomposition level, and the noise is opposite, so that the higher the level is, the smaller the threshold value is, so as to fully reserve the high-frequency coefficient of the actual signal. The larger the contrast, the larger the energy difference between the noise energy and the actual signal, and therefore the larger the threshold is selected to remove as much noise as possible. The addition of M takes into account the statistical properties of the high frequency coefficients.

The corresponding threshold processing function adopts a soft threshold, and the specific formula is as follows:

${\hat{w}}_{j,k}=\left\{\begin{array}{cc}\mathrm{sgn}\left(w_{j,k}\right)\left(\right|w_{j,k}|-T),|& \left|w_{j,k}\right|\&GreaterEqual;T\\ 0,& \left|w_{j,k}\right|<T\end{array}\right.$

where T is a threshold value, w_j,kIs the observed high-frequency coefficient of the signal,is an estimate of the high frequency coefficient.

The method comprises the following specific steps:

1) and carrying out two-dimensional discrete orthogonal wavelet decomposition on the image.

2) And calculating a threshold value of the high-frequency coefficient in each direction of each layer.

3) Threshold collapse is performed on the high frequency coefficients using a soft threshold.

Step 1.2, the image is subjected to scale atrophy

Compliance at high frequency coefficients of signalNormal distribution N (0, sigma)²) Noise obeys normal distributionUnder the assumption that the actual signal is estimated according to the minimum mean square error:

$\hat{X}\left(k\right)=\frac{{\mathrm{\&sigma;}}^2\left(k\right)}{{\mathrm{\&sigma;}}^2\left(k\right)+{\mathrm{\&sigma;}}_n^2}Y\left(k\right)$

wherein sigma²(k) Is the variance of the high frequency coefficients without noise interference,is the variance of the noise, and Y (k) is the high frequency coefficient of the noisy image. Sigma²(k) Is unknown, under the assumption that the variance of each point is uniformly changed, it can be estimated through the neighborhood, and the estimation formula is as follows:

${\mathrm{\&sigma;}}^2\left(k\right)=\max (0,\frac{1}{M}\underset{k\&Element;\mathrm{\&Omega;}\left(k\right)}{\&Sigma;}{Y}^2(k)-{\mathrm{\&sigma;}}_n^2)$

where Ω (k) is the neighborhood window, M is the number of high frequency coefficients in the neighborhood, Y (k) is the high frequency coefficients in the neighborhood,is the noise variance.

The derivation of the shrinkage coefficient requires the variance of the noise, which is unknown in practical applications, and thus the variance of the noise needs to be estimated. In the estimation of the noise variance, the invention adopts the following estimation method:

${\mathrm{\&sigma;}}_n^2=M_n^2/2$

whereinRepresents the mean of the pure background portion of the modulo-squared image, which is mainly based on the property that the noise of the modulo-squared image obeys non-central chi-square distribution (see document [ 9 ]) nowak rd].IEEETrans,onImageProcessing,10(10):1408-1419,1999.)。

The modulus squared image is obtained by squaring the modulus image. And extracting the pure background part, namely performing edge detection on the image by adopting a canny operator, and then scanning and communicating the detected edges, wherein the part outside the obtained communicated region is the pure background region of the image.

The method comprises the following specific steps:

1) and carrying out two-dimensional discrete orthogonal wavelet decomposition on the image.

2) Determining high-frequency shrinking coefficients of all points, and shrinking the high-frequency coefficients.

Step 1.3 fusion of threshold atrophy and proportional atrophy

From the single denoising effect, the image noise obtained based on the soft threshold atrophy method is filtered, but the image noise is not thorough enough, which means that the high-frequency coefficient is not sufficiently shrunk in the shrinking process; the noise of the de-noised image obtained by shrinkage of the scale is filtered thoroughly, but the problems of image blurring and excessive detail loss exist, which means that the high-frequency coefficient is excessively shrunk in the shrinking process. Therefore, the two can be endowed with different weights for fusion, thereby obtaining an ideal denoising effect.

In the determination of the weights of the two methods, because the proportional shrinkage method is premised on that the variance of the high-frequency coefficient of an actual signal is uniformly changed, gradient information is calculated on the estimated variance by using a sobel operator so as to judge the accuracy of the high-frequency coefficient after shrinkage. At places with larger gradients, the scale shrinkage precondition is not satisfied, so the coefficient reliability is not high, and the weight of the high-frequency coefficient estimated by the coefficient is low at the moment, while at places with uniform changes, the reliability is high and the weight is high.

The method comprises the following specific steps:

1) on the basis of the first two steps, the high-frequency coefficient matrix processed by the threshold collapsing method is recorded as A, and the high-frequency coefficient matrix processed by the proportional collapsing method is recorded as B.

2) And (3) calculating gradient information of the variance of the high-frequency coefficient of each point by using a sobel operator, normalizing the gradient information, wherein if the gradient value is greater than a threshold value T (empirical value), a corresponding position 0 is arranged in a matrix C (the matrix with the same size as A and used for storing the scale shrinkage credibility condition), and if not, the corresponding position is 1.

3) Calculating the high-frequency coefficient after combination:

R＝α*(B*C)+β*(A*C)+γ*(B*D)+*(A*D)

where α + β ═ 1, γ + ═ 1, D ═ 1-C, α, γ are empirical values, and the above x represents the dot product of the corresponding elements of the matrix.

4) And performing inverse transformation on the image to obtain a denoised image.

On the basis of a number of experiments, we chose T0.1, α 0.75, γ 0.45 and a neighborhood window of 3 x 3 for the empirical values that occurred in the above procedure.

Step 1.4 denoising Process of the present invention

The process of denoising a magnetic resonance image in the invention is shown in fig. 1, and the specific steps are as follows:

STEP 1: a magnetic resonance image containing noise is input.

STEP 2: and carrying out two-dimensional discrete orthogonal wavelet transform on the image containing the noise.

STEP 3: two sets of high-frequency coefficients are obtained by performing threshold shrinking treatment and proportional shrinking treatment on the high-frequency coefficients after STEP 2.

STEP 4: determining a fusion weight, and fusing the two high-frequency coefficients after STEP3 to obtain a new high-frequency coefficient.

STEP 5: and performing two-dimensional discrete orthogonal wavelet inverse transformation to obtain a denoised image.

STEP 6: and outputting the denoised magnetic resonance image.

Fig. 2 is a schematic diagram of denoising effect at 3% noise, in which (a) a magnetic resonance simulation image with 3% noise added; (b) denoising the image by a proportional shrinkage method; (c) denoising the image by a threshold shrinkage method; (d) and improving the denoised image of the algorithm.

Fig. 3 is a schematic diagram of denoising effect at 5% noise, wherein (a) a magnetic resonance simulation image with 5% noise added; (b) denoising the image by a proportional shrinkage method; (c) denoising the image by a threshold shrinkage method; (d) and improving the denoised image of the algorithm.

As can be seen from the comparison of the denoised images in fig. 2 and fig. 3, several methods can remove the noise to some extent. In contrast, the threshold shrinkage method is insufficient in denoising, partial noise (particularly obvious under 5% of noise) remains after denoising, the scale shrinkage method is over-blurred in denoising, more image detail information is lost, and the improved method not only enables the denoising to be sufficient, but also is ideal for preserving image details.

Table 1 shows the signal-to-noise ratio (SNR) of the denoising effect, and the calculation formula of the SNR is as follows:

$SNR=10{\log }_{10}\left(\frac{\mathrm{var}\left(x\right)}{\mathrm{var}(\hat{x}-x)}\right)$

where x is a simulated image without noise,the image is a noisy image or a denoised image.

From the point of view of the signal-to-noise ratio, the improved algorithm has a higher signal-to-noise ratio than the unmodified two algorithms.

TABLE 1 Signal-to-noise ratio (SNR) of denoising Effect

SNR (dB)/noise intensity	3％	5％
			Post-noise image	37.20	26.38
De-noised image after scale shrinkage	42.91	35.94
			Denoised image by threshold shrinkage method	42.96	30.50
Improved algorithm denoised image	44.82	37.32

Portions of the invention not disclosed in detail are well within the skill of the art.

Although illustrative embodiments of the present invention have been described above to facilitate the understanding of the present invention by those skilled in the art, it should be understood that the present invention is not limited to the scope of the embodiments, and various changes may be made apparent to those skilled in the art as long as they are within the spirit and scope of the present invention as defined and defined by the appended claims, and all matters of the invention which utilize the inventive concepts are protected.

Claims (2)

1. A magnetic resonance image denoising method based on wavelet shrinkage is characterized by comprising the following steps:

step 1.1, threshold atrophy of images

In the selection of the threshold function, the following formula is adopted:

$T_{ji}=\frac{1}{2^j-1}*{\mathrm{\&lambda;}}_{ji}*M_{ji}$

where j represents the number of wavelet layers, i represents the direction,the contrast of the image is represented by sigma, the standard deviation of the high-frequency coefficient is represented by mu, the mean value of the high-frequency coefficient is represented by mu, and M is the absolute median of the high-frequency coefficient; the threshold function is derived from the propagation characteristics of the noise and the actual signal; the energy of the actual signal is increased along with the increase of decomposition levels, and the noise is just opposite, so that the higher the level is, the smaller the threshold value is, and the high-frequency coefficient of the actual signal is fully reserved; the larger the contrast is, the larger the energy difference between the noise energy and the actual signal is, so that a larger threshold is selected to remove the noise as much as possible; the addition of M takes the statistical characteristics of the high-frequency coefficient into consideration;

the corresponding threshold processing function adopts a soft threshold, and the specific formula is as follows:

${\hat{w}}_{j,k}=\left\{\begin{array}{cc}\mathrm{sgn}\left(w_{j,k}\right)\left(\right|w_{j,k}|-T),& \left|w_{j,k}\right|\&GreaterEqual;T\\ 0,& \left|w_{j,k}\right|<T\end{array}\right.$

wherein T is a threshold value, w_j,kIs the observed high-frequency coefficient of the signal,is an estimate of the high frequency coefficient;

the method comprises the following specific steps:

1) performing two-dimensional discrete orthogonal wavelet decomposition on the image;

2) calculating a threshold value of the high-frequency coefficient in each layer and each direction;

3) carrying out threshold atrophy on the high-frequency coefficient by using a soft threshold;

step 1.2, the image is subjected to scale atrophy

Obeying normal distribution N (0, sigma) at high frequency coefficient of signal²) Noise obeys normal distributionUnder the assumption that the actual signal is estimated according to the minimum mean square error:

$\hat{X}\left(k\right)=\frac{{\mathrm{\&sigma;}}^2\left(k\right)}{{\mathrm{\&sigma;}}^2\left(k\right)+{\mathrm{\&sigma;}}_n^2}Y\left(k\right)$

wherein σ²(k) Is the variance of the high frequency coefficients without noise interference,is the variance of the noise, and Y (k) is the high-frequency coefficient of the noise-containing image; sigma²(k) Is unknown, under the assumption that the variance of each point is uniformly changed, it can be estimated through the neighborhood, and the estimation formula is as follows:

${\mathrm{\&sigma;}}^2\left(k\right)=\max (0,\frac{1}{H}\underset{k\&Element;\mathrm{\&Omega;}\left(k\right)}{\&Sigma;}{Y}^2(k)-{\mathrm{\&sigma;}}_n^2)$

wherein Ω (k) is the neighborhood window, M is the number of high frequency coefficients in the neighborhood, Y (k) is the high frequency coefficients in the neighborhood, σ_n ²Is the noise variance;

the variance of noise is required for solving the shrinkage coefficient, and in practical application, the variance of noise is unknown, so that the variance of noise needs to be estimated; on the estimation of the noise variance, the following estimation method is adopted:

${\mathrm{\&sigma;}}_n^2=M_n^2/2$

wherein,the mean value represents the pure background part of the modular square image and is mainly obtained according to the characteristic that the noise of the modular square image obeys non-central chi-square distribution;

the module square image can be obtained by squaring the module image; extracting a pure background part, namely, performing edge detection on the image by adopting a canny operator, and then scanning and communicating the detected edge to obtain a part outside a communicated region, namely the pure background region of the image;

the method comprises the following specific steps:

1) performing two-dimensional discrete orthogonal wavelet decomposition on the image;

2) determining high-frequency shrinking coefficients of all points, and shrinking the high-frequency coefficients;

step 1.3 fusion of threshold atrophy and proportional atrophy

From the single denoising effect, the image noise obtained based on the soft threshold atrophy method is filtered, but the image noise is not thorough enough, which means that the high-frequency coefficient is not sufficiently shrunk in the shrinking process; the noise of the de-noised image obtained by shrinkage of the proportion is filtered thoroughly, but the problems of image blur and excessive detail loss exist, which means that in the shrinkage process, the high-frequency coefficient is excessively shrunk, so that the de-noised image and the high-frequency coefficient can be endowed with different weights for fusion, and an ideal de-noising effect is obtained;

in the determination of the weights of the two methods, because the proportional shrinkage method is premised on that the variance of the high-frequency coefficient of the actual signal is uniformly changed, gradient information is calculated on the estimated variance by using a sobel operator so as to judge the accuracy of the high-frequency coefficient after shrinkage; in places with larger gradients, the scale shrinkage precondition is not satisfied, so the reliability of the coefficient is not high, and the weight of the high-frequency coefficient estimated by the coefficient is low, while in places with uniform changes, the reliability is high and the weight is high;

the method comprises the following specific steps:

1) on the basis of the first two steps, recording a high-frequency coefficient matrix processed by a threshold collapsing method as A, and recording a high-frequency coefficient matrix processed by a proportional collapsing method as B;

2) calculating gradient information of the high-frequency coefficient variance of each point by using a sobel operator, and normalizing the gradient information, wherein the gradient value is greater than a threshold value T, the T is an empirical value, the corresponding position of the matrix C is 0, otherwise, the matrix C is 1, and the matrix C has the same size as the matrix A and is used for storing the scale shrinkage credibility condition;

3) calculating the fused high-frequency coefficient:

R＝α*(B*C)+β*(A*C)+γ*(B*D)+*(A*D)

where α + β ═ 1, γ + ═ 1, D ═ 1-C, α, γ are empirical values, and the above x represents the dot product of the corresponding elements of the matrix;

4) and performing inverse transformation on the image coefficient to obtain a denoised image.

2. The method of claim 1, wherein T is 0.1, α is 0.75, γ is 0.45, and the neighborhood window is 3 x 3, based on a large number of experiments.