WO2014082441A1 - Noise elimination method and apparatus - Google Patents

Abstract

Embodiments of the present invention provide a noise elimination method and apparatus. The method comprises: based on a mixed noise model, acquiring a parameter estimated value of a noise standard deviation function of a first signal of the noise to be eliminated, so as to obtain an estimated noise standard deviation function; performing variance stabilizing transformation on the first signal according to the estimated noise standard deviation function, so as to obtain a second signal of the noise serving as the signal-independent noise; performing denoising on the second signal; performing inverse transformation to the variance stabilizing transformation on the denoised second signal, so as to complete noise elimination on the first signal. The noise elimination method and apparatus according to the embodiments of the present invention can implement effective denosing on the mixed noise comprising a signal-dependent noise component and a signal-independent noise component.

Description

 The present invention claims priority to Chinese Patent Application No. 201210504221.3, entitled "Noise Cancellation Method and Apparatus", filed on November 30, 2012, the entire contents of which are incorporated by reference. In this application. Technical field

 The present invention relates to image processing technology, and in particular, to a noise cancellation method and apparatus, and belongs to the field of communication technologies. Background technique

 Image denoising has always been a hot research issue in the field of image processing.

 Traditional image denoising methods are mostly based on the idea of spatial local filtering, such as mean filtering, Gaussian filtering, and bilateral filtering. The spatial local filtering is based on the assumption that the pixel points adjacent to the spatial position generally have relatively similar gray values, and the noise is removed by weighting the gray value of the pixel points and the gray value of the neighboring pixel points. The assumption that the pixel points adjacent to the spatial position generally have relatively similar gray values is only applicable to the smooth portion of the image, but not to the detail portion of the image (edge, strong texture region, etc.), so the local filtering method easily leads to image details. Lost.

 In 2005, Buades et al. proposed a non-local mean image denoising method. The non-local mean denoising algorithm mainly uses the redundant information of a large number of self-similar blocks in the digital image, and calculates the similarity measure of the neighborhood of the pixel to be denoised and the pixel of the search area to calculate the pixel of the search area. The similarity weight with the pixel to be denoised, and then weighted and averaged the pixels in the search area, thereby calculating a new gray value of the pixel to be denoised. The algorithm has very good effects on denoising performance and image texture and edge information preservation, but it is based on signal-independent Gaussian noise assumption.

Existing denoising techniques are mostly based on signal-independent Gaussian noise models. In some practical applications, image noise can be mixed noise of Gaussian noise and signal-related noise. For example, for sensor imaging, the noise model has both signal-independent Gaussian noise components and signal-related The Poisson noise component, therefore, directly using the denoising method based on the Gaussian noise hypothesis, can not achieve effective denoising of such mixed noise including both signal-related noise components and signal-independent noise components. SUMMARY OF THE INVENTION In view of the deficiencies in the prior art, embodiments of the present invention provide a noise cancellation method and apparatus for implementing effective denoising of mixed noise including both signal-related noise components and signal-independent noise components.

 In a first aspect, a noise cancellation method is provided, including:

 Obtaining a parameter estimation value of a noise standard deviation function of the first signal of the noise to be cancelled based on the mixed noise model to obtain an estimated noise standard deviation function;

 Performing a variance stabilization transformation on the first signal according to the estimated noise standard deviation function to obtain a second signal whose noise is signal-independent noise;

 Denoising the second signal;

 Performing an inverse transform of the variance stabilization transform on the denoised second signal to complete noise cancellation on the first signal.

In a first possible implementation manner of the first aspect, the variance stabilization transform is implemented by using the following formula:

 Where (for the estimated noise standard deviation function, c is the transformed constant standard deviation, the current pixel gray value before the transform, ^s^) is the current pixel gray value after the transform.

 In a first possible implementation manner of the first aspect, the acquiring, by the hybrid noise model, a parameter estimation value of a noise standard deviation function of the first signal of the noise to be cancelled, to obtain an estimated noise standard deviation function, includes:

 Performing wavelet domain analysis on the first signal to obtain a (χ, σ) scatter plot;

 A random sampling consistency algorithm RANSAC is used to perform curve fitting on the (χ, σ) scatter plot to obtain a first noise parameter a and a second noise parameter b, and:

The original noise-free signal corresponding to the first signal is the estimated noise standard deviation function. In conjunction with the first aspect or the first or second possible implementation manner of the first aspect, in a third possible implementation manner of the first aspect, the performing the demodulating the second signal, Traversing the pixels of the second signal:

 Determining, by a ratio of a grayscale mean of a neighborhood of the pixel point i to be denoised and a neighborhood of the pixel point j of the search area, whether the difference from 1 is less than or equal to a preset difference; and determining the pixel point i and Whether the angle of the gradient direction of the pixel point j is less than or equal to a preset angle; wherein i and j are both natural numbers;

 If at least one of the two is judged as no, the neighborhood similarity of the pixel point i and the pixel point j is determined to be 0;

 If both are judged as YES, the neighborhood similarity between the pixel point i and the pixel point j is calculated according to a preset formula;

 The denoised gray value of the pixel point i is calculated according to the neighborhood similarity and the gray value of the pixel point j.

 In conjunction with the first aspect or the first or second possible implementation manner of the first aspect, in a fourth possible implementation manner of the first aspect, the performing, Traversing the pixels of the second signal:

 The neighboring window of the pixel point i to be denoised and the pixel point j of the search area is downsampled; wherein i and j are natural numbers;

 Calculating a neighborhood similarity between the pixel point i and the pixel point j according to a gray value of a downsampled neighborhood of the pixel point i and a gray value of a downsampled neighborhood of the pixel point j;

 The denoised gray value of the pixel point i is calculated according to the neighborhood similarity and the gray value of the pixel point j.

 In a second aspect, a noise canceling apparatus is provided, including:

 An estimation module, configured to acquire a parameter estimation value of a noise standard deviation function of the first signal of the noise to be cancelled based on the mixed noise model, to obtain an estimated noise standard deviation function;

 a variance stabilization transformation module, configured to perform a variance stabilization transformation on the first signal according to the estimated noise standard deviation function, to obtain a second signal whose noise is signal-independent noise; and a denoising module, configured to The two signals are denoised;

The variance stabilization inverse transform module performs inverse transformation of the variance stabilization transform on the denoised second signal to complete noise cancellation on the first signal. In a first possible implementation manner of the second aspect, the variance stabilization transformation is implemented by using the following formula:

Where ( ^x ) is the estimated noise standard deviation function, c is the transformed constant standard deviation, is the current pixel gray value before the transform, and ^s^) is the transformed current pixel gray value.

 In a second possible implementation manner of the second aspect, the estimating module is configured to: perform wavelet domain analysis on the first signal, and obtain (X, ^ scatter diagram;

 A random sampling consistency algorithm RANSAC is used to perform curve fitting on the (χ, σ) scatter plot to obtain a first noise parameter a and a second noise parameter b, and:

 Wherein, the original noiseless signal corresponding to the first signal is the estimated noise standard deviation function.

 With reference to the second aspect, or the first or the second possible implementation manner of the second aspect, in a third possible implementation manner of the second aspect, the the denoising module is configured to traverse the second signal in the following manner Each pixel point:

 Determining, by a ratio of a grayscale mean of a neighborhood of the pixel point i to be denoised and a neighborhood of the pixel point j of the search area, whether the difference from 1 is less than or equal to a preset difference; and determining the pixel point i and Whether the angle of the gradient direction of the pixel point j is less than or equal to a preset angle; wherein i and j are both natural numbers;

 If at least one of the two is judged to be no, the neighborhood similarity of the pixel point i and the pixel point j is determined to be 0;

 If both are judged as YES, the neighborhood similarity between the pixel point i and the pixel point j is calculated according to a preset formula;

 The denoised gray value of the pixel point i is calculated according to the neighborhood similarity and the gray value of the pixel point j.

 With reference to the second aspect, or the first or the second possible implementation manner of the second aspect, in a fourth possible implementation manner of the second aspect, the the denoising module is configured to traverse the second signal in the following manner Each pixel point:

Downsampling the neighborhood window of the pixel point i to be denoised and the pixel point j of the search area; wherein i and j are both natural numbers; Calculating a neighborhood similarity between the pixel point i and the pixel point j according to a gray value of a downsampled neighborhood of the pixel point i and a gray value of a downsampled neighborhood of the pixel point j;

 The denoised gray value of the pixel point i is calculated according to the neighborhood similarity and the gray value of the pixel point j.

 According to the noise cancellation method and apparatus provided by the embodiment of the present invention, the noise standard deviation function is estimated by using the mixed noise model, and the noise standard deviation function is used to perform the variance stabilization transformation to remove the noise signal. It is converted to a signal with signal-independent noise so that it can be denoised using any denoising method based on signal-independent noise assumptions. Therefore, an effective cancellation of mixed noise including both signal-dependent noise components and signal-independent noise components is achieved. DRAWINGS

 In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the embodiments or the description of the prior art will be briefly described below, and obviously, the attached in the following description The drawings are only some of the embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive labor.

 1 is a schematic flow chart of a noise cancellation method according to an embodiment of the present invention;

 2 is a schematic diagram showing the principle of noise parameter estimation based on wavelet domain analysis;

 Figure 3 is an example of a (χ, σ) scatter plot;

 4 is a schematic diagram showing the results of curve fitting directly to the (χ, σ) scatter plot shown in FIG. 3; FIG. 5 is a schematic diagram showing the result of curve fitting using the RANSAC pair (χ, σ) scatter plot; A schematic structural diagram of a noise canceling apparatus according to an embodiment of the present invention;

 FIG. 7 is a schematic structural diagram of a noise canceling apparatus according to another embodiment of the present invention. detailed description

 FIG. 1 is a schematic flow chart of a noise cancellation method according to an embodiment of the present invention. As shown in Figure 1, the noise cancellation method includes:

 101. Acquire, according to the mixed noise model, a parameter estimation value of a noise standard deviation function of the first signal of the noise to be cancelled, to obtain an estimated noise standard deviation function;

102. Perform variance reconstruction on the first signal according to an estimated noise standard deviation function. Transforming to obtain a second signal whose noise is signal-independent noise;

 103. Perform denoising on the second signal.

 104. Perform inverse transformation of the variance stabilization transform on the denoised second signal to complete noise cancellation on the first signal.

 In the following, the above-mentioned steps 101-104 will be described in detail by taking the mixed noise model of sensor imaging as an example.

 Specifically, the mixed noise model for sensor imaging is a Poisson-Gaussian mixed noise model, as shown in the following equation (1):

y = x + v _p (x) + v _G Formula ( ₁₎ where y is the signal containing the mixed noise (ie the first signal), eg the image imaged by the sensor, X is the original noiseless signal, is signal dependent The Poisson noise component is a signal-independent Gaussian noise component.

 For the Poisson noise component, the following formulas (2) and (3) are satisfied:

 —(x + Vp) ~ P (—x)

 a a formula (2)

Var^) = ax Equation (3) where, v _ar (o is the noise variance of the Poisson noise component. For Gaussian noise components, the following equations (4) and (5) are satisfied.

 Formula (4)

^o) = b Equation ( ₅₎ where Var ^ is the noise variance of the Gaussian noise component.

 Therefore, the noise variance of the mixed noise satisfies the following formula (6):

σ ² =αχ + δ Equation (6) where σ ² is the noise variance of the mixed noise; a and b are noise parameters, which can be implemented in any way, for example, based on wavelet domain analysis for noise parameter estimation.

Figure 2 is a schematic diagram of the principle of noise parameter estimation based on wavelet domain analysis. As shown in FIG. 2, a two-dimensional discrete wavelet transform is performed on the image y to obtain four different wavelet coefficient subgraphs of W _A , W _H , W _v and W _D , where W _A is an approximate coefficient and W _H is a horizontal detail. W _v is the vertical detail, W _D is the diagonal detail; the gradient of each pixel in the \\^ subgraph is compared with the preset gradient value, and the pixel with the larger gradient in the W _A subgraph is removed, leaving Smooth area; image to be included in the smooth area The prime points are divided into N level sets S ₂ , ..., S _N according to the gray scale, wherein different level sets correspond to different gray scale ranges; in each level set (i=[ l, N] ) Using W _{A to} estimate x, using W _D estimation and using curve fitting, the obtained (χ, σ) scatter plot is fitted to obtain the noise parameter _a and .

 In addition, since the noise parameters a and b are fixed for the fixed sensor, the noise parameters a and b can be estimated based on some images taken by the sensor, and the estimated noise parameters a and b are stored for execution. Called directly after subsequent image denoising.

 Based on the estimated noise parameters a and b, the estimated noise standard deviation function as shown in the following formula (7) can be obtained:

 σ(χ) = fax + b Equation (7) After obtaining the noise standard deviation function, the image y is subjected to Variance Stabilizing Transformation (VST). Among them, the effect of the variance stabilization transformation is to convert the noise of signal correlation and variance variation into noise with signal independence and constant variance through nonlinear mapping of gray values, and the noise of the transformed image (ie, the second signal) can be It is considered to be Gaussian noise.

 The variance stabilization transformation is achieved, for example, by the following formula (8):

, VST ( ⁼ " -dx

ά^ Equation (8) where ^χ) is the estimated noise standard deviation function; c is the transformed constant standard deviation, which can be any value greater than 0, for example set to 0.01; t is the current pixel point, execution variance Stabilize the gray value before the transform; / _VST (0 is the current pixel point, and the gray value after the variance is stabilized.

 After performing the variance stabilization transformation, any denoising method based on Gaussian noise can be used for denoising, for example, an image denoising method using non-local mean.

 After the denoising is completed, the inverse transform of the variance stabilization transform is used to map the grayscale values back, thereby obtaining an image after the noise is cancelled for the image y, and thus, the denoising processing for the image y with the mixed noise is completed.

 Wherein, the inverse transformation of the variance stabilization transformation and the variance stabilization transformation satisfy the following formula (9):

IVST( ^ /VST( Equation ( 9 ) For the Poisson-Gaussian mixed noise model, the inverse transformation of the variance stabilization transformation is eg The following formulas (10) and (11 formulas (10)

Formula (11) It should be noted that in the above process, although the Poisson-Gaussian mixed noise of the sensor imaging is taken as an example, the specific process of the noise canceling method of the above embodiment is explained, but those skilled in the art can It is understood that noise cancellation can be performed by the noise cancellation method of the above embodiment for any mixed noise that satisfies the monotonic function whose noise variance is the signal strength.

 In addition, the specific implementation of the variance stabilization transformation may also take other manners different from the formula (8), as long as the signal correlation variance can be changed to a signal-independent variance, which is not limited in the embodiment of the present invention, for example, the following formula is adopted. Implement the variance stabilization transformation:

According to the noise canceling method of the above embodiment, the noise standard deviation function is estimated by using the mixed noise model, and the noise standard deviation function is used to perform the variance stabilization transform, and the signal to be cancelled is converted into a signal. A noise-independent signal that can be denoised using any denoising method based on signal-independent noise assumptions. Therefore, an effective cancellation of mixed noise including both signal-dependent noise components and signal-independent noise components is achieved.

 Further, on the basis of the noise canceling method of the above embodiment, the process of obtaining the parameter estimated value of the noise standard deviation function of the first signal to be cancelled is optimized. Specifically, the noise parameter estimation is performed based on the wavelet domain analysis, and after obtaining the (χ, σ) scattergram, the random sampling consistency algorithm (RANSAC) is used to perform curve fitting on the (χ, σ) scatter plot to obtain Noise parameters a and b.

 More specifically, curve fitting a (χ, σ) scatter plot using RANSAC includes the following steps:

 Step 1. Randomly select two points and fit the noise model curve;

 Step 2: Check whether the fitted noise model curve satisfies "≥0 and 6≥0. If not, repeat step 1;

Step 3. In all points, the point where the distance from the fitted curve is less than the given threshold is calculated. The number of such points is referred to as the inner point in the text. If the number of internal points exceeds the number of internal points obtained by the previous iteration, it is recorded as the maximum number of internal points.

 Step 4. Calculate the number of iterations required based on the maximum number of interior points.

 Step 5. Iterate steps 1 to 4 until the number of iterations is sufficient.

 Step 6. Re-model the model with all the interior points in the noise model curve with the largest number of points.

 Figure 3 is an example of a (χ, σ) scatter plot; Figure 4 is a schematic diagram of the result of curve fitting directly to the (χ, σ) scatter plot shown in Figure 3; Figure 5 is a RANSAC pair (χ, σ) Schematic diagram of the results of curve fitting for scatter plots. It can be seen from Fig. 3-5 that the curve fitting of the (χ, σ) scatter plot by RANSAC significantly improves the robustness of the noise parameter estimation, thus effectively improving the noise cancellation performance.

 Further, on the basis of the noise canceling method of the above embodiment, the process of performing the denoising of the signal (i.e., the second signal) after performing the variance stabilization conversion is optimized.

 Specifically, the existing non-local mean image denoising method is used to perform denoising on the signal after performing the variance stabilization transformation, by establishing a neighborhood of the pixel to be denoised and the pixel of the search area. The similarity measure calculates the similarity weight of each pixel in the search area and the pixel to be denoised, and then performs weighted averaging on the pixels in the search area, thereby calculating a new gray value of the pixel to be denoised, that is, going Gray value after noise.

 The gray value after denoising the pixel to be denoised is calculated by the following formula (12):

 NL(v)( =∑w( , 7)v(i)

 ^ Equation (12) where pixel i is the pixel to be denoised; pixel j is the pixel in the search region; I is the pixel of the image after performing the variance-constant transformation; ν (for pixel j Gray value

 The gray value after denoising the pixel point i; κ is the neighborhood similarity of the pixel point i with the pixel point j in the J pixel points of the search area; ^) is calculated by the following formula (13):

₁ ||ν(Ν _; ·)— v(N ||l.

w(i, j) = e ^hl

 z« Equation (13) where N is the neighborhood of pixel i; ^ is the neighborhood of pixel j; h is the filter depth control parameter, which is generally determined by the noise variance; Z (0 is the normalization constant, satisfied) The following formula (14): Z() = ∑w( , 7)

1 formula (14) In the embodiment of the present invention, the following two schemes for optimizing the image denoising method based on the non-local mean are provided:

 Option 1: Optimize the number of times the neighborhood similarity is calculated;

 Determining, by the ratio of the grayscale mean of the neighborhood of the pixel point i to be denoised and the neighborhood of the pixel j of the search area, whether the difference from 1 is less than or equal to the preset difference; and determining the pixel point i and the pixel point j Whether the angle of the gradient direction is less than or equal to the preset angle;

 If at least one of the two is judged as no, the neighborhood similarity between the pixel point i and the pixel point j is determined to be 0;

 If both are judged as YES, the neighborhood similarity between the pixel point i and the pixel point j is calculated according to a preset formula;

 The denoised gray value of the pixel point i is calculated according to the neighborhood similarity and the gray value of the pixel point j.

 Specifically, in the first scheme, the ratio of the gray mean value of the neighborhood of the pixel point i to the neighborhood of the pixel point j is defined, as in the following formula (15):

 w(N,)

 Equation (15) where N is the neighborhood of pixel i; ^ is the neighborhood of pixel j; N,) is the gray mean of the neighborhood of pixel i; ^.) is the neighborhood of pixel j Gray mean value; ^, is the ratio of the gray mean of ^ and ^;

Also defining the angle between the pixel point i and the gradient direction of the pixel point j, as expressed by the following formula (16)

Equation ( ₁₆₎ where G (0 is the gradient direction of the pixel point i; is the gradient direction of the pixel point j; ^·, ·) is the angle between the pixel point i and the gradient direction of the pixel point j.

For a pixel with a large difference in gray mean values in the neighborhood or a large difference in the gradient direction, the neighborhood similarity is not calculated and is directly considered to be 0. Based on this idea, combined with the above definition, the neighborhood similarity calculation formula shown in equation (13) is reduced to the following formula (17): when _ηι ≤ ι^, ≤ η ₂ Ά θ , β ≤ ζ

Other formulas ( 17) Where η ;; ₂ and ^ are both values set in advance according to needs, where ^ is a positive number less than 1, for example set to 0.9; a value greater than 1, for example set to 1.1; an angle less than or equal to 90 degrees The value, for example, is set to 60 degrees.

 Through the foregoing scheme 1, the number of times of calculating the neighborhood similarity can be effectively reduced.

 Option 2: Optimize the complexity of calculating neighborhood similarity;

 Decoding the denoised pixel point i and the neighborhood window of the pixel j of the search area; according to the gray value of the downsampled neighborhood of the pixel point i, and the gray value of the downsampled neighborhood of the pixel point j, The neighborhood similarity between the pixel point i and the pixel point j is calculated.

 Specifically, in the second scheme, the calculation is performed by downsampling of the neighborhood window. Due to the continuity of the image, the weighted distance of the downsampled neighborhood approximates the weighted distance of the original neighborhood.

 Based on this idea, the neighborhood similarity calculation formula shown in equation (13) is reduced to the following formula (18):

||ν( -Ν, )-ν( -Ν, )|| _α

Where Ν represents the downsampling of the neighborhood window.

 Through the above scheme 2, the computational complexity of the neighborhood similarity can be greatly reduced.

 The foregoing solution 1 and solution 2 may be used alone or in combination, and are limited in the embodiment of the present invention.

 FIG. 6 is a schematic structural diagram of a noise canceling apparatus according to an embodiment of the present invention. As shown in FIG. 6, the noise canceling device 60 includes:

 The estimating module 61 is configured to obtain, according to the mixed noise model, a parameter estimation value of a noise standard deviation function of the first signal of the noise to be cancelled, to obtain an estimated noise standard deviation function;

 The variance stabilization conversion module 62 is configured to perform a variance stabilization transformation on the first signal according to the estimated noise standard deviation function to obtain a second signal whose noise is signal-independent noise; and a denoising module 63, configured to Decoding the second signal;

 The variance stabilization inverse transform module 64 performs inverse transformation of the variance stabilization transform on the denoised second signal to complete noise cancellation on the first signal.

 The specific flow of the noise canceling apparatus of the above embodiment for eliminating noise is the same as that of the noise canceling method of the above embodiment, and therefore will not be described herein.

The noise canceling apparatus according to the above embodiment, based on the mixed noise model, has a mixture The noise-compensated signal is used to estimate the noise standard deviation function, and the noise standard deviation function is used to perform the variance stabilization transformation, and the signal to be cancelled is converted into a signal with signal-independent noise, thereby being able to utilize any signal-independent noise-based denoising. The method performs denoising. Therefore, an effective cancellation of mixed noise including both signal-dependent noise components and signal-independent noise components is achieved.

 Further, in the noise canceling apparatus of the above embodiment, the variance stabilization conversion is realized by the following formula:

Where ( ^x ) is the estimated noise standard deviation function, c is the transformed constant standard deviation, and is the current pixel gray value before the transformation, and is the current pixel gray value after the transformation.

 Further, in the noise canceling apparatus of the above embodiment, the estimating module is configured to: perform wavelet domain analysis on the first signal, and acquire a (χ, σ) scattergram;

 A random sampling consistency algorithm RANSAC is used to perform curve fitting on the (χ, σ) scatter plot to obtain a first noise parameter a and a second noise parameter b, and:

 Wherein, the original noiseless signal corresponding to the first signal is the estimated noise standard deviation function.

 Further, in the noise canceling apparatus of the above embodiment, the denoising module is configured to traverse the pixels of the second signal in the following manner:

 Determining, by the ratio of the grayscale mean of the neighborhood of the pixel point i to be denoised and the neighborhood of the pixel j of the search area, whether the difference from 1 is less than or equal to the preset difference; and determining the pixel point i and the pixel point j Whether the angle of the gradient direction is less than or equal to a preset angle; wherein i and j are natural numbers;

 If at least one of the two is judged as no, the neighborhood similarity between the pixel point i and the pixel point j is determined to be 0;

 If both are judged as YES, the neighborhood similarity between the pixel point i and the pixel point j is calculated according to a preset formula;

 The denoised gray value of the pixel point i is calculated according to the neighborhood similarity and the gray value of the pixel point j.

Further, in the noise canceling apparatus of the above embodiment, the denoising module is configured to traverse the pixels of the second signal in the following manner: Downsampling the neighborhood window of the pixel point i to be denoised and the pixel point j of the search area; wherein i and j are both natural numbers;

 Calculating the neighborhood similarity between the pixel point i and the pixel point j according to the gray value of the downsampling neighborhood of the pixel point i and the gray value of the downsampling neighborhood of the pixel point j;

 The denoised gray value of the pixel point i is calculated according to the neighborhood similarity and the gray value of the pixel point j.

 FIG. 7 is a schematic structural diagram of a noise canceling apparatus according to another embodiment of the present invention. As shown in Figure 7, the noise canceling device 70 includes a memory 71 and a processor 72, wherein:

 A set of program codes is stored in the memory 71, and the processor 72 is configured to call the program code stored in the memory 71 for performing the following operations:

 Obtaining a parameter estimation value of a noise standard deviation function of the first signal of the noise to be cancelled based on the mixed noise model to obtain an estimated noise standard deviation function;

 Performing a variance stabilization transformation on the first signal according to the estimated noise standard deviation function to obtain a second signal whose noise is signal-independent noise;

 Denoising the second signal;

 Performing an inverse transform of the variance stabilization transform on the denoised second signal to complete noise cancellation on the first signal.

 The specific flow of the noise canceling apparatus of the above embodiment for eliminating noise is the same as that of the noise canceling method of the above embodiment, and therefore will not be described herein.

 According to the noise canceling apparatus of the above embodiment, the noise standard deviation function is estimated by the mixed noise model based on the mixed noise model, and the variance standardization function is performed by using the noise standard deviation function to convert the signal to be cancelled into a signal. A noise-independent signal that can be denoised using any denoising method based on signal-independent noise assumptions. Therefore, an effective cancellation of mixed noise including both signal-dependent noise components and signal-independent noise components is achieved.

It will be understood by those skilled in the art that all or part of the steps of implementing the above method embodiments may be performed by hardware related to the program instructions. The aforementioned program can be stored in a computer readable storage medium. The program, when executed, performs the steps including the foregoing method embodiments; and the foregoing storage medium includes: a medium that can store program codes, such as a ROM, a RAM, a magnetic disk, or an optical disk. It should be noted that the above embodiments are only for explaining the technical solutions of the present invention, and are not intended to be limiting; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those skilled in the art that: The technical solutions described in the foregoing embodiments are modified, or some of the technical features are equivalently replaced. The modifications and substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

Claim

A noise cancellation method, comprising:

 Obtaining a parameter estimation value of a noise standard deviation function of the first signal of the noise to be cancelled based on the mixed noise model to obtain an estimated noise standard deviation function;

 Performing a variance stabilization transformation on the first signal according to the estimated noise standard deviation function to obtain a second signal whose noise is signal-independent noise;

 Denoising the second signal;

 Performing an inverse transform of the variance stabilization transform on the denoised second signal to complete noise cancellation on the first signal.

2. The noise canceling method according to claim 1, wherein the variance stabilization transform is implemented by the following formula:

 Wherein, for the estimated noise standard deviation function, c is the transformed constant standard deviation, which is the current pixel gray value before the transformation, and is the current pixel gray value after the transformation.

 The noise cancellation method according to claim 1, wherein the parameter estimation value of the noise standard deviation function of the first signal of the noise to be cancelled is obtained based on the mixed noise model to obtain an estimated noise standard deviation function. , including:

 Performing a wavelet domain analysis on the first signal to obtain a (χ, σ) scattergram;

 A random sampling consistency algorithm RANSAC is used to perform curve fitting on the (χ, σ) scatter plot to obtain a first noise parameter a and a second noise parameter b, and:

 Wherein, the original noiseless signal corresponding to the first signal is the estimated noise standard deviation function.

 The noise canceling method according to any one of claims 1 to 3, wherein the denoising the second signal comprises traversing each pixel of the second signal in the following manner:

Determining, by a ratio of a grayscale mean of a neighborhood of the pixel point i to be denoised and a neighborhood of the pixel point j of the search area, whether the difference from 1 is less than or equal to a preset difference; and determining the pixel point i and Whether the angle of the gradient direction of the pixel point j is less than or equal to a preset angle; wherein i and j are both natural Number

 If at least one of the two is judged as no, the neighborhood similarity of the pixel point i and the pixel point j is determined to be 0;

 If both are judged as YES, the neighborhood similarity between the pixel point i and the pixel point j is calculated according to a preset formula;

 The denoised gray value of the pixel point i is calculated according to the neighborhood similarity and the gray value of the pixel point j.

 The noise canceling method according to any one of claims 1-3, wherein the denoising the second signal comprises traversing each pixel of the second signal in the following manner:

 The neighboring window of the pixel point i to be denoised and the pixel point j of the search area is downsampled; wherein i and j are natural numbers;

 Calculating a neighborhood similarity between the pixel point i and the pixel point j according to a gray value of a downsampled neighborhood of the pixel point i and a gray value of a downsampled neighborhood of the pixel point j;

 The denoised gray value of the pixel point i is calculated according to the neighborhood similarity and the gray value of the pixel point j.

 6. A noise canceling device, comprising:

 An estimation module, configured to acquire a parameter estimation value of a noise standard deviation function of the first signal of the noise to be cancelled based on the mixed noise model, to obtain an estimated noise standard deviation function;

 a variance stabilization transformation module, configured to perform a variance stabilization transformation on the first signal according to the estimated noise standard deviation function to obtain a second signal whose noise is signal-independent noise;

 a denoising module, configured to perform denoising on the second signal;

 The variance stabilization inverse transform module performs inverse transformation of the variance stabilization transform on the denoised second signal to complete noise cancellation on the first signal.

7. The noise canceling apparatus according to claim 6, wherein the variance stabilization conversion is implemented by the following formula:

 Wherein, for the estimated noise standard deviation function, c is the transformed constant standard deviation, which is the current pixel gray value before the transformation, and is the current pixel gray value after the transformation.

8. The noise canceling apparatus according to claim 6, wherein said estimating mode The block is used to:

 Perform wavelet domain analysis on the first signal to obtain a (χ, σ) scatter plot;

 A random sampling consistency algorithm RANSAC is used to perform curve fitting on the (χ, σ) scatter plot to obtain a first noise parameter a and a second noise parameter b, and:

Wherein, the original noise-free signal corresponding to the first signal, ( ^x ) is the estimated noise standard deviation function.

 The noise canceling apparatus according to any one of claims 6-8, wherein the denoising module is configured to traverse each pixel of the second signal in the following manner:

 Determining, by a ratio of a grayscale mean of a neighborhood of the pixel point i to be denoised and a neighborhood of the pixel point j of the search area, whether the difference from 1 is less than or equal to a preset difference; and determining the pixel point i and Whether the angle of the gradient direction of the pixel point j is less than or equal to a preset angle; wherein i and j are both natural numbers;

 If at least one of the two is judged to be no, the neighborhood similarity of the pixel point i and the pixel point j is determined to be 0;

 If both are judged as YES, the neighborhood similarity between the pixel point i and the pixel point j is calculated according to a preset formula;

 The denoised gray value of the pixel point i is calculated according to the neighborhood similarity and the gray value of the pixel point j.

 The noise canceling apparatus according to any one of claims 6-8, wherein the denoising module is configured to traverse each pixel of the second signal in the following manner:

 The neighboring window of the pixel point i to be denoised and the pixel point j of the search area is downsampled; wherein i and j are natural numbers;

 Calculating a neighborhood similarity between the pixel point i and the pixel point j according to a gray value of a downsampled neighborhood of the pixel point i and a gray value of a downsampled neighborhood of the pixel point j;

 The denoised gray value of the pixel point i is calculated according to the neighborhood similarity and the gray value of the pixel point j.