CN104504652A - Image denoising method capable of quickly and effectively retaining edge and directional characteristics - Google Patents

Abstract

The invention discloses an image denoising method capable of quickly and effectively retaining edge and directional characteristics. After an image is subjected to wavelet transform, a low-frequency subimage centralizes most energy noise in an original image and is subjected to secondary Wiener filtering to highlight the details of a fusion image; since the image noise is mainly centralized in wavelet high-frequency subimage parts in three different directions, the coefficient of the image noise is small, denoising processing is carried out by singular value decomposition, a high-frequency diagonal subimage is subjected to the singular value decomposition with a high-frequency subimage in line direction or a high-frequency subimage in column direction after rotating to the line direction or the column direction, and meanwhile, the high-frequency subimage is subjected to edge extraction and retaining in order to avoid losing edge detail information; and finally, the denoised low-frequency and high-frequency subimages are subjected to inverse wavelet transform to reconstruct a final denoising image. Singular value numbers required for singular value reconstruction images are jointly determined through a peak signal to noise ratio of the image and a traditional method.

Description

A kind of effective preserving edge and directional characteristic image de-noising method fast

Technical field

The invention belongs to Signal and Information Processing field, particularly a kind of image de-noising method being applicable to the noisy image that image occurs in generation or transmitting procedure.

Background technology

Image generation or transmitting procedure in because being subject to interference and the impact of various noise, inevitably go out item to degrade phenomenon, there is the problem such as comparative poor of edge fog in various degree, local and entirety, this process to successive image (as segmentation, compressing and image understanding etc.) can have a negative impact.Therefore carrying out denoising to image and improve picture quality, is a basis in image procossing and important work.

Wavelet transformation have employed the method for multiresolution, have low entropy, decorrelation and select the dirigibility of base, and the noise information of image mainly concentrates on the HFS of its wavelet field, and wavelet theory therefore can be utilized signal and noise to be separated.The basic thought of wavelet shrinkage method denoising is: image has different distribution characters through the wavelet coefficient that multi-resolution decomposition obtains, noise and detailed information mainly at high band, the wavelet coefficient that corresponding absolute value is less, and noise has identical amplitude; Wavelet transformation is by continuous layering, signal is made to split into various frequency range (determining according to employing frequency), and this process will use low-pass filter and Hi-pass filter, and Wavelet Denoising Method is exactly change digital quantity at HFS (because usual white noise appears at HFS), use some algorithms to remove some mixed noisy numerals, and then use reconstruction low pass filter and Hi-pass filter that the frequency range of just layering is added up.

The feature of Db (Daubechies) wavelets is along with order increases, vanishing moment exponent number is larger, frequency band division effect is better, but time domain compactly supported can be made to weaken, calculated amount increases greatly simultaneously, and real-time is deteriorated. therefore, when carrying out order and selecting, not only to focus on the effect of algorithm itself, also should take into account the efficiency of algorithm.

And the useful information of image concentrates on low-frequency range, the wavelet coefficient that corresponding absolute value is larger.Therefore select a suitable threshold value, threshold process is carried out to wavelet coefficient, just can reach and remove noise and the object that retains useful signal.But while denoising, also lost some useful edge detail information.And local edge is detailed information the most useful in image, be image to the most important feature of vision, therefore, the edge feature of image should be retained while carrying out image denoising as far as possible.

Svd (Singular Value Decomposition, SVD) is a kind of nonlinear filtering, has good numerical robustness.The singular value of image array and feature space thereof reflect heterogeneity in image and feature, it is generally acknowledged larger singular value and characteristic of correspondence vector representation picture signal thereof, and noise are reflected on less singular value and characteristic of correspondence vector thereof.According to certain selection thresholding, singular value zero setting (blocking) is obtained lower than this thresholding, then carry out denoising by these singular values and characteristic of correspondence vector reconstruction image thereof, not only can process dissimilar image and noise, and without the need to the priori about noise.Consider the local stationary of image, also been proposed image block svd denoise algorithm.But general simple pattern recognition method does not consider the directivity feature of pattern recognition, and the noise of image is only distributed in the HFS of wavelet transformation frequency field, again because these HFSs have level, vertical, diagonal line (45 °/135 °) directivity characteristics, therefore can consider that three direction HFSs after to wavelet transformation carry out svd to reach the object of filtering and noise reduction.

In addition, because the singular value number in traditional svd denoising process needed for reconstructed image depends on Conventional wisdom value, or determined by the picture quality repeatedly calculating more different singular value number reconstructed image, have influence on the denoising effect of image like this, also have a strong impact on the ageing of algorithm.

The present invention is based on above-mentioned analysis, be intended to propose a kind of image de-noising method being applicable to the noisy image that image occurs in generation or transmitting procedure, improve rapidity and the denoising quality of image restoration.

Summary of the invention

Object of the present invention, a kind of effective preserving edge and directional characteristic motion blur image restoration method are fast provided, it can be the trajectory planning that unmanned plane realizes automatic obstacle avoiding provides data volume little and effective barrier point identification extracting method, can improve real-time and the security of unmanned aerial vehicle flight path planning.

In order to reach above-mentioned purpose, solution of the present invention is: a kind of effective preserving edge and directional characteristic moving image denoising method fast, i.e. motion blur image restoration method, image is after wavelet transformation, low frequency subgraph image set has suffered most of energy noise of original image, carries out the details that secondary Wiener filtering is beneficial to outstanding fused images; Because picture noise mainly concentrates on the small echo high frequency subgraph part of three different directions, its coefficient is less, svd can be utilized to carry out denoising, high frequency diagonal line subgraph is rotated to ranks direction, svd is carried out together with the subgraph of high frequency row, column direction, for avoiding losing edge detail information, also edge extracting and reservation are carried out to high frequency subgraph simultaneously; Finally the low frequency after denoising and high frequency subgraph are carried out inverse wavelet transform and reconstruct final denoising image.Singular value number wherein needed for singular value reconstructed image is determined jointly by the Y-PSNR of image and classic method.

The present invention is in order to solve the problems of the technologies described above by the following technical solutions: the present invention devises a kind of preserving edge and directional characteristic image de-noising method, comprises the steps:

Step 001, image wavelet transform, adopts Db3 small echo to carry out 2 layers of wavelet decomposition to the image with white Gaussian noise, obtains low frequency LL subgraph, the horizontal L of one deck high frequency ₁h ₁subgraph, one deck frequency vertical H ₁l ₁subgraph, one deck high frequency diagonal line H ₁h ₁subgraph, two floor heights horizontal L frequently ₂h ₂subgraph, two layers of frequency vertical H ₂l ₂subgraph and two floor heights diagonal line H frequently ₂h ₂subgraph;

It is level and smooth that step 002. pair low frequency LL subgraph carries out secondary Wiener filtering;

Step 003. carries out the svd denoising of utilization orientation characteristic to all high frequency subgraphs;

Step 004. also adopts direction edge detection operator to extract edge gray scale Sub-Image Feature to all high frequency subgraphs;

Step 005. is got greatly to the svd denoising high frequency subgraph obtained and corresponding edge extracting high frequency gray scale subgraph one_to_one corresponding, the pixel value of namely corresponding with edge gray table picture in denoising image pixel edge gray scale subgraph replaces, the denoising high frequency subgraph that the pixel obtaining edge gray scale subgraph retains.

The low frequency LL subimage that the Wiener filtering that step 006. pair step 002 obtains is level and smooth and all high frequency subgraphs that step 005 obtains carry out wavelet inverse transformation, reconstruct final denoising image.

As a preferred technical solution of the present invention: in described step 003, specifically comprise the steps:

The high frequency subgraph of step 00301. pair horizontal direction and vertical direction directly carries out svd matrix A=USV ^t, wherein, for two-dimentional original image B is subject to the image after noise X pollution, the order of matrix A is m, rank (A)=m (m≤l ₂), $U=(u_1,u_2\·\·\·u_{l1})\∈R^{l_1\×l_2}$ With $V=(v_1,v_2\·\·\·v_{l2})\∈R^{l_1\×l_2}$ Be respectively left singular matrix and the right singular matrix of A, the l of U and V ₂column vector is respectively the left singular vector of A and right singular vector, for singular value battle array, its diagonal entry λ ₁>=λ ₂>=...>=λ _r> 0 is the non-zero singular value of matrix, λ _ifor i-th singular value of matrix A.

R is all pixels of two-dimensional digital image matrix, l ₁and l ₂for two-dimensional digital image matrix ranks value at once with column vector (comprising A, B, X, U, V and s-matrix).

The horizontal direction that step 00302. utilizes step 00301 to obtain and vertical direction high frequency subgraph singular value are carried out Image Reconstruction and are obtained denoising image (λ _i>=γ), the singular value number k (k≤R) of k needed for reconstruct, the γ singular value threshold value needed for reconstruct, obtains denoising image .

U _iand v _i(i=1,2,3 ... k) the i-th row singular value vector of U and V is respectively.

Diagonal high frequency subgraph is rotated 45 ° or 135 ° to horizontal direction (former 45 ° or 135 ° to horizontal direction) by step 00303., obtain the larger image that comprises original image, the average gray value of non-original image part original image is wherein filled, then carries out svd A=USV ^t, wherein, for two-dimentional original image B is subject to the image after noise X pollution, the order of matrix A is m, rank (A)=m (m≤l ₂), $U=(u_1,u_2\·\·\·u_{l1})\∈R^{l_1\×l_2}$ With $V=(v_1,v_2\·\·\·v_{l2})\∈R^{l_1\×l_2}$ Be respectively left singular matrix and the right singular matrix of A, the front l of U and V ₂column vector is respectively the left singular vector of A and right singular vector, for singular value battle array, its diagonal entry λ ₁>=λ ₂>=...>=λ _r> 0 is the non-zero singular value of matrix, λ _ifor i-th singular value of matrix A.

After the rotation that step 00304. utilizes step 00303 to obtain, the singular value of diagonal high frequency subgraph carries out Image Reconstruction (λ _i>=γ), the singular value number k (k≤R) of k needed for reconstruct, the γ singular value threshold value needed for reconstruct, obtains denoising image .

Step 00304. reduces out original image part in diagonal high frequency reconstruction subgraph after rotation;

Step 00305. is reduced the original image part obtained and is obtained the diagonal high frequency imaging after denoising through reverse rotation.

As a preferred technical solution of the present invention: in described step 004, specifically comprise the steps:

Step 00401. utilizes sobel horizontal direction $\mathrm{Ss}=\left[\begin{array}{ccc}-1& -2& -1\\ 1& 0& 0\\ 1& 2& 1\end{array}\right]$ And vertical direction $t=\left[\begin{array}{ccc}-1& 0& 1\\ -2& 0& 2\\ -1& 0& 1\end{array}\right]$ The marginal information of edge detection operator difference detection level direction and vertical direction high frequency subgraph, obtains the high frequency subgraph edge binary images of horizontal direction and vertical direction;

Step 00402. utilizes Roberts45 ° of diagonal $s=\left[\begin{array}{ccc}0& 0& 0\\ 0& 1& 0\\ 0& 0& -1\end{array}\right]$ With 135 ° of diagonals $t=\left[\begin{array}{ccc}0& 0& 0\\ 0& 1& 0\\ -1& 0& 0\end{array}\right]$ Edge detection operator detects the marginal information of diagonal high frequency subgraph, obtains diagonal high frequency subgraph edge binary images;

Horizontal direction, vertical direction are reverted to the new edge gray table consistent with original image gray scale with the high frequency subgraph edge binary images of diagonal by step 00403., when edge image pixel value is 1, recover its gray-scale value, otherwise zero setting.Computing formula is $g(x,y)=\left\{\begin{array}{cc}f(x,y)& g(x,y)=1\\ 0& g(x,y)=0\end{array}\right.,$ Wherein g, f are respectively edge image and original image, and x, y are pixel coordinate.

Singular value number wherein needed for singular value reconstructed image is determined jointly by the Y-PSNR of image and classic method.

As a preferred technical solution of the present invention: in described step 00302, comprise the defining method of the required number k of singular value reconstruct and reconstruct singular value threshold gamma, specifically comprise the steps:

Step 0030201. utilizes Conventional wisdom determine the upper limit of singular value threshold gamma, wherein l ₁and l ₂for the ranks value of image, σ is noise variance;

Step 0030202. has carried out l ~ i (γ≤λ to image in the γ upper limit _i, i is singular value number corresponding to singular value threshold value) and reconstruct, i singular value before each reconstruct uses, calculates the Y-PSNR PSNR of each reconstructed image, $\begin{array}{c}\mathrm{PSNR}=10{\log }_{10}\left(\frac{{255}^2}{\mathrm{MSE}}\right)\\ \mathrm{MSE}=\frac{\underset{1}{\overset{M}{\mathrm{\Σ}}}\underset{1}{\overset{N}{\mathrm{\Σ}}}{(f(x,y)-f_0(x,y))}^2}{l_1\×l_2}\end{array},$ Wherein f (x, y) is the grey scale pixel value of image after denoising, f ₀(x, y) represents the grey scale pixel value of original image, and x, y are pixel coordinate.

The reconstruct singular value number that step 0030203. chooses that width image of the highest PSNR of Y-PSNR in all reconstructed images corresponding is required reconstruct singular value number k.

A kind of image de-noising method of the present invention adopts above technical scheme compared with prior art, has following beneficial effect:

(1) in the image de-noising method that the present invention designs, after utilizing wavelet transformation by the present invention's design, the directivity characteristics of high frequency imaging carries out the singular value denoising of different directions to high frequency subgraph, do not consider that svd only has for the method for ranks directivity characteristics relative to common, the method for the present invention's design has the non-horizontal that retains original image and non-vertical direction information, advantage that denoising quality is higher;

(2) in the image de-noising method that the present invention designs, problem is lost for normal image denoising method edge details, by extracting the marginal information of high frequency subgraph all directions and remaining in final final denoising image, effective reservation and outstanding image edge details information, further increase the denoising effect of image;

(3) in the image de-noising method that the present invention designs, singular value number in the method determination singular value reconstruct adopting Conventional wisdom and image quality evaluation effectively to combine, first determine the upper limit of singular value threshold value, determine according to picture quality again in limited range, not only increase process ageing, and improve the processing accuracy of algorithm.

Daubechies wavelets (Db3) efficiency that the present invention selects exponent number little can improve greatly.

In a word, image de-noising method of the present invention is for not considering the direction characteristic of image and the denoising effect that causes is not good and marginal information is lost problem in common prediction filtering denoising process, effectively can retain direction and the edge feature of image, improve denoising quality, and general general experience combines with real image quality assessment by the extraction of singular value reconstructed image number, be also quick and effective relative to single classic method or based on the method for image Y-PSNR.

Accompanying drawing explanation

Fig. 1 is Wavelet domain image denoising overview flow chart of the present invention;

Fig. 2 is high frequency subgraph svd denoising process flow diagram;

Fig. 3 is high frequency subgraph edge extracting process flow diagram;

Fig. 4 is Image Reconstruction singular value number determination schematic diagram.

Embodiment

The invention provides a kind of effective preserving edge and directional characteristic image de-noising method fast, its main thought be distribute according to the noise energy of image after wavelet transformation, direction message and local edge, directly adopt secondary Wiener filtering to be beneficial to the details of outstanding fused images to the less low frequency subgraph of noise energy distribution; The high frequency subgraph very large to noise energy distribution adopts prediction filtering and extracts the high frequency denoising image that edge obtains preserving edge; Low frequency after denoising, high frequency subgraph are carried out inverse wavelet transform reconstruct, obtains final denoising image, to improve denoising speed and processing accuracy.Below with reference to accompanying drawing, technical scheme of the present invention is described in detail.

Wavelet domain image denoising overview flow chart as shown in Figure 1, image is after wavelet transformation, and low frequency subgraph image set has suffered most of energy noise of original image, carries out the details that secondary Wiener filtering is beneficial to outstanding fused images; Because picture noise mainly concentrates on the small echo high frequency subgraph part of three different directions, its coefficient is less, svd can be utilized to carry out denoising, high frequency diagonal line subgraph is rotated to ranks direction, svd is carried out together with the subgraph of high frequency row, column direction, for avoiding losing edge detail information, also edge extracting and reservation are carried out to high frequency subgraph simultaneously; Finally the low frequency after denoising and high frequency subgraph are carried out inverse wavelet transform and reconstruct final denoising image.

Corresponding therewith, a kind of preserving edge and directional characteristic image de-noising method, comprise the steps (can carry out following step by MATLAB):

Step 001. adopts Db3 small echo to carry out 2 layers of wavelet decomposition to the image with white Gaussian noise, obtains low frequency LL subgraph, the horizontal L of one deck high frequency ₁h ₁subgraph, one deck frequency vertical H ₁l ₁subgraph, one deck high frequency diagonal line H ₁h ₁subgraph, two floor heights horizontal L frequently ₂h ₂subgraph, two layers of frequency vertical H ₂l ₂subgraph and two floor heights diagonal line H frequently ₂h ₂subgraph;

It is level and smooth that step 002. pair low frequency LL subimage carries out secondary Wiener filtering;

Step 003. carries out the svd denoising of utilization orientation characteristic on the one hand to all high frequency subimages;

Step 004. adopts direction edge detection operator to extract edge to all high frequency subimages on the other hand;

Step 005. is got greatly to the svd denoising high frequency subimage obtained and corresponding edge extracting high frequency grayscale sub-image one_to_one corresponding, the pixel value of namely corresponding with edge gray table picture in denoising image pixel edge gray table picture replaces, and obtains the denoising high frequency subgraph that edge retains.

All high frequency subgraphs that the Wiener filtering that step 006. pair step 002 obtains level and smooth low frequency LL subimage and step 005 obtain carry out wavelet inverse transformation, reconstruct final denoising image.

As a preferred technical solution of the present invention: in described step 003, as shown in Figure 2, specifically comprise the steps:

The high frequency subgraph of step 00301. pair horizontal direction and vertical direction directly carries out svd A=USV ^t, wherein, for two-dimentional original image B is subject to the image after noise X pollution, order rank (A)=m (m≤l of matrix A ₂), $U=(u_1,u_2\·\·\·u_{l1})\∈R^{l_1\×l_2}$ With $V=(v_1,v_2\·\·\·v_{l2})\∈R^{l_1\×l_2}$ Be respectively left singular matrix and the right singular matrix of A, the front l of U and V ₂column vector is respectively the left singular vector of A and right singular vector, for singular value battle array, its diagonal entry λ ₁>=λ ₂>=...>=λ _r> 0 is the non-zero singular value of matrix, λ _ifor i-th singular value of matrix A.

The horizontal direction that step 00302. utilizes step 00301 to obtain and vertical direction high frequency subgraph singular value carry out Image Reconstruction (λ _i>=γ), the singular value number k (k≤R) of k needed for reconstruct, the γ singular value threshold value needed for reconstruct, obtains denoising image .

Diagonal high frequency subgraph is rotated 45 ° (135 °) to horizontal direction by step 00303., obtain the larger image that comprises original image, the average gray value of non-original image part original image is wherein filled, then carries out svd A=USV ^t, wherein, for two-dimentional original image B is subject to the image after noise X pollution, order rank (A)=R (R≤l of matrix A ₂), $U=(u_1,u_2\·\·\·u_{l1})\∈R^{l_1\×l_2}$ With $V=(v_1,v_2\·\·\·v_{l2})\∈R^{l_1\×l_2}$ Be respectively left singular matrix and the right singular matrix of A, the front l of U and V ₂column vector is respectively the left singular vector of A and right singular vector, for singular value battle array, its diagonal entry λ ₁>=λ ₂>=...>=λ _r> 0 is the non-zero singular value of matrix, λ _ifor i-th singular value of matrix A.

After the rotation that step 00304. utilizes step 00303 to obtain, the singular value of diagonal high frequency subgraph carries out Image Reconstruction (λ _i>=γ), the singular value number k (k≤R) of k needed for reconstruct, the γ singular value threshold value needed for reconstruct, obtains denoising image .

Step 00304. reduces out original image part in diagonal high frequency reconstruction subgraph after rotation;

Step 00305. is reduced the original image part obtained and is obtained the diagonal high frequency imaging after denoising through reverse rotation.

As a preferred technical solution of the present invention: in described step 004, as shown in Figure 3, specifically comprise the steps:

Step 00401. utilizes sobel horizontal direction $s=\left[\begin{array}{ccc}-1& -2& -1\\ 1& 0& 0\\ 1& 2& 1\end{array}\right]$ And vertical direction $t=\left[\begin{array}{ccc}-1& 0& 1\\ -2& 0& 2\\ -1& 0& 1\end{array}\right]$ The marginal information of edge detection operator difference detection level direction and vertical direction high frequency subgraph, obtains the high frequency subgraph edge binary images of horizontal direction and vertical direction;

Step 00402. utilizes Roberts45 ° of diagonal $s=\left[\begin{array}{ccc}0& 0& 0\\ 0& 1& 0\\ 0& 0& -1\end{array}\right]$ With 135 ° of diagonals $t=\left[\begin{array}{ccc}0& 0& 0\\ 0& 1& 0\\ -1& 0& 0\end{array}\right]$ Edge detection operator detects the marginal information of diagonal high frequency subgraph, obtains diagonal high frequency subgraph edge binary images;

Horizontal direction, vertical direction are reverted to the new edge gray table consistent with original image gray scale with the high frequency subgraph edge binary images of diagonal by step 00403., when edge image pixel value is 1, recover its gray-scale value, otherwise zero setting.Computing formula is $g(x,y)=\left\{\begin{array}{cc}f(x,y)& g(x,y)=1\\ 0& g(x,y)=0\end{array}\right.,$ Wherein g, f are respectively edge image and original image, and x, y are pixel coordinate.

As a preferred technical solution of the present invention: in described step 00302, as described in Figure 4, comprise the defining method of the required number k of singular value reconstruct and reconstruct singular value threshold gamma, specifically comprise the steps:

Step 0030201. utilizes Conventional wisdom determine the upper limit of singular value threshold gamma, wherein l ₁and l ₂for the ranks value of image, σ is noise variance;

Step 0030202. has carried out l ~ i (γ≤λ to image in the γ upper limit _i, i is singular value number corresponding to singular value threshold value) and reconstruct, i singular value before each reconstruct uses, calculates the Y-PSNR PSNR of each reconstructed image, $\begin{array}{c}\mathrm{PSNR}=10{\log }_{10}\left(\frac{{255}^2}{\mathrm{MSE}}\right)\\ \mathrm{MSE}=\frac{\underset{1}{\overset{M}{\mathrm{\Σ}}}\underset{1}{\overset{N}{\mathrm{\Σ}}}{(f(x,y)-f_0(x,y))}^2}{l_1\×l_2}\end{array},$ , wherein f (x, y) is the grey scale pixel value of image after denoising, f ₀(x, y) represents the grey scale pixel value of original image, and x, y are pixel coordinate.

The reconstruct singular value number that step 0030203. chooses that width image of the highest PSNR of Y-PSNR in all reconstructed images corresponding is required reconstruct singular value number k.

To sum up, by setting up and implementing effective fast preserving edge of the present invention's design and directional characteristic image de-noising method, effectively significantly can improve denoising effect, image edge details information obtains effective reservation, particularly image non-horizontal (non-perpendicular) direction detailed information obtains effective reservation, improve data processing speed and processing accuracy, certain basis is established to aspects such as further image characteristics extraction, target detection and pattern-recognitions, there is wide market application foreground and economic worth.

Above embodiment is only and technological thought of the present invention is described, can not limit protection scope of the present invention with this, and every technological thought proposed according to the present invention, any change that technical scheme basis is done, all falls within scope.

Claims (6)

1. quick effectively preserving edge and a directional characteristic image de-noising method, is characterized in that: image is after wavelet transformation, and low frequency subgraph image set has suffered most of energy noise of original image, carries out the details that secondary Wiener filtering gives prominence to fused images; Because picture noise mainly concentrates on the small echo high frequency subgraph part of three different directions, its coefficient is less, svd is utilized to carry out denoising, high frequency diagonal line subgraph is rotated to ranks direction, svd is carried out together with the subgraph of high frequency row, column direction, for avoiding losing edge detail information, also edge extracting and reservation are carried out to high frequency subgraph simultaneously; Finally the low frequency after denoising and high frequency subgraph are carried out inverse wavelet transform and reconstruct final denoising image.

2. as claimed in claim 1 effectively preserving edge and directional characteristic image de-noising method fast, is characterized in that: comprise the steps: that the singular value number needed for described singular value reconstructed image is determined jointly by the Y-PSNR of image and classic method.

3. effective preserving edge and directional characteristic image de-noising method fast as claimed in claim 1, is characterized in that, comprise the steps:

Step 001. adopts Db3 small echo to carry out 2 layers of wavelet decomposition to the image with white Gaussian noise, obtains low frequency LL subgraph, the horizontal L of one deck high frequency ₁h ₁subgraph, one deck frequency vertical H ₁l ₁subgraph, one deck high frequency diagonal line H ₁h ₁subgraph, two floor heights horizontal L frequently ₂h ₂subgraph, two layers of frequency vertical H ₂l ₂subgraph and two floor heights diagonal line H frequently ₂h ₂subgraph;

It is level and smooth that step 002. pair low frequency LL subimage carries out secondary Wiener filtering;

Step 003. carries out the svd denoising of utilization orientation characteristic to all high frequency subimages;

Step 004. also adopts direction edge detection operator to extract edge to all high frequency subimages;

Step 005. is got greatly to the svd denoising high frequency subimage obtained and corresponding edge extracting high frequency grayscale sub-image one_to_one corresponding, the pixel value of namely corresponding with edge grayscale sub-image in denoising high frequency subimage pixel edge gray table picture replaces, and obtains the denoising high frequency subimage that edge retains;

All high frequency subgraphs that the Wiener filtering that step 006. pair step 002 obtains level and smooth low frequency LL subimage and step 005 obtain carry out wavelet inverse transformation, reconstruct final denoising image.

4. effective preserving edge and directional characteristic image de-noising method fast as claimed in claim 3, is characterized in that, in described step 003, specifically comprise the steps:

The high frequency subgraph of step 00301. pair horizontal direction and vertical direction directly carries out svd A=USV ^t, wherein, for two-dimentional original image B is subject to the image after noise X pollution, order rank (A)=m (m≤l of matrix A ₂), $U=(u_1,u_2,...,u_{l1})\∈R^{l_1\×l_2}$ With $V=(v_1,v_2,...,v_{l1})\∈R^{l_1\×l_2}$ Be respectively left singular matrix and the right singular matrix of A, the l of U and V ₂column vector is respectively the left singular vector of A and right singular vector, for singular value battle array, its diagonal entry λ ₁>=λ ₂>=...>=λ _r> 0 is the non-zero singular value of matrix, λ _ifor i-th singular value of matrix A; l ₁for the row value of image;

The horizontal direction that step 00302. utilizes step 00301 to obtain and vertical direction high frequency subgraph singular value carry out Image Reconstruction (λ _i>=γ), the singular value number k (k≤m) of k needed for reconstruct, the γ singular value threshold value needed for reconstruct, obtains denoising image

Diagonal high frequency subgraph is rotated 45 ° (135 °) to horizontal direction by step 00303., obtain the larger image that comprises original image, the average gray value of non-original image part original image is wherein filled, then carries out svd A=USV ^t, wherein, for two-dimentional original image B is subject to the image after noise X pollution, order rank (A)=m (m≤l of matrix A ₂), $U=(u_1,u_2,...,u_{l1})\∈R^{l_1\×l_2}$ With $V=(v_1,v_2,...,v_{l2})\∈R^{l_1\×l_2}$ Be respectively left singular matrix and the right singular matrix of A, the l of U and V ₂column vector is respectively the left singular vector of A and right singular vector, for singular value battle array, its diagonal entry λ ₁>=λ ₂>=...>=λ _r> 0 is the non-zero singular value of matrix, λ _ifor i-th singular value of matrix A.

After the rotation that step 00304. utilizes step 00303 to obtain, the singular value of diagonal high frequency subgraph carries out Image Reconstruction (λ _i>=γ), the singular value number k (k≤m) of k needed for reconstruct, the γ singular value threshold value needed for reconstruct, obtains denoising image

Step 00304. reduces out original image part in diagonal high frequency reconstruction subgraph after rotation;

Step 00305. is reduced the original image part obtained and is obtained the diagonal high frequency imaging after denoising through reverse rotation.

5. effective preserving edge and directional characteristic image de-noising method fast as claimed in claim 3, is characterized in that, in described step 004, specifically comprise the steps:

Step 00401. utilizes sobel horizontal direction $s=\left[\begin{array}{ccc}-1& -2& -1\\ 1& 0& 0\\ 1& 2& 1\end{array}\right]$ And vertical direction $t=\left[\begin{array}{ccc}-1& 0& 1\\ -2& 0& 2\\ -1& 0& 1\end{array}\right]$ The marginal information of edge detection operator difference detection level direction and vertical direction high frequency subimage, obtains the edge binary images of the high frequency subimage of horizontal direction and vertical direction;

Step 00402. utilizes Roberts45 ° of diagonal $s=\left[\begin{array}{ccc}0& 0& 0\\ 0& 1& 0\\ 0& 0& -1\end{array}\right]$ With 135 ° of diagonals $t=\left[\begin{array}{ccc}0& 0& 0\\ 0& 1& 0\\ -1& 0& 0\end{array}\right]$ Edge detection operator detects the marginal information of diagonal high frequency subgraph, obtains diagonal high frequency subgraph edge binary images;

Horizontal direction, vertical direction are reverted to the new edge gray table consistent with original image gray scale with the high frequency subgraph edge binary images of diagonal by step 00403., when edge image pixel value is 1, recover its gray-scale value, otherwise zero setting; Computing formula is $g(x,y)=\left\{\begin{array}{cc}f(x,y)& g(x,y)=1\\ 0& g(x,y)=0\end{array}\right.,$ Wherein g, f are respectively edge image and original image, and x, y are pixel coordinate.

6. effective preserving edge and directional characteristic image de-noising method fast as claimed in claim 4, is characterized in that, in described step 00302, comprises the defining method of the required number k of singular value reconstruct and reconstruct singular value threshold gamma, specifically comprise the steps:

Step 0030201. utilizes Conventional wisdom determine the upper limit of singular value threshold gamma, wherein l ₁and l ₂for the row, column value of image, σ is noise variance;

Step 0030202. has carried out l ~ i reconstruct to image in the γ upper limit, γ≤λ _n, i is singular value number corresponding to singular value threshold value; I singular value before each reconstruct uses, calculates the Y-PSNR PSNR of each reconstructed image, $\begin{array}{c}\mathrm{PSNR}=10{\log }_{10}\left(\frac{{255}^2}{\mathrm{MSE}}\right)\\ \mathrm{MSE}=\frac{\underset{1}{\overset{M}{\mathrm{\Σ}}}\underset{1}{\overset{N}{\mathrm{\Σ}}}{(f(x,y)-f_0(x,y))}^2}{l_1\×l_2}\end{array}0,$ Wherein f (x, y) is the grey scale pixel value of image after denoising, f ₀(x, y) represents the grey scale pixel value of original image, and x, y are pixel coordinate;

The reconstruct singular value number that step 0030203. chooses that width image of the highest PSNR of Y-PSNR in all reconstructed images corresponding is required reconstruct singular value number k.