CN103020918B - Shape-adaptive neighborhood mean value based non-local mean value denoising method - Google Patents

Abstract

The invention discloses a shape-adaptive neighborhood mean value based non-local mean value denoising method which mainly solves the problem of similarity weight calculation inaccuracy in an existing non-local mean value denoising method. An implementation process includes: (1) calculating a shape-adaptive neighborhood of each pixel point in an inputted natural image by the aid of an SA-DCT (shape adaptive-discrete cosine transform) method; (2) calculating a mean value of the shape-adaptive neighborhoods of the pixel points so as to obtain a new image; (3) selecting a central pixel block in the new image and selecting a similar block in a search area; (4) calculating a similarity weight between the central pixel block and the similar block; (5) using the similarity weight and the pixel points in the inputted image for weighted averaging so that a pixel point recovery value is obtained; and (6) calculating a recovery value of each pixel point in the inputted image and using each recovery value for substituting for an original grey value, so that a denoised image is obtained. The shape-adaptive neighborhood mean value based non-local mean value denoising method effectively suppresses noise interference, improves similarity weight calculation accuracy, smoothes noise while keeps detail information of the natural image, and can be used for denoising natural images.

Description

The non-local mean denoising method of Shape-based interpolation adaptive neighborhood average

Technology neighborhood

The invention belongs to image processing techniques neighborhood, relate to image de-noising method, can be used for the denoising to natural image.

Background technology

In the last few years, universal along with all kinds of digital implementation and digital product, Digital Image Processing became a study hotspot.Image is always inevitably subject to the interference of various noise source in the process obtaining, transmit and store.Therefore image denoising be in image procossing neighborhood one substantially and very crucial technology, paid attention to widely always.

Traditional image de-noising method roughly can be divided into two classes: a class is the method based on spatial domain, grey scale pixel value in local window is mainly utilized to carry out gray proces to current pixel, to reach the object of suppression or stress release treatment, airspace filter method has non-local mean filtering method, the image de-noising method etc. under rarefaction representation preferably at present.Image de-noising method under rarefaction representation adopts sparse approximate in redundant dictionary of image, realizes noise remove, as KSVD dictionary learning denoising method; Another kind of is method based on transform domain.

Owing to including bulk redundancy information in natural image, the people such as Buades propose a kind of non-local mean denoising method.The method, centered by current pixel point, utilizes the similarity of the redundancy structure information in image and current pixel point peripheral region to carry out filtering.Facts have proved, the accuracy that the method judges due to the disturbing effect similarity of noise, the similarity weights of acquisition are not accurate enough, final denoising result usually the edge of image or texture information give fuzzy fall.

Non-local mean method is one of very outstanding method of image denoising neighborhood, and after it proposes, a lot of scholar is studied it and improves.As bayesian non-local mean denoising method, it is by arranging penalty, obtains the optimal estimation value making penalty minimum, obtain denoising result according to partial differential equation; PPB filtering method obtains weighted average formula under maximal possibility estimation framework, and progressively revises prior imformation by alternative manner, finally converges to optimal result, obtains denoising result.These methods improve the similarity measurement of non-local mean method, but be still subject to noise effect when similarity judges, reduce the accuracy of judgement, cause the edge of image and details to a certain extent by fuzzy or filtering, make us occur deviation to the subsequent analysis process of image detail.

Summary of the invention

The object of the invention is to the deficiency overcoming above-mentioned prior art, propose a kind of non-local mean denoising method of Shape-based interpolation adaptive neighborhood average, to reduce the impact that noise judges similarity, improve the accuracy that similarity judges, and then improve image denoising effect.

Realizing the object of the invention technical scheme is: improve the accuracy of similarity measurement computing method in image, by in conjunction with SA-DCT method, calculate the similar neighborhood average of form adaptive, realize taking into account edge in natural image denoising and smooth region, concrete steps comprise as follows:

(1) input the natural image x that a width treats denoising, according to SA-DCT method, line by line scan and obtain the pixel x of t position in natural image x _tshape-adaptive neighborhood S (x in n × n size area _t);

(2) the pixel x in natural image x is calculated _tat its shape-adaptive neighborhood S (x _t) in average , traversal view picture natural image x, obtains the image x ' that a width is new;

(3) line by line scan in the image x ' newly obtained, choose the pixel x ' of i-th position _ifor pixel to be restored, with pixel x ' to be restored _icentered by, choose r × r size square center block of pixels V (x ' _i);

(4) with the pixel x ' to be restored chosen _icentered by, choose the square aearch window of f × f size, line by line scan in search window, choose a pixel x ' of a jth position _jas pixel x ' to be restored _isimilitude, with this similitude x ' _jcentered by, choose one with center pixel block V (x ' _i) equal-sized block of pixels as similar piece of V (x ' _j);

(5) according to center pixel block V (x ' _i) to selected by step (4) similar piece of V (x ' _j), calculate pixel x ' to be restored _iwith similitude x ' _jbetween similarity weight w (x ' _i, x ' _j):

$w(x_i^{\′},x_j^{\′})=\exp (-\frac{{\left|\right|V\left(x_i^{\′}\right)-V\left(x_j^{\′}\right)\left|\right|}_2^2}{h^2}),$

Wherein, V (x ' _i) and V (x ' _j) be respectively center pixel block selected by step (3) and similar piece selected by step (4), h is level and smooth controling parameters, and σ is the Gaussian noise standard deviation in natural image, h=k σ, k is a constant, similarity weight w (x ' _i, x ' _j) satisfy condition: 0≤w (x ' _i, x ' _j)≤1, and ∑ w (x ' _i, x ' _j)=1;

(6) with similarity weight w (x ' _i, x ' _j) with the pixel x of a jth position in the natural image x treating denoising _jbe weighted average, calculate pixel x ' to be restored _irecovery value

(7) with pixel x ' to be restored _irecovery value replace and treat i-th position pixel x in the natural image x of denoising _igray-scale value, obtain the image after denoising.

The present invention has the following advantages compared with prior art:

1. the present invention is owing to combining SA-DCT method, can judge the similarity in the natural image of Noise between pixel more accurately, and then can calculate the gray-scale value of pixel to be restored more accurately;

2. send out the gray-scale value that more accurately can calculate pixel to be restored, and then while better smooth noise, can keep and recover edge and the grain details of natural image.

Accompanying drawing explanation

Fig. 1 is process flow diagram of the present invention;

Fig. 2 is the test pattern that the present invention uses;

Fig. 3 is the noisy image Lena that the present invention uses;

Fig. 4 uses existing non-local mean method to the result figure of Fig. 3 denoising;

Fig. 5 uses existing bayesian non-local mean method to the result figure of Fig. 3 denoising;

Fig. 6 uses existing PPB filtering method to the result figure of Fig. 3 denoising;

Fig. 7 uses existing KSVD method to the result figure of Fig. 3 denoising;

Fig. 8 uses the inventive method to the result figure of Fig. 3 denoising.

Embodiment

With reference to Fig. 1, the present invention is based on the non-local mean denoising method of shape-adaptive neighborhood average, comprise the steps:

Step 1, inputs the natural image x that a width treats denoising, according to SA-DCT method, lines by line scan and obtain the pixel x of t position in natural image x _tshape-adaptive neighborhood S (x in n × n size area _t), wherein t is the position of pixel in natural image x, and t=1,2..., m × m, m is the diameter of input natural image x, and n is shape-adaptive neighborhood S (x _t) diameter, choose the shape-adaptive neighborhood that size is 3 × 3 in the present invention.

Described SA-DCT method is proposed in Pointwise Shape-Adaptive DCT for High-Quality Denoising and Deblocking of Grayscale and Color Images article by people such as Alessandro Foi, by judging that whether pixel is consistent with the heterogeneity of all pixels in its neighborhood, a shape-adaptive neighborhood of this pixel can be determined.

Step 2, calculates the pixel x of t position in natural image x _tat its shape-adaptive neighborhood S (x _t) in average , traversal view picture natural image x, obtains the image x ' that a width is new.

Step 3, in the image x ' newly obtained, choose r × r size square center block of pixels V (x ' _i).

3.1) line by line scan in the image x ' newly obtained and choose the pixel x ' of i-th position _ifor pixel to be restored, wherein i is the position of pixel in the image x ' newly obtained, and i=1,2..., m × m, m is the diameter of input natural image x;

3.2) with pixel x ' to be restored _icentered by, choose r × r size square center block of pixels V (x ' _i), wherein r be odd sized center pixel block V (x ' _i) diameter, with pixel x ' in the present invention _icentered by, choose a size be 7 × 7 center pixel block V (x ' _i).

Step 4, with the pixel x ' to be restored chosen _icentered by, choose similar piece of V (x ' _j).

4.1) with the pixel x ' to be restored chosen _icentered by, choose the square aearch window of f × f size, line by line scan in search window, choose a pixel x ' of a jth position _jas pixel x ' to be restored _isimilitude, wherein j is the position of similitude in square aearch window, and j=1,2..., f × f, f is the square aearch window diameter of odd sized, and be greater than center pixel block V (x ' _i) diameter r, choosing a size in the present invention is the search window of 21 × 21;

4.2) with this similitude x ' _jcentered by, choose one with center pixel block V (x ' _i) equal-sized block of pixels, as similar piece of V (x ' _j).

Step 5, according to center pixel block V (x ' _i) to selected by step (4) similar piece of V (x ' _j), calculate pixel x ' to be restored _iwith similitude x ' _jbetween similarity weight w (x ' _i, x ' _j):

$w(x_i^{\′},x_j^{\′})=\exp (-\frac{{\left|\right|V\left(x_i^{\′}\right)-V\left(x_j^{\′}\right)\left|\right|}_2^2}{h^2}),$

Wherein, V (x ' _i) and V (x ' _j) being respectively center pixel block selected by step (3) and similar piece selected by step (4), h is level and smooth controling parameters, and σ is the Gaussian noise standard deviation in natural image, h=k σ, k is a constant, chooses k=7 in the present invention, similarity weight w (x ' _i, x ' _j) satisfy condition: 0≤w (x ' _i, x ' _j)≤1, and ∑ w (x ' _i, x ' _j)=1.

Step 6, with similarity weight w (x ' _i, x ' _j) with the pixel x of a jth position in the natural image x treating denoising _jbe weighted average, calculate pixel x ' to be restored _irecovery value

$\hat{v}\left(x_i\right)=\frac{1}{C\left(x_j^{\′}\right)}\underset{i,j=1}{\overset{f\×f}{\mathrm{\Σ}}}w(x_i^{\′},x_j^{\′})v\left(x_j\right),$

Wherein, v (x _j) be the pixel x of a jth position in natural image x _jgray-scale value, j=1,2..., f × f, f is the square aearch window diameter of odd sized, C (x ' _j) be normalization constant:

$C\left(x_j^{\′}\right)=\underset{i,j=1}{\overset{f\×f}{\mathrm{\Σ}}}w(x_i^{\′},x_j^{\′}).$

Step 7, according to step 3 to step 6, calculates the recovery value of each pixel, with pixel x ' to be restored _irecovery value replace and treat i-th position pixel x in the natural image x of denoising _igray-scale value, obtain the image after denoising, wherein i is that pixel is newly obtaining the position in image x ', and i=1,2..., m × m, m is the diameter of input natural image x.

Effect of the present invention can be confirmed further by following experiment:

One. experiment condition and content

Experiment condition: the input picture that experiment uses as shown in Figures 2 and 3, wherein, Fig. 2 (a) is test pattern Lena, Fig. 2 (b) is test pattern House, Fig. 2 (b) is test pattern Cameraman, Fig. 2 (b) is test pattern Peppers, Fig. 3 is add to Fig. 2 (a) the noisy Lena image that Gauss's additive white noise standard deviation is 20.Non-local mean filtering method with all adopt similar piece of 7 × 7 sizes based on bayesian non-local mean method, region is sought in the search of 21 × 21, PPB filtering method adopts 4 iteration, similar piece of 7 × 7 sizes, region is sought in the search of 21 × 21, and KSVD method uses the dictionary of 64 × 256 sizes, and the inventive method uses the shape-adaptive neighborhood of 3 × 3 sizes, 7 × 7 similar piece, the square of 21 × 21 searches window.

Experiment content:

Emulation 1: under these experimental conditions, use existing non-local mean filtering method to carry out denoising to Fig. 3, experimental result is as Fig. 4; Use and existingly carry out denoising based on bayesian non-local mean filtering method to Fig. 3, experimental result is as Fig. 5; Use existing PPB filtering method to carry out denoising to Fig. 3, experimental result is as Fig. 6; Use existing KSVD filtering method to carry out denoising to Fig. 3, experimental result is as Fig. 7; Use the inventive method to carry out denoising to Fig. 3, experimental result is as Fig. 8.

As can be seen from Figure 4, the noise inhibiting ability of non-local mean filtering method is limited, and edge and details exist fuzzy;

As can be seen from Figure 5, the stability based on bayesian non-local mean filtering method noise inhibiting ability is better than non-local mean filtering method, but it exists edge and the fuzzy problem of details equally;

As can be seen from Figure 6, the noise inhibiting ability of PPB filtering method is better than non-local mean filtering method, but edge and Hemifusus ternatanus degree all not good enough;

As can be seen from Figure 7, KSVD method has good noise inhibiting ability, and also comparatively front several method is better in the maintenance of edge and details;

As can be seen from Figure 8, the image homogeneous region denoising of the inventive method is abundant, and brightness keeps effect better, and the edge of image, details have also been obtained good maintenance, obtains comparatively ideal effect.

Emulation 2: adding noise criteria difference to all test patterns in Fig. 2 is 10,20,30, Gauss's additive white noise of 40, use existing non-local mean filtering method respectively, based on bayesian non-local mean filtering method, PPB filtering method, KSVD filtering method and the inventive method carry out denoising to noisy image, and experimental result is as shown in table 1.

Table 1 denoising result contrasts

From in table 1, for most test pattern, the inventive method all achieves good result in different noise grade.Existing non-local mean filtering method, PPB filtering method and the inventive method, be and carry out a Recovery processing to central pixel point, and bayesian non-local mean method make use of polymerization technique, utilizes center pixel block to carry out Recovery processing.From PSNR, the present invention to the image of low-level noise, PSNR substantially higher than non-local mean filtering method and PPB filtering method, a little less than bayesian non-local mean method, and the KSVD method of rarefaction representation; For the noise of higher level, noise suppression effect of the present invention is better, and PSNR has had certain raising compared with the method for block and sparse representation method, is greatly improved relative to similar some restoration methods.It can also be seen that from table 1, along with the increase of picture noise standard deviation, the advantage of the inventive method is also more and more obvious.

From visual effect, noise inhibiting ability of the present invention is fine, and homogeneous region is smoother, and it is better that the brightness of image also keeps, and the edge of image and details have also been obtained good maintenance.

Above experimental result shows, the present invention is better than most existing denoising method on overall performance, can better keep edge and the detailed information of natural image while level and smooth homogeneous region.

Claims (1)

1. a non-local mean denoising method for Shape-based interpolation adaptive neighborhood average, comprises the steps:

(1) input the natural image x that a width treats denoising, according to SA-DCT method, line by line scan and obtain the pixel x of t position in natural image x _tshape-adaptive neighborhood S (x in n × n size area _t);

(2) the pixel x in natural image x is calculated _tat its shape-adaptive neighborhood S (x _t) in average traversal view picture natural image x, obtains the image x ' that a width is new;

(3) line by line scan in the image x ' newly obtained, choose the pixel x ' of i-th position _ifor pixel to be restored, with pixel x ' to be restored _icentered by, choose r × r size square center block of pixels V (x ' _i);

(4) with the pixel x ' to be restored chosen _icentered by, choose the square aearch window of f × f size, line by line scan in search window, choose a pixel x ' of a jth position _jas pixel x ' to be restored _isimilitude, with this similitude x ' _jcentered by, choose one with center pixel block V (x ' _i) equal-sized block of pixels as similar piece of V (x ' _j);

(5) according to center pixel block V (x ' _i) to selected by step (4) similar piece of V (x ' _j), calculate pixel x ' to be restored _iwith similitude x ' _jbetween similarity weight w (x ' _i, x ' _j):

Wherein, V (x ' _i) and V (x ' _j) be respectively center pixel block selected by step (3) and similar piece selected by step (4), h is level and smooth controling parameters, and σ is the Gaussian noise standard deviation in natural image, h=k σ, k is a constant, similarity weight w (x ' _i, x ' _j) satisfy condition: 0≤w (x ' _i, x ' _j)≤1, and ∑ w (x ' _i, x ' _j)=1;

(6) with similarity weight w (x ' _i, x ' _j) with the pixel x of a jth position in the natural image x treating denoising _jbe weighted average, calculate pixel x ' to be restored _irecovery value

Wherein, f × f is the size of square aearch window, v (x _j) be the pixel x of a jth position in natural image x _jgray-scale value, C (x ' _j) be normalization constant, be by following formulae discovery:

(7) with pixel x ' to be restored _irecovery value replace and treat i-th position pixel x in the natural image x of denoising _igray-scale value, obtain the image after denoising.