CN101853496A - Hybrid image filtering method based on target scale - Google Patents

Abstract

The invention relates to a hybrid image filtering method based on target scale, which is mainly used for self-adaptive median filtering and anisotropic diffusion. The method is characterized in that: first anisotropic diffusion is carried out to a processed image, thereby inhibiting most noise other than pulse noise; then the target scale of all pixels of the diffused image are calculated; and finally the calculated target scale carries out self-adaptive median filtering to the diffused image so as to remove the rest pulse noise after diffusion and obtain the final result. The method can remove the noise of images with various gray levels, and simultaneously keeps the details of the images while removing the noise.

Description

A kind of hybrid image filtering method of based target yardstick

Affiliated technical field

The present invention relates to a kind of hybrid image filtering method of based target yardstick, relate in particular to the adaptive median filter of based target yardstick and the hybrid image filtering method of anisotropy diffusion, can be applied to all kinds of image pretreatment systems.

Background technology

Be subjected to the influence of hardware, environment, factor such as artificial, always there is noise inevitably in image, and these noises have influenced the truth of image detail largely, and picture quality is reduced.In order to improve picture quality, improve the validity and the reliability of subsequent treatment such as image segmentation, feature extraction, Target Recognition, must use the interference noise in the wave filter removal image.

Traditional linear image filtering method does not have adaptation function, so always noise is being carried out having blured edge of image in the filtering owing to can not consider the local characteristics of image.The anisotropy diffusion is a kind of nonlinear adaptive filtering method based on partial differential equation comparatively commonly used at present, and it can protect the bigger edge of gradient when removing noise.Impulsive noise is because have a bigger gradient usually, thereby can be considered as the edge by the anisotropy diffusion and can not be removed.Spread this defective at anisotropy, existingly improved one's methods two kinds: the first, the anisotropy diffusion is improved itself, enable to remove impulsive noise, but the method computation complexity after improving improved a lot, calculated amount is big; Second, after the anisotropy diffusion, add medium filtering again and remove impulsive noise, but existing median filter all can not be only impulsive noise by intelligent which pixel of identification, often all is that entire image is all carried out filtering, causes some unnecessary details losses.Therefore, must seek a kind of simple filtering method of the original details of image of effectively protecting to greatest extent again simultaneously and remedy the effectively defective of suppressor pulse noise of anisotropic diffusion.

Summary of the invention

The technical matters that solves

For fear of the deficiencies in the prior art part, the present invention proposes a kind of adaptive median filter of based target yardstick and the hybrid image filtering method of anisotropy diffusion.

Basic thought of the present invention is: because target scale reflection is that the localized target structural form at each pixel place is learned the size on the meaning, that is: less at details abundant zone or boundary vicinity target scale, bigger at homogeneity smooth region internal object yardstick, therefore, adaptive median filter among the present invention utilizes this characteristic of target scale, target scale size with each pixel serves as according to adaptive identification pulse noise spot, and the impulsive noise point of selecting is carried out medium filtering.Adaptive median filter among the present invention is owing to only carry out medium filtering to the one part of pixel point, not only reduced calculated amount but also reduced the loss of image detail, cooperates the anisotropy diffusion can suppress various types of point-like noises substantially.

Technical scheme

A kind of hybrid image filtering method of based target yardstick is characterized in that step is as follows:

Step 1: adopt the P-M diffusion equation to carry out the anisotropy diffusion for a width of cloth gray level image, the equation expression formula is:

$\frac{\∂I(x,y,t)}{\∂t}=\mathrm{div}\left(g\right|\▿I(x,y,t)|\▿I(x,y,t))$

Wherein: x, the coordinate of y represent pixel point in processed image, t is a number of iterations, I (t) the processed image of expression is at the image of iteration t after the step for x, y, and div () is the divergence function,

Represent gradient operator, g () is the coefficient of diffusion function, and described coefficient of diffusion function is a Tukey coefficient of diffusion function, and expression formula is:

When | x|≤T, Otherwise g (x)=0;

Wherein T is a Grads threshold, and the computing method of described T are: obtain the histogram of gradients of processed image, the Grad at histogram of gradients 90% place be multiply by Be Grads threshold T;

Step 2: the target scale of each pixel in the image after calculation procedure 1 is handled: $D_{\mathrm{xy}}=R_{\mathrm{xy}}+\frac{U_{\mathrm{xy}}(R_{\mathrm{xy}}+1)}{T_s},$

Wherein: D _Xy(x y) locates the target scale of pixel at coordinate for processed image; R _Xy(x y) locates the integer target scale of pixel at coordinate for processed image; T _sFor threshold parameter equals 0.85; U _Xy() is processed image, and (x y) locates the similarity function in pixel and its field boundary zone at coordinate;

Described R _XyFor: $\begin{array}{c}{R}_{\mathrm{xy}}=\underset{r\&Element;Z,r>1}{\arg \max }\{U_{\mathrm{xy}}\left(r\right)\&GreaterEqual;T_s\}\\ s.t.\&ForAll;R\&Element;Z(0\&le;R<R_{\mathrm{xy}}-1),U_{\mathrm{xy}}(r-R)\&GreaterEqual;T_s\end{array},$

Described U _Xy() is: $U_{\mathrm{xy}}\left(R\right)=\frac{\underset{(i,j)\&Element;B_{\mathrm{xy}}\left(R\right)}{\mathrm{\&Sigma;}}\exp [-\frac{{(I_{(x,y)}-I_{(i,j)})}^2}{2{\mathrm{\&sigma;}}_{\mathrm{\&mu;}}^2}]}{\left|B_{\mathrm{xy}}\left(R\right)\right|},$

Wherein: I _{(x, y)}(x y) locates gray values of pixel points, I at coordinate for processed image _{(i, j)}For processed image in that (i j) locates gray values of pixel points, σ _μBe the statistical property parameter that the reflection image gradient distributes, B _Xy(R) be processed image coordinate (x y) locates the borderline region that the pixel radius is the neighborhood of R, | B _Xy(R) | be the number of pixel in the borderline region;

Described σ _μAsk for: calculate the histogram of gradients of processed image, then the gradient of high 20% part in the histogram of gradients is removed, remaining gradient is averaged equals σ _μ

Described B _Xy(R) expression formula is:

B _xy(R)＝{(i，j)|(i，j)∈N _xy(R)-N _xy(R-1)}

Wherein: N _Xy(R) be that (x y) locates the neighborhood that the pixel radius is R to processed image at coordinate;

Described N _XyThe expression formula of () is:

N _xy(R)＝{(x，y)||x-i?|≤R，|y-i?|≤R}；

Step 3: utilize the image after the target scale of obtaining in the step 2 is handled step 1 to carry out adaptive median filter, the expression formula of adaptive median filter is:

Work as D _XyDuring＜T, M _{X, y}=median{f _{X-k, x-l}(k, l) ∈ W},

Work as D _XyDuring 〉=T, M _{X, y}=I _{X, y},

Wherein: M _{X, y}(x y) locates gray values of pixel points at coordinate for filtered image; I _{X, y}(x y) locates gray values of pixel points at coordinate for filtered image; W is that (x y) locates the pixel neighborhood of a point to filtered image, and the number of neighborhood interior pixel point is an odd number at coordinate; K is the capable sequence number of each pixel in the W, and l is the row sequence number of each pixel in the W; f _{X-k, x-1}(x-k y-l) locates gray values of pixel points at coordinate for filtered image; Median{} is that to get median be element value in the middle of getting after all elements in the pair set sorts by size; T is that the threshold value of target scale is 0.3.

Between the gray area of described gray level image [0,1] or [0,255].

Beneficial effect

In the hybrid image filtering method of the adaptive median filter of the based target yardstick that the present invention proposes and anisotropy diffusion, most noises that the anisotropy diffusion of at first using can suppress except that impulsive noise also keep edge of image simultaneously; The adaptive median filter of Shi Yonging then was according to selecting afterpulse noise spot that anisotropy diffusion fails to suppress and with its removal with the target scale afterwards, in the overwhelming majority uses, impulsive noise point generally only accounts for the sub-fraction of the total pixel of image, therefore the adaptive median filter among the present invention can make the most pixels in the processed image avoid handling when removing impulsive noise, has not only reduced calculated amount but also has protected the details of image; Two kinds of wave filters among the present invention not only all possess the ability that keeps the image border but also have complementarity, so the method that the present invention proposes can obtain good effect to all kinds point-like Noise Suppression.

Description of drawings

Fig. 1: the basic flow sheet of the inventive method

Fig. 2: use the inventive method to finish the example of infrared image filtering emulation experiment

(a) original image

(b) add the image of making an uproar

(c) to adding the image after the image of making an uproar uses linear gaussian filtering

(d) to adding the image after the image of making an uproar uses the anisotropy diffusion

(e) to adding the image after the image use anisotropy diffusion of making an uproar adds traditional medium filtering

(f) to adding the image after the image of making an uproar uses the inventive method

Embodiment

Now in conjunction with the embodiments, accompanying drawing is further described the present invention:

The hardware environment that is used to implement is: Duo 2 double-core 2G computing machines, 2GB internal memory, 512M video card, the software environment of operation is: Matlab7.10 and Windows XP.We have carried out emulation experiment with Matlab software to the method that the present invention proposes.Original image has been selected an infrared picture that resolution is 343*288 for use, shown in Fig. 2 (a), in stack average on the original image is that 0 variance is to obtain adding the image of making an uproar after 100 Gaussian noise and noise density are 0.01 salt-pepper noise, shown in Fig. 2 (b), the method that adopts this law proposition then is to Fig. 2 (L) filtering.

The concrete enforcement of the present invention is as follows:

1, anisotropy diffusion: adopt following P-M diffusion equation that Fig. 2 (b) is carried out the anisotropy diffusion:

$\frac{\&PartialD;I(x,y,t)}{\&PartialD;t}=\mathrm{div}\left(g\right|\&dtri;I(x,y,t)|\&dtri;I(x,y,t))$

Coefficient of diffusion function in the formula is selected Tukey coefficient of diffusion function for use:

$g\left(x\right)=\left\{\begin{array}{cc}\frac{1}{2}{[1-{(x/T)}^2]}^2& \left|x\right|\&le;T\\ 0& \mathrm{otherwise}\end{array}\right.$

T in the formula can calculate in the following way: obtain the histogram of gradients of Fig. 2 (b), the Grad at histogram of gradients 90% place be multiply by

Compose and give T.

The discrete iteration solution formula of described P-M diffusion equation is:

$I_{i,j}^{t+1}=I_{i,j}^t+\mathrm{\&lambda;}(C_N^t\&CenterDot;{\&dtri;}_N{I}_{i,j}^t+C_S^t\&CenterDot;{\&dtri;}_S{I}_{i,j}^t+C_W^t\&CenterDot;{\&dtri;}_W{I}_{i,j}^t+C_E^t\&CenterDot;{\&dtri;}_E{I}_{i,j}^t)$

Described

Represent I (x, y, t) coordinate (i j) locates gray values of pixel points,

(t+1) (i j) locates gray values of pixel points at coordinate for x, y to represent I;

Described Represent respectively I (x, y, t) coordinate (i, the pixel of j) locating are in the difference of upper and lower, left and right four direction, and computing formula is:

${\&dtri;}_N{I}_{i,j}^t=I_{i-1,j}^t-I_{i,j}^t$

${\&dtri;}_S{I}_{i,j}^t=I_{i+1,j}^t-I_{i,j}^t$

${\&dtri;}_W{I}_{i,j}^t=I_{i,j+1}^t-I_{i,j}^t$

${\&dtri;}_E{I}_{i,j}^t=I_{i,j-1}^t-I_{i,j}^t$

Described Represent respectively I (X, y, t) coordinate (i, the pixel of j) locating are at the coefficient of diffusion of upper and lower, left and right four direction, and computing formula is:

$C_N^t=g\left(\right|{\&dtri;}_N{I}_{i,j}^t\left|\right)$

$C_S^t=g\left(\right|{\&dtri;}_S{I}_{i,j}^t\left|\right)$

$C_W^t=g\left(\right|{\&dtri;}_W{I}_{i,j}^t\left|\right)$

$C_E^t=g\left(\right|{\&dtri;}_E{I}_{i,j}^t\left|\right)$

Described λ is a constant in the iterative computation, and span is 0≤λ≤1/4.Get λ=1/4 in this example, in four steps of iteration, Fig. 2 (d) is the image after the diffusion of Fig. 2 (b) use anisotropy.

2, calculate target scale: the target scale D of each pixel in the calculating chart 2 (d) _XyComputing formula is as follows:

$D_{\mathrm{xy}}=R_{\mathrm{xy}}+\frac{U_{\mathrm{xy}}(R_{\mathrm{xy}}+1)}{T_s}$

T _sBe a threshold parameter, be set at 0.85.U in the following formula _Xy(R) and R _XyComputing formula as follows:

$U_{\mathrm{xy}}\left(R\right)=\frac{\underset{(i,j)\&Element;B_{\mathrm{xy}}\left(R\right)}{\mathrm{\&Sigma;}}\exp [-\frac{{(I_{(x,y)}-I_{(i,j)})}^2}{2{\mathrm{\&sigma;}}_{\mathrm{\&mu;}}^2}]}{\left|B_{\mathrm{xy}}\left(R\right)\right|}$

$R_{\mathrm{xy}}=\underset{r\&Element;Z,r>1}{\arg \max }\{U_{\mathrm{xy}}\left(r\right)\&GreaterEqual;T_s\}$

$s.t.\&ForAll;R\&Element;Z(0\&le;R<R_{\mathrm{xy}}-1),U_{\mathrm{xy}}(r-R)\&GreaterEqual;T_s$

σ in the following formula _μIt is the statistical property parameter that the reflection image gradient distributes, can calculate in the following way: the histogram of gradients of at first obtaining Fig. 2 (d), then the gradient of high 20% part in the histogram of gradients is removed, remaining gradient is averaged, at last the average of obtaining is composed to σ _μB in the following formula _Xy(R) be:

B _xy(R)＝{(i，j)|(i，j)∈N _xy(R)-N _xy(R-1)}

N _xy(R)＝{(x，y)||x-i|≤R，|y-i?|≤R}

| B _Xy(R) | represent B _Xy(R) number of pixel in.

3, adaptive median filter: the D that utilizes previous step to obtain _XyTo Fig. 2 (d) execution carrying out adaptive median filter.The expression formula of adaptive median filter is as follows:

$M_{x,y}=\left\{\begin{array}{cc}\mathrm{median}\{f_{x-k,x-l}(k,l)\&Element;W\},& D_{\mathrm{xy}}<T\\ I_{x,y},& D_{\mathrm{xy}}\&GreaterEqual;T\end{array}\right.$

Median{} is for getting median, that is: all elements in the pair set sorts by size, the element value of size output in the middle of getting; T represents the threshold value of target scale, is set at 0.3.Fig. 2 (f) is Fig. 2 (d) is carried out image behind the adaptive median filter.

Method proposed by the invention and additive method are compared the result that Fig. 2 (b) handles, and evaluation result is as shown in table 1.The size of image PSNR (spike signal to noise ratio (S/N ratio)) has shown the degree of closeness of filtered image and original image, and PSNR is big more, and filtered image illustrates that more near original image the effect of filtering is good more.

Table 1 filtering evaluation of result

Claims (2)

1. the hybrid image filtering method of a based target yardstick is characterized in that step is as follows:

Step 1: adopt the P-M diffusion equation to carry out the anisotropy diffusion for a width of cloth gray level image, the equation expression formula is:

$\frac{\&PartialD;I(x,y,t)}{\&PartialD;t}=\mathrm{div}\left(g\right|\&dtri;I(x,y,t)|\&dtri;I(x,y,t))$

Wherein: x, the coordinate of y represent pixel point in processed image, t is a number of iterations, I (t) the processed image of expression is at the image of iteration t after the step for x, y, and div () is the divergence function, Represent gradient operator, g () is the coefficient of diffusion function, and described coefficient of diffusion function is a Tukey coefficient of diffusion function, and expression formula is:

When | x|≤T,

Otherwise g (x)=0;

Wherein T is a Grads threshold, and the computing method of described T are: obtain the histogram of gradients of processed image, the Grad at histogram of gradients 90% place be multiply by

Be Grads threshold T;

Step 2: the target scale of each pixel in the image after calculation procedure 1 is handled: $D_{\mathrm{xy}}=R_{\mathrm{xy}}+\frac{U_{\mathrm{xy}}(R_{\mathrm{xy}}+1)}{T_s},$

Wherein: D _Xy(x y) locates the target scale of pixel at coordinate for processed image; R _Xy(x y) locates the integer target scale of pixel at coordinate for processed image; T _sFor threshold parameter equals 0.85; U _Xy() is processed image, and (x y) locates the similarity function in pixel and its field boundary zone at coordinate;

Described R _XyFor: $\begin{array}{c}{R}_{\mathrm{xy}}=\underset{r\&Element;Z,r>1}{\arg \max }\{U_{\mathrm{xy}}\left(r\right)\&GreaterEqual;T_s\}\\ s.t.\&ForAll;R\&Element;Z(0\&le;R<R_{\mathrm{xy}}-1),U_{\mathrm{xy}}(r-R)\&GreaterEqual;T_s\end{array},$

Described U _Xy() is: $U_{\mathrm{xy}}\left(R\right)=\frac{\underset{(i,j)\&Element;B_{\mathrm{xy}}\left(R\right)}{\mathrm{\&Sigma;}}\exp [-\frac{{(I_{(x,y)}-I_{(i,j)})}^2}{2{\mathrm{\&sigma;}}_{\mathrm{\&mu;}}^2}]}{\left|B_{\mathrm{xy}}\left(R\right)\right|},$

Wherein: I _{(x, y)}(x y) locates gray values of pixel points, I at coordinate for processed image _{(i, j)}For processed image in that (i j) locates gray values of pixel points, σ _μBe the statistical property parameter that the reflection image gradient distributes, B _Xy(R) be processed image coordinate (x y) locates the borderline region that the pixel radius is the neighborhood of R, | B _Xy(R) | be the number of pixel in the borderline region;

Described σ _μAsk for: calculate the histogram of gradients of processed image, then the gradient of high 20% part in the histogram of gradients is removed, remaining gradient is averaged equals σ _μ

Described B _Xy(R) expression formula is:

B _xy(R)＝{(i，j)|(i，j)∈N _xy(R)-N _xy(R-1)}

Wherein: N _Xy(R) be that (x y) locates the neighborhood that the pixel radius is R to processed image at coordinate;

Described N _XyThe expression formula of () is:

N _xy(R)＝{(x，y)||x-i|≤R，|y-i?|≤R}；

Step 3: utilize the image after the target scale of obtaining in the step 2 is handled step 1 to carry out adaptive median filter, the expression formula of adaptive median filter is:

Work as D _XyDuring＜T, M _{X, y}=median{f _{X-k, x-l}, (k, l) ∈ W},

Work as D _XyDuring 〉=T, M _{X, y}=I _{X, y},

Wherein: M _{X, y}(x y) locates gray values of pixel points at coordinate for filtered image; I _{X, y}(x y) locates gray values of pixel points at coordinate for filtered image; W is that (x y) locates the pixel neighborhood of a point to filtered image, and the number of neighborhood interior pixel point is an odd number at coordinate; K is the capable sequence number of each pixel in the W, and l is the row sequence number of each pixel in the W; f _{X-k, x-l}(x-k y-l) locates gray values of pixel points at coordinate for filtered image; Median{} is that to get median be element value in the middle of getting after all elements in the pair set sorts by size; T is that the threshold value of target scale is 0.3.

2. the hybrid image filtering method of a kind of based target yardstick according to claim 1 is characterized in that: be between the gray area of described gray level image [0,1] or [0,255].

Images