CN104992436A - Image segmentation method for natural scene - Google Patents

Abstract

The invention relates to an image segmentation method for a natural scene, and the method comprises the following steps: 1, inputting a to-be-processed image; 2, judging whether the image is a gray scale image or not; 3, carrying out initialization; 4, initializing a clustering center V or a membership grade U; 5, solving an Euclidean distance between a neighborhood point and the clustering center V for (iter-th) iteration; 6, calculating the membership grade Uiter of the (iter-th) iteration again if the solved one is the clustering center Viter of the (iter-th) iteration, and calculating the clustering center Viter of the (iter-th) iteration again if the solved one is the membership grade Uiter of the (iter-th) iteration; 7, completing the image segmentation and outputting the segmented image if the difference between the membership grades before and after the (iter-th) iteration is less than an iteration stop threshold value epsilon or the number iter of iteration times is greater than the maximum number maxIter of iteration times. Compared with the prior art, the method greatly improves the effect of image segmentation.

Description

Image partition method in a kind of natural scene

Technical field

The present invention relates to a kind of image partition method, especially relate to the image partition method in a kind of natural scene.

Background technology

Along with developing rapidly of Internet technology and ecommerce, digital picture is all increasing every day with surprising rapidity, in a large amount of images, how to find the image interested to user to become the popular problem of current research.Iamge Segmentation is exactly image is divided into several are specific, have peculiar property region and proposes technology and the process of interesting target.It is by the committed step of image procossing to graphical analysis.Existing image partition method mainly divides following a few class: the dividing method based on threshold value, the dividing method based on region, the dividing method based on edge and the dividing method etc. based on particular theory.The various aspects such as FCM Algorithms (i.e. FCM) is a kind of partitioning algorithm based on particular theory, is widely used in MRI brain image segmentation at present, text identification, the process of Car license recognition and large data.

FCM is proposed by Dunn and is promoted by Bezdek, and it with the addition of membership function on the basis of Hard clustering, makes each sample point no longer belong to a certain class determined, but is under the jurisdiction of different classes in certain proportion.This, with regard to making traditional FCM comparatively Hard clustering, can keep more image information.But traditional FCM does not consider the information of surrounding neighbors, make FCM algorithm to noise sensitivity very.FCM_S algorithm considers the information of surrounding neighbors on the basis of FCM, makes the classification of central point by the impact of surrounding neighbors point classification, this greatly enhances the robustness of algorithm to noise and singular point.But algorithm all will calculate surrounding neighbors point in each iterative process, makes algorithm very consuming time.And when noise increases, FCM_S algorithm well can not remove noise, often occurs the situation of misclassification class.

Summary of the invention

Object of the present invention be exactly in order to overcome above-mentioned prior art exist defect and the image partition method in a kind of natural scene is provided.

Object of the present invention can be achieved through the following technical solutions:

An image partition method in natural scene, is characterized in that, comprises the following steps:

Step 1: input a pending picture;

Step 2: judge whether this picture is gray level image, if so, will carry out step 3); Otherwise carry out step 3 after picture being converted into gray level image);

Step 3: setting Fuzzy Exponential m, iteration stopping threshold epsilon and maximum iteration time maxIter, and initialization cluster number c and neighborhood window W;

Step 4: initialization cluster centre V or degree of membership U; Be cluster centre if initialized, then calculate the degree of membership U of the 0th iteration ₀after carry out step 5); Be degree of membership if initialized, then calculate the cluster centre V of the 0th iteration ₀after carry out step 5);

Step 5: for i-th ter time iteration, iter=1,2 ..., maxIter, obtains the Euclidean distance of neighborhood point and cluster centre; And according to degree of membership U _iter-1recalculate the cluster centre V of i-th ter time iteration _iteror according to cluster centre V _iter-1recalculate the degree of membership U of i-th ter time iteration _iter;

Step 6: if be the cluster centre V of i-th ter time iteration required by step 5 _iter, then the degree of membership U of i-th ter time iteration is recalculated _iterand carry out step 7); If required by step 5 is the degree of membership U of i-th ter time iteration _iter, then the cluster centre V of i-th ter time iteration is newly calculated _iterand carry out step 7);

Step 7: if the degree of membership value difference before and after i-th ter time iteration is less than iteration stopping threshold epsilon or iterations iter when exceeding maximum iteration time maxIter, then complete Iamge Segmentation and image after exporting segmentation; Otherwise repeat step 5 and step 6 to carry out next iteration and calculate degree of membership and cluster centre till meeting this condition.

Described step 4) calculate the degree of membership U of the 0th iteration ₀be specially, U ₀for the matrix of C × N, each element is calculated as follows:

$u_{ki}=\frac{{\left({||x_i-v_k||}^2+\underset{\begin{array}{c}r\∈N\\ i\≠r\end{array}}{\Σ}{\left(1-u_{kr}\right)}^m{||x_r-v_k||}^2\right)}^{-1/\left(m-1\right)}}{{\Σ}_{j=1}^C{\left({||x_i-v_j||}^2+\underset{\begin{array}{c}r\∈N\\ i\≠r\end{array}}{\Σ}{\left(1-u_{kr}\right)}^m{||x_r-v_k||}^2\right)}^{-1/\left(m-1\right)}}$

Wherein u _kii-th pixel x in image _ibelong to the degree of kth class, x _ii-th pixel, wherein i ∈ [1,2 ..., N], the number of pixel in N representative picture, C is also will be the number of cluster centre by the number of picture segmentation, x _rrepresent x _ineighborhood point, v _krepresent a kth cluster centre, wherein k ∈ [1,2 ..., C], v _jrepresent a jth cluster centre, m is weights coefficients.

Described step 4) calculate the cluster centre V of the 0th iteration ₀be specially, V ₀for the matrix of C × 1, each element is calculated as follows:

$v_k=\frac{{\Σ}_{i=1}^N{u}_{ki}^m\left(x_i+{\Σ}_{\begin{array}{c}r\∈N\\ i\≠r\end{array}}{\left(1-u_{kr}\right)}^m{x}_r\right)}{{\Σ}_{i=1}^N\left(1+{\Σ}_{\begin{array}{c}r\∈N\\ i\≠r\end{array}}{\left(1-u_{kr}\right)}^m\right)u_{ki}^m}$

Wherein v _krepresent a kth cluster centre, u _kii-th pixel x in image _ibelong to the degree of kth class, the number of pixel in N representative picture, C is also will be the number of cluster centre by the number of picture segmentation, u _krr pixel x in image _ibelong to the degree of kth class, x _rrepresent x _ineighborhood point.

The correctness SA of the method segmentation carrys out the superiority-inferiority of assessment algorithm.

Described SA computing formula is specific as follows:

$SA=\underset{i=1}{\overset{C}{\Σ}}\frac{A_i\∩C_i}{\underset{j=1}{\overset{c}{\Σ}}{C}_j}$

C is the number of cluster, A _irepresentative uses partitioning algorithm to be partitioned into the point belonging to the i-th class, C _irepresent the original point belonging to the i-th class, C _jrepresent the original point belonging to jth class.

Compared with prior art, the present invention adopts the FCM algorithm of improvement, the degree of membership of neighborhood point self is added on respective neighborhood point, the neighborhood point changing FCM_S algorithm has identical impact to central point, add non-noise point to the impact of central point, reduce the impact of noise spot on central point.This improvement adds the robustness of algorithm to noise, substantially increases image segmentation.Synthesising picture and the experiment of natural picture under different noise, demonstrate validity of the present invention.

Accompanying drawing explanation

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

Fig. 2 (a) is original image, and Fig. 2 (b) is the result of FCM algorithm process, the result that Fig. 2 (c) is FCM_S, and Fig. 2 (d) is result of the present invention;

To be three kinds of algorithms giving segmentation accuracy Fig. 2 (a) added under different Gaussian noise to Fig. 3;

Fig. 4 (a) is original image, Fig. 4 (b) is the result of FCM algorithm process, the result that Fig. 4 (c) is FCM_S, Fig. 4 (d) is result of the present invention, and Fig. 4 (e) is the excavator colour picture of reduction;

To be three kinds of algorithms giving segmentation accuracy Fig. 4 (a) added under different salt-pepper noise to Fig. 5.

Embodiment

Below in conjunction with the drawings and specific embodiments, the present invention is described in detail.

Embodiment

The present invention proposes a kind of FCM algorithm of improvement, is not only joined in algorithm by the half-tone information of surrounding neighbors, and the degree of membership information of surrounding neighbors is also joined in algorithm.FCM_S algorithm has the factor controlling neighborhood information impact, and this factor will be debugged manually according to different pictures, to the effect reached.But this just constrains the practical use of this algorithm greatly.And the degree of membership of the present invention's neighborhood point replaces this Graph One factor, this factor will change along with different situations oneself like this, and does not need artificial regulatory, and the present invention can be used in real life.And this improvement increases the robustness of algorithm to noise, realizes better segmentation effect.

1. image segmentation algorithm

The early stage of 1.1 algorithms is theoretical

A. FCM Algorithms (FCM)

FCM algorithm is a kind of clustering algorithm based on dividing, its thought be exactly make to be divided into same cluster object between similarity maximum, and similarity between different bunches is minimum.If X=is (x ₁, x ₂..., x _n) be the data set that will carry out the picture split.The number of pixel in N representative picture, C is also will be the number of cluster centre by the number of picture segmentation, then objective function is such as formula 1.V _krepresent a kth cluster centre, wherein k ∈ [1,2 ..., C].U _kirepresent degree of membership, i.e. i-th pixel x in image _ibelong to the degree of kth class, 0≤u _ki≤ 1 and m is a weights coefficient, generally gets 2.

$J_m=\underset{i=1}{\overset{N}{\Σ}}\underset{k=1}{\overset{C}{\Σ}}{u}_{ki}^m||x_i-v_k||---\left(1\right)$

According to the method for Lagrange multiplier, obtain cluster centre v _kwith degree of membership u _ki, as follows:

$v_k=\frac{{\Σ}_{i=1}^N{u}_{ki}^m{x}_i}{{\Σ}_{i=1}^N{u}_{ki}^m}---\left(2\right)$

$u_{ki}=\frac{{\left({||x_i-v_k||}^2\right)}^{-1/\left(m-1\right)}}{{\Σ}_{j=1}^C{\left({||x_i-v_j||}^2\right)}^{-1/\left(m-1\right)}}---\left(3\right)$

B. there is the fuzzy clustering (FCM_S) of space constraint

FCM algorithm is highstrung to noise, this is because FCM algorithm is the classification based on single pixel, does not consider neighborhood information.The people such as Ahmed add neighborhood information on the basis of FCM, propose a kind of FCM algorithm of improvement, are referred to as FCM_S.The objective function of FCM_S is as follows:

$J_m=\underset{i=1}{\overset{N}{\Σ}}\underset{k=1}{\overset{C}{\Σ}}{u}_{ki}^m{||x_i-v_k||}^2+\frac{\mathrm{\α}}{N_R}\underset{i=1}{\overset{N}{\Σ}}\underset{k=1}{\overset{C}{\Σ}}{u}_{ki}^m\underset{r\∈N_i}{\Σ}{||x_r-v_k||}^2---\left(4\right)$

Wherein x _rrepresent x _ineighborhood point, N _rbe the number of neighborhood point, α is a controling parameters, represents neighborhood information to current pixel point x _iinfluence degree, it regulates algorithm to the robustness of noise and the quality of balance that keeps picture detail information.If α is excessive, namely too rely on neighborhood information, this will fuzzy fall the details of picture; If but α is too small, namely do not consider neighborhood information, so it just becomes the same with FCM algorithm, and at this moment algorithm will be more responsive to noise ratio.So the selection of α is concerning most important FCM_S algorithm.By minimizing formula (4), we can obtain cluster centre v _kwith degree of membership u _ki.

$v_k=\frac{{\Σ}_{i=1}^N{u}_{ki}^m(x_i+\frac{\mathrm{\α}}{N_R}{\mathrm{\Σ}}_{r\∈N_i}{x}_r)}{\left(1+\mathrm{\α}\right){\Σ}_{i=1}^N{u}_{ki}^m}---\left(5\right)$

$u_{ki}=\frac{{\left({||x_i-v_k||}^2+\frac{\mathrm{\α}}{N_R}{\Σ}_{r\∈N_i}{||x_r-v_k||}^2\right)}^{-1/\left(m-1\right)}}{{\Σ}_{j=1}^C{\left({||x_i-v_j||}^2+\frac{\mathrm{\α}}{N_R}{\Σ}_{r\∈N_i}{||x_r-v_j||}^2\right)}^{-1/\left(m-1\right)}}---\left(6\right)$

1.2 algorithms propose

FCM algorithm only considered half-tone information, and in image, each point is independently point, between points not any contact separately, so it is to noise sensitivity very.Concerning piece image, the pixel of neighborhood often has identical feature, and has the pixel of same characteristic features often to belong to same class, so neighborhood information has great importance concerning Iamge Segmentation.FCM_S algorithm introduces neighborhood information, so FCM_S algorithm enhances the robustness of algorithm to noise to a certain extent, but does not reach our expected effect.And FCM_S algorithm all will calculate surrounding neighbors point in each iterative process, makes algorithm very consuming time.FCM_S algorithm has a parameter alpha, and this parameter needs according to different pictures, and the situations such as different noises manually set.This makes FCM_S be difficult to be used in real life.

The present invention proposes a new algorithm, and the degree of membership of the method neighborhood point self replaces the parameter alpha in FCM_S, automatically regulates algorithm to the robustness of noise and the quality of balance to image detail retentivity.The objective function of this algorithm is as follows:

$J_m=\underset{i=1}{\overset{N}{\Σ}}\underset{k=1}{\overset{C}{\Σ}}{u}_{ki}^m\left({||x_i-v_k||}^2+\underset{\begin{array}{c}r\∈N_i\\ i\≠r\end{array}}{\Σ}{(1-u_{kr})}^m{||x_r-v_k||}^2\right)---\left(7\right)$

In FCM_S algorithm, each neighborhood point is to central point x _iimpact be identical, be irrational in fact like this.If the neighborhood point point that to be noise spot or right and wrong similar, so it is to central point x _iimpact should be a little bit smaller; Otherwise, if neighborhood point and central point x _ibelong to similar, so it is to x _iimpact should be more greatly.So the degree of membership of the present invention's neighborhood point controls the impact of neighborhood information on central point.It is parameterless that this makes the present invention be not only, and greatly strengthen the robustness of algorithm.

Identical with FCM, FCM_S, the cluster centre v of algorithm can be obtained _kwith degree of membership u _ki, as follows:

$v_k=\frac{{\Σ}_{i=1}^N{u}_{ki}^m\left(x_i+{\Σ}_{\begin{array}{c}r\∈N_i\\ i\≠r\end{array}}{\left(1-u_{kr}\right)}^m{x}_r\right)}{{\Σ}_{i=1}^N\left(1+{\Σ}_{\begin{array}{c}r\∈N_i\\ i\≠r\end{array}}{\left(1-u_{kr}\right)}^m\right)u_{ki}^m}---\left(8\right)$

$u_{ki}=\frac{{\left({||x_i-v_k||}^2+\underset{\begin{array}{c}r\∈N_i\\ i\≠r\end{array}}{\Σ}{\left(1-u_{kr}\right)}^m{||x_r-v_k||}^2\right)}^{-1/\left(m-1\right)}}{{\Σ}_{j=1}^C{\left({||x_i-v_j||}^2+\underset{\begin{array}{c}r\∈N_i\\ i\≠r\end{array}}{\Σ}{\left(1-u_{kr}\right)}^m{||x_r-v_k||}^2\right)}^{-1/\left(m-1\right)}}---\left(9\right)$

1.3 algorithm steps

As shown in Figure 1, algorithm steps of the present invention comprises as lower part:

Step a: input a pending picture, this picture may contain noise, the interference such as dust storm.

Step b: judge whether this picture is gray level image, if so, will carry out next step; Gray level image is converted into if not by picture.

Step c: setting Fuzzy Exponential m, iteration stopping threshold epsilon and maximum iteration time maxIter, and initialization cluster number c and neighborhood window W;

Steps d: initialization cluster centre V or degree of membership U.Be cluster centre if initialized, then obtain the degree of membership U of the 0th iteration according to formula 9 ₀; Be degree of membership if initialized, then obtain the cluster centre V of the 0th iteration according to formula 8 ₀.

Step e: for i-th ter time iteration, iter=1,2 ..., maxIter, obtains the Euclidean distance of neighborhood point and cluster centre.And according to degree of membership U _iter-1recalculate the cluster centre V of i-th ter time iteration _iteror according to cluster centre V _iter-1recalculate the degree of membership U of i-th ter time iteration _iter.

Step f: if be the cluster centre V of i-th ter time iteration required by step e _iter, then the degree of membership U of i-th ter time iteration is recalculated _iter; If required by step e is the degree of membership U of i-th ter time iteration _iter, then the cluster centre V of i-th ter time iteration is newly calculated _iter.

Step g: the degree of membership value difference judging before and after i-th ter time iteration is less than iteration stopping threshold epsilon or iterations iter when exceeding maximum iteration time maxIter, then complete Iamge Segmentation and image after exporting segmentation; Otherwise repeat step e and step f to carry out next iteration and calculate degree of membership and cluster centre till meeting this condition.

2. experiment test and analysis

In order to prove the superiority of algorithm, the present invention will and FCM, FCM_S compare.The present invention tests with the picture under two natural scenes respectively.To all algorithms, the present invention arranges m=2, ξ=0.01 (algorithm end condition), N _r=9 (number of neighborhood point, the i.e. windows of 3 × 3 sizes), concerning FCM_S, the value of parameter alpha needs to arrange according to different pictures, by description of test below.

The correctness (SA) of the present invention's segmentation carrys out the superiority-inferiority of assessment algorithm.Its formula is as follows:

$SA=\underset{i=1}{\overset{C}{\Σ}}\frac{A_i\∩C_i}{\underset{j=1}{\overset{c}{\Σ}}{C}_j}---\left(10\right)$

Wherein C is the number of cluster, A _irepresentative uses partitioning algorithm to be partitioned into the point belonging to the i-th class, C _irepresent the original point belonging to the i-th class.

Fig. 2 (a) is the width picture that Iamge Segmentation often uses, and it is by petal, leaf and background three part composition, but due to shooting and store time noise interference, picture thickens unclear.FCM, FCM_S and the present invention will be used for splitting this picture.Cluster centre C=3, FCM_S parameter alpha is set to 10.Experimental result is as Fig. 2 (b), and 2 (c), shown in 2 (d).FCM algorithm cannot distinguish noise spot and non-noise point, so it puts on an equal footing noise spot and non-noise point, when picture is by noise pollution, it just can not remove noise.But non-noise point it but can realize correct classification, namely as the part fickle in love in picture, the pixel do not substituted by noise spot it can be correct classification; Although FCM_S introduces the information of neighborhood, eliminate partial noise, surrounding neighbors is all the same on the impact of central pixel point, when noise ratio is larger time, it will be used as non-noise point as noise remove and fall, and namely as shown in Fig. 2 (c), part fickle in love has been removed.And in the present invention, neighborhood information decides the impact on center pixel according to the degree of membership of self, noise spot and non-noise point can be distinguished preferably.What Fig. 3 represented is add certain Gaussian noise to Fig. 2 (a), the result that three algorithms obtain.The robustness that the present invention has had noise under natural scene and stability can be found out.

Fig. 4 (a) is the excavator in natural scene, can find out, receives the interference of dust storm when the shooting of this pictures, and picture background under natural scene is very complicated, and sky, very large interference all can be caused to segmentation in soil and meadow.In this experiment, picture is divided into two classes by the present invention, and excavator is a class, and other is as an other class.The parameter alpha of FCM_S is set to 5.Fig. 4 (b), 4 (c), 4 (d) is FCM, FCM_S and result of the present invention respectively.From this experiment, obviously can find out that FCM algorithm is single pixel, be not subject to the impact of neighborhood information, what color was close is just attributed to same class, and what color was different is exactly an other class, so the partially light sky of excavator and color has just belonged to same class.And the introducing of FCM_S algorithm neighborhood information makes experimental result be deteriorated on the contrary, as can be seen from Fig. 4 (c), excavator, sky, soil can not be split out, and segmentation is in block distribution.And while invention removes most of noise, correct has extracted the part of excavator.But because the present invention has also introduced the information of neighborhood, so cutting state also presents block distribution.Fig. 5 shows the segmentation situation of Fig. 4 (a) under different salt-pepper noise.As can be seen from Figure 5 parameter alpha be the FCM_S of 10 when noise increases, its Clustering Effect is not as good as FCM.And when parameter alpha is 0.01, when namely neighborhood information affects hardly on central point, can see that the cluster result of FCM_S is similar with FCM.As can be seen from this width picture, neighborhood point is deteriorated on the identical segmentation result that may make of central point impact.And although the present invention adds neighborhood information, it can change along with self degree of membership, adjusts oneself generic, improves accuracy.

3. conclusion

In FMC_S algorithm, neighborhood point has identical effect to central point, and this just makes the noise spot in neighborhood point and non-noise point have identical effect to central point, thus makes FCM_S algorithm lack enough robustnesss to noise.The present invention proposes a FCM algorithm improved, the degree of membership of neighborhood point self is added on respective neighborhood point, the neighborhood point changing FCM_S algorithm has identical impact to central point, adds non-noise point to the impact of central point, reduces the impact of noise spot on central point.This improvement adds the robustness of algorithm to noise, substantially increases image segmentation.Synthesising picture and the experiment of natural picture under different noise, demonstrate validity of the present invention.

Claims (5)

1. the image partition method in natural scene, is characterized in that, comprises the following steps:

Step 1: input a pending picture;

Step 2: judge whether this picture is gray level image, if so, will carry out step 3); Otherwise carry out step 3 after picture being converted into gray level image);

Step 3: setting Fuzzy Exponential m, iteration stopping threshold epsilon and maximum iteration time maxIter, and initialization cluster number c and neighborhood window W;

Step 4: initialization cluster centre V or degree of membership U; Be cluster centre if initialized, then calculate the degree of membership U of the 0th iteration ₀after carry out step 5); Be degree of membership if initialized, then calculate the cluster centre V of the 0th iteration ₀after carry out step 5);

Step 5: for i-th ter time iteration, iter=1,2 ..., maxIter, obtains the Euclidean distance of neighborhood point and cluster centre; And according to degree of membership U _iter-1recalculate the cluster centre V of i-th ter time iteration _iteror according to cluster centre V _iter-1recalculate the degree of membership U of i-th ter time iteration _iter;

Step 6: if be the cluster centre V of i-th ter time iteration required by step 5 _iter, then the degree of membership U of i-th ter time iteration is recalculated _iterand carry out step 7); If required by step 5 is the degree of membership U of i-th ter time iteration _iter, then the cluster centre V of i-th ter time iteration is newly calculated _iterand carry out step 7);

Step 7: if the degree of membership value difference before and after i-th ter time iteration is less than iteration stopping threshold epsilon or iterations iter when exceeding maximum iteration time maxIter, then complete Iamge Segmentation and image after exporting segmentation; Otherwise repeat step 5 and step 6 to carry out next iteration and calculate degree of membership and cluster centre till meeting this condition.

2. the image partition method in a kind of natural scene according to claim 1, is characterized in that, described step 4) calculate the degree of membership U of the 0th iteration ₀be specially, U ₀for the matrix of C × N, each element is calculated as follows:

$u_{ki}=\frac{{\left(\right||x_i-v_k|{|}^2+\underset{i\≠r}{\underset{r\∈N}{\mathrm{\Σ}}}{(1-u_{kr})}^m\left|\right|x_r-v_k\left|{|}^2\right)}^{-1/(m-1)}}{{\mathrm{\Σ}}_{j=1}^C{\left(\right||x_i-v_j|{|}^2+\underset{i\≠r}{\underset{r\∈N}{\mathrm{\Σ}}}{(1-u_{kr})}^m\left|\right|x_r-v_k\left|{|}^2\right)}^{-1/(m-1)}}$

Wherein u _kii-th pixel x in image _ibelong to the degree of kth class, x _ii-th pixel, wherein i ∈ [1,2 ..., N], the number of pixel in N representative picture, C is also will be the number of cluster centre by the number of picture segmentation, x _rrepresent x _ineighborhood point, v _krepresent a kth cluster centre, wherein k ∈ [1,2 ..., C], v _jrepresent a jth cluster centre, m is weights coefficients.

3. the image partition method in a kind of natural scene according to claim 1, is characterized in that, described step 4) calculate the cluster centre V of the 0th iteration ₀be specially, V ₀for the matrix of C × 1, each element is calculated as follows:

$v_k=\frac{{\mathrm{\Σ}}_{i=1}^N{u}_{ki}^m(x_i+{\Σ}_{{}_{i\≠r}^{r\∈N}}{(1-u_{kr})}^m{x}_r)}{{\mathrm{\Σ}}_{i=1}^N(1+{\Σ}_{{}_{i\≠r}^{r\∈N}}{(1-u_{kr})}^m)u_{ki}^m}$

Wherein v _krepresent a kth cluster centre, u _ki-th pixel x in i image _ibelong to the degree of kth class, the number of pixel in N representative picture, C is also will be the number of cluster centre by the number of picture segmentation, u _krr pixel x in image _ibelong to the degree of kth class, x _rrepresent x _ineighborhood point.

4. the image partition method in a kind of natural scene according to claim 1, is characterized in that, the correctness SA of the method segmentation carrys out the superiority-inferiority of assessment algorithm.

5. the image partition method in a kind of natural scene according to claim 4, is characterized in that, described SA computing formula is specific as follows:

$SA=\underset{i=1}{\overset{C}{\Σ}}\frac{A_i\∩C_i}{\underset{j=1}{\overset{c}{\Σ}}{C}_j}$

C is the number of cluster, A _irepresentative uses partitioning algorithm to be partitioned into the point belonging to the i-th class, C _irepresent the original point belonging to the i-th class, C _jrepresent the original point belonging to jth class.