CN101727657A - Image segmentation method based on cloud model - Google Patents

Abstract

The invention discloses an image segmentation method based on a cloud model. The method comprises the following steps of: firstly, utilizing cloud transformation to realize the transformation of a gray level column diagram of an image and the extraction of lower layer cloud of the image; secondly, utilizing cloud integration to realize cloud merging and zooming; and finally, utilizing the maximum judging method to realize image judgment and image segmentation. The method takes fuzziness and randomness in image segmentation as well as the uncertain relevance between the fuzziness and the randomness into consideration, thus, compared with the traditional image segmentation method, the method has better image segmentation effect.

Description

A kind of image partition method based on cloud model

Technical field

The present invention relates to image segmentation, Target Recognition in the Flame Image Process, particularly a kind of image partition method based on cloud model.

Background technology

Image segmentation is the major issue of graphical analysis and processing, is the technical process (Zhang Yujin, 2001) that extracts interesting target on image.Most of image partition methods are based on Deterministic Methods, yet existence is a lot of uncertain in the image information, and the uncertainty in the image segmentation becomes important research direction (Arnaud martin, Hicham laanaya, 2006).Ambiguity and randomness are probabilistic two importances, and present uncertain image partition method mainly sets about from these two aspects analyzing.Image partition method based on fuzzy theory is mainly considered ambiguity, judges the division (Weiling Cai et al., 2006) of pixel class for the degree of membership of each class by calculating pixel as the fuzzy C-means clustering algorithm.Image partition method based on probability theory is mainly considered randomness, from statistical angle digital picture is carried out modeling, and each gray values of pixel points in the image is regarded as the stochastic variable with certain probability distribution.Wherein, researchist's attention (Cheng Bing etc., 2004 have been obtained based on the image partition method of markov random field (MRF); Yan Gang, 2006).Yet this method can not be handled the ambiguity in the image segmentation effectively.Randomness and ambiguity are closely-related, the two can not be isolated and come.Cloud model and correlation technique thereof had both been considered randomness, had also considered ambiguity, and had considered the relevance between the two, were to analyze probabilistic strong instrument (Li Deyi etc., 1995).As the how theory and the method for this consideration ambiguity of cloud model, randomness and the relevance between the two, the new method that works out uncertain image segmentation is an important probing direction of uncertain image segmentation, will produce some valuable new methods.

Summary of the invention

The present invention is directed to most of uncertain image partition methods and often only consider ambiguity, perhaps only consider the problem of randomness, utilize cloud model to take all factors into consideration the advantage of ambiguity, randomness and the relevance between the two, propose a kind of image partition method based on cloud model.

In the image partition method based on cloud model of the present invention, realize image segmentation by following steps:

Utilize the cloud conversion to realize that the conversion of image grey level histogram realizes that image bottom cloud extracts;

Utilize cloud to realize that comprehensively cloud merges and rises to;

Utilize very big criterion to realize image discriminating and image segmentation.

Wherein, described bottom cloud is made of numerous water dusts.

Specifically, the image partition method based on cloud model of the present invention is realized image segmentation by following steps:

(1), generates the grey level histogram of image to be split according to the image pixel statistical information of former figure;

(2) utilize the cloud conversion that the grey level histogram of image to be split is transformed into a series of bottom cloud C (Ex _i, En _i, He _i), wherein, Ex _i, En _iAnd He _iBe expectation, entropy and the super entropy of i cloud;

(3) progressively merge the nearest bottom cloud of all bottom cloud middle distances, a plurality of altostratus that obtain specifying number are realized by bottom cloud rising to altostratus;

(4) utilize very big criterion to carry out image pixel and be subordinate to differentiation, realize image segmentation;

Wherein, by comparing the expectation Ex of each bottom cloud _iValue, obtain described nearest bottom cloud.

Wherein, the frequency f (x) according to the image pixel gray scale converts image grey level histogram to a series of bottom cloud C (Ex _i, En _i, He _i), its mathematic(al) representation is:

$f\left(x\right)\→\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i*C({\mathrm{Ex}}_i,{\mathrm{En}}_i,{\mathrm{He}}_i)$

In the formula, a _iBe the range coefficient of the size of the water dust number that comprises of each bottom cloud of reflection, n is the number of the discrete notion that generates after the conversion.

Wherein, described step (3) comprising: two nearest bottom clouds of combined distance at first; Put together merging later cloud and other cloud, form the cloud that rises to; Then, merge the nearest cloud of cloud middle distance that rises to, the rest may be inferred, progressively merges, and obtains a plurality of altostratus.

Wherein, the number of described altostratus is specified according to actual conditions by the user.

Wherein, each the bottom cloud in described a series of bottom cloud is made of the included numerous water dusts of this bottom cloud.

Wherein, the step that grey level histogram is converted to a series of bottom clouds comprises:

(2-1) image is carried out statistics of histogram, obtain gradation of image data frequency function g (x);

(2-2) seek the position at the crest value place of DATA DISTRIBUTION function g (x), with its peak value as expectation Ex _iAdd up then with expectation Ex _iBe the gray scale frequency distribution of the neighborhood at center, go match with Ex by the entropy (giving an initial value earlier) of adjusting the match cloud _iBe the gray scale frequency curve of the neighborhood at center, obtain the entropy En of the bottom cloud of match _i, and match is adapted to the DATA DISTRIBUTION function g of described gray scale frequency distribution _i(x);

(2-3) according to a first given concrete value, obtain super entropy He according to experimental result adjustment then _i

(2-4) utilize and the expectation Ex that obtains (2-3) in step (2-2) _i, entropy En _iSuper entropy He _i, obtain the bottom cloud C (Ex of match _i, En _i, He _i);

(2-5) from gradation of image data frequency distribution g (x), deduct the DATA DISTRIBUTION g of known cloud model _i(x), obtain new DATA DISTRIBUTION function g ' (x), and repeating step (2-2) obtains the bottom cloud C (Ex of a plurality of matches to step (2-4) on this basis _i, En _i, He _i).

Wherein, utilizing very big criterion to carry out image pixel is subordinate to and judges and image segmentation step comprises:

(4-1) calculate the degree of certainty that each pixel in the image is under the jurisdiction of each altostratus;

(4-2) according to the principle of degree of certainty maximum just each pixel of entire image differentiate the altostratus of degree of membership maximum;

(4-3) all pixels that will be under the jurisdiction of each altostratus are differentiated and are certain specific image category.

Said method of the present invention had both been considered the ambiguity in the image segmentation, also considered randomness, and considered the uncertainty of the relevance of the two, thereby with respect to traditional image partition method, had the better pictures segmentation effect.

Below in conjunction with accompanying drawing method of the present invention is elaborated.

Description of drawings

Fig. 1 has shown the grey level histogram of experimental image;

Fig. 2 has shown and utilizes the cloud conversion to obtain level image cloud notion;

Fig. 3 has shown and utilizes cloud comprehensively to obtain the altostratus notion;

Fig. 4 has shown and utilizes very big criterion to realize image segmentation;

Fig. 5 a1-Fig. 5 b4 has shown the part experiment of image segmentation;

Fig. 6 a-6c has shown the former figure of experiment of experiment comparative analysis, and wherein, accompanying drawing 6a is people's image pattern picture, and interested target is a hair; Accompanying drawing 6b is a cell image, and interested target is a cell body; Accompanying drawing 6c drops on epiphyllous image for dragonfly Yan, and interested target is the health of dragonfly Yan;

Fig. 7 a-7c has shown the reference diagram of experiment comparative analysis, has represented a kind of optimal segmentation result near truth, can be used as the standard of cutting apart quality analysis.Wherein, accompanying drawing 7a is the reference diagram of people's image pattern picture; Accompanying drawing 7b is the reference diagram of cell image; Accompanying drawing 7c is the reference diagram of dragonfly Yan image.

Embodiment

The cloud that the present invention relates to comprises the cloud between bottom cloud, altostratus and bottom cloud and the altostratus.In general, cloud is made up of numerous water dusts that it comprised, and water dust has following characteristics: a water dust is qualitativing concept once realization quantitatively, and water dust is many more, can reflect the global feature of this qualitativing concept more; The probability that water dust occurs is big, and the degree of certainty of water dust is big, and then water dust is big to the contribution of notion.

Cloud comes notion of general token with expectation Ex (Expected value), entropy En (Entropy) and 3 numerical characteristics of super entropy He (Hyper entropy).

Expectation Ex: water dust is in the expectation of domain space distribution.Generally, be exactly the point that can represent qualitativing concept, or perhaps the typical sample that quantizes of this notion.

Entropy En: the uncertainty measure of qualitativing concept is determined jointly by the randomness and the ambiguity of notion.En is the tolerance of qualitativing concept randomness on the one hand, has reflected the dispersion degree that can represent the water dust of this qualitativing concept; Be again the tolerance of being this or that property of qualitativing concept on the other hand, reflected in the domain space span of the water dust that can be accepted by notion.Reflect randomness and ambiguity with same numerical characteristic, also must reflect the relevance between them.

Super entropy He: be the uncertainty measure of entropy, i.e. the entropy of entropy.Randomness and ambiguity by entropy determine jointly.

Say that from the general extent the uncertainty of notion can be represented with a plurality of numerical characteristics.Can think: the expectation in the probability theory, variance and High Order Moment are a plurality of numerical characteristics of reflection randomness, but do not touch ambiguity; Degree of membership is an accuracy method of ambiguity, does not but consider randomness; Rough set be use based on two under the accurate knowledge background accurately set measure probabilisticly, but ignored the uncertainty of background knowledge.In cloud method, except expectation, entropy, super entropy, can also go to portray the uncertainty of notion with the entropy of high-order more, can infinitely chase after deeply in theory.

Image partition method based on cloud model of the present invention comprises following three steps:

The first, thus utilize the cloud conversion to realize that the conversion of image grey level histogram realizes that the image bottom extracts;

The second, utilize cloud to realize that comprehensively cloud merges and rises to;

The 3rd, utilize very big criterion to realize image discriminating and image segmentation.

Be elaborated below:

(1) utilize the cloud conversion to realize the extraction of image bottom notion

The grey level histogram of image is the statistical information of gradation of image, based on this, utilizes the cloud conversion that the grey level histogram of image is decomposed into a series of normal state cloud.

Grey level histogram with image shown in Figure 1 is an example, the gray-scale value of abscissa axis X representative image pixel, axis of ordinates is represented the frequency values of gray-scale value, it has reflected the distribution situation of the frequency (occurrence number) of all grey scale pixel values of image, and frequency f wherein (x) is the function of the gray scale of this grey level histogram.Utilize cloud maker (being the normal function generator), grey level histogram can be converted to (resolving into) a series of normal state cloud atlas, promptly a series of bottom clouds.

The cloud conversion is a kind of a kind of method from quantitative to qualitative conversion, each cloud C (Ex _i, En _i, He _i) have discrete, character qualitatively.

The mathematic(al) representation of cloud conversion is:

$f\left(x\right)\→\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i*C({\mathrm{Ex}}_i,{\mathrm{En}}_i,{\mathrm{He}}_i)$

In the formula, a _iBe range coefficient; N is the number of the discrete notion that generates after the conversion;

The step that grey level histogram is converted to (resolving into) a series of bottom clouds can may further comprise the steps:

1, image is carried out statistics of histogram, obtain gradation of image data frequency function g (x);

2, seek the position at the crest value place of DATA DISTRIBUTION function g (x), its peak value is defined as the centre of gravity place of cloud, promptly expect Ex _i(i=0 ..., m-1); Add up then with expectation Ex _iGray scale frequency distribution for the neighborhood at center has the expectation Ex as crest in this gray scale frequency distribution _iAnd the trough of these crest value both sides, the entropy (give earlier an initial value) of progressively adjusting the match cloud model goes the distribution curve between match crest and the trough to obtain the entropy En of match cloud _i, match simultaneously is adapted to the DATA DISTRIBUTION function g of the cloud of gray scale frequency distribution _i(x);

3, super entropy determines.Because when carrying out curve fitting, be that expectation curve and the data frequency distribution curve with cloud model relatively carries out match, the information of super entropy is not used.Therefore, when carrying out the cloud conversion, can adjust according to experimental result then for the direct given concrete value (as 0.1) of super entropy.

4, obtain the bottom cloud model of match.Promptly obtain the expectation Ex of match cloud model according to the calculating of step 2 and step 3 _i, entropy En _iSuper entropy He _i, obtain the bottom cloud C (Ex of match _i, En _i, He _i).

5, from gradation of image data frequency distribution g (x), deduct the DATA DISTRIBUTION g of known cloud model _i(x), obtain new DATA DISTRIBUTION function g ' (x).And repeating step 2 obtains a plurality of DATA DISTRIBUTION function g based on cloud to step 4 on this basis _i(x) be the bottom cloud C (Ex of match _i, En _i, He _i).

(2) utilize cloud comprehensively to realize the merging of image bottom notion and rise to

Two or many bottom clouds that n bottom cloud middle distance is nearest carry out comprehensively just can generating a new altostratus.So all nearest two or many bottom clouds in n the bottom cloud are merged respectively, just can realize cloud model rising to from the bottom cloud to altostratus.

(3) utilize very big criterion to realize that pixel is subordinate to the differentiation and the image segmentation of notion

Comprehensively be met by cloud after the altostratus notion of requirement, each cloud notion is represented an image-region or image type.Utilize very big criterion,, calculate the degree of certainty that it is under the jurisdiction of these altostratus respectively, this pixel is differentiated corresponding cloud notion, thereby realize image segmentation according to the principle of degree of certainty maximum for each pixel.

Utilize very big criterion, to each pixel, according to its gray-scale value, calculate the degree of certainty μ i that it is under the jurisdiction of each altostratus, principle according to the degree of certainty maximum is judged corresponding altostratus with this pixel, thereby each pixel of entire image is judged corresponding altostratus, the corresponding a kind of image type of each altostratus, represent with unified gray-scale value, thereby realize image segmentation.

At a specific embodiment, the image partition method based on cloud model of the present invention is elaborated below.

Step (1) at first generates the grey level histogram of image to be split according to the image pixel statistical information of former figure;

Step (2) utilizes the cloud conversion that the grey level histogram of image to be split is transformed into a series of bottom cloud C (Ex then _i, En _i, He _i), wherein, Ex _i, En _iAnd He _iBe expectation, entropy and the super entropy of i cloud;

Step (3) then progressively merges the nearest bottom cloud of all bottom cloud middle distances, and a plurality of altostratus that obtain specifying number are realized by bottom cloud rising to altostratus;

Step (4) is utilized very big criterion to carry out image pixel at last and is subordinate to differentiation, realizes image segmentation.

Wherein, by comparing the expectation Ex of each bottom cloud _iValue, obtain described nearest bottom cloud.

Wherein, the frequency f (x) according to the image pixel gray scale converts image grey level histogram to a series of bottom cloud C (Ex _i, En _i, He _i), its mathematic(al) representation is:

$f\left(x\right)\→\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i*C({\mathrm{Ex}}_i,{\mathrm{En}}_i,{\mathrm{He}}_i)$

In the formula, a _iBe the range coefficient of the size of the water dust number that comprises of each bottom cloud of reflection, n is the number of the discrete notion that generates after the conversion.

Described step (3) comprising: two nearest bottom clouds of combined distance at first; Put together merging later cloud and other cloud, form the cloud that rises to; Then, merge the nearest cloud of cloud middle distance that rises to, the rest may be inferred, progressively merges, and obtains a plurality of altostratus.

Wherein, the number of described altostratus can be specified according to actual conditions by the user.

Wherein, each the bottom cloud in the described bottom cloud is made of the included numerous water dusts of this bottom cloud.

Wherein, the step that grey level histogram is converted to a series of bottom clouds comprises:

Step (2-1) is carried out statistics of histogram to image, obtains gradation of image data frequency function g (x);

Step (2-2) from frequency distribution function g (x) figure, is sought the position at the crest value place of DATA DISTRIBUTION function g (x), with its peak value as expectation Ex _iAdd up then with expectation Ex _iBe the gray scale frequency distribution of the neighborhood at center, the neighborhood distribution curve of described gray scale frequency as the basis, is progressively adjusted the entropy En that entropy (giving an initial value earlier) obtains the bottom cloud of match _i, and match is adapted to the DATA DISTRIBUTION function g of described gray scale frequency distribution _i(x);

Step (2-3) is according to a first given concrete value, obtain super entropy He according to experimental result adjustment then _i

Step (2-4) is utilized in step (2-2) and the expectation Ex that obtains (2-3) _i, entropy En _iSuper entropy He _i, obtain the bottom cloud C (Ex of match _i, En _i, He _i);

Step (2-5) deducts the DATA DISTRIBUTION g of known cloud model from gradation of image data frequency distribution g (x) _i(x), obtain new DATA DISTRIBUTION function g ' (x), and repeating step (2-2) obtains the bottom cloud C (Ex of a plurality of matches to step (2-4) on this basis _i, En _i, He _i).

Wherein, utilizing very big criterion to carry out image pixel is subordinate to and judges and image segmentation step comprises:

Step (4-1), calculate the degree of certainty that each pixel in the image is under the jurisdiction of each altostratus, because each pixel has its gray-scale value, and each altostratus also has corresponding gray-scale value, utilize and to obtain the degree of certainty μ that each pixel is under the jurisdiction of each altostratus near principle _i

Step (4-2), according to the principle of degree of certainty maximum just each pixel of entire image differentiate the altostratus of degree of membership maximum;

Step (4-3), all pixels differentiations that will be under the jurisdiction of each altostratus are certain specific image category, thereby realize image segmentation.

Fig. 1-Fig. 4 has shown that the cloud conversion that utilizes cloud model, cloud are comprehensive, greatly gordian technique such as criterion realizes the detailed process of image partition method.Fig. 5 a1 to Fig. 5 b4 reality the part experiment of image segmentation, show that this method has obtained effect preferably, wherein, Fig. 5 a1 is former figure, Fig. 5 b1 is the figure of cutting apart at Fig. 5 a1, it extracts Aircraft Target preferably.Fig. 5 a2 is former figure, and Fig. 5 b2 is the figure of cutting apart at Fig. 5 a2, and it extracts people, mountain peak and cloud preferably.Fig. 5 a3 is former figure, and Fig. 5 b3 is the figure of cutting apart at Fig. 5 a3, and it extracts automobile, numeral and alphabetical preferably.Fig. 5 a4 is former figure, and Fig. 5 b4 is the figure of cutting apart at Fig. 5 a4, and it extracts trees preferably.

The experimental image of having selected often to use in the document of 3 width of cloth image segmentation is carried out the quantitative evaluation of image segmentation quality.Test former figure as shown in Figure 6.Wherein, accompanying drawing 6a is people's image pattern picture, and interested target is a hair; Accompanying drawing 6b is a cell image, and interested target is a cell body; Accompanying drawing 6c drops on epiphyllous image for dragonfly Yan, and interested target is the health of dragonfly Yan.The reference diagram of experimental image can relatively obtain mistake branch rate with resulting segmentation result of this method and reference diagram with reference diagram as correct segmentation result shown in Fig. 7 a, 7b, 7c.Simultaneously, utilize the common method in the image segmentation respectively: the K mean cluster is cut apart, fuzzy C-means clustering is cut apart, the optimal threshold dividing method carries out the image segmentation experiment to three width of cloth experimental image in the accompanying drawing 6, and the branch rate is as experimentize result's comparison of evaluation index by mistake, and comparative result is as shown in table 1.

Table 1 mistake is divided the rate contrast table

From the experiment contrast as can be seen, the mistake branch rate of K mean cluster image partition method is the highest; The mistake branch rate of fuzzy C-means clustering dividing method generally is lower than K mean cluster image partition method, this is because the fuzzy C-means clustering dividing method has been considered ambiguity, cut apart with respect to the K mean cluster, considered the ambiguity in the uncertainty, this method science more is reliable.Image partition method based on cloud model proposed by the invention is because utilized the cloud model analysis-by-synthesis and handled the advantage of ambiguity, randomness and the relevance between the two, to probabilistic analysis with handle science more, therefore obtained the better pictures segmentation effect.

Although above the present invention is had been described in detail, the invention is not restricted to this, those skilled in the art of the present technique can carry out various modifications according to principle of the present invention.Therefore, all modifications of doing according to the principle of the invention all should be understood to fall into protection scope of the present invention.

Claims (10)

1. image partition method based on cloud model may further comprise the steps:

(1) utilize the cloud conversion to realize that the conversion of image grey level histogram realizes that image bottom cloud extracts;

(2) utilize cloud to realize that comprehensively cloud merges and rises to;

(3) utilize very big criterion to realize image discriminating and image segmentation.

2. method according to claim 1, wherein said bottom cloud is made of numerous water dusts.

3. image partition method based on cloud model may further comprise the steps:

(1), generates the grey level histogram of image to be split according to the image pixel statistical information of former figure;

(2) utilize the cloud conversion that the grey level histogram of image to be split is transformed into a series of bottom cloud C (Ex _i, En _i, He _i), wherein, Ex _i, En _iAnd He _iBe expectation, entropy and the super entropy of i cloud;

(3) progressively merge the nearest bottom cloud of all bottom cloud middle distances, a plurality of altostratus that obtain specifying number are realized by bottom cloud rising to altostratus;

(4) utilize very big criterion to carry out image pixel and be subordinate to differentiation, realize image segmentation.

4. method according to claim 3 is wherein by comparing the expectation Ex of each bottom cloud _iValue, obtain described nearest bottom cloud.

5. method according to claim 4, wherein the frequency f (x) according to the image pixel gray scale converts image grey level histogram to a series of bottom cloud C (Ex _i, En _i, He _i), its mathematic(al) representation is:

$f\left(x\right)\→\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{a}_i*C(E{x}_i,E{n}_i,H{e}_i)$

In the formula, a _iBe the range coefficient of the size of the water dust number that comprises of each bottom cloud of reflection, n is the number of the discrete notion that generates after the conversion.

6. according to claim 3 or 4 described methods, wherein, described step (3) comprising:

Two nearest bottom clouds of combined distance at first;

Put together merging later cloud and other cloud, form the cloud that rises to;

Then, merge the nearest cloud of cloud middle distance that rises to, the rest may be inferred, progressively merges, and obtains a plurality of altostratus.

7. method according to claim 6, the number of wherein said altostratus is specified according to actual conditions by the user.

8. according to claim 3 or 4 or 5 described methods, each the bottom cloud in wherein said a series of bottom clouds is made of the included numerous water dusts of this bottom cloud.

9. method according to claim 3, the step that wherein grey level histogram is converted to a series of bottom clouds comprises:

(2-1) image is carried out statistics of histogram, obtain gradation of image data frequency function g (x);

(2-2) seek the position at the crest value place of DATA DISTRIBUTION function g (x), with its peak value as expectation Ex _iAdd up then with expectation Ex _iBe the gray scale frequency distribution of the neighborhood at center, progressively adjust entropy (giving an initial value earlier) and remove the neighborhood gray scale curve of frequency distribution of the described gray scale frequency of match, obtain the entropy En of the bottom cloud of match _i, and match is adapted to the DATA DISTRIBUTION function g of described gray scale frequency distribution _i(x);

(2-3) according to a first given concrete value, obtain super entropy He according to experimental result adjustment then _i

(2-4) utilize and the expectation Ex that obtains (2-3) in step (2-2) _i, entropy En _iSuper entropy He _i, obtain the bottom cloud C (Ex of match _i, En _i, He _i);

(2-5) from gradation of image data frequency distribution g (x), deduct the DATA DISTRIBUTION g of known cloud model _i(x), obtain new DATA DISTRIBUTION function g ' (x), and repeating step (2-2) obtains the bottom cloud C (Ex of a plurality of matches to step (2-4) on this basis _i, En _i, He _i).

10. method according to claim 3, wherein utilize very big criterion to carry out image pixel and be subordinate to and judge and image segmentation step comprises:

(4-1) calculate the degree of certainty that each pixel in the image is under the jurisdiction of each altostratus;

(4-2) each pixel of entire image is differentiated the altostratus of degree of membership maximum according to the principle of degree of certainty maximum;

(4-3) all pixels that will be under the jurisdiction of each altostratus are differentiated and are certain specific image category.

Images