CN101030297A - Method for cutting complexity measure image grain - Google Patents

Abstract

A method for dividing image texture of complicated measure includes carrying out binaryzation on image by utilizing automatic threshold with square invariable and making scan on texture image by using Hilbert curve to convert 2-D image information to be 1-D sequence, carrying out window-adding treatment on 1-D sequence then picking up KC and Co complicated texture character to obtain 4-D character vector describing texture complicated character and using support vector computer to obtain texture-divided image.

Description

The image texture dividing method that a kind of complicacy is estimated

Technical field

The invention belongs to the Image Information Processing field, relate to a kind of image texture dividing method, particularly relate to a kind of a kind of dividing method based on nonlinear theory (complicacy is estimated) feature description image texture and support vector machine.

Background technology

Study Of Segmentation Of Textured Images to the Study Of Segmentation Of Textured Images Study on Technology, has important theory and realistic meaning as a basic problem in Target Recognition, image understanding and the computer vision.Be subjected to domestic and international researcher's attention in the past few decades, obtained significant progress.Method about Texture Segmentation emerges in an endless stream, and some method has also obtained gratifying segmentation effect, but generally speaking, not high, unsupervised segmentation effect difference of robustness and time loss height.

A kind of typical texture segmentation algorithm mainly comprises two stages: first stage is the feature extraction to texture, and second stage is to utilize the classification of mode identification technology to feature.

Feature is the key of Texture Segmentation.Cutting apart of general gray level image is based on the gray-scale value consistance, proximity characterizes regional consistance, thereby realize cutting apart, the consistance in zone is to be represented by the consistance of some feature of texture in the zone in texture image, cuts apart to be to carry out on the basis of certain or some features.According to the difference of texture structure, some zones of extraction or split image are to understand the formation of image.How extracting the characteristic quantity of each pixel of expression or representation space intensity profile in image, carry out regional extraction and cut apart according to the texture structure information that is extracted then, is exactly subject matter to be solved, i.e. texture feature extraction in the Texture Segmentation.

In feature extraction phases, utilize certain algorithm, obtain an eigenvector of each pixel in the image or each zonule.In this process, the texture information that lies in the image is converted into more easy-operating vector.Through graphical analysis researchers years of researches, in texture feature extraction, formed several common methods at present, be divided into four classes: based on statistical method, based on structural approach, based on model method with based on space/frequency field method.

(1) based on statistical method

Statistical method is one of method of using in texture analysis, and statistical method is more, the prevailing a kind of method of studying at present, and it utilizes the statistical property of image to obtain eigenwert, cuts apart based on the image feature space consistance.Mainly comprise by autocorrelation function, gray level co-occurrence matrixes and wait the eigenwert of calculating texture image.

Because a key character of texture is repeating of certain pattern, so the autocorrelation function of image is used to assess this regularity the earliest.Autocorrelation function calculates its difference (or other certain computings) and is designated as autocorrelation function by certain specific texture primitive is carried out translation as template in the texture image scope.The shortcoming of autocorrelation function is that method is described the texture with repetition fine structure.

Gray level co-occurrence matrixes describe a pair of at a distance of certain distance, become certain orientation, have the probability of occurrence (frequency) of the pixel of specific gray value respectively, reflected that gradation of image is distributed in the integrated information of direction, local neighborhood and amplitude of variation.In general, there are big, the shortcomings such as segmentation precision is poor, anti-noise ability difference of calculated amount in the feature based on statistics.

(2) based on structural approach

With the queueing discipline between one group of texture primitive and the primitive textural characteristics is described.If analogy done in the character in texture primitive and the language, spatial relationship just can be with being to meet organize grammatical corresponding with description character between the primitive so.Therefore, character combines with the syntax and has just constituted the linguistic model of texture.The synthetic of language can be finished by setting up grammatical character string in linguistics, and the identification of language is carried out syntactic analysis to character string exactly.Similarly, the synthetic and also available similar methods of identification of texture is finished.

The structure analysis method of texture is mainly formed with deduction queueing discipline two parts by extracting texture primitive.Generally only be applicable to the artificial texture that systematicness is stronger, therefore application is very limited.

(3) based on model method

The many models of method research that utilize model to carry out Study Of Segmentation Of Textured Images mainly are probabilistic model (Markov random field models, Gibbs random field models) and fractal model.

Morkov random field (MRF) model is widely used in cutting apart of texture image.MRF is a kind of conditional probability model, can describe the correlativity of each pixel and its neighborhood in the image.The general parameter estimation that adopts is taken turns iteration initialization condition model parameter, cuts apart on this basis, utilizes the further estimation model parameter of segmentation result, and then cuts apart, up to satisfying the condition of convergence.

Fractal is a kind of set that is based upon on the fractal dimension.Natural texture can well be described with fractal.Fractal method mainly contains four steps: the fractal dimension of estimating a certain block of pixels of figure; The histogram of forming fractal dimension; The low ebb place peak-to-peak at histogram is divided into some with histogram; Piece according to different fractal dimensions is determined texture region.

(4) based on space/frequency field method

Can overcome the deficiency of traditional frequency domain analytical approach to a great extent based on graphical representation space/frequency field Conjoint Analysis method.This is a kind of shortcoming of analyzing at Fourier, the new method between spatial domain and frequency field are represented that grows up in signal analysis and the research of human vision mechanism recently.People's vision physiological and psychological research show, human vision is simultaneously to position and spatial frequency sensitivity, and visual cortex is the local information of capture space position and spatial frequency simultaneously.These class methods are similar to people's vision mechanism, can simultaneously obtain the better localization characteristic in spatial domain and frequency field, and its complexity is from by Marr, and three to four frequency channels that Crick and Poggio propose are to the whole spectral coverage of frequency-domain analyser.Frequency-domain analyser can adopt the Wigner function, and Gabor function or section Gauss smooth function are realized.Wherein the Gabor function is to use the most universally at present, and effect is the best way also.

The process that second stage of Texture Segmentation is pattern-recognition and tagsort.The classification of feature is to carry out on the basis of feature extraction.The task of tagsort is that the proper vector of presentation video pixel characteristic is classified by certain similarity criterion.For the Texture Segmentation that supervision is arranged, in the pattern-recognition stage, the eigenvector of the various known textures that we at first calculate according to the phase one is set up a model, then the eigenvector of the known texture in the eigenvector of the unknown texture to be measured and each model is compared, thereby determine that this unknown texture should belong to any texture; For unsupervised Texture Segmentation, the pattern-recognition stage can be regarded the clustering problem of eigenvector in a certain hyperspace as.By cluster, finally reach cutting apart of different texture in the image to eigenvector.

Traditional cluster analysis strictly is divided into a certain class to each sample, belongs to the category of hard division.In fact, the attribute that sample is not strict, they are existing intermediary aspect condition and the generic.Along with the proposition of fuzzy set theory, hard cluster is extended to fuzzy clustering.In fuzzy clustering, each sample no longer belongs to a certain class, but belongs to each class with certain degree of membership.Promptly by the tradition set being gone up two codomains { 0 of fundamental function, 1} expands to interval [0 1], the uncertainty of judging by the fuzzy of the caused notion extension of things intermediary transition and identification with quarterization, thus the fundamental characteristics that information showed in natural things and people's the cognitive process showed effectively.Therefore, by means of fuzzy set theory, can reflect ambiguity and the human ambiguity that in graphical analysis and understanding process, exists that image itself has.

(Support Vector Machine SVM) is the new tool that occurs in recent years to support vector machine in pattern-recognition and machine learning field.The mid-90 in 20th century, a kind of theory of studying machine learning rule under the small sample situation, promptly Statistical Learning Theory (Statistical Learning Theory) begins to be subjected to attention more and more widely.Statistical Learning Theory is based upon a cover than on the solid theory, provides a united frame for solving limited sample learning problem.SVM avoids crossing the problem that traditional classifications such as study, dimension disaster, local minimum exist in the classical learning method based on Statistical Learning Theory effectively, still has good model ability under condition of small sample, therefore has been subjected to paying close attention to widely.The SVM successful Application has arrived multiple fields such as recognition of face, text classification, genetic analysis, speech recognition.Recent years, support vector machine also has been used in the texture image classification, has obtained good effect.

Summary of the invention

Purpose of the present invention is exactly at the deficiencies in the prior art, can from image texture, excavate nonlinear characteristic information as much as possible, all have the image texture dividing method that favorable actual application is worth to realize robustness, splitting speed and a segmentation effect.

In order to realize above purpose, image texture dividing method of the present invention mainly comprises 4 steps:

(1) adopt the constant automatic threshold method of square that image is carried out binaryzation;

(2) utilize the Hilbert curve that texture image is scanned, make two-dimentional image information convert the sequence s of one dimension to;

(3) one-dimensional sequence in (2) is carried out windowing process after, extract KC complicacy textural characteristics and C ₀The complicacy textural characteristics obtains describing one 4 dimensional feature vector of texture complexity features;

(4) adopt support vector machine method, obtain the image after the Texture Segmentation.

The principle that the automatic threshold method that square in the step (1) is constant adopts the preceding third moment of binaryzation front and back image to equate is selected threshold value.

The fundamental element of Hilbert curve comprises " cup " and " connection " in the middle Hilbert curved scanning method of step (2).Described " cup " is the untight rectangle in one side; Described " connection " is the one section directed line segment that connects two " cups ".The Hilbert curve is to be generated by the fundamental element iteration, detailed process is: the Hilbert curve of single order is a rectangle region that covers 2 * 2 zones, the curve of second order replaces the rectangle region of single order by three " connection " with four big " cups " such as grade, and the like, the curve of following single order can utilize that each " cup " in the first order curve obtains in four littler " cups " and three " connection " replacements.

KC complicacy textural characteristics in the step (3) comprises KC value matrix A ₁With KC mean square deviation matrix A ₂, the extracting method concrete steps comprise:

A. to after the one dimension time series s windowing in the step (2), carry out the KC complexity calculations;

B. travel through entire image, all windows are carried out above-mentioned processing, will obtain KC value matrix A ₁

C. utilize ${\mathrm{\μ}}_{\mathrm{kc}}=\frac{1}{M\×N}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\underset{j=i}{\overset{N}{\mathrm{\Σ}}}{A}_1\left[i\right]\left[j\right]$ Obtain KC average μ _Kc, wherein M is the number of the window on the line direction, N is the number of the window on the column direction;

D. utilize ${\mathrm{\σ}}_{i,j}^2={\left(A_1\right[i\left]\right[j]-{\mathrm{\μ}}_{\mathrm{kc}})}^2$ Obtain each corresponding window KC meansquaredeviation _{I, j} ², with all σ _{I, j} ²Form KC mean square deviation matrix A ₂

The KC complexity calculations adopts existing maturation method, for example Lempel and Ziv algorithm.

C in the step (3) ₀The complicacy textural characteristics comprises C ₀Value matrix A ₃And C ₀The mean square deviation matrix A ₄, the extracting method concrete steps comprise:

A. to after the one dimension time series s windowing in the step (2), carry out C ₀Complexity calculations;

B. travel through entire image, all windows are carried out above-mentioned processing, obtain C ₀Value matrix A ₃

C. utilize ${\mathrm{\μ}}_{c0}=\frac{1}{M\×N}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{A}_3\left[i\right]\left[j\right]$ Obtain C ₀Average μ _C0, wherein M is the number of the window on the line direction, N is the number of the window on the column direction;

D. utilize ${\mathrm{\σ}}_{i,j}^2={\left(A_3\right[i\left]\right[j]-{\mathrm{\μ}}_{c0})}^2$ Obtain the C of each window correspondence ₀Meansquaredeviation _{I, j} ², with all σ _{I, j} ²Form C ₀The mean square deviation matrix A ₄

C ₀Complexity calculations adopts existing maturation method, for example turn round and look at all and etc. the algorithm that proposes.

The inventive method has been described the complicacy of texture image from the angle in non-linear field, has disclosed the chaotic characteristic of texture image, has realized the Study Of Segmentation Of Textured Images that complicacy is estimated.The present invention compares with existing many image segmentation algorithms, is comparing under the suitable situation of its segmentation effect with method before, and its robustness and splitting speed improve greatly, and vast potential for future development will be arranged in actual applications.

Description of drawings

The Texture Segmentation frame diagram of Fig. 1 the inventive method;

Fig. 2 is the synoptic diagram of complexity features extraction step among Fig. 1;

The synoptic diagram of Fig. 3 Hilbert curve iteration generative process.

Embodiment

Describe the inventive method in detail below in conjunction with accompanying drawing.

The image texture dividing method that a kind of complicacy is estimated as depicted in figs. 1 and 2.Mainly comprise 4 steps: (1) carries out the coarse pre-service based on the constant automatic threshold method binaryzation of square to image; (2) Hilbert curve texture image scanning; (3) image texture characteristic of estimating based on complicacy extracts; (4) adopt support vector machine method, can obtain image after the Texture Segmentation; Below it is made introductions all round.

Step 1: adopt the constant automatic threshold method of square to be used for the binaryzation of image, its basic thought is: make before and after the Threshold Segmentation, the preceding third moment of image remains unchanged.The constant automatic threshold method of square can be regarded as a kind of image transformation, and it is transformed into ideal image with original blurred picture.

Each rank square m of two dimensional image _iBe defined as: $\left\{\begin{array}{c}{m}_0=1\\ m_i=\underset{j}{\overset{l}{\mathrm{\Σ}}}{p}_j{\left(z_j\right)}^{i}k=\mathrm{1,2},\·\·\·\end{array}\right.$

Z wherein _jBe gray-scale value, l is the total number of greyscale levels of image, p _jFor gray scale in the image is Z _jPixel ratio.For image segmentation, cut apart if carry out two-value, have only Z after then cutting apart ₀And Z ₁Two gray levels, and Z ₀＜Z ₁The pixel ratio that is lower than threshold value is used p respectively with the pixel ratio that is higher than threshold value ₀And p ₁Expression, the preceding third moment of image after then cutting apart:

$m_i^{\′}=\underset{j=0}{\overset{1}{\mathrm{\Σ}}}{P}_j{\left(z_j\right)}^i$ i＝1，2，3. (2)

For the correct threshold value of dividing target and background, i.e. optimum threshold value, the preceding third moment of the image before and after should keeping cutting apart equates.Promptly have:

m _i′＝m _i i＝1，2，3. (3)

Notice simultaneously,

p ₀+p ₁＝1 (4)

Then can obtain following system of equations:

$\left\{\begin{array}{c}{p}_0{Z}_0^0+P_1{Z}_1^0=m_0\\ p_0{Z}_0^1+p_1{Z}_1^1=m_1\\ p_0{Z}_0^2+p_1{Z}_1^2=m_2\\ p_0{Z}_0^3+p_1{Z}_1^3=m_3\end{array}\right.---\left(5\right)$

In order to find the threshold value T of hope, need from above-mentioned system of equations, to solve p earlier ₀:

$p_0=\frac{G-m_1}{{(c_1^2-{4c}_0)}^{1/2}}---\left(6\right)$

Wherein,

$c_0=\frac{m_1{m}_3-m_2^2}{m_2-m_1^2},$ $c_1=\frac{m_1{m}_2-m_3}{m_2-m_1^2},$ $G=\frac{1}{2}[{(c_1^2-{4c}_0)}^{1/2}-c_1]---\left(7\right)$

Obtain p ₀After on the original image histogram, select suitable T to make it satisfied again:

$p_0\underset{i\≤t}{\mathrm{\Σ}}{p}_i---\left(8\right)$

Then T is exactly the segmentation threshold of being asked.Do thresholding and satisfy p when can not find accurate gray-scale value ₀The time, select the most approaching gray-scale value as segmentation threshold T.

Step 2: Hilbert curve texture image scanning

Usually, because of the past nonlinear characteristic is used for analyzing one-dimensional signal more, so will convert two-dimensional image sequence to one-dimensional sequence.As adopt raster scanning, i.e. horizontal scanning or vertical scanning, defective has been only to utilize the correlativity of a direction in the two dimensional image (level or vertical) like this.Italian mathematician G.Peano had constructed a kind of space filling curve in 1890, it not self intersection pass through each point in the space.Afterwards, Germany mathematics man D.Hilbert has constructed the simple two-dimensional space filling curve of a class in 1891, be referred to as the Hilbert curve.

The fundamental element of Hilbert curve is one side untight rectangle in untight limit (being called " cup ") and the one section directed line segment (being called " connection ") that is connected two " cups ".According to different entrance and exit directions, four kinds difform " cup " arranged.

The Hilbert curve of single order is a rectangle region that covers 2 * 2 zones, and the curve of second order replaces the rectangle region of single order by three " connection " with four big rectangle regions (four " cup ") such as grade.And the like, the curve of following single order can utilize that each " cup " in the first order curve obtains in four littler " cups " and three " connection " replacements, and the generation iterative process of Hilbert curve is as shown in Figure 3.

Adopt the advantage of Hilbert curved scanning image to be: 1) less cost just can realize the extraction of the adjacent image point point of similar brightness; 2) having kept the cohesive characteristic of two dimensional image, is that the optimal spatial that can keep characteristics of image in all scanning curve modes is filled curve.

Step 3: the image texture characteristic of estimating based on complicacy extracts

Because KC complicacy, C ₀Complicacy, fluctuation complexity are a kind of descriptions to the power complicacy in a zone, so if when calculating the complicacy of a signal (time series or two dimensional image) and distribute, need be to the complicacy behind the calculating sampling behind the original signal zoning.The selection in its zone also has requirement, and promptly the zone can not be chosen too for a short time, otherwise the complicacy that comes out can not be as the representative of this kind texture; The zone can not be chosen excessive, otherwise the calculated amount of algorithm can be too big, is unfavorable for the application of algorithm.Here we have adopted the method for a kind of " big window represent fenestella ", promptly the size of the texture image selection piece of m * n are represented the central window 16 * 16 of this window or the texture complicacy of central window 8 * 8 with 32 * 32 the complexity of calculation result of window institute.Through after the edge treated, the texture image of a width of cloth m * n is sampled as

Or

Individual complicacy distribution matrix, each matrix are that 32 * 32 picture elements constitute.

After the windowing process of image was finished, we just can carry out the calculating of KC complicacy, C0 complicacy respectively.

The KC complexity is at first proposed in nineteen sixty-five by Kolmogorov: it is exactly the bit number [56,57] that produces the minimum computer program of given " 0,1 " sequence that complicacy is estimated.Lempel and Ziv have proposed to realize the algorithm of this complicacy subsequently.The KC complexity is that a kind of randomness is estimated, and it has reflected that a time series speed of new model occurs with the growth of its length, and representation sequence has reflected the architectural characteristic of symbol sebolic addressing to a certain extent near at random degree, rather than the characteristic of dynamic system.

The KC complexity calculations adopts existing maturation method Lempel and Ziv algorithm, and specific algorithm is as follows: if any a sequence S=(s ₁, s ₂, s ₃S _n), S is divided substring by certain rule define.Forming S=(s ₁, s ₂, s ₃S _n) after, add one or a string character Q (Q=s again _N+1Or Q=(s _N+1s _N+2S _N+k)), obtain SQ, make that SQv is that a string character SQ deducts a last character, see again whether Q belongs to existing in the SQv character string " words and expressions ".If occurred, this character is added in the back be referred to as " duplicating (copy) " so; If do not occur, then be referred to as " inserting (insert) ".With one " " front and back are separated when " insertion ".Next step then regards all characters of " " front as S, repeats as above step again.

For example, the complexity of sequence 0010 can be got by the following step:

First symbol is to insert forever, because S=is φ, Q=0, SQ=0, SQ π=φ, Q do not belong to " the words and expressions φ " of words and expressions SQ π substring, and then Q is an insertion, and note inserts → 0;

S=0, Q=0, SQ=00, SQ π=0, Q belongs to words and expressions SQ π, note copy → 00;

S=0, Q=01, SQ=001, SQ π=00, Q does not belong to words and expressions SQ π, and note inserts → 001;

S=001, Q=0, SQ=0010, SQ π=001, Q belongs to words and expressions SQ π, note copy → 0010;

At this moment, C (4)=3.

As sequence 00000... should be the simplest, and its form is 00000..., C (n)=2; Symbol rank 01010101... should be 01010101..., C (n)=3;

As mentioned above, obtain the character string with the " " section of being divided into, the number of the section of being divided into is exactly " complexity C (n) " that we will calculate, i.e. KC complicacy.

The KC complicacy textural characteristics of image comprises KC value matrix A ₁With KC mean square deviation matrix A ₂, the extracting method concrete steps comprise:

A. to after the one dimension time series s windowing in the step (2), carry out the KC complexity calculations;

B. travel through entire image, all windows are carried out above-mentioned processing, will obtain KC value matrix A ₁

C. utilize ${\mathrm{\μ}}_{\mathrm{kc}}=\frac{1}{M\×N}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{A}_1\left[i\right]\left[j\right]$ Obtain KC average μ _Kc, wherein M is the number of the window on the line direction, N is the number of the window on the column direction;

D. utilize ${\mathrm{\σ}}_{i,j}^2={\left(A_1\right[i\left]\right[j]-{\mathrm{\μ}}_{\mathrm{kc}})}^2$ Obtain each corresponding window KC meansquaredeviation _{I, j} ², with all σ _{I, j} ²Form KC mean square deviation matrix A ₂, the extracting method of KC complicacy textural characteristics

C ₀Complicacy thinks that the dynamic behavior that power system showed may be very complicated, but also regular following in the complexity.Briefly compound movement is mixed by regular motion and random motion.The share that random motion is shared is exactly C ₀Complexity.C ₀The architectural characteristic of the also is-symbol sequence of complicated reflection, rather than the characteristic of dynamic system.C ₀Complexity calculations adopts existing professor Gu Fanji to wait the maturation method that proposes, and concrete steps are as follows:

(1) utilize fast fourier transform to calculate power spectrum and the mean value x of original time series x (t):

x(k)＝F[x(t)] (9)

$\stackrel{\‾}{x}=\frac{1}{N}\underset{k=1}{\overset{N}{\mathrm{\Σ}}}x\left(k\right)---\left(10\right)$

In the formula, k is the frequency domain variable, and N is the length of x (k).

(2) the amplitude wave spectrum composition bigger than mean value kept, remaining all is changed to 0, forms new wave spectrum x ' (k):

$x^{\′}\left(k\right)=\left\{\begin{array}{ccc}x\left(k\right)& \mathrm{if}& \mathrm{xx}\left(k\right)>\stackrel{\‾}{x}\\ 0& \mathrm{if}& x\left(k\right)\≤\stackrel{\‾}{x}\end{array}\right.---\left(11\right)$

(3) this new wave spectrum is carried out inverse fourier transform, obtain a new time series, the regular motion composition x of this time series as original time series ₁(t).And original time series is random motion composition x (t)-x with the difference that rule becomes to divide ₁(t).

x ₁(t)＝F ^-1[x′(k)] (12)

(4) area A of random motion composition ₁With the original time series area A ₀Ratio be designated as C ₀Complexity.

$A_0{\∫}_0^{\∞}\left|x\left(t\right)\right|\mathrm{dt}---\left(13\right)$

$A_1={\∫}_0^{\∞}\left|x\left(t\right)x_1\left(t\right)\right|\mathrm{dt}---\left(14\right)$

$C_0=\underset{t\→\∞}{\lim }\frac{A_1}{A_0}---\left(15\right)$

Obviously, work as x ₁When (t) portion is very big in x (t), C ₀→ 0, the dynamic behavior of illustrative system almost is regular, does not contain random element.Otherwise, work as x ₁When (t) the very little and share that random motion part-time sequence is shared of portion is very big, C ₀→ 1, the dynamics of illustrative system almost is completely random.So, along with C ₀Increase, mean that the random element in the dynamics increases.

The C of image ₀The complicacy textural characteristics comprises C ₀Value matrix A ₃And C ₀The mean square deviation matrix A ₄, the extracting method concrete steps comprise:

A. to after the one dimension time series s windowing in the step (2), carry out C ₀Complexity calculations;

B. travel through entire image, all windows are carried out above-mentioned processing, obtain C ₀Value matrix A ₃

C. utilize ${\mathrm{\μ}}_{c0}=\frac{1}{M\×N}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{A}_3\left[i\right]\left[j\right]$ Obtain C ₀Average μ _C0, wherein M is the number of the window on the line direction, N is the number of the window on the column direction;

D. utilize ${\mathrm{\σ}}_{i,j}^2={\left(A_3\right[i\left]\right[j]-{\mathrm{\μ}}_{c0})}^2$ Obtain the C of each window correspondence ₀Meansquaredeviation _{I, j} ², with all σ _{I, j} ²Form C ₀The mean square deviation matrix A ₄

We have obtained the eigenvalue matrix A of 4 M * N sizes ₁, A ₂, A ₃, A ₄, 4 eigenwerts on the matrix same position are formed one 4 proper vector of tieing up respectively, as the texture feature vector of central window corresponding on the former figure.

Step 4: the Texture Segmentation of carrying out that adopts support vector machine

Support vector machine is similar to artificial neural network, needs a training process, and the texture in Brotadz texture storehouse of selecting for use equally and the Uni-Bonn texture storehouse experimentizes.Earlier every class texture image is carried out the feature extraction that complicacy is estimated, as training set, and texture image feature to be split is as test set with the feature of each single texture image.In training set, choose the point of (not repeating) some in each characteristic image and train the acquisition support vector as test sample book, then test set is imported support vector machine fully, can get split image to the end.

Core concept of the present invention is the complicacy that the angle in non-linear field has been described texture image, has disclosed the chaotic characteristic of texture image, realizes the Study Of Segmentation Of Textured Images technology estimated based on complicacy, has expanded the new method of Texture Segmentation.The operand that this method needs is few, and segmentation precision and accuracy meet the demands, and important practical reference value is arranged.

Practice shows that the inventive method segmentation effect is good, and accurately. We have contrasted existing typical case Texture description method and their segmentation result. Texture Segmentation resolution chart to one 512 * 512 Picture uses respectively co-occurrence matrix, Gabor filtering, wavelet transformation, fractal dimension and complexity to survey Degree is cut apart, and adds up the required time. Can be found out by segmentation effect and required time: base Owing to will add up a lot of characteristic parameters, consuming time more, segmentation effect is not in the texture model of co-occurrence matrix Good; Based on the texture model of Gabor filtering owing to need to obtain small on frequency and direction of texture Change information causes the number of required wave filter very big, and elapsed time is more; Based on wavelet transformation Texture Segmentation Methods owing to used the fast algorithm of dyadic wavelet, segmentation effect is better, the time consumption Take also relatively less; Can well the simulating nature texture based on the texture model of fractal dimension, process The Texture Segmentation of refinement is effective, needs bigger operand but calculate fractal dimension; Relatively and Speech, the operand that the texture segmentation algorithm based on Complexity Measurement that the present invention proposes needs is few, is dividing Cut precision and accuracy and meet the demands under the prerequisite, computational speed improves greatly.

Claims (5)

1, the image texture dividing method estimated of a kind of complicacy is characterized in that these method concrete steps comprise:

(1) adopt the constant automatic threshold method of square that image is carried out binaryzation;

(2) utilize the Hilbert curve that texture image is scanned, make the image information of two dimension convert one-dimensional sequence s to;

(3) one-dimensional sequence in the step (2) is carried out windowing process after, extract KC complicacy textural characteristics and C ₀The complicacy textural characteristics obtains describing one 4 dimensional feature vector of texture complexity features;

(4) adopt support vector machine method, obtain the image after the Texture Segmentation.

2, the image texture dividing method estimated of a kind of complicacy as claimed in claim 1 is characterized in that the constant automatic threshold method of square in the step (1) adopts the principle that the preceding third moment of image equates before and after the binaryzation to select threshold value.

3, the image texture dividing method estimated of a kind of complicacy as claimed in claim 1 is characterized in that in the step (2) that the fundamental element of Hilbert curve comprises " cup " and " connection " in the Hilbert curved scanning method; Described " cup " is the untight rectangle in one side, and described " connection " is the one section directed line segment that connects two " cups ";

The Hilbert curve is to be generated by the fundamental element iteration, detailed process is: the Hilbert curve of single order is a rectangle region that covers 2 * 2 zones, the curve of second order replaces the rectangle region of single order by three " connection " with four big " cups " such as grade, and the like, the curve of following single order utilizes that each " cup " in the first order curve obtains in four littler " cups " and three " connection " replacements.

4, the image texture dividing method estimated of a kind of complicacy as claimed in claim 1 is characterized in that the KC complicacy textural characteristics in the step (3) comprises KC value matrix A ₁With KC mean square deviation matrix A ₂, the extracting method concrete steps comprise:

A. to after the one dimension time series s windowing in the step (2), carry out the KC complexity calculations;

B. travel through entire image, all windows are carried out above-mentioned processing, obtain KC value matrix A ₁

C. utilize ${\mathrm{\μ}}_{\mathrm{kc}}=\frac{1}{M\×N}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{A}_1\left[i\right]\left[j\right]$ Obtain KC average μ _Kc, wherein M is the number of the window on the line direction, N is the number of the window on the column direction;

D. utilize ${\mathrm{\σ}}_{i,j}^2={\left(A_1\right[i\left]\right[j]-{\mathrm{\μ}}_{\mathrm{kc}})}^2$ Obtain each corresponding window KC meansquaredeviation _{I, j} ², with all σ _{I, j} ²Form KC mean square deviation matrix A ₂

5, the image texture dividing method estimated of a kind of complicacy as claimed in claim 1 is characterized in that the C in the step (3) ₀The complicacy textural characteristics comprises C ₀Value matrix A ₃And C ₀The mean square deviation matrix A ₄, the extracting method concrete steps comprise:

A. to after the one dimension time series s windowing in the step (2), carry out C ₀Complexity calculations;

B. travel through entire image, all windows are carried out above-mentioned processing, obtain C ₀Value matrix A ₃

C. utilize ${\mathrm{\μ}}_{c0}=\frac{1}{M\×N}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{A}_3\left[i\right]\left[j\right]$ Obtain C ₀Average μ _C0, wherein M is the number of the window on the line direction, N is the number of the window on the column direction;

D. utilize ${\mathrm{\σ}}_{i,j}^2={\left(A_3\right[i\left]\right[j]-{\mathrm{\μ}}_{c0})}^2$ Obtain the C of each window correspondence ₀Meansquaredeviation _{I, j} ², with all σ _{I, j} ²Form C ₀The mean square deviation matrix A ₄

Images