CN105741264A - Two-phase image segmentation method based on semi-local texture features - Google Patents

Abstract

The invention discloses a two-phase image segmentation method based on semi-local texture features, mainly comprising the following steps: segmenting an image into non-overlapping small blocks with M*M pixels; extracting the texture features based on a Beltrami framework and semi-local information of each block, clustering the blocks through use of a K-means algorithm, and dividing the image into four areas; determining the optimal position of a target based on the principle of photographic composition, extracting the target, and completing coarse segmentation; and finely segmenting the extracted target through use of a geometric active contour model to get a more accurate segmentation result. The beneficial effects are as follows: a segmentation method in which coarse segmentation is carried out first to extract a target object and then fine segmentation is carried out is put forward, and the coarse-to-fine segmentation strategy can be used to segment an image with a fuzzy or complex background; novel texture features based on the Beltrami framework and semi-local information are defined, the features have stronger discriminating power than a single feature and have very strong anti-noise ability, and a more accurate experimental result can be obtained when the texture features are used in clustering and image segmenting.

Description

A kind of two-stage image partition method based on half Local textural feature

Technical field

The present invention relates to image partition method, image procossing, pattern recognition, artificial intelligence field, be specifically related to a kind of two-stage image partition method based on half Local textural feature.

Background technology

Texture Segmentation is a problem the most challenging in image segmentation.Human eye can be easy to tell different textures, but is difficult to be defined texture from the aspect of mathematical term, also therefore is difficult to describe it with a model.Additionally, texture may cause the loss of important edges and the uneven of intensity profile.The definition being widely recognized as at present is: texture is a series of fine characteristic details with periodicity and oscillatory.

Normally used characteristics of image includes: feature, feature based on half local message and feature based on Beltrami framework of based on filtering.Wherein feature such as gradient filtering based on filtering, the filtering of small echo porous have been applied in feature extraction and image segmentation, and the feature extracted by Gabor or Morlet wavelet transformation is to discriminate between the important evidence of the texture of different directions and yardstick；The main method obtaining feature based on half local message is: extracted by the half-tone information of the block adjacent with existing pixel, thus obtain half local message of each pixel, the idea of this block-based characteristic vector proposes first when introducing textures synthesis, later, Buades et al. based on block diversity and non local average proposition to the idea of image noise reduction, Gilboa and Osher framework based on change proposes non local noise reduction model, finally, the diversity that Bresson and Chan is equally based between block proposes a kind of variable unsupervised dividing method；Feature based on Beltrami framework is a kind of new geometric representation of image, characteristics of image is regarded as the Riemann manifold being embedded in higher dimensional space by it, the advantage of this method is that it allows to use different geometry instruments to carry out different image procossing (such as noise reduction and segmentation), and can process the image of any N-dimensional, shortcoming is the most sensitive to noise.Therefore, latter two feature is combined by the present invention, the feature based on half local message extracted is introduced under Beltrami framework by mapping, thus obtains a strongest new textural characteristics of noise immunity.

Being widely known by the people most in image is split and most successful model be exactly the movable contour model first proposed by Kass, Witkin et al., it is successfully applied in medical imaging extraction anatomical structure.But the shortcoming of this model is that the initial position to active contour is very sensitive and bad in the depression boundary convergence of target.Movable contour model is improved by Caselles subsequently, kimmel et al., proposes geometric active contour model (GAC), and this model can obtain preferable segmentation result in image is split, and therefore the segmentation jog section of the present invention just uses this kind of model.

The present invention proposes a kind of dividing method being different from conventional segmentation thinking, is broadly divided into coarse segmentation and two stages are cut in segmentation.First, divide the image into into the non-overlapped block of M × M pixel, extract each piece based on Beltrami framework and the textural characteristics of half local message, then carry out block cluster with K-means algorithm, picture be divided into four regions, according to photography structureFigurePrinciple determines the optimum position of target, thus extracts target, completes coarse segmentation；Finally with geometric active contour model (GAC) target extracted is finely divided and cuts, thus obtain more accurate segmentation result.

Summary of the invention

Present invention is primarily targeted at, propose a kind of first coarse segmentation to go out target object and be finely divided the thought cut again, and combine a kind of based on Beltrami framework with the new textural characteristics of half local message, apply it to cluster and during image splits, obtain segmentation result more accurately.

The object of the invention to solve the technical problems is to the technical scheme is that a kind of two-stage image partition method based on half Local textural feature, including herein below:

If P_{x , y}Be with pixel (x, y) centered by, size is the block of τ × τ, then have

The following X that maps is used to be incorporated into by textural characteristics under Beltrami framework:

X:(x, y) → (X₁=x, X₂=y, X₃=P_{x , y}(I)) (2)

This mapping contains local message (locus) and half topography's information (value of the block of pixels around center pixel).Assume the textured pattern that a given width is complicated, by map the geometry manifold being mounted to higher dimensional space of (2) with we observed to texture be consistent (this hypothesis is all establishment to most natural image).This means that the metric tensor of manifold of identical texture region is identical, metric tensor is used to a variable of the distance in measurement manifold between 2, when the manifold of certain specific region is almost in a plane, in this region, the distance between any two points is all equal, in conjunction with the manifold obtained by mapping (2) knowable to half topography's information almost in plane, therefore there is identical texture in this region.Map corresponding metric tensor in (2) to be defined as:

Finally, textural characteristics describes sub-F and is defined as

$F=\exp (-\frac{\det \left(g_{xy}\right)}{{\mathrm{\σ}}^2})---\left(4\right)$

Wherein σ ＞ 0 is scale parameter, uses gaussian kernel function as low-pass filtering, controls to describe the degree of image detail.

It addition, for coloured image, derivation is the simplest.Making coloured image is I=(I₁, I₂..., I_k), wherein k is the dimension of image, and corresponding half local is mapped as:

X:(x, y) → (X₁=x, X₂=y, X₃=P_{x , y}(I₁) ..., X_2+k=P_{x , y}(I_k)) (5)

Corresponding metric tensor is write as following form:

It addition, geometric active contour model can be converted into following minimization problem:

$\begin{array}{cc}\underset{C}{min}& \{E_{G4C}\left(C\right)={\∫}_0^{L\left(C\right)}g\left(\right|\▿I_0\left(C\left(s\right)\right)\left|\right)ds\}\end{array}---\left(7\right)$

Wherein ds is the length of Euclid's element,Length for curve C.Accordingly, it is capable to functional (7) is, by ds, the function g comprising object boundary information is integrated a new length obtaining in fact, function g is that edge indicator function is for eliminating such asSuch object edge, I₀Being original image, β is any normal number.Function E is can get by the calculus of variations_GACEuler-Lagrange equation formula, gradient descent method can minimize E as quickly as possible_GAC:

Wherein,T is the time parameter of artificial regulation, and k, N are respectively curvature and the normal of curve C, the EVOLUTION EQUATION existence and unique solution of the active contour defined in formula (8).The Level Set Method that Osher and Sethian proposes efficiently solves profile extension problem and has processed change in topology problem, and equation (8) can be written as level set form:

Wherein, the active contour C that φ is embedded in constantly developing is (such as C (t)={ x ∈ R^N| the level set function of φ (x, t)=0}), based on hyperbolic conservation law, partial differential equation (9) may be used on multiple quantizing and solve, and obtain fairly precise segmentation result.

Compared with prior art, advantages of the present invention and effect are: 1) propose and a kind of first carry out being finely divided after coarse segmentation extracts target object the dividing method cut, this be can be used to split have by the thick segmentation strategy to essence obscure or the picture of complex background.2) define a kind of based on Beltrami framework and half local messageNovelTextural characteristics, has the resolving ability more higher than single feature, and has the strongest anti-noise ability, is used for by this textural characteristics in cluster and image segmentation, it is possible to obtain more accurate experimental result.

Accompanying drawing explanation

Figure 1A kind of two-stage image partition method flow process based on half Local textural featureFigure

Detailed description of the invention

Such as figure 1Shown in, overall procedure of the present invention is as follows: first, divides the image into into the non-overlapped fritter of M × M pixel；Secondly, extract each piece based on Beltrami framework and the textural characteristics of half local message, then carry out block cluster with K-means algorithm, picture be divided into four regions；Then according to photography structureFigurePrinciple determines the optimum position of target, thus extracts target, completes coarse segmentation；Finally with geometric active contour model (GAC) target extracted is finely divided and cuts, thus obtain more accurate segmentation result.

The present invention specifically comprises the following steps that

Step S101: divide the image into into the non-overlapped fritter of M × M pixel

Considering the effect of textural characteristics and the time complexity of algorithm, all pictures are standardized as 126 × 189 or 189 × 126, the tile size split is 3 × 3, and therefore every standardized image includes 2646 fritters.

Step S102: extract each piece based on Beltrami framework and the textural characteristics of half local message, then carry out block cluster with K-means algorithm, picture be divided into four regions

1) choose with pixel (x, y) centered by, size is the block P of τ × τ_{x , y}:

2) the following X that maps is used to be incorporated into by textural characteristics under Beltrami framework:

X:(x, y) → (X₁=x, X₂=y, X₃=P_{x , y}(I)) (11)

This mapping contains local message (locus) and half topography's information (value of the block of pixels around center pixel).Assuming the textured pattern that a given width is complicated, be consistent by mapping the geometry manifold being mounted to higher dimensional space of (11) observed with us to texture, in mapping (11), metric tensor is defined as accordingly:

3) extract each piece based on Beltrami framework and the textural characteristics of half local message

Finally, obtain textural characteristics and describe sub-F be

$F=\exp (-\frac{\det \left(g_{xy}\right)}{{\mathrm{\σ}}^2})---\left(13\right)$

Wherein σ ＞ 0 is scale parameter, uses gaussian kernel function as low-pass filtering, controls to describe the degree of image detail.

4) cluster by K-means method, image is polymerized to four classes

Step S103: according to photography structureFigurePrinciple determines the optimum position of target, thus extracts target, completes coarse segmentation

1) result obtained in the previous step is merged

One secondary given image finally has only to be divided into foreground area and background area, merges the region meeting certain similarity in image obtained in the previous step, and similarity measurement is defined as:

$sim(R_i,R_j)=-d(R_i,R_j);d(R_i,R_j)=\left|\right|{\stackrel{\→}{R}}_i-{\stackrel{\→}{R}}_j\left|\right|$

I, j=0,1 ..., N_r, i ≠ j (14)

Wherein, R_iFor texture feature vector, d (R_i, R_j) it is vector R_iAnd R_jBetween distance, similarity measurement is inversely proportional to distance, therefore merges the region with maximum comparability (minimum range), and the characteristic vector of new combined region to recalculate, until image in leave behind two regions.

2) according to photography structureFigurePrinciple determines the optimum position of target

Photography structureFigureDetermine that typically there are two kinds of methods optimum position: three points of structuresFigureMethod and dynamic symmetry method.Three points of structuresFigureMethod refers to laterally and vertically be respectively divided into image trisection, and in image, the position of four intersection points is the optimum position of target, namely foreground area.Dynamic symmetry method refers to make a diagonal of image, then makees vertical line to this diagonal respectively from two other angle, and in image, the position of two intersection points is the optimum position of target, and other regions are then background area.

3) target object is extracted

First the bianry image in foreground and background region is respectively obtained, then by three points of structuresFigureThe mask table of method or dynamic symmetry method does with two width bianry images and operates respectively, and the bianry image maximum to pixel count carries out Objective extraction, and another width is defaulted as background area.

Step S104: with geometric active contour model the target extracted is finely divided and cuts, thus obtain more accurate segmentation result

1) geometric active contour model (GAC)

Geometric active contour model (GAC) can be converted into following minimization problem:

$\begin{array}{cc}\underset{C}{min}& \{E_{G4C}\left(C\right)={\∫}_0^{L\left(C\right)}g\left(\right|\▿I_0\left(C\left(s\right)\right)\left|\right)ds\}\end{array}---\left(15\right)$

Wherein ds is the length of Euclid's element,Length for curve C.Accordingly, it is capable to functional (15) is, by ds, the function g comprising object boundary information is integrated a new length obtaining in fact, function g is that edge indicator function is for eliminating such asSuch object edge, I₀Being original image, β is any normal number.Function E is can get by the calculus of variations_GACEuler-Lagrange equation formula, gradient descent method can minimize E as quickly as possible_GAC:

Wherein,T is the time parameter of artificial regulation, and k, N are respectively curvature and the normal of curve C, the EVOLUTION EQUATION existence and unique solution of the active contour defined in formula (16).

2) level set function is rewritten

The Level Set Method that Osher and Sethian proposes efficiently solves profile extension problem and has processed change in topology problem, and equation (16) can be written as level set form:

Wherein, the active contour C that φ is embedded in constantly developing is (such as C (t)={ x ∈ R^N| the level set function of φ (x, t)=0}), based on hyperbolic conservation law, partial differential equation (17) may be used on multiple quantizing and solve, and obtain fairly precise segmentation result.

3) target object extracting coarse segmentation is finely divided and cuts, and obtains segmentation result.

Claims (5)

1. a two-stage image partition method based on half Local textural feature, it is characterised in that comprise the following steps:

Step 1, divide the image into into the non-overlapped fritter of M × M pixel；

Step 2, extract each piece based on Beltrami framework and the textural characteristics of half local message, then carry out block cluster with K-means algorithm, picture be divided into four regions；

Step 3, according to photography structureFigurePrinciple determines the optimum position of target, thus extracts target, completes coarse segmentation；

Step 4, with geometric active contour model the target extracted is finely divided and cuts, thus obtain more accurate segmentation result.

2.According to claimTwo-stage image partition method based on half Local textural feature described in 1, in described step 1, divides the image into into the non-overlapped fritter of M × M pixel, it is characterised in that:

Considering the effect of textural characteristics and the time complexity of algorithm, all pictures are standardized as 126 × 189 or 189 × 126, the tile size split is 3 × 3, and therefore every standardized image includes 2646 fritters.

3.According to claimTwo-stage image partition method based on half Local textural feature described in 1, in described step 2, extract each piece based on Beltrami framework and the textural characteristics of half local message, then carry out block cluster with K-means algorithm, picture is divided into four regions, it is characterised in that:

1) choose with pixel (x, y) centered by, size is the block P of τ × τ_{x , y}:

2) the following X that maps is used to be incorporated into by textural characteristics under Beltrami framework:

X:(x, y) → (X₁=x, X₂=y, X₃=P_{x , y}(I)) (2)

This mapping contains local message (locus) and half topography's information (value of the block of pixels around center pixel).Assuming the textured pattern that a given width is complicated, be consistent by mapping the geometry manifold being mounted to higher dimensional space of (2) observed with us to texture, in mapping (2), metric tensor is defined as accordingly:

3) extract each piece based on Beltrami framework and the textural characteristics of half local message

Finally, obtain textural characteristics and describe sub-F be

Wherein σ ＞ 0 is scale parameter, uses gaussian kernel function as low-pass filtering, controls to describe the degree of image detail.

4) cluster by K-means method, image is polymerized to four classes.

4.According to claimTwo-stage image partition method based on half Local textural feature described in 1, in described step 3, according to photography structureFigurePrinciple determines the optimum position of target, thus extracts target, completes coarse segmentation, it is characterised in that:

1) result obtained in the previous step is merged

One secondary given image finally has only to be divided into foreground area and background area, merges the region meeting certain similarity in image obtained in the previous step, and similarity measurement is defined as:

Wherein, R_iFor texture feature vector, d (R_i, R_j) it is vector R_iAnd R_jBetween distance, similarity measurement is inversely proportional to distance, therefore merges the region with maximum comparability (minimum range), and the characteristic vector of new combined region to recalculate, until image in leave behind two regions.

2) according to photography structureFigurePrinciple determines the optimum position of target

Photography structureFigureDetermine that typically there are two kinds of methods optimum position: three points of structuresFigureMethod and dynamic symmetry method.Three points of structuresFigureMethod refers to laterally and vertically be respectively divided into image trisection, and in image, the position of four intersection points is the optimum position of target, namely foreground area.Dynamic symmetry method refers to make a diagonal of image, then makees vertical line to this diagonal respectively from two other angle, and in image, the position of two intersection points is the optimum position of target, and other regions are then background area.

3) target object is extracted

First the bianry image in foreground and background region is respectively obtained, then by three points of structuresFigureThe mask table of method or dynamic symmetry method does with two width bianry images and operates respectively, and the bianry image maximum to pixel count carries out Objective extraction, and another width is defaulted as background area.

5.According to claimTwo-stage image partition method based on half Local textural feature described in 1, in described step 4, is finely divided the target extracted with geometric active contour model and cuts, thus obtains more accurate segmentation result, it is characterised in that:

1) geometric active contour model (GAC)

Geometric active contour model (GAC) can be converted into following minimization problem:

Wherein ds is the length of Euclid's element,Length for curve C.Accordingly, it is capable to functional (6) is, by ds, the function g comprising object boundary information is integrated a new length obtaining in fact, function g is that edge indicator function is for eliminating such asSuch object edge, I₀Being original image, β is any normal number.Function is can get by the calculus of variationsEuler-Lagrange equation formula, gradient descent method can minimize as quickly as possible

Wherein,T is the time parameter of artificial regulation, and k, N are respectively curvature and the normal of curve C, the EVOLUTION EQUATION existence and unique solution of the active contour defined in formula (7).

2) level set function is rewritten

The Level Set Method that Osher and Sethian proposes efficiently solves profile extension problem and has processed change in topology problem, and equation (7) can be written as level set form:

Wherein, the active contour C that φ is embedded in constantly developing is (such as C (t)={ x ∈ R^N| the level set function of φ (x, t)=0}), based on hyperbolic conservation law, partial differential equation (8) may be used on multiple quantizing and solve, and obtain fairly precise segmentation result.

3) target object extracting coarse segmentation is finely divided and cuts, and obtains segmentation result.