CN102855624B - A kind of image partition method based on broad sense data fields and Ncut algorithm - Google Patents

Abstract

The invention provides a kind of image partition method based on broad sense data fields and Ncut algorithm, first, feature space is divided into level grid, ground floor grid and second layer grid, form grid search-engine space Ω corresponding with it simultaneously _sand Ω _b, every 8 the adjacent small grid of ground floor form a macrolattice of the second layer; Then obtained the gesture Distribution value of ground floor grid based on second layer grid computing by utilization GDF algorithm.Based on gesture Distribution value, then ground floor grid is carried out cluster, cluster result is mapped on image, thus the initial segmentation realized for piece image operates, and is divided into the different regions mutually disjointed; Finally, based on image initial segmentation result build undirected weighted graph after, use the Ncut algorithm based on region to come the identical region of amalgamation property, until reach optimized image segmentation result.The present invention has fast on Iamge Segmentation, simply, advantage accurately.

Description

A kind of image partition method based on broad sense data fields and Ncut algorithm

Technical field

The invention belongs to technical field of image processing, particularly a kind of image partition method based on broad sense data fields and Ncut algorithm.

Background technology

Iamge Segmentation is exactly piece image is divided into region that is meaningful and that do not overlap, and every block region almost has identical character, and this is link important in image procossing research, is also research topic important in computer vision simultaneously; Target detection, feature extraction, target identification all rely on image Segmentation Technology accurately, and because image Segmentation Technology is as an element task in image procossing, therefore obtain and apply comparatively widely, various partitioning algorithm is suggested in succession.

In numerous partitioning algorithm, non parameter modeling is wherein the simplest and a kind of image segmentation algorithm of the widest model of application, and non parameter modeling method roughly can be divided into two classes: hierarchical clustering and density Estimation, distance between hierarchical clustering technical basis data point is classified, so often cause higher computational complexity, and can not directly define a significant stopping criterion for data clusters, this means that different data sets needs to arrange different stopping criterions, the ultimate principle of the non parameter modeling that density based is estimated is in feature space, describe the feature distribution of data set with empirical probability density function, the local maximum (i.e. summit) of the corresponding density function of the close quarters in feature space, once determine the position on summit, just can according to the partial structurtes determination cluster result of feature space, such as, mean shift (MS) is a kind of nonparametric image clustering algorithm, but the independent selection of MS algorithm to window width is very sensitive, namely for different optimum configurations, the segmentation result of this algorithm is variant very large, and be a kind of partitioning algorithm very consuming time, therefore, in practice, too much cut zone may be there is in the segmentation result of this algorithm, the segmentation of mistake and the time of costing a lot of money in segmentation process.

In the algorithm that some propose, in order to improve the segmentation result of MS, be integrated with the dividing method based on figure, method based on figure is also a very important class in Iamge Segmentation, such as there is normalized cuts (Ncut), average association, minimum cut etc., in these methods, each pixel is regarded as a summit, connected by a limit between adjacent point, and the dissimilarity measurement of two points is as the weight on limit, thus a structure undirected weighted graph, with other based on figure partitioning algorithm compared with, the application of Ncut algorithm is comparatively extensive, in order to overcome the shortcoming of MS image segmentation algorithm, MS algorithm and recurrence Ncut algorithm are be combined with each other, be called MS-Ncut, first MS-Ncut algorithm obtains by MS algorithm the initial segmentation image comprising a lot of fragment, then a undirected weighted graph is set up according to the block of these over-segmentations, adopt Ncut algorithm correction initial segmentation result, in Ncut calculating process, although each node generates auxiliary child node and optimizes segmentation result further to a certain extent, but produced problem in MS-Ncut is not fundamentally solved, and considerably increase the time complexity of algorithm.

Summary of the invention

In order to obtain better segmentation result, avoid occurring above-mentioned problem, the present invention proposes the image partition method of a kind of new image segmentation algorithm-based on broad sense data fields and Ncut algorithm, the method by integrated broad sense data fields GDF and Ncut two kinds of algorithms, simply, rapidly, exactly by piece image can be divided into significant region in logic.

For solving the problems of the technologies described above, the present invention adopts following technical scheme:

Based on an image partition method for broad sense data fields and Ncut algorithm, comprise the following steps,

The division of step 1, level grid and gesture value are estimated, specifically comprise the following steps,

Step 1.1, the RGB color feature space of image is converted to L*u*v* or L*a*b* color feature space, L*u*v* or L*a*b* color feature space Ω is divided into 2N × 2N × 2N small grid as ground floor grid, calculate the average of data point in each small grid, and in this, as the eigenwert of this small grid, form a new feature space Ω _s;

Eight neighborhood small grid merges by step 1.2 becomes a macrolattice as second layer grid, forms a new feature space Ω _band corresponding mesh space coordinate figure;

Step 1.3 calculates the gesture value of each small grid in ground floor grid according to gesture value estimation formulas

At feature space Ω _bin, denotation coordination is (i _th, j _th, k _th) grid, (i _th, j _th, k _th) quality of grid, $f_{i_u{j}_u{k}_u}=(f_{i_u{j}_u{k}_u}^{\left(1\right)},f_{i_u{j}_u{k}_u}^{\left(2\right)},f_{i_u{j}_u{k}_u}^{\left(3\right)}),$ $w_{i_u{j}_u{k}_u}=Q_{i_u{j}_u{k}_u}\×e^{-{\left(\frac{x-x_{i_u{j}_u{k}_u}}{{\mathrm{\σ}}_X}\right)}^2+{\left(\frac{y-y_{i_u{j}_u{k}_u}}{{\mathrm{\σ}}_Y}\right)}^2},$ be positioned at grid (i _th, j _th, k _th) quantity of interior data point, correspond to grid (i _th, j _th, k _th) spatial value, σ _xand σ _ythe spacial influence factor, σ=ch=c (h ₁, h ₂, h ₃) ^t, σ _j=ch _j, j=1,2,3, c is proportionality constant, h=(h ₁, h ₂, h ₃) ^tbe the window width of Density Estimator, K (x) is unit potential function;

Step 2, gesture Distribution value according to small grid, to small grid cluster, cluster result is mapped to image, image is divided into the different regions mutually disjointed, and concrete steps are:

Step 2.1. asks local derviation to the formula in step 1.3, obtains formula:

utilize it to calculate the local derviation of each small grid in ground floor grid, and determine all mesh of vertices with this, describe cluster { C by six neighbo r pattern combination mesh of vertices _k} _{k=1 ..., v}, wherein C _kat least comprise a mesh of vertices;

Step 2.2 for each k=1,2 ..., v, cluster C _kin mesh of vertices as initial data point, namely along the direction search grid that Grad rises, until Grad no longer rises till, the small grid that Jiang Yanlu searches is divided into cluster C _kin;

After step 2.3 is searched for, for each cluster C _k, k=1,2 ..., v, by Mapping of data points all in each cluster on image, and be incorporated in the image fragment that spatially data point number is less than the individual point of M (20≤M≤100), piece image is divided into the prime area Ω that R block does not overlap _i, i=1,2 ..., R;

Step 3, the Ncut algorithm based on region is used to merge the region of over-segmentation; The formula of wherein used calculating weight matrix W is:

$w(i,j)=\left\{\begin{array}{c}\frac{1}{n_i}\underset{f\∈{\mathrm{\Ω}}_1}{\mathrm{\Σ}}{e}^{-{\left[\frac{\left|\right|f-F_1\left|\right|}{{\mathrm{\σ}}_s}\right]}^2}+\frac{1}{n_j}\underset{f\∈{\mathrm{\Ω}}_j}{\mathrm{\Σ}}{e}^{-{\left[\frac{\left|\right|f-F_1\left|\right|}{{\mathrm{\σ}}_s}\right]}^2}\\ 0\end{array}\right.$

2, the image partition method based on broad sense data fields and Ncut algorithm according to claim 1, is characterized in that: described step 3 specifically comprises the following steps,

Step 3.1 based on obtain in step 2.3 not heavy and prime area build a undirected weighted graph G=(V, E, W), V is the summit of image, and E is the set on the limit of connect Vertex, and W is weight matrix, according to formula $w(i,j)=\left\{\begin{array}{c}\frac{1}{n_i}\underset{f\∈{\mathrm{\Ω}}_1}{\mathrm{\Σ}}{e}^{-{\left[\frac{\left|\right|f-F_1\left|\right|}{{\mathrm{\σ}}_s}\right]}^2}+\frac{1}{n_j}\underset{f\∈{\mathrm{\Ω}}_j}{\mathrm{\Σ}}{e}^{-{\left[\frac{\left|\right|f-F_1\left|\right|}{{\mathrm{\σ}}_s}\right]}^2}\\ 0\end{array}\right.$

Calculate weight matrix W;

Step 3.2 calculates diagonal matrix D by weight matrix W,

Wherein D (i, i)=∑ _jw (i, j);

Step 3.3 is solved an equation (D-W) y=λ Dy, obtains eigenwert and corresponding proper vector, determines the second little proper vector;

Step 3.4 is according to formula finding out cut-point, some when namely Ncut value is minimum, with the summit of the second little proper vector bipartite graph, is two subgraphs by Iamge Segmentation;

Step 3.5 splits for two the subgraph obtained, and calculates weight matrix respectively, and repeats step 3.2 to 3.4;

Step 3.6 repeats step 3.5, until Ncut value exceeds given threshold values.

Respectively the present invention's theory used or principle are introduced below:

Be subject to the inspiration of physical field, the interaction between the particulate of material and description are introduced in abstract art of mathematics, so define data fields; Its data capacity is radiated whole parent space from sample space by data radiation by data, accept data capacity and the space that covers by data radiation, be called data fields; Data fields can be considered a space being full of data capacity, data by the data fields of oneself, to another data emitted energy in field; The mutual emittance of meeting between data point in data fields, these energy superpose the gesture forming data fields mutually; Known spatial in comprise the data set D={x of n object ₁, x ₂..., x _n.Think that each object has particle or the nucleon of certain mass, there is an applied field around it, any object being positioned at field all will be subject to the synergy of other objects, spatially determine a data fields thus whole.In data fields, the gesture value of any point x ∈ Ω can be expressed as

Wherein K (x) is unit potential function.σ is used for the interaction range between control object, is called factor of influence.Use for reference the condition that Density Estimator center meets, K (x) should meet: $\&Integral;K\left(x\right)\mathrm{dx}=1,\&Integral;\mathrm{xK}\left(x\right)\mathrm{dx}=\mathrm{0,0}<R\left(K\right)=\&Integral;K{\left(x\right)}^2\mathrm{dx}<\&infin;.$ Quality m _i, m _i>=0 is object X _iquality, suppose to meet normalizing condition and certain convergence, namely have ${\mathrm{\&Sigma;}}_{i=1}^n{m}_i=1,m_i\&GreaterEqual;0\mathrm{and}{\lim }_{n\&RightArrow;\&infin;}n{\sup }_{1\&le;i\&le;n}\left\{m_i\right\}=1.$

In multidimensional data field, the potential function in whole data fields estimates that in (1), factor of influence σ value in different dimensional is identical, this means that the energy distribution of each observation data point is evenly scattered to all directions.But under normal circumstances, the data of different dimensional have different attributes.σ should be anisotropic, and namely the value of σ is different in different directions.In addition, when data have different variability in different directions, or when data are almost positioned on a low dimensional manifold, think that the estimation that all directions have same yardstick to obtain is often not ideal.Therefore, in multidimensional data field, estimate to obtain better data fields potential function, we replace factor of influence σ by matrix H, give broad sense data field potential Function Estimation.Its formula is:

Wherein, H is the p × p positive definite constant matrix relevant with factor of influence, and P represents the dimension of hyperspace, P=1,2,3; Conveniently, get H=σ A, wherein σ > 0, | A|=1.Potential function K is real-valued multidimensional data field potential function.For the ease of calculating, getting H is positive definite triangular matrix, and the one based on (3) formula simplifies potential function and estimates

Wherein σ _jfor the factor of influence of jth dimension.Such as, if data object is two-dimentional, then, j=1,2.

The present invention proposes a kind of novel clustering algorithm, this algorithm is estimated based on graded mesh structure gesture value, and this algorithm effectively can improve arithmetic speed, is referred to as down-hill method; Contrast climbing method cluster process, the first step is the summit (maximum point) finding that gesture value is estimated, merge the summit in six fields that are positioned at, as the cluster centre of each class, then find cluster along cluster centre search grid with this, finally all grids containing data point are assigned in a cluster; The cluster process that the present invention adopts puts from the hilltop to set out constantly to move down, until gradient no longer increases; The summit of gesture value is positioned at the Grad of 0, namely

The segmentation of one sub-picture can be selected to perform under various different color space, for the algorithm proposed, is necessary selection one most suitable color space segmentation image, to reach optimum segmentation effect.At present, the most generally adopt L*u*v* and L*a*b* two kinds of color spaces in Iamge Segmentation field, this is because consistent with the heterochromia expressed by geometric distance European in feature space by the heterochromia of L*u*v* and L*a*b* space display; In above-mentioned two situations, L* represents lightness dimension, and unique difference is that chromaticity coordinate is different.For the algorithm newly put forward, the result that two kinds of color spaces obtain is significantly difference not, therefore can be chosen on arbitrary color space and split image.The present invention adopts L*u*v* color space to complete image segmentation process as feature space.

1. level stress and strain model

Input data point, under the prerequisite of not knowing cluster number in advance, is divided into multiple class based on the inherent distance in sample between data point by Unsupervised clustering algorithm, that is, those data points close in distance are most possibly classified as same class; Therefore, in order to reduce the complexity of algorithm, we are at feature space grid division, and the point in grid is regarded as in advance belongs to same class; Specifically, form the network of multidimensional at feature space, each characteristic number strong point is put in one of them small grid determined.Such as, a three-dimensional color space, data object can be mapped to N ₁× N ₂× N ₃grid matrix in; The data point being positioned at same grid is considered to belong to a class.

Feature space is divided into two-grid by the present invention, and wherein the division of second layer network is the result based on first time stress and strain model.First feature space is divided into volume be the small grid structure of 2N × 2N × 2N as ground floor stress and strain model, more each adjacent eight neighborhood small grid merged become a macrolattice, and in this, as second layer grid.

2. gesture value is estimated

Utilize the graded mesh structure of feature space, the present invention proposes a new gesture value method of estimation.Feature space Ω is divided into 2N × 2N × 2N small grid object, calculates the average of data point in each small grid, and in this, as the eigenwert of this small grid, so feature based space Ω defines a new feature space Ω _s, the arithmetic mean of corresponding solid axes is as the spatial value of grid; Then merge eight neighborhood small grid and become a macrolattice as second layer grid, so obtain a new feature space Ω _band corresponding mesh space coordinate figure; For feature space Ω _b, use denotation coordination is (i _th, j _th, k _th) grid, wherein $f_{i_u{j}_u{k}_u}=(f_{i_u{j}_u{k}_u}^{\left(1\right)},f_{i_u{j}_u{k}_u}^{\left(2\right)},f_{i_u{j}_u{k}_u}^{\left(3\right)}),$ (i _th, j _th, k _th) quality of grid.Therefore for arbitrarily volume coordinate is (x, y), and gesture value estimation formulas is:

Wherein, $w_{i_u{j}_u{k}_u}=Q_{i_u{j}_u{k}_u}\&times;e^{-{\left(\frac{{x-x}_{i_u{j}_u{k}_u}}{{\mathrm{\&sigma;}}_x}\right)}^2+{\left(\frac{{y-y}_{i_u{j}_u{k}_u}}{{\mathrm{\&sigma;}}_Y}\right)}^2},$ be positioned at grid (i _th, j _th, k _th) quantity of interior data point, corresponding to grid (i _th, j _th, k _th) spatial value, σ _xand σ _yit is the spacial influence factor.Obtain K () by the gesture Distribution value of analogies center field of science, K () is proportional with gaussian kernel function; In order to improve the accuracy of this algorithm, factor of influence σ should arrange different values for different dimensional; Polynary gesture value function K is defined as the product of 3 one dimension gesture value functions.

Because gesture value is estimated to be similar to Density Estimator, can improve algorithm performance by arranging factor of influence σ, σ is the multiple of mesh width h, that is: σ=ch=c (h ₁, h ₂, h ₃) ^t, wherein c is proportionality constant, h=(h ₁, h ₂h ₃) ^tit is the window width of Density Estimator; User can regulate the value of c adaptively, thus obtains the image segmentation result of different levels; H value can obtain by using Sheather-Jones insertion.

3. clustering algorithm

The present invention proposes a kind of new clustering algorithm, this algorithm, based on the gesture Distribution value of network, effectively can improve arithmetic speed, be referred to as down-hill method.Contrast climbing method cluster process, the first step is the summit and maximum point that find that gesture value is estimated, and then do not have each summit to search for this to find cluster downwards, finally all small grid are by cluster.The cluster of multi-peak and arbitrary shape is the particular attribute of feature space, and the cluster process that the present invention proposes sets out from the hilltop constantly to move down, until Grad no longer increases, the summit of gesture value is the Grad being positioned at O point, namely

From any one summit, the direction increased according to Grad is constantly searched for, and summit and searched small grid are classified as a cluster.In down-hill method process, unlike climbing method, do not duplicate the small grid that search is identical, because this simplify the process of search.

4. based on the Ncut algorithm in region

Iamge Segmentation can be counted as the optimum segmentation of a figure.An image constructs a undirected weighting G=(V, E, W), and V is the summit of image, and E is the set on the limit of connect Vertex, and W is weight matrix; Weight w (u, v) on every bar limit is summit u and v similarity measurement function, and the figure of this structure is by minimizing cut value, and be divided into two disjoint subgraph A and B, cut is defined as foloows:

cut(A，B)＝∑ _{u∈A，v∈B}w(u，v).

But above-mentioned formula minimizes cut standard tends to mark off many isolated summits, in order to overcome this shortcoming, the cut of standard, namely Ncut algorithm is suggested, and is defined as follows:

$\mathrm{Ncut}(A,B)=\frac{\mathrm{cut}(A,B)}{\mathrm{assoc}(A,V)}+\frac{\mathrm{cut}(A,B)}{\mathrm{assoc}(B,V)}$

Minimize above-mentioned Ncut formula, obtain optimal dividing, the form of above-mentioned formula matrix is expressed as again:

${\min }_x\mathrm{Ncut}\left(x\right)=\frac{y^T(D-W)y}{y^T\mathrm{Dy}}$

Wherein D is diagonal matrix, D (i, i)=∑ _jw (i, j), the original block design of graphics G=(V that the present invention utilizes GDF algorithm to obtain, E, W), each block is regarded as a node, often pair of adjacent node is connected by a limit, on limit, weight reflects the similarity of two pieces of area attributes, namely belongs to the possibility of same target in image; Assuming that a sub-picture is divided into N number of region Ω do not overlapped _i(i=1,2 ..., R), this region contains n _iindividual characteristic number strong point, F _i(i=1,2 ..., R) and corresponding to each region Ω _iin the mean value of data point, the weighted value on every bar limit can be obtained by the similarity calculating adjacent area, and the weight of i block and j block is

Wherein, || .|| is the value asking vector, σ _tit is fixing factor of influence.

Compared with prior art, the image partition method that the broad sense data fields GDF that the present invention adopts is combined with Ncut algorithm, it adopts stress and strain model to carry out cluster, and adopt block-based Ncut algorithm to segment the image into the block with different characteristic meaning, reduce time complexity, substantially increase the travelling speed of segmentation, make the splitting speed of image sooner, more accurate.

Accompanying drawing explanation

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

Fig. 2 is the original graph of the embodiment of the present invention;

Fig. 3 of the present inventionly arranges surface chart;

Fig. 4 is the L*u*v* feature space set figure of the embodiment of the present invention;

Fig. 5 is the set figure of the feature space based on grid of the embodiment of the present invention;

Fig. 6 is the cluster result figure by using GDF algorithm to complete of the embodiment of the present invention;

Fig. 7 is the cluster result figure by using GDF algorithm to complete of the embodiment of the present invention;

Fig. 8 is the initial image segmentation result figure of the embodiment of the present invention;

Fig. 9 is the weight map that the Ncut algorithm of the scope domain hierarchy of the embodiment of the present invention produces;

Figure 10 is the final image segmentation result figure of the embodiment of the present invention;

Figure 11 is the contrast test figure of the present invention and other classical image segmentation algorithms;

Figure 12 is the present invention and other classical image segmentation algorithms comparing result when running.

Embodiment

Below in conjunction with embodiment shown in the drawings, the invention will be further described.

For accompanying drawing 2, before segmentation image, need the preliminary work done: first, the coordinate (x, y) of coloured image kingfisher.jpg data point is processed into RGB tristimulus values, and preserve data point RGB coordinate figure with text form; Then read text, obtain the RGB color value of all data points of this image, and this value is mapped to image space, form complete image and show, the number at statistical number strong point is 122500 simultaneously.The size of original image is 350 × 350; Finally input algorithm parameter, under default situations, monolateral small grid number 2N=12, macrolattice number of dimensions=2, feature space Simga-L=C=2.1, feature space Simga-U=C=2.1, feature space Simga-V=C=2.1, coordinate space Simga-X=70, coordinate space Simga-Y=70, the level and smooth threshold value M=40 of cluster, piecemeal Simga-Matrix=15, auxiliary node number=3, Ncut segmentation threshold=0.25, * three values of (times of) are respectively 4.8054,2.6593,4.8318, represent h respectively ₁, h ₂, h ₃, as shown in Figure 3.

Concrete dividing method comprises the following steps:

The division of step 1, level grid and gesture value are estimated, specifically comprise the following steps,

Step 1.1, the RGB color feature space of image is converted to L*u*v* color feature space, by the priming color value (R of all for this image data points, G, B) corresponding color feature space value (L is converted to, U, V), Fig. 4 is the space distribution of the data point after conversion, L*u*v* color feature space Ω is divided into 12 × 12 × 12 small grid as ground floor grid, calculate the average of data point in each small grid, and in this, as the eigenwert of this small grid, form a new feature space Ω _s, as shown in Figure 5;

Eight neighborhood small grid merges by step 1.2 becomes a macrolattice as second layer grid, forms a new feature space Ω _band corresponding mesh space coordinate figure;

Step 1.3 calculates the gesture value of each small grid in ground floor grid according to gesture value estimation formulas

At feature space Ω _bin, denotation coordination is (i _th, j _th, k _th) grid, (i _th, j _th, k _th) quality of grid, $f_{i_u{j}_u{k}_u}=(f_{i_u{j}_u{k}_u}^{\left(1\right)},f_{i_u{j}_u{k}_u}^{\left(2\right)},f_{i_u{j}_u{k}_u}^{\left(3\right)}),$ $W_{i_u{j}_u{k}_u}=Q_{i_u{j}_u{k}_u}\&times;e^{-{\left(\frac{x-x_{i_u{j}_u{k}_u}}{{\mathrm{\&sigma;}}_X}\right)}^2+{\left(\frac{y-y_{i_u{j}_u{k}_u}}{{\mathrm{\&sigma;}}_Y}\right)}^2,}$ be positioned at grid (i _th, j _th, k _th) quantity of interior data point, correspond to grid (i _th, j _th, k _th) spatial value, σ _xand σ _ythe spacial influence factor, σ in embodiment _xand σ _yget fixed value 70, σ=ch=c (h ₁, h ₂, h ₃) ^t, σ _j=ch _j, j=1,2,3, c is proportionality constant, and the span of c is [2.0,2.5], and in the present embodiment, value is 2.1, h=(h ₁, h ₂, h ₃) ^tbe the window width of Density Estimator, h value obtains by using Sheather-Jones insertion, and K (x) is unit potential function;

Step 2, gesture Distribution value according to small grid, to small grid cluster, cluster result is mapped to image, image is divided into the different regions mutually disjointed, and concrete steps are:

Step 2.1. asks local derviation to the formula in step 1.3, obtains formula:

utilize it to calculate the local derviation of each small grid in ground floor grid, and determine all mesh of vertices with this, describe cluster (C by six neighbo r pattern combination mesh of vertices _k} _{k=1 ..., v}, wherein C _kat least comprise a mesh of vertices;

Step 2.2 for each k=1,2 ..., v, cluster C _kin mesh of vertices as initial data point, namely along the direction search grid that Grad rises, until Grad no longer rises till, the small grid that Jiang Yanlu searches is divided into cluster C _kin;

After step 2.3 is searched for, for each cluster C _k, k=1,2 ..., v, by Mapping of data points all in each cluster on image, and is incorporated in the image fragment that spatially data point number is less than 40 points, and piece image is divided into R the prime area Ω do not overlapped _i, i=1,2 ..., R;

For every one dimension of (L, U, V) color feature space, pick out all small grid of the local maximum at this Directional partial derivative, so define three set, calculate three intersection of sets collection alternatively vertex set, if two candidate cluster comprise identical small grid in six neighborhoods, then merge this two candidate vertices set successively, until all candidate vertices set are all processed, which ensure that in any two set there is no identical element, the initial cluster center of new vertex set as class will be obtained, according to down-hill method search grid, carry out cluster, cluster result as shown in Figure 6 and Figure 7, wherein Fig. 7 is the projection of L*u*v* plane, 15 clusters are created in the present embodiment, and present with five kinds of colors and three kinds of symbols, namely

According to the cluster result produced, for the data point in each small grid and small grid distributes class-mark, record all data points that those participate in cluster calculation simultaneously, 121406 some records are had in the present embodiment, loss point 1094 before level and smooth, account for sum 0.0089, loss point is assigned in that class nearest with the Euclidean distance of its color, cluster result is mapped to plane space, obtain disjoint piece, then smoothing operation, comprise by those block that data point is less than level and smooth threshold value (M=40) and be considered as fragment, these fragments are all assigned in that block maximum in surrounding neighbors, and change the class-mark of data point in fragment, result as shown in Figure 8.Step 3, use the Ncut algorithm based on region to merge the region of over-segmentation, concrete steps are:

Step 3.1 based on obtain in step 2.3 not heavy and prime area build a undirected weighted graph G=(V, E, W), V is the summit of image, and E is the set on the limit of connect Vertex, and W is weight matrix, according to formula w (i, j)=

Calculate weight matrix W;

Step 3.2 calculates diagonal matrix D by weight matrix W,

Wherein D (i, i)=∑ _jw (i, j);

Step 3.3 is solved an equation (D-W) y=λ Dy, obtains eigenwert and corresponding proper vector, determines the second little proper vector;

Step 3.4 is according to formula finding out cut-point, some when namely Ncut value is minimum, with the summit of the second little proper vector bipartite graph, is two subgraphs by Iamge Segmentation;

Step 3.5 splits for two the subgraph obtained, and calculates weight matrix respectively, and repeats step 3.2 to 3.4;

Step 3.6 repeats step 3.5, until Ncut value exceeds given threshold values.

Claims (1)

1., based on an image partition method for broad sense data fields and Ncut algorithm, it is characterized in that: comprise the following steps,

The division of step 1, level grid and gesture value are estimated, specifically comprise the following steps,

Step 1.1, the RGB feature space of image is converted to L*u*v* or L*a*b* feature space, L*u*v* or L*a*b* feature space Ω is divided into 2N × 2N × 2N small grid as ground floor grid, calculate the average of data point in each small grid, and in this, as the eigenwert of this small grid, form a new feature space Ω _s;

Eight neighborhood small grid merges by step 1.2 becomes a macrolattice as second layer grid, forms a new feature space Ω _band corresponding mesh space coordinate figure;

Step 1.3 calculates the gesture value of each small grid in ground floor grid according to gesture value estimation formulas

At feature space Ω _bin, denotation coordination is (i _th, j _th, k _th) grid, (i _th, j _th, k _th) quality of grid, x, y are the variablees of volume coordinate, be positioned at grid (i _th, j _th, k _th) quantity of interior data point, correspond to grid (i _th, j _th, k _th) spatial value, σ _x, σ _ythe spacial influence factor, for the interaction force between control object, σ=ch=c (h ₁, h ₂, h ₃) ^t, σ _t=ch _t, t=1,2,3, c is proportionality constant, h=(h ₁, h ₂, h ₃) ^tbe the window width of Density Estimator, K (x) is unit potential function, meets: ∫ K (x) dx=1, ∫ xK (x) dx=0,0 < R (K)=∫ K (x) ²dx < ∞;

Step 2, employing down-hill method are to small grid cluster, according to the gesture Distribution value of small grid, based on graded mesh structure, cluster process is from any one summit, and the direction increased according to Grad is constantly searched for, summit and searched small grid are classified as a cluster, until gradient no longer increases, the small grid that Jiang Yanlu searches is divided in cluster, and cluster result is mapped to image, image is divided into the different regions mutually disjointed, concrete steps are:

Step 2.1. asks local derviation to the formula in step 1.3, obtains formula:

, utilize it to calculate the local derviation of each small grid in ground floor grid, and determine all mesh of vertices with this, describe cluster { C by six neighbo r pattern combination mesh of vertices _k, k=1,2 ..., v, v are the summit of image, wherein C _kat least comprise a mesh of vertices;

Step 2.2 for each k=1,2 ..., v, cluster C _kin mesh of vertices as initial data point, namely along the direction search grid that Grad rises, until Grad no longer rises till, the small grid that Jiang Yanlu searches is divided into cluster C _kin;

After step 2.3 is searched for, for each cluster C _k, k=1,2 ..., v, by Mapping of data points all in each cluster on image, and be incorporated in spatially data point number be less than M point image fragment,

20≤M≤100; Piece image is divided into the prime area Ω that R block does not overlap _i, i=1,2 ..., R;

Step 3, the Ncut algorithm based on region is used to merge the region of over-segmentation; The formula of wherein used calculating weight matrix W is:

N _i, n _jΩ respectively _i, Ω _jthe quantity at the characteristic number strong point comprised, i, j=1,2 ..., R and i ≠ j,

F _i, F _jcorrespond respectively to Ω _i, Ω _jthe mean value of middle data point, i, j=1,2 ..., R and i ≠ j, σ _iit is fixing factor of influence.