CN105631416A - Method for carrying out face recognition by using novel density clustering - Google Patents

Abstract

The present invention provides a method for carrying out face recognition by using novel density clustering. The method comprises the steps of reading a face image, converting an image matrix into characteristic vectors, taking a characteristic vector set f as the input of a data set P to be clustered, calculating the distance between two face characteristic vector points, calculating the average value of all points, finding a point closest to an average value point in the P as a density center point, calculating the density of all points, carrying out iterative searching of a density center point, taking a final convergence point as a density convergence point, a point whose density value is larger than t in the density convergence point as a local density center point, carrying out clustering of all points in a set LPS according to a neighbor algorithm, making a category label, initializing the category of all other non-local density center points p as -1, classifying the points into the same category of a convergence center point, marking the points with the category label of -1 as outliers, and finally outputting a clustering result. The method has high accuracy, complex data can be identified, and the accuracy of face identification is improved.

Description

Novel density is adopted to cluster the method carrying out recognition of face

Technical field

The present invention relates to and adopt novel density to cluster the method carrying out recognition of face.

Background technology

In recent years, recognition of face becomes a popular computer technology research field. Face recognition technology is as the one of biological identification technology, and it combines multiple research fields such as image procossing, computer graphics, pattern recognition. Clustering method is one of most important components of face recognition technology. Due to the usual complex distributions of face image data, conventional clustering method cannot identify the irregularly shaped classification of complexity well. How to be applied in recognition of face by clustering method accurate, healthy and strong, efficient is current urgent problem.

Summary of the invention

It is an object of the invention to propose a kind of method adopting novel density cluster to carry out recognition of face, with other based on compared with the clustering method of barycenter, there is accuracy height, can recognise that the advantage being distributed the complex datas such as unordered, aspherical, improve the accuracy of recognition of face.

The present invention is a kind of adopts novel density to cluster the method carrying out recognition of face, including following several steps:

Step 1, reading facial image:

Facial image used is the gray level image being sized to M �� N, and by each pixel as a characteristic point, reading K opens facial image and obtains image characteristic matrix A_i(M��N), wherein i=1,2 ..., K; By image characteristic matrix A_i(M��N)Be converted to characteristic vector f_i(_1��MN), wherein f_i(1:N)=A_i(1,1:N), f_i(N+1:2N)=A_i(2,1:N), the rest may be inferred, to characteristic vector f_iEach dimension do 0-1 normalization, namelyWherein j=1,2 ..., MN, then update characteristic vector f_i;

Step 2, computed range matrix:

Inputting as data set P to be clustered with set of eigenvectors f, the number of its face feature vector point is designated as N, and in data set P to be clustered, i-th face feature vector point is designated as p_i, calculate each two face feature vector point p in data set P to be clustered_iBetween distance, generate distance matrix (d_i,j)_N��N, wherein d_i,jRepresent some p_iTo p_jDistance;

Step 3, find out all face feature vector point p_iDensity center point:

With r as sweep radius to each face feature vector point p_iBe scanned, by data set P to be clustered with p_iThe distance face feature vector point less than r join p_iR-Neighbor gather Nei_i={ p_j|d_i,jIn��r}, calculate set Nei_iIn all face feature vector point p_iMeansigma methodsWherein k is set Nei_iTotal number of middle face feature vector point, finds distance average Mean in data set P to be clustered_iNearest face feature vector point is as p_iDensity center point c_i, namelyWherein | | p-mean_i||₂Represent some p and mean_iGeometric distance;

Step 4, calculating face vector characteristics point p_jDensity:

Nei will be gathered_iThe number of middle element is as face feature vector point p_iDensity value ��_i=size (Nei_i), namely ${\mathrm{\ρ}}_{i,r}=\underset{k=1}{\overset{N}{\Σ}}l(d_{i,k}\≤r),$ Wherein $l\left(x\right)=\left\{\begin{array}{cc}1,& xistrue\\ 0,& else\end{array}\right.;$

Step 5, find out all face feature vector point p_jDensity convergent point:

Each face feature vector point p is found by iterative computation_jLocal density's central point, to the last restrain, convergence point is called the density convergent point of this point:

Find p_iDensity center point c_i=p_j, find p_jDensity center point c_j=p_kIf, p_kDensity center point remain c_k=p_k, then iteration is stopped, by p_kAs p_i��p_j��p_kDensity convergent point, be denoted as cp (i)=cp (j)=cp (k)=p_k;

Step 6, find out all face feature vector point p_jLocal density's peak point:

If face feature vector point p_iDensity convergent point be himself, i.e. cp (i)=p_i, and p_iDensity value ��_iMore than given density threshold t, then add it to, in local density peak point set LPS, be denoted as LPS={p_i| (cp (i)=p_i)��(��_i��t)}

Step 7, all face feature vector point pi in local density peaks point set LPS are clustered by nearest algorithm, each classification is exactly a density core, and the classification of local density central point lp each in local density peak point set LPS is designated as cl (lp);

Step 8, the classification of other all non local density peaks point p is initialized as-1, incorporate into again and its convergence center point identical category, i.e. cl (p)=cl (cp (p)), is designated as outlier the point of cl (p)=-1;

Step 9, output cluster result cl (p):

Cl (p) represents the classification of this face, if cl (p1)=cl (p2), represents that two facial images representated by p1 and p2 are same people, if cl (p)=-1 represents that this face classification is unidentified.

Owing to the present invention is clustered by searching density core but not single barycenter, the density core of one classification is made up of the loose convergent point being bound up that some density are of a relatively high, other data point focuses on these some peripheries by a fixed structure. thus traditional clustering algorithm can be overcome and be difficult to the data of non-norm spherical structure, improve the accuracy of recognition of face.

Detailed description of the invention

The present invention is a kind of adopts novel density to cluster the method carrying out recognition of face, including following several steps:

Step 1, reading facial image:

Facial image used is the gray level image being sized to M �� N, and by each pixel as a characteristic point, reading K opens facial image and obtains image characteristic matrix A_i(M��N), wherein i=1,2 ..., K; Calculate in order to convenient, by image characteristic matrix A_i(M��N)Be converted to characteristic vector f_i(1��MN), wherein f_i(1:N)=A_i(1,1:N), f_i(N+1:2N)=A_i(2,1:N), the rest may be inferred, to characteristic vector f_iEach dimension do 0-1 normalization, namelyWherein j=1,2 ..., MN, then update characteristic vector f_i;

Step 2, computed range matrix:

Inputting as data set P to be clustered with set of eigenvectors f, the number of its face feature vector point is designated as N, and in data set P to be clustered, i-th face feature vector point is designated as p_i, calculate each two face feature vector point p in data set P to be clustered_iBetween distance, generate distance matrix (d_i,j)_N��N, wherein d_i,jRepresent some p_iTo p_jDistance;

Step 3, find out all face feature vector point p_iDensity center point:

With r as sweep radius to each face feature vector point p_iIt is scanned, with i-th face feature vector point p_iFor example, by P with p_iThe distance face feature vector point less than r join p_iR-Neighbor gather Nei_i={ p_j|d_i,jIn��r}, calculate set Nei_iIn the meansigma methods of all face feature vector pointsWherein k is set Nei_iTotal number of middle face feature vector point, finds distance average Mean in data set P to be clustered_iNearest face feature vector point is as p_iDensity center point c_i, namelyWherein | | p-mean_i||₂Represent some p and mean_iGeometric distance;

Step 4, calculating face vector characteristics point p_jDensity:

Nei will be gathered_iThe number of middle element is as face feature vector point p_iDensity value ��_i=size (Nei_i), namely ${\mathrm{\ρ}}_{i,r}=\underset{k=1}{\overset{N}{\Σ}}l(d_{i,k}\≤r),$ Wherein $l\left(x\right)=\left\{\begin{array}{cc}1,& xistrue\\ 0,& else\end{array}\right.;$

Step 5, find out all face feature vector point p_jDensity convergent point (convergentpoint), namely find each face feature vector point p by iterative computation_jLocal density's central point, to the last restrain, convergence point is called the density convergent point of this point:

Find p_iDensity center point be c_i=p_j, find p_jDensity center point be c_j=p_kIf, p_kDensity center point remain c_k=p_k, then iteration is stopped, by p_kAs p_i��p_j��p_kDensity convergent point, be denoted as cp (i)=cp (j)=cp (k)=p_k;

Step 6, find out all face feature vector point p_jLocal density's peak point (Localdensitypeak):

If a face feature vector point p_iDensity convergent point be himself, i.e. cp (i)=p_i, and p_iDensity value ��_iMore than given density threshold t, then add it to, in local density peak point set LPS, be denoted as LPS={p_i| (cp (i)=p_i)��(��_i��t)}

Step 7, all face feature vector point pi in local density peaks point set LPS are clustered by nearest algorithm (NearestNeighborClustering), each classification is exactly a density core, and the classification of local density central point lp each in local density peak point set LPS is designated as cl (lp);

Step 8, the classification of other all non local density peaks point p is initialized as-1, incorporate into again and its convergence center point identical category, i.e. cl (p)=cl (cp (p)), is designated as outlier the point of cl (p)=-1;

Step 9, output cluster result cl (p):

Cl (p) represents the classification of this face, if cl (p1)=cl (p2), represents that two facial images representated by p1 and p2 are same people, if cl (p)=-1 represents that this face classification is unidentified.

The above, not impose any restrictions the technical scope of the present invention, therefore every any trickle amendment, equivalent variations and modification above example made according to the technical spirit of the present invention, all still fall within the scope of technical solution of the present invention.

Claims (1)

1. one kind adopts novel density to cluster the method carrying out recognition of face, it is characterised in that include following several step:

Step 1, reading facial image:

Facial image used is the gray level image being sized to M �� N, and by each pixel as a characteristic point, reading K opens facial image and obtains image characteristic matrix A_i(M��N), wherein i=1,2 ..., K; By image characteristic matrix A_i(M��N)Be converted to characteristic vector f_i(1��MN), wherein f_i(1:N)=A_i(1,1:N), f_i(N+1:2N)=A_i(2,1:N), the rest may be inferred, to characteristic vector f_iEach dimension do 0-1 normalization, namely $f_i\left(j\right)=\frac{f_i\left(j\right)-\min \left(f_{1:K}\left(j\right)\right)}{\max \left(f_{1:K}\left(j\right)\right)-\min \left(f_{1:K}\left(j\right)\right)},$ Wherein j=1,2 ..., MN, then update characteristic vector f_i;

Step 2, computed range matrix:

Inputting as data set P to be clustered with set of eigenvectors f, the number of its face feature vector point is designated as N, and in data set P to be clustered, i-th face feature vector point is designated as p_i, calculate each two face feature vector point p in data set P to be clustered_iBetween distance, generate distance matrix (d_i,j)_N��N, wherein d_i,jRepresent some p_iTo p_jDistance;

Step 3, find out all face feature vector point p_iDensity center point:

With r as sweep radius to each face feature vector point p_iBe scanned, by data set P to be clustered with p_iThe distance face feature vector point less than r join p_iR-Neighbor gather Nei_i={ p_j|d_i,jIn��r}, calculate set Nei_iIn all face feature vector point p_iMeansigma methodsWherein k is set Nei_iTotal number of middle face feature vector point, finds distance average Mean in data set P to be clustered_iNearest face feature vector point is as p_iDensity center point c_i, namelyWherein | | p-mean_i||₂Represent some p and mean_iGeometric distance;

Step 4, calculating face vector characteristics point p_jDensity:

Nei will be gathered_iThe number of middle element is as face feature vector point p_iDensity value ��_i=size (Nei_i), namely ${\mathrm{\ρ}}_{i,r}=\underset{k=1}{\overset{N}{\Σ}}l(d_{i,k}\≤r),$ Wherein $l\left(x\right)=\left\{\begin{array}{cc}1,& xistrue\\ 0,& else\end{array}\right.;$

Step 5, find out all face feature vector point p_jDensity convergent point:

Each face feature vector point p is found by iterative computation_jLocal density's central point, to the last restrain, convergence point is called the density convergent point of this point:

Find p_iDensity center point c_i=p_j, find p_jDensity center point c_j=p_kIf, p_kDensity center point remain c_k=p_k, then iteration is stopped, by p_kAs p_i��p_j��p_kDensity convergent point, be denoted as cp (i)=cp (j)=cp (k)=p_k;

Step 6, find out all face feature vector point p_jLocal density's peak point:

If face feature vector point p_iDensity convergent point be himself, i.e. cp (i)=p_i, and p_iDensity value ��_iMore than given density threshold t, then add it to, in local density peak point set LPS, be denoted as LPS={p_i| (cp (i)=p_i)��(��_i��t)}

Step 7, all face feature vector point pi in local density peaks point set LPS are clustered by nearest algorithm, each classification is exactly a density core, and the classification of local density central point lp each in local density peak point set LPS is designated as cl (lp);

Step 8, the classification of other all non local density peaks point p is initialized as-1, incorporate into again and its convergence center point identical category, i.e. cl (p)=cl (cp (p)), is designated as outlier the point of cl (p)=-1;

Step 9, output cluster result cl (p):

Cl (p) represents the classification of this face, if cl (p1)=cl (p2), represents that two facial images representated by p1 and p2 are same people, if cl (p)=-1 represents that this face classification is unidentified.