CN104881634A - Illumination face recognition method based on completed local convex-and-concave pattern - Google Patents

Abstract

The invention relates to an illumination face recognition method based on a completed local convex-and-concave pattern, and belongs to the pattern recognition field. The method includes the steps of dividing an image; carrying out bilinear interpolation for each piece of the image; encoding the symbol characteristic and the amplitude characteristic of a local difference of each pixel point in each piece of the image to obtain a symbol characteristic matrix and an amplitude characteristic matrix of each piece of the image; encoding pixel points of each piece of the image to obtain a central pixel characteristic matrix of each piece of the image, extracting the histogram characteristics of the three characteristic matrixes to obtain three characteristic vectors, and successively connecting the three characteristic vectors to obtain histogram characteristic vectors of all pieces of the image; and finally connecting the histogram characteristic vectors of all pieces of the image to obtain a histogram characteristic vector of the original image, sending the characteristic vector to the nearest neighboring classifier to be classified, and verifying the identity of an original face image. The method is an image texture description method based on second-order differential, and is capable of effectively identifying human faces in an illumination environment.

Description

A kind of illumination face recognition method based on complete Local Convex diesinking

Technical field

The present invention relates to a kind of illumination face recognition method based on complete Local Convex diesinking, belong to mode identification technology.

Background technology

Local binary patterns (Local binary pattern, LBP) [L.Wang and D.C.He, " Texture classification usingtexture spectrum ", Pattern Recognition, vol.23, pp.905-910,1990.] be a kind of important image characteristics extraction operator, there is the little and effective feature of calculated amount.Although LBP is successful in computer vision and area of pattern recognition, its working mechanism still has worth improvements.Dominant local binary patterns (Dominant local binary patterns, DLBP) [S.Liao, M.W.K.Law, and A.C.S.Chung, " Dominant local binary patterns for texture classification; " IEEETrans.Image Process., vol.18, no.5, pp.1107 – 1118, May 2009.] on all pattern bases of the LBP of statistical picture, filter out the pattern of upper frequency, and the high frequency mode that cumulative frequency reaches 80% is formed final proper vector.LBP only considers the symbolic information of center pixel and surrounding pixel difference, complete local binary patterns (Completed local binary pattern, CLBP) [Z.Guo, L.Zhang and D.Zhang, " A completed modeling of local binary pattern operator for textureclassification; " IEEE Trans.Image Process., vol.19, no.6, pp.1657-1663,2010.] not only consider symbolic information, also contemplate the amplitude information of difference and the feature of central pixel point.What LBP extracted is the first differential information of image, infinitesimal pattern (Local derivative pattern, LDP) [B.Zhang, Y.Gao, S.Zhao, and J.Liu, " Local derivative pattern versuslocal binary pattern:Face recognition with higher-order local pattern descriptor; " IEEE Trans.Image Process., vol.19, no.2, pp.533 – 544, Feb.2010.] improve LBP algorithm, be extracted the second-order differential information of image.In order to reduce the number of pattern in LBP algorithm, researchist proposes Central Symmetry infinitesimal pattern (Center-Symmetric Local derivativePattern, CS-LDP) [G.Xue, L.Song, J.Sun, M.Wu, Hybrid Center-Symmetric Local Pattern for DynamicBackground Subtraction, ICME, Barcelona, Spain (2011), pp.1 – 6, July 2011.] and Central Symmetry local binary patterns algorithm (Center-symmetric local binary pattern, CS-LBP) [Marko H, Matti P, Cordelia S.Description of interestregions with center-symmetric local binary pattern [C] //Conference on Computer Vision Graphics and ImageProcessing.2006, 4338:58-69].Local binary counting (Local binary count, LBC) [Zhao Y, Huang D S, Jia W, " Completed local binary count for rotation invariant texture classification, " IEEE Trans.Image Process., vol.21, no.10, pp.4492-4497,2012.] only consider that in binary pattern, pattern is the number of " 1 ".Unified local binary patterns decreases model number, decreases calculated amount [T.Ojala, M. t. " Gray scale and rotation invariant textureclassification with local binary patterns; " in:D.Vernon (Ed.), Proceedings of the Sixth European Conference onComputer Vision (ECCV2000), Dublin, Ireland, pp.404 – 420,2000.].In order to strengthen LBP algorithm extract the distinctive of texture, LBP algorithm also combines [Zhang W C with Gabor filter and some Data Dimensionality Reduction Algorithms, Shan S G, Gao W, et a1.Local Gabor Binary Pattern Histogram Sequence. (LGBPHS): A Novel Non-Statistical Model for FaceRepresentation and Recognition [C] Proc of the 10th IEEE Int ' l Conf on Computer Vision, 2005:786-791.; B.Zhang, S.Shan, X.Chen, and W.Gao, " Histogram of Gabor Phase Patterns (HGPP): A novel object representationapproach for face recognition, " IEEE Trans.Image Process., vol.16, no.1, pp.57 – 68,2007.].

LBP only considers the first differential information of image, the object of the present invention is to provide a kind of facial image second-order differential textural characteristics describing method based on image complete local convexo-concave characteristic (Completed Local convex-and-concave Pattern, CLCCP).

Summary of the invention

The invention provides a kind of illumination face recognition method based on complete Local Convex diesinking, for solution photoenvironment human face identification problem.Only can the defect of Description Image first differential for local binary patterns, the complete Local Convex diesinking that the present invention proposes can the second-order differential feature of effective Description Image.Complete Local Convex diesinking not only considers the symbolic information of local difference, and considers the amplitude characteristic of difference, also take into account the distinctive of central pixel point.

The illumination face recognition method that the present invention is based on complete Local Convex diesinking is achieved in that and first carries out piecemeal to image; Then bilinear interpolation is carried out to each block image, make each pixel in image can build 8 symmetry directions, then to calculate in block image each pixel along local, 8 directions difference; The symbolic feature CLCCP_S of this local difference of then encoding and amplitude characteristic CLCCP_M; Each pixel of each image block is encoded, obtains the center pixel feature CLCCP-C of each image block; Next the CLCCP_S to each block image, CLCCP_M and CLCCP_C eigenmatrix extracts histogram feature vector, connect this block image CLCCP_S successively, the histogram feature vector of CLCCP_S and CLCCP_C feature obtains the histogram feature vector of each block image; The histogram feature vector finally connecting each block image obtains the histogram feature vector of this original image, this proper vector is sent into and classifies based on the nearest neighbor classifier of chi amount, identify the identity of original facial image.

The concrete steps of the described illumination face recognition method based on complete Local Convex diesinking are as follows:

Step1, first image is carried out piecemeal: image I ^(l)evenly be divided into the zero lap square of 4 × 4,16 pieces altogether, be expressed as

Step2, bilinear interpolation computing is carried out to each block image, make each pixel can build about point-symmetric 8 directions of this pixel, then calculate the local difference of each pixel along different directions, this local difference is decomposed into symbolic component and and amplitude portion;

As shown in Figure 2, pixel P ₁and P ₂between can increase pixel Q by interpolation ₁.Interpolation method as shown in Figure 4, wherein P ₁₁, P ₁₂, P ₂₁, P ₂₂be four neighbor pixels original in image, insert by interpolation method the pixel Q that makes new advances ₀.Interpolation formula is as follows:

$I_{R_1}\≈\frac{x_2-x}{x_2-x_1}{I}_{P_{11}}+\frac{x-x_1}{x_2-x_1}{I}_{P_{21}},$

$I_{R_2}\≈\frac{x_2-x}{x_2-x_1}{I}_{P_{12}}+\frac{x-x_1}{x_2-x_1}{I}_{P_{22}},$

$I_{Q_0}\≈\frac{y_2-y}{y_2-y_1}{I}_{R_1}+\frac{y-y_1}{y_2-y_1}{I}_{R_2};$

Wherein with represent R respectively ₁, R ₂with the pixel value of position, x ₁, x and x ₂represent pixel P respectively ₁₁, R ₁and P ₂₁the horizontal ordinate at place, y ₁, y and y ₂represent pixel P respectively ₁₁, Q ₀and P ₁₂the ordinate at place.Fig. 3 represents pixel X in original image ₀there is P around ₀, P ₁, P ₂, P ₃, P ₄, P ₅, P ₆and P ₇8 Neighbor Points, only can form four about pixel X ₀symmetry direction.Fig. 2 represents pixel X after interpolation ₀there is Q around ₀, Q ₁, Q ₂, Q ₃, Q ₄, Q ₅, Q ₆and Q ₇8 interpolation points, therefore pixel X after interpolation ₀one co-exist in 16 Neighbor Points around, 8 can be obtained about pixel X ₀symmetry direction.Owing to adding interpolation point, the resolution of image is enhanced;

Pixel X in image block ₀partial error along 8 directions is divided into with wherein i=0,1,2,3 and j=0,1,2,3;

Step3, symbolic component and amplitude portion carried out respectively to corresponding Local Convex diesinking coding, obtain symbolic feature CLCCP-S and the amplitude characteristic CLCCP-M of each block image, wherein pixel X ₀symbolic feature and the coding formula of amplitude characteristic be respectively:

$\mathrm{CLCCP}-S_{\mathrm{1,8}}{\left(X_0\right)}_D=\underset{i=0}{\overset{3}{\mathrm{\Σ}}}[f(I_{X_0}-0.5*(I_{P_i}+I_{P_{i+4}}))*2^{i*2}]+\underset{j=0}{\overset{3}{\mathrm{\Σ}}}[f(I_{X_0}-0.5*(I_{Q_j}+I_{Q_{j+4}}))*2^{j*2+1}]$

$\mathrm{CLCCP}-M_{\mathrm{1,8}}{\left(X_0\right)}_D=\underset{i=0}{\overset{3}{\mathrm{\Σ}}}[f(\mathrm{abs}(I_{X_0}-0.5*(I_{P_i}+I_{P_{i+4}}))-c)*2^{i*2}]+\underset{j=0}{\overset{3}{\mathrm{\Σ}}}[f(\mathrm{abs}(I_{X_0}-0.5*(I_{Q_j}+I_{Q_{j+4}}))-c)*2^{j*2+1}]$

Wherein, CLCCP-S _1,8(X ₀) _drepresent pixel X ₀the Local Convex concavity symbolic feature at place, CLCCP-M _1,8(X ₀) _drepresent pixel X ₀the Local Convex concavity amplitude characteristic at place, $f\left(x\right)=\left\{\begin{array}{cc}0,\mathrm{if}& x\≤\mathrm{threshold}\\ 1,\mathrm{if}& x>\mathrm{threshold}\end{array}\right.,$ represent X in image ₀the pixel value at place, with represent P in image _iand P _i+4the pixel value at place, with represent Q in image _jand Q _j+4the pixel value at place, abs () expression asks for signed magnitude arithmetic(al), CLCCP-S _1,8(X ₀) _dand CLCCP-M _1,8(X ₀) _dmiddle subscript " 1 " represents the pixel distance X that calculating convex-concave characteristic is used ₀distance be 1, namely yardstick is 1, subscript " 8 " represent calculate through pixel X ₀the convexo-concave characteristic in 8 directions, it is decimal system amount that subscript " D " represents, threshold is the threshold value pre-set;

Step4, to encode to each pixel of each image block, obtain the center pixel feature CLCCP-C of each image block, coding formula is: here c _irepresent the mean value of entire image, represent X in image ₀the pixel value at place, $f\left(x\right)=\left\{\begin{array}{cc}0,\mathrm{if}& x\≤\mathrm{threshold}\\ 1,\mathrm{if}& x>\mathrm{threshold}\end{array}\right.,$ CLCCP-C _1,8(X ₀) _dmiddle subscript " 1 " represents the pixel distance X that calculating convex-concave characteristic is used ₀distance be 1, namely yardstick is 1, subscript " 8 " represent calculate through pixel X ₀the convexo-concave characteristic in 8 directions, subscript " D " represents decimal system amount, and threshold is the threshold value pre-set;

Step5, through step Step2, Step3 and Step4, be extracted image block complete local convexo-concave characteristic, comprise and meet feature, amplitude characteristic and center pixel feature, work as image block in pixel X ₀when traveling through whole image block, obtain each block image the eigenmatrix of CLCCP-S, CLCCP-M, CLCCP-C, be respectively

The histogram feature vector of Step6, the next each image block of extraction three eigenmatrixes, image block three eigenmatrixes histogram feature vector be expressed as: connect this three histogram feature vectors successively, obtain image block histogram feature vector ${H_i^{\left(l\right)}}_{\mathrm{CLCCP}}=[{H_i^{\left(l\right)}}_{\mathrm{CLCCP}-S},{H_i^{\left(l\right)}}_{\mathrm{CLCCP}-M},{H_i^{\left(l\right)}}_{\mathrm{CLCCP}-C}],$ Herein subscript CLCCP represent complete Local Convex diesinking feature, it comprises symbolic feature CLCCP-S, amplitude characteristic CLCCP-M and center pixel point patterns CLCCP-C;

Step7, connect each image block histogram feature vector, the complete Local Convex diesinking histogram feature vector obtaining original image is: ${I^{\left(l\right)}}_{\mathrm{CLCCP}}=[{H_0^{\left(l\right)}}_{\mathrm{CLCCP}},{H_1^{\left(l\right)}}_{\mathrm{CLCCP}},...,{H_{15}^{\left(l\right)}}_{\mathrm{CLCCP}}];$

Step8, this proper vector send into classify based on the nearest neighbor classifier of chi amount, identify the identity of original facial image;

In described step Step8, when the nearest neighbor classifier based on chi amount is classified, first computer card side statistic; Set two width facial image I ⁽⁰⁾and I ⁽¹⁾complete Local Convex diesinking histogram feature vector be respectively: ${I^{\left(0\right)}}_{\mathrm{CLCCP}}=[{H_0^{\left(0\right)}}_{\mathrm{CLCCP}},{H_1^{\left(0\right)}}_{\mathrm{CLCCP}},...,{H_{15}^{\left(0\right)}}_{\mathrm{CLCCP}}]$ With ${I^{\left(1\right)}}_{\mathrm{CLCCP}}=[{H_0^{\left(1\right)}}_{\mathrm{CLCCP}},{H_1^{\left(1\right)}}_{\mathrm{CLCCP}},...,{H_{15}^{\left(1\right)}}_{\mathrm{CLCCP}}],$ Distance then between these two vectors, namely chi span is from the following formulae discovery of employing:

${\mathrm{\χ}}^2({I^{\left(0\right)}}_{\mathrm{CLCCP}},{I^{\left(1\right)}}_{\mathrm{CLCCP}})=\underset{i=1}{\overset{K^{\′}}{\mathrm{\Σ}}}\frac{{({I^{\left(0\right)}}_{\mathrm{CLCCP}}\left(i\right)-{I^{\left(1\right)}}_{\mathrm{CLCCP}}\left(i\right))}^2}{({I^{\left(0\right)}}_{\mathrm{CLCCP}}\left(i\right)+{I^{\left(1\right)}}_{\mathrm{CLCCP}}\left(i\right)+\mathrm{eps})}$

Wherein I ⁽⁰⁾ _cLCCP(i) and I ⁽¹⁾ _cLCCPi () represents texture feature vector I respectively ⁽⁰⁾ _cLCCPand I ⁽¹⁾ _cLCCPi-th element, K' represents the length of texture, and eps is a fixed value, is positive number minimum in Matlab, represents one and makes I ⁽⁰⁾ _cLCCP(i)+I ⁽¹⁾ _cLCCPthe very little normal number of (i)+eps ≠ 0.

The invention has the beneficial effects as follows:

1, the complete Local Convex diesinking facial image Texture Segmentation Algorithm of the present invention's structure is a kind of texture description operator based on image second order differential characteristics, and overcoming LBP can only the defect of Description Image first differential information;

2, the complete Local Convex diesinking facial image Texture Segmentation Algorithm of the present invention's structure not only describes the symbolic information of image local convexo-concave characteristic, also describe the amplitude characteristic of image local convexo-concave characteristic, and consider the distinguishing ability of picture centre pixel, merge the distinctive that this three improves texture;

3, this method not only considers the concavity of facial image local grain, also contemplates the size of image local texture concavity.When the experiment of recognition of face aspect shows that this algorithm carries out illumination recognition of face, computation complexity is low, and accuracy of identification is high, has insensitivity to illumination;

4, in the match cognization stage, the present invention adopts chi amount (Chi square statistic) as the distance metric between two texture feature vectors, employing nearest neighbor classifier is classified, and algorithm is simple, convenience of calculation, can accomplish real-time images match identification.

Accompanying drawing explanation

Fig. 1 is image complete Local Convex diesinking characteristic extraction step schematic block diagram in the present invention;

Fig. 2 is present invention pixel point X ₀8 symmetry direction schematic diagram;

Fig. 3 is pixel 4 symmetry direction schematic diagram in image of the present invention;

Fig. 4 is bilinear interpolation schematic diagram in the present invention;

Fig. 5 is 64 sample images of a people in embodiment of the present invention the extended used YaleB face database illumination subset;

Fig. 6 is local binary patterns in the present invention, unified local binary patterns (Uniform local binary pattern, UniformLBP), complete local binary patterns and the cumulative matching characteristic curve of this method on the extended Yale B database;

Fig. 7 is local binary patterns in the present invention, the correct recognition rata curve of unified local binary patterns, complete local binary patterns and this method on theextended Yale B database.

Embodiment

Embodiment 1: as shown in figs. 1-7, a kind of illumination face recognition method based on complete Local Convex diesinking, first carries out piecemeal to image; Then bilinear interpolation is carried out to each block image, make each pixel in image can build 8 symmetry directions, then to calculate in block image each pixel along local, 8 directions difference; The symbolic feature of this local difference of then encoding and amplitude characteristic; Each pixel of each image block is encoded, obtains the center pixel feature of each image block; The eigenmatrix of the following symbolic feature to each block image, amplitude characteristic, center pixel feature extracts histogram feature vector, connects the vectorial histogram feature vector obtaining each block image of histogram feature of this block image symbolic feature, amplitude characteristic, center pixel feature successively; The histogram feature vector finally connecting each block image obtains the histogram feature vector of this original image, this proper vector is sent into and classifies based on the nearest neighbor classifier of chi amount, identify the identity of original facial image.

The concrete steps of the described illumination face recognition method based on complete Local Convex diesinking are as follows:

Step1, first image is carried out piecemeal: image I ^(l)evenly be divided into the zero lap square of 4 × 4,16 pieces altogether, be expressed as

Step2, bilinear interpolation computing is carried out to each block image, make each pixel can build about point-symmetric 8 directions of this pixel, then calculate the local difference of each pixel along different directions, this local difference is decomposed into symbolic component and and amplitude portion;

Step3, symbolic component and amplitude portion carried out respectively to corresponding Local Convex diesinking coding, obtain symbolic feature CLCCP-S and the amplitude characteristic CLCCP-M of each block image, wherein pixel X ₀symbolic feature and the coding formula of amplitude characteristic be respectively:

$\mathrm{CLCCP}-S_{\mathrm{1,8}}{\left(X_0\right)}_D=\underset{i=0}{\overset{3}{\mathrm{\Σ}}}[f(I_{X_0}-0.5*(I_{P_i}+I_{P_{i+4}}))*2^{i*2}]+\underset{j=0}{\overset{3}{\mathrm{\Σ}}}[f(I_{X_0}-0.5*(I_{Q_j}+I_{Q_{j+4}}))*2^{j*2+1}]$

$\mathrm{CLCCP}-M_{\mathrm{1,8}}{\left(X_0\right)}_D=\underset{i=0}{\overset{3}{\mathrm{\Σ}}}[f(\mathrm{abs}(I_{X_0}-0.5*(I_{P_i}+I_{P_{i+4}}))-c)*2^{i*2}]+\underset{j=0}{\overset{3}{\mathrm{\Σ}}}[f(\mathrm{abs}(I_{X_0}-0.5*(I_{Q_j}+I_{Q_{j+4}}))-c)*2^{j*2+1}]$

Wherein, CLCCP-S _1,8(X ₀) _drepresent pixel X ₀the Local Convex concavity symbolic feature at place, CLCCP-M _1,8(X ₀) _drepresent pixel X ₀the Local Convex concavity amplitude characteristic at place, $f\left(x\right)=\left\{\begin{array}{cc}0,\mathrm{if}& x\≤\mathrm{threshold}\\ 1,\mathrm{if}& x>\mathrm{threshold}\end{array}\right.,$ represent X in image ₀the pixel value at place, with represent P in image _iand P _i+4the pixel value at place, with represent Q in image _jand Q _j+4the pixel value at place, abs () expression asks for signed magnitude arithmetic(al), CLCCP-S _1,8(X ₀) _dand CLCCP-M _1,8(X ₀) _dmiddle subscript " 1 " represents the pixel distance X that calculating convex-concave characteristic is used ₀distance be 1, namely yardstick is 1, subscript " 8 " represent calculate through pixel X ₀the convexo-concave characteristic in 8 directions, it is decimal system amount that subscript " D " represents, threshold is the threshold value pre-set;

Step4, to encode to each pixel of each image block, obtain the center pixel feature CLCCP-C of each image block, coding formula is: here c _irepresent the mean value of entire image, represent X in image ₀the pixel value at place, $f\left(x\right)=\left\{\begin{array}{cc}0,\mathrm{if}& x\≤\mathrm{threshold}\\ 1,\mathrm{if}& x>\mathrm{threshold}\end{array}\right.,$ CLCCP-C _1,8(X ₀) _dmiddle subscript " 1 " represents the pixel distance X that calculating convex-concave characteristic is used ₀distance be 1, namely yardstick is 1, subscript " 8 " represent calculate through pixel X ₀the convexo-concave characteristic in 8 directions, subscript " D " represents decimal system amount, and threshold is the threshold value pre-set;

Step5, through step Step2, Step3 and Step4, be extracted image block complete local convexo-concave characteristic, comprise and meet feature, amplitude characteristic and center pixel feature, work as image block in pixel X ₀when traveling through whole image block, obtain each block image the eigenmatrix of CLCCP-S, CLCCP-M, CLCCP-C, be respectively

The histogram feature vector of Step6, the next each image block of extraction three eigenmatrixes, image block three eigenmatrixes histogram feature vector be expressed as: connect this three histogram feature vectors successively, obtain image block histogram feature vector ${H_i^{\left(l\right)}}_{\mathrm{CLCCP}}=[{H_i^{\left(l\right)}}_{\mathrm{CLCCP}-S},{H_i^{\left(l\right)}}_{\mathrm{CLCCP}-M},{H_i^{\left(l\right)}}_{\mathrm{CLCCP}-C}],$ Herein subscript CLCCP represent complete Local Convex diesinking feature, it comprises symbolic feature CLCCP-S, amplitude characteristic CLCCP-M and center pixel point patterns CLCCP-C;

Step7, connect each image block histogram feature vector, the complete Local Convex diesinking histogram feature vector obtaining original image is: ${I^{\left(l\right)}}_{\mathrm{CLCCP}}=[{H_0^{\left(l\right)}}_{\mathrm{CLCCP}},{H_1^{\left(l\right)}}_{\mathrm{CLCCP}},...,{H_{15}^{\left(l\right)}}_{\mathrm{CLCCP}}];$

Step8, this proper vector send into classify based on the nearest neighbor classifier of chi amount, identify the identity of original facial image.

In described step Step8, when the nearest neighbor classifier based on chi amount is classified, first computer card side statistic; Set two width facial image I ⁽⁰⁾and I ⁽¹⁾complete Local Convex diesinking histogram feature vector be respectively: ${I^{\left(0\right)}}_{\mathrm{CLCCP}}=[{H_0^{\left(0\right)}}_{\mathrm{CLCCP}},{H_1^{\left(0\right)}}_{\mathrm{CLCCP}},...,{H_{15}^{\left(0\right)}}_{\mathrm{CLCCP}}]$ With ${I^{\left(1\right)}}_{\mathrm{CLCCP}}=[{H_0^{\left(1\right)}}_{\mathrm{CLCCP}},{H_1^{\left(1\right)}}_{\mathrm{CLCCP}},...,{H_{15}^{\left(1\right)}}_{\mathrm{CLCCP}}],$ Distance then between these two vectors, namely chi span is from the following formulae discovery of employing:

${\mathrm{\χ}}^2({I^{\left(0\right)}}_{\mathrm{CLCCP}},{I^{\left(1\right)}}_{\mathrm{CLCCP}})=\underset{i=1}{\overset{K^{\′}}{\mathrm{\Σ}}}\frac{{({I^{\left(0\right)}}_{\mathrm{CLCCP}}\left(i\right)-{I^{\left(1\right)}}_{\mathrm{CLCCP}}\left(i\right))}^2}{({I^{\left(0\right)}}_{\mathrm{CLCCP}}\left(i\right)+{I^{\left(1\right)}}_{\mathrm{CLCCP}}\left(i\right)+\mathrm{eps})}$

Wherein I ⁽⁰⁾ _cLCCP(i) and I ⁽¹⁾ _cLCCPi () represents texture feature vector I respectively ⁽⁰⁾ _cLCCPand I ⁽¹⁾ _cLCCPi-th element, K' represents the length of texture, and eps is a fixed value, is positive number minimum in Matlab.

Embodiment 2: as shown in figs. 1-7, a kind of illumination face recognition method based on complete Local Convex diesinking, first carries out piecemeal to image; Then bilinear interpolation is carried out to each block image, make each pixel in image can build 8 symmetry directions, then to calculate in block image each pixel along local, 8 directions difference; The symbolic feature of this local difference of then encoding and amplitude characteristic; Each pixel of each image block is encoded, obtains the center pixel feature of each image block; The eigenmatrix of the following symbolic feature to each block image, amplitude characteristic, center pixel feature extracts histogram feature vector, connects the vectorial histogram feature vector obtaining each block image of histogram feature of this block image symbolic feature, amplitude characteristic, center pixel feature successively; The histogram feature vector finally connecting each block image obtains the histogram feature vector of this original image, this proper vector is sent into and classifies based on the nearest neighbor classifier of chi amount, identify the identity of original facial image.

The concrete steps of the described illumination face recognition method based on complete Local Convex diesinking are as follows:

Step1, first image is carried out piecemeal: image I ^(l)evenly be divided into the zero lap square of 4 × 4,16 pieces altogether, be expressed as

Step2, bilinear interpolation computing is carried out to each block image, make each pixel can build about point-symmetric 8 directions of this pixel, then calculate the local difference of each pixel along different directions, this local difference is decomposed into symbolic component and and amplitude portion;

As shown in Figure 2, pixel P ₁and P ₂between can increase pixel Q by interpolation ₁.Interpolation method as shown in Figure 4, wherein P ₁₁, P ₁₂, P ₂₁, P ₂₂be four neighbor pixels original in image, insert by interpolation method the pixel Q that makes new advances ₀.Interpolation formula is as follows:

$I_{R_1}\≈\frac{x_2-x}{x_2-x_1}{I}_{P_{11}}+\frac{x-x_1}{x_2-x_1}{I}_{P_{21}},$

$I_{R_2}\≈\frac{x_2-x}{x_2-x_1}{I}_{P_{12}}+\frac{x-x_1}{x_2-x_1}{I}_{P_{22}},$

$I_{Q_0}\≈\frac{y_2-y}{y_2-y_1}{I}_{R_1}+\frac{y-y_1}{y_2-y_1}{I}_{R_2};$

Wherein with represent R respectively ₁, R ₂with the pixel value of position, x ₁, x and x ₂represent pixel P respectively ₁₁, R ₁and P ₂₁the horizontal ordinate at place, y ₁, y and y ₂represent pixel P respectively ₁₁, Q ₀and P ₁₂the ordinate at place.Fig. 3 represents pixel X in original image ₀there is P around ₀, P ₁, P ₂, P ₃, P ₄, P ₅, P ₆and P ₇8 Neighbor Points, only can form four about pixel X ₀symmetry direction.Fig. 2 represents pixel X after interpolation ₀there is Q around ₀, Q ₁, Q ₂, Q ₃, Q ₄, Q ₅, Q ₆and Q ₇8 interpolation points, therefore pixel X after interpolation ₀one co-exist in 16 Neighbor Points around, 8 can be obtained about pixel X ₀symmetry direction.Owing to adding interpolation point, the resolution of image is enhanced;

Pixel X in image block ₀partial error along 8 directions is divided into with wherein i=0,1,2,3 and j=0,1,2,3;

Step3, symbolic component and amplitude portion carried out respectively to corresponding Local Convex diesinking coding, obtain symbolic feature CLCCP-S and the amplitude characteristic CLCCP-M of each block image, wherein pixel X ₀symbolic feature and the coding formula of amplitude characteristic be respectively:

$\mathrm{CLCCP}-S_{\mathrm{1,8}}{\left(X_0\right)}_D=\underset{i=0}{\overset{3}{\mathrm{\Σ}}}[f(I_{X_0}-0.5*(I_{P_i}+I_{P_{i+4}}))*2^{i*2}]+\underset{j=0}{\overset{3}{\mathrm{\Σ}}}[f(I_{X_0}-0.5*(I_{Q_j}+I_{Q_{j+4}}))*2^{j*2+1}]$

$\mathrm{CLCCP}-M_{\mathrm{1,8}}{\left(X_0\right)}_D=\underset{i=0}{\overset{3}{\mathrm{\Σ}}}[f(\mathrm{abs}(I_{X_0}-0.5*(I_{P_i}+I_{P_{i+4}}))-c)*2^{i*2}]+\underset{j=0}{\overset{3}{\mathrm{\Σ}}}[f(\mathrm{abs}(I_{X_0}-0.5*(I_{Q_j}+I_{Q_{j+4}}))-c)*2^{j*2+1}]$

Wherein, CLCCP-S _1,8(X ₀) _drepresent pixel X ₀the Local Convex concavity symbolic feature at place, CLCCP-M _1,8(X ₀) _drepresent pixel X ₀the Local Convex concavity amplitude characteristic at place, $f\left(x\right)=\left\{\begin{array}{cc}0,\mathrm{if}& x\≤\mathrm{threshold}\\ 1,\mathrm{if}& x>\mathrm{threshold}\end{array}\right.,$ represent X in image ₀the pixel value at place, with represent P in image _iand P _i+4the pixel value at place, with represent Q in image _jand Q _j+4the pixel value at place, abs () expression asks for signed magnitude arithmetic(al), CLCCP-S _1,8(X ₀) _dand CLCCP-M _1,8(X ₀) _dmiddle subscript " 1 " represents the pixel distance X that calculating convex-concave characteristic is used ₀distance be 1, namely yardstick is 1, subscript " 8 " represent calculate through pixel X ₀the convexo-concave characteristic in 8 directions, it is decimal system amount that subscript " D " represents, threshold is the threshold value pre-set;

Step4, to encode to each pixel of each image block, obtain the center pixel feature CLCCP-C of each image block, coding formula is: here c _irepresent the mean value of entire image, represent X in image ₀the pixel value at place, $f\left(x\right)=\left\{\begin{array}{cc}0,\mathrm{if}& x\≤\mathrm{threshold}\\ 1,\mathrm{if}& x>\mathrm{threshold}\end{array}\right.,$ CLCCP-C _1,8(X ₀) _dmiddle subscript " 1 " represents the pixel distance X that calculating convex-concave characteristic is used ₀distance be 1, namely yardstick is 1, subscript " 8 " represent calculate through pixel X ₀the convexo-concave characteristic in 8 directions, subscript " D " represents decimal system amount, and threshold is the threshold value pre-set;

Step5, through step Step2, Step3 and Step4, be extracted image block complete local convexo-concave characteristic, comprise and meet feature, amplitude characteristic and center pixel feature, work as image block in pixel X ₀when traveling through whole image block, obtain each block image the eigenmatrix of CLCCP-S, CLCCP-M, CLCCP-C, be respectively

The histogram feature vector of Step6, the next each image block of extraction three eigenmatrixes, image block three eigenmatrixes histogram feature vector be expressed as: connect this three histogram feature vectors successively, obtain image block histogram feature vector ${H_i^{\left(l\right)}}_{\mathrm{CLCCP}}=[{H_i^{\left(l\right)}}_{\mathrm{CLCCP}-S},{H_i^{\left(l\right)}}_{\mathrm{CLCCP}-M},{H_i^{\left(l\right)}}_{\mathrm{CLCCP}-C}],$ Herein subscript CLCCP represent complete Local Convex diesinking feature, it comprises symbolic feature CLCCP-S, amplitude characteristic CLCCP-M and center pixel point patterns CLCCP-C;

Step7, connect each image block histogram feature vector, the complete Local Convex diesinking histogram feature vector obtaining original image is: ${I^{\left(l\right)}}_{\mathrm{CLCCP}}=[{H_0^{\left(l\right)}}_{\mathrm{CLCCP}},{H_1^{\left(l\right)}}_{\mathrm{CLCCP}},...,{H_{15}^{\left(l\right)}}_{\mathrm{CLCCP}}];$

Step8, this proper vector send into classify based on the nearest neighbor classifier of chi amount, identify the identity of original facial image.

In described step Step8, when the nearest neighbor classifier based on chi amount is classified, first computer card side statistic; Set two width facial image I ⁽⁰⁾and I ⁽¹⁾complete Local Convex diesinking histogram feature vector be respectively: ${I^{\left(0\right)}}_{\mathrm{CLCCP}}=[{H_0^{\left(0\right)}}_{\mathrm{CLCCP}},{H_1^{\left(0\right)}}_{\mathrm{CLCCP}},...,{H_{15}^{\left(0\right)}}_{\mathrm{CLCCP}}]$ With ${I^{\left(1\right)}}_{\mathrm{CLCCP}}=[{H_0^{\left(1\right)}}_{\mathrm{CLCCP}},{H_1^{\left(1\right)}}_{\mathrm{CLCCP}},...,{H_{15}^{\left(1\right)}}_{\mathrm{CLCCP}}],$ Distance then between these two vectors, namely chi span is from the following formulae discovery of employing:

${\mathrm{\χ}}^2({I^{\left(0\right)}}_{\mathrm{CLCCP}},{I^{\left(1\right)}}_{\mathrm{CLCCP}})=\underset{i=1}{\overset{K^{\′}}{\mathrm{\Σ}}}\frac{{({I^{\left(0\right)}}_{\mathrm{CLCCP}}\left(i\right)-{I^{\left(1\right)}}_{\mathrm{CLCCP}}\left(i\right))}^2}{({I^{\left(0\right)}}_{\mathrm{CLCCP}}\left(i\right)+{I^{\left(1\right)}}_{\mathrm{CLCCP}}\left(i\right)+\mathrm{eps})}$

Wherein I ⁽⁰⁾ _cLCCP(i) and I ⁽¹⁾ _cLCCPi () represents texture feature vector I respectively ⁽⁰⁾ _cLCCPand I ⁽¹⁾ _cLCCPi-th element, K' represents the length of texture, and eps is a fixed value, is positive number minimum in Matlab.

In order to prove the beneficial effect of described method, proved by statistics this method and the discrimination of other related algorithms in illumination face database, the match cognization rate that adds up comparing with other algorithms;

First add up the discrimination of this method in illumination face database, and compare with related algorithm, draw corresponding recognition performance curve.The present embodiment adopts MATLAB software environment, in the present embodiment, threshold gets 0, in the present embodiment, face picture used is the illumination subset of the extended YaleB face database, this subset has 38 individuals, everyone takes 64 photos in different light situation, 2432 photos altogether, photo size is 64x64, is 64 samples pictures of a people in this database as shown in Figure 5.This database can in all face picture cut of the upper download of this database website (http://vision.ucsd.edu/ ~ leekc/ExtYaleDatabase/ExtYaleB.html).In this embodiment, correct recognition rata and cumulative matching properties (the cumulative match characteristic) curve of this method, local binary patterns, unified local binary patterns and complete local binary patterns four kinds of algorithms is calculated.Adopt nearest neighbor classifier to calculate discrimination, when calculating discrimination, training sample set is by everyone difference 1,2,3,4, and 5 compositions of sample, remaining image is used as test.Test sample book and all training samples compare, if with test sample book apart from the identity of minimum training sample and test sample book consistent, then think that it is correct for identifying.The sample number of all correct identification is correct recognition rata divided by all test sample book numbers.

Also calculate the cumulative matching properties curve of this method, local binary patterns, complete local binary patterns, unified local binary patterns in addition.Gallery picture library collection and Probe picture library collection is needed when calculating cumulative matching properties curve.Gallery picture library collection is by the extended Yale B database, everyone provides a pictures to form, and everyone other 63 pictures remaining form Probe picture library collection.Supposing Gallery picture library to concentrate number of pictures to be L, P be a length is the full null vector of L.Picture library Probe concentrates a pictures I and Gallery to concentrate all pictures to carry out distance coupling, obtains a distance vector D={d ₁, d ₂..., d _l, if Probe concentrates picture I and Gallery to concentrate the distance between common identity picture to be d, then d must be an element of vectorial D, if by arranging D vector from small to large, now suppose that d is arranged in D position l, then the element value on the l position of vectorial P adds 1.So picture is concentrated to repeat once Probe picture library, then by the length of each element of vectorial P divided by vectorial P, then " order (rank) 1 " discrimination is exactly first element value of vectorial P, " order (rank) 2 " discrimination is exactly second element value of vectorial P, the like.In the present embodiment, when calculating cumulative matching properties curve, from everyone photo, one is selected to form Gallery picture library collection at random, 63 formation Probe collection that everyone is left.Local binary patterns, unified local binary patterns, complete local binary patterns, the cumulative match curve of this method under this Gallery and Probe picture library collection are as shown in Figure 6;

As can be seen from Figure 6, the Performance comparision of local binary patterns and unified local binary patterns is close, but is all weaker than complete local binary patterns, and complete local binary patterns is substantially weaker than this method.Along with " order (rank) " be increased to concentrate number of pictures close to Gallery picture library time, several algorithm performance is close, but now few of engineering practice has been worth.

Also simulate the correct recognition rata of each algorithm in different training sample number situation in the present embodiment.We, by emulation repetition 5 times, calculate average correct recognition rata and standard deviation, and result are drawn in the figure 7; As can be seen from Figure 7, the performance of this method is better than other several algorithms greatly, when number of training is 5, by calculating local binary patterns based on the nearest neighbor classifier of chi amount, unifying the average recognition rate of local binary patterns, complete local binary patterns and this method be: 66.09%, 62.34%, 70.92%, 76.76%.Wherein this method will exceed 10.67% than local binary patterns algorithm discrimination, exceeds 14.42% than unified local binary patterns, exceeds 5.84% than complete local binary patterns, and this illustrates that this method is one illumination face recognition method very efficiently.

By reference to the accompanying drawings the specific embodiment of the present invention is explained in detail above, but the present invention is not limited to above-mentioned embodiment, in the ken that those of ordinary skill in the art possess, various change can also be made under the prerequisite not departing from present inventive concept.

Claims (3)

1. based on an illumination face recognition method for complete Local Convex diesinking, it is characterized in that: first piecemeal is carried out to image; Then bilinear interpolation is carried out to each block image, make each pixel in image can build 8 symmetry directions, then to calculate in block image each pixel along local, 8 directions difference; The symbolic feature of this local difference of then encoding and amplitude characteristic; Each pixel of each image block is encoded, obtains the center pixel feature of each image block; The eigenmatrix of the following symbolic feature to each block image, amplitude characteristic, center pixel feature extracts histogram feature vector, connects the vectorial histogram feature vector obtaining each block image of histogram feature of this block image symbolic feature, amplitude characteristic, center pixel feature successively; The histogram feature vector finally connecting each block image obtains the histogram feature vector of this original image, this proper vector is sent into and classifies based on the nearest neighbor classifier of chi amount, identify the identity of original facial image.

2. the illumination face recognition method based on complete Local Convex diesinking according to claim 1, is characterized in that: the concrete steps of the described illumination face recognition method based on complete Local Convex diesinking are as follows:

Step1, first image is carried out piecemeal: image I ^(l)evenly be divided into the zero lap square of 4 × 4,16 pieces altogether, be expressed as $I_i^{\left(l\right)}(i=\mathrm{0,1,2},...,15);$

Step2, bilinear interpolation computing is carried out to each block image, make each pixel can build about point-symmetric 8 directions of this pixel, then calculate the local difference of each pixel along different directions, this local difference is decomposed into symbolic component and and amplitude portion;

Step3, symbolic component and amplitude portion carried out respectively to corresponding Local Convex diesinking coding, obtain symbolic feature CLCCP-S and the amplitude characteristic CLCCP-M of each block image, wherein pixel X ₀symbolic feature and the coding formula of amplitude characteristic be respectively:

$\mathrm{CLCCP}-S_{\mathrm{1,8}}{\left(X_0\right)}_D=\underset{i=0}{\overset{3}{\mathrm{\Σ}}}[f(I_{X_0}-0.5*(I_{P_i}+I_{P_{i+4}}))*2^{i*2}]+\underset{j=0}{\overset{3}{\mathrm{\Σ}}}[f(I_{X_0}-0.5*(I_{Q_j}+I_{Q_{j+4}}))*2^{j*2+1}]$

$\mathrm{CLCCP}-M_{\mathrm{1,8}}{\left(X_0\right)}_D=\underset{i=0}{\overset{3}{\mathrm{\Σ}}}[f(\mathrm{abs}(I_{X_0}-0.5*(I_{P_i}+I_{P_{i+4}}))-c)*2^{i*2}]+\underset{j=0}{\overset{3}{\mathrm{\Σ}}}[f(\mathrm{abs}(I_{X_0}-0.5*(I_{Q_j}+I_{Q_{j+4}}))-c)*2^{j*2+1}]$

Wherein, CLCCP-S _1,8(X ₀) _drepresent pixel X ₀the Local Convex concavity symbolic feature at place, CLCCP-M _1,8(X ₀) _drepresent pixel X ₀the Local Convex concavity amplitude characteristic at place, $f\left(x\right)=\left\{\begin{array}{c}0,\mathrm{ifx}\≤\mathrm{threshold}\\ 1,\mathrm{ifx}>\mathrm{threshold}\end{array}\right.,$ represent X in image ₀the pixel value at place, with represent P in image _iand P _i+4the pixel value at place, with represent Q in image _jand Q _j+4the pixel value at place, abs () expression asks for signed magnitude arithmetic(al), CLCCP-S _1,8(X ₀) _dand CLCCP-M _1,8(X ₀) _dmiddle subscript " 1 " represents the pixel distance X that calculating convex-concave characteristic is used ₀distance be 1, namely yardstick is 1, subscript " 8 " represent calculate through pixel X ₀the convexo-concave characteristic in 8 directions, it is decimal system amount that subscript " D " represents, threshold is the threshold value pre-set;

Step4, to encode to each pixel of each image block, obtain the center pixel feature CLCCP-C of each image block, coding formula is: here c _irepresent the mean value of entire image, represent X in image ₀the pixel value at place, $f\left(x\right)=\left\{\begin{array}{c}0,\mathrm{ifx}\≤\mathrm{threshold}\\ 1,\mathrm{ifx}>\mathrm{threshold}\end{array}\right.,$ CLCCP-C _1,8(X ₀) _dmiddle subscript " 1 " represents the pixel distance X that calculating convex-concave characteristic is used ₀distance be 1, namely yardstick is 1, subscript " 8 " represent calculate through pixel X ₀the convexo-concave characteristic in 8 directions, subscript " D " represents decimal system amount, and threshold is the threshold value pre-set;

Step5, through step Step2, Step3 and Step4, be extracted image block complete local convexo-concave characteristic, comprise and meet feature, amplitude characteristic and center pixel feature, work as image block in pixel X ₀when traveling through whole image block, obtain each block image the eigenmatrix of CLCCP-S, CLCCP-M, CLCCP-C, be respectively

The histogram feature vector of Step6, the next each image block of extraction three eigenmatrixes, image block three eigenmatrixes histogram feature vector be expressed as: connect this three histogram feature vectors successively, obtain image block histogram feature vector ${H_i^{\left(i\right)}}_{\mathrm{CLCCP}}=[{H_i^{\left(l\right)}}_{\mathrm{CLCCP}-S},{H_i^{\left(l\right)}}_{\mathrm{CLCCP}-M},{H_i^{\left(l\right)}}_{\mathrm{CLCCP}-C}],$ Herein subscript CLCCP represent complete Local Convex diesinking feature, it comprises symbolic feature CLCCP-S, amplitude characteristic CLCCP-M and center pixel point patterns CLCCP-C;

Step7, connect each image block histogram feature vector, the complete Local Convex diesinking histogram feature vector obtaining original image is: ${I^{\left(l\right)}}_{\mathrm{CLCCP}}=[{H_0^{\left(l\right)}}_{\mathrm{CLCCP}},{H_1^{\left(l\right)}}_{\mathrm{CLCCP}},...,{H_{15}^{\left(L\right)}}_{\mathrm{CLCCP}}];$

Step8, this proper vector send into classify based on the nearest neighbor classifier of chi amount, identify the identity of original facial image.

3. the illumination face recognition method based on complete Local Convex diesinking according to claim 1, is characterized in that: in described step Step8, when the nearest neighbor classifier based on chi amount is classified, and first computer card side statistic; Set two width facial image I ⁽⁰⁾and I ⁽¹⁾complete Local Convex diesinking histogram feature vector be respectively: ${I^{\left(0\right)}}_{\mathrm{CLCCP}}={H_0^{\left(0\right)}}_{\mathrm{CLCCP}},{H_1^{\left(0\right)}}_{\mathrm{CLCCP}},...,{H_{15}^{\left(0\right)}}_{\mathrm{CLCCP}}]$ With ${I^{\left(1\right)}}_{\mathrm{CLCCP}}={H_0^{\left(1\right)}}_{\mathrm{CLCCP}},{H_1^{\left(1\right)}}_{\mathrm{CLCCP}},...,{H_{15}^{\left(1\right)}}_{\mathrm{CLCCP}}],$ Distance then between these two vectors, namely chi span is from the following formulae discovery of employing:

${\mathrm{\χ}}^2({I^{\left(0\right)}}_{\mathrm{CLCCP}},{I^{\left(1\right)}}_{\mathrm{CLCCP}})=\underset{i=1}{\overset{K^{\′}}{\mathrm{\Σ}}}\frac{{({i^{\left(0\right)}}_{\mathrm{CLCCP}}\left(i\right)-{I^{\left(1\right)}}_{\mathrm{CLCCP}}\left(i\right))}^2}{({I^{\left(0\right)}}_{\mathrm{CLCCP}}\left(i\right)+{I^{\left(1\right)}}_{\mathrm{CLCCP}}\left(i\right)+\mathrm{eps})}$

Wherein I ⁽⁰⁾ _cLCCP(i) and I ⁽¹⁾ _cLCCPi () represents texture feature vector I respectively ⁽⁰⁾ _cLCCPand I ⁽¹⁾ _cLCCPi-th element, K' represents the length of texture, and eps is a fixed value, is positive number minimum in Matlab.