CN105550657B - Improvement SIFT face feature extraction method based on key point - Google Patents
Improvement SIFT face feature extraction method based on key point Download PDFInfo
- Publication number
- CN105550657B CN105550657B CN201510977092.3A CN201510977092A CN105550657B CN 105550657 B CN105550657 B CN 105550657B CN 201510977092 A CN201510977092 A CN 201510977092A CN 105550657 B CN105550657 B CN 105550657B
- Authority
- CN
- China
- Prior art keywords
- face
- pixel
- crucial
- layer
- feature
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/168—Feature extraction; Face representation
- G06V40/171—Local features and components; Facial parts ; Occluding parts, e.g. glasses; Geometrical relationships
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/172—Classification, e.g. identification
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- General Health & Medical Sciences (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
Improvement SIFT face feature extraction method based on key point, this method use the method improving SIFT face characteristic and extracting based on key point.By five crucial pixels in locating human face, and this five key points are described using the direction histogram in SIFT method, to form the facial image feature vector of robust.The similarity score value between two face feature vectors is calculated in conjunction with bilinearity similarity function and mahalanobis distance.Two-value classification is carried out to similarity score value using KELM classifier, the higher a kind of face picture of score value, two face pictures are judged to from the same person, and the lower a kind of face picture of score value, two face pictures are judged to from different people.During carrying out recognition of face on the basis of face feature vector, bilinearity similarity function and mahalanobis distance are combined to calculate the similarity score value of two feature vectors, enhances the distinguishability between class.
Description
Technical field
The present invention relates to a kind of improvement SIFT (Scale invariant features transform) face feature extraction method based on key point,
Belong to field of face identification.
Background technique
Recognition of face is a kind of biological identification technology for carrying out identification based on facial feature information of people.With other lifes
Object feature is compared, and face characteristic has many advantages, such as natural sex, convenience and untouchable, make its security monitoring, authentication,
Human-computer interaction etc. has huge application prospect.Therefore, face recognition technology has researching value very much.In general,
Face recognition process is divided into two processes: face characteristic extracts and human face similarity degree score value calculates.Face characteristic extraction process
It is some key features formation face feature vector for extracting face picture, human face similarity degree score value calculating process is calculating two
Similarity between a face feature vector, the similarity the high, shows two face pictures more possibly from same
People, conversely, then more showing two face pictures from different people.In some cases, more concerned be that face characteristic mentions
Take part.
Existing face feature extraction method includes the mutation method etc. of LBP (local binary patterns) method and it, these
Local textural feature extracting method forms histogram vectors by carrying out block statistics to whole face picture, and by each piece
Histogram vectors cascade ultimately forms face feature vector.Since this method is to carry out Local textural feature to whole face to mention
It takes, therefore, it is bigger for being formed by feature vector dimension, and wherein contains some redundancies.In addition, this
Variation not robust of the mode for expression under complex environment or posture.
SIFT feature extracting method has been widely used in the identification of general object, its main thought is to find image to exist
Key point under different scale, and direction histogram is used to describe key point as the feature vector of image.However, SIFT method
When for facial image, the key point in face cannot be accurately positioned, because it is primarily adapted for use in higher contrast
The identification of general object, and similitude with higher between facial image.
Summary of the invention
The main object of the present invention is to provide a kind of improvement SIFT face feature extraction method based on key point.
It is different from traditional feature extracting method based on whole face, specific innovative point of the invention is to use base
In the method for key point improving SIFT face characteristic and extracting.By five crucial pixels in locating human face, and utilize
Direction histogram in SIFT method describes this five key points, to form the facial image feature vector of robust.It is described
Five crucial pixels are respectively the pixel in left eye middle position, the pixel in right eye middle position, supratip pixel,
The pixel of the pixel of the left corners of the mouth and the right corners of the mouth.
Technical solution of the present invention mainly includes following technology contents specifically:
1, using five crucial pixel (pictures in left eye middle position in three layer depth convolutional networks cascade locating human face
Vegetarian refreshments, the pixel in right eye middle position, supratip pixel, the pixel of the pixel of the left corners of the mouth and the right corners of the mouth).
2, it improves SIFT feature extracting method: replacing SIFT feature extracting method with five crucial pixels in face
The key point itself detected reduces characteristic dimension, rejects the redundancy in characteristic.
3, by the maps feature vectors of face into the subspace intra-personal, guarantee the different faces of the same person
There is invariance in class between picture.
4, the similarity calculated in conjunction with bilinearity similarity function and mahalanobis distance between two face feature vectors obtains
Score value.The the score value the high, shows two face pictures more may be from the same person, conversely, the more low then table of the score value
Bright two face pictures more may be from different people.
5, two-value classification, score value are carried out to similarity score value using KELM classifier (extreme learning machine based on core)
Higher one kind face picture, two face pictures are judged to from the same person, and the lower a kind of face of score value
Picture, two face pictures are judged to from different people.
Flow chart of the invention is as shown in Figure 1, implementing procedure is as follows:
Step 1 reads face picture, and utilizes five keys on three layer depth convolutional networks cascade locating human face's picture
Pixel (pixel in left eye middle position, the pixel in right eye middle position, supratip pixel, the pixel of the left corners of the mouth
The pixel of point and the right corners of the mouth).
The cascade of depth convolutional network used in the step includes three layers.First layer closes five using depth convolutional network
Key pixel is accurately positioned, and is reaffirmed using positioning result of the convolutional network to first layer for other two layers.In order to
The accuracy for guaranteeing positioning, in every layer the positioning result of each depth convolutional network be merged together take it is average as finally
Positioning result.Deep layer convolutional network includes four convolutional layers, pond layer and two full articulamentums, and initial layers obtain face picture
Global context information, since convolutional network is to five crucial pixels while to predict, so each key pixel
Between relative position convolutional network training while be also carried out coding, and then weaken expression shape change, illumination variation with
And it is influenced caused by other environmental factors.
Step 2, five crucial pixels for extracting in step 1, carry out feature description using SIFT method, extract people
The feature of face picture.
SIFT feature extracting method is a kind of method for detecting local feature, not only has scale invariability, has simultaneously
Rotational invariance.This method generally comprises Four processes: (1) constructing scale space, detect key point;(2) it rejects unstable
Key point;It (3) is key point assignment directioin parameter;(4) description of key point is generated.In general, SIFT feature extraction side
Method is suitable for the identification of the general object with higher contrast, and face picture has lower contrast and skirt response,
The key point in face picture cannot be accurately positioned due to SIFT feature extracting method, with the crucial pixel of positioning in step 1
Step (1) and step (2) in the method replacement SIFT feature extracting method of point.
Then, for five obtained in step 1 crucial pixels, some pixels in each crucial pixel field are taken
Point calculates gradient modulus value and the direction of each pixel.The coordinate for defining some pixel is P (x, y):
θ (x, y)=tan-1((P(x,y+1)-P(x,y-1))/(P(x+1,y)-P(x-1,y)))
Wherein, m (x, y) is the gradient modulus value of the pixel, and θ (x, y) is the gradient direction of the pixel.
According to the calculated result of above formula, the gradient direction of pixel in statistics with histogram field is utilized.Shadow is mutated to reduce
It rings, needs to carry out smoothly histogram with Gaussian function.So, the peak value of histogram represents crucial pixel field pixel
Gradient principal direction, namely the direction of crucial pixel.
In order to keep rotational invariance, reference axis is rotated to be to the direction of crucial pixel, then centered on key point,
The field window for taking 16 × 16 sizes calculates the histogram of 8 gradient directions in the grid of every 4 × 4 size, ultimately forms 4
The SIFT face feature vector of × 4 × 8=128 dimension.
Step 3, by maps feature vectors obtained in step 2 into the subspace intra-personal.
In this step, in order to weaken the influence of noise, it is necessary first to which feature vector obtained in step 2 is utilized PCA
Method (principal component analysis) carries out dimensionality reduction, forms eigenface.Its covariance matrix expression formula is as follows:
Wherein, n is face sample size, xiIndicate that face vector, m are the mean value of n face vector.Due to covariance square
Battle array describes the correlation between vector, therefore the feature vector of above-mentioned covariance matrix forms mapping matrix, according to mapping square
Battle array maps face image data, can form eigenface.Then, in order to guarantee the different face picture of the same person it
Between class in invariance, eigenface is mapped in the subspace intra-personal, covariance matrix expression formula is as follows:
Wherein, S indicates the face picture set of the same person, xiAnd xjIt indicates in the face picture set of the same person not
The face vector of two same face pictures.∧={ λ1..., λkAnd V={ v1..., vkRespectively indicate above formula covariance square
The preceding k characteristic value and feature vector of battle array.Similarly, preceding k feature vector forms mapping matrix, and features described above face data are led to
It crosses the mapping matrix to be mapped, to guarantee invariance in the class between the face picture of the same person.If CSIt is reversible
, then, eigenface is mapped to the subspace intra-personal with following formula to express:
Wherein, V is the mapping matrix formed by above-mentioned k feature vector,It is special by above-mentioned k
Diagonal matrix composed by value indicative,I.e. finally formed eigenmatrix.
Step 4, calculated using bilinearity similarity function and mahalanobis distance it is similar between two face feature vectors
Spend score value.
Mahalanobis distance has been widely used for field of face identification, but its recognition effect is not especially good, and grinding in recent years
Study carefully and shows that bilinearity similarity function has obtained good effect in picture similarity search field.Therefore, in this method,
The similarity score value between two face feature vectors is calculated in conjunction with bilinearity similarity function and mahalanobis distance, is expressed
Formula is as follows:
Wherein,It indicatesWithBilinearity similarity function between eigenmatrix,It indicates
WithMahalanobis distance between eigenmatrix.G and M is the matrix of k × k size, needs that suitable M and G is trained to protect as far as possible
Guarantee the maximum distinguishability between class while demonstrate,proving invariance in class.Therefore, by the subspace intra-personal similarity measure
The expression formula of study is defined as following form:
Wherein, S and D denotes like face to (i.e. two face pictures of the same person) and dissimilar face pair
The label of (i.e. two face pictures of different people).||.||FBe the F norm of matrix, the i.e. quadratic sum of matrix element absolute value again
Evolution.It acts on 2 norms for being similar to vector, therefore, expression formulaGuaranteeing invariance in class
Simultaneously effective prevent over-fitting.ξtThe loss function that experience differentiates, minimize the parameter can be enhanced between class can
Discrimination property.It can be seen that ξtIt ensure that the maximum distinguishability between class, andIt ensure that constant in class
Property, and positive number γ is used to coordinate to influence brought by the two expression formulas.For the inequality in above-mentioned expression formula, as a pair of of people
When face picture is from the same person, yij=1, and ξijValue it is smaller, thenValue can be as big as possible.And work as
When a pair of of face picture is from different people, yij=-1, and ξijValue it is smaller, thenValue can be as far as possible
It is small.Therefore, whenWhen being worth larger, then show a pair of of face picture from the same person, conversely, then showing a pair
Face is from different people.
Whether step 5 judges two pictures from the same person using KELM classifier.
ELM is the neural network for only including a hidden layer and an output layer.Its most significant feature is that it is implied
Layer parameter does not need to be debugged, but set at random, there is stronger generalization ability.Assuming that the hidden layer of ELM includes L
Node, then its output function is as follows:
Wherein, x ∈ Rd, y ∈ RC, β indicate hidden layer L node and output layer between weight, h (x) indicate L save
Relationship between point and input x, it is a nonlinear excitation function (such as sigmoid function), in fact, its effect is
D dimension data is mapped in L dimension data space, wiIndicate the connection weight of hidden layer i-th of node and input layer, biIndicate hidden
The deviation of i-th of node containing layer.
On the basis of ELM, a kind of ELM method based on kernel function, i.e. KELM method are had also been proposed.This method is to use core
Function hides the original excitation function H of ELM, to preferably improve the generalization ability of algorithm.For a certain sample xi,
It is as follows that output function expresses formula:
Y=[y1;...;yn]∈Rn×c
Wherein, C is a regression coefficient.
Using similarity score value obtained in step 4 as the input of KELM classifier, obtained output is if 1, then
Show two face pictures from the same person, if 0, then shows two face pictures from different people.
Compared with prior art, the present invention has the advantage that
Different from traditional feature extracting method based on whole face, specific innovative point of the invention is to use base
In the method for key point improving SIFT face characteristic and extracting.By the pixel in face left eye middle position, right eye middle position
Pixel, supratip pixel, five passes of the pixel of the pixel of the left corners of the mouth and the right corners of the mouth as whole face
Key pixel, and feature vector of the SIFT feature as face of this five key points is extracted, not only reducing intrinsic dimensionality
On the basis of improve operation efficiency, and weaken since illumination, expression shape change and other environmental factors etc. are to recognition effect
Influence.In addition, combining bilinearity similarity function during carrying out recognition of face on the basis of face feature vector
The similarity score value that two feature vectors are calculated with mahalanobis distance enhances the distinguishability between class, meanwhile, in order to consider class
Similarity invariance between interior, face feature vector has been mapped in the subspace intra-personal.It is tested by intersecting
Confirmation is tested, and the retrievable face recognition accuracy rate of the present invention is up to 80.56%.
Detailed description of the invention
Fig. 1 is specific flow chart of the present invention.
Specific embodiment
The basic procedure of improvement SIFT face feature extraction method based on key point of the invention is as shown in Figure 1, specific
The following steps are included:
1) data in face database are divided into 10 groups of carry out cross-validation experiments, wherein 9 groups of data are as training number
Test data is used as according to, remaining 1 group of data, and face picture in every group comprising 300 pairs from the same person and 300 pairs come from not
With the face picture of people.For every face picture, positioning five crucial pixels of three layer depth convolutional networks cascade are utilized
Coordinate position (pixel in left eye middle position, the pixel in right eye middle position, supratip pixel, the picture of the left corners of the mouth
The pixel of vegetarian refreshments and the right corners of the mouth).
It include three depth convolutional networks in the first layer of three layer depth convolutional network cascade structures.First depth
The input of convolutional network is whole face, exports the position of five crucial pixels;The input of second depth convolutional network is
The top half of face exports the position of eyes and nose shape key pixel;The input of third depth convolutional network is
The lower half portion of face exports the position of nose and mouth position key pixel.Finally, by these three depth convolutional networks
Output result is averagely obtained the final output of first layer.The cascade second layer of deep layer convolutional network and third layer take previous
Field in layer output result around crucial pixel is used as input, reaffirms to the coordinate position of crucial pixel,
It is the supplement to first layer output result.
2) 14 × 14 neighborhood is taken centered on each crucial pixel for five crucial pixels in step 1)
Window size calculates gradient modulus value and the direction of each pixel in the window.Then the histogram in statistical gradient direction.In order to
The influence for reducing mutation is smoothed histogram with the Gaussian function that parameter σ is 1.5 × 14.Finally with each key
The gradient direction distribution characteristic of neighborhood of pixel points pixel determines the direction of each crucial pixel.
Behind the direction that crucial pixel has been determined, reference axis is rotated to be to the direction of crucial pixel, to ensure to rotate not
Denaturation.Then 16 × 16 neighborhood window size is taken centered on crucial pixel, is then counted on the fritter of every 4 × 4 size
The histogram of gradients in 8 directions is calculated, finally, each feature vector forms description of 4 × 4 × 8=128 dimension.Due to the present invention
In the crucial pixel number of every face picture be 5, therefore, the feature vector of every face picture is that 5 × 128=640 is tieed up,
This not only greatly reduces intrinsic dimensionality, but also caused by weakening expression shape change, illumination variation and other environmental factors etc.
It influences, enhances the robustness of face characteristic.
3) feature vector obtained in step 2) is subjected to dimensionality reduction using PCA (principal component analysis) method, obtains first 400
Main variables form the eigenface of 400 dimensions.Then, in order to guarantee in the class between the different face picture of the same person not
Denaturation, eigenface is mapped in the subspace intra-personal, wherein the intrinsic dimensionality in the space intra-personal takes
300。
4) the M parameter and G parameter in bilinearity similarity function and mahalanobis distance are obtained by 9 groups of training data training.
The phase in training data and test data between each pair of face picture is calculated using the obtained M parameter of training process and G parameter
Like degree score value.
5) judge two pictures whether from the same person using KELM classifier.In KELM classifier, the present invention
Select Radial basis kernel function (RBF) as kernel function, regression coefficient C=1024.By two face pictures obtained in step 4)
Between input of the similarity score value as KELM classifier, obtained classifier output then shows two faces if 1
Picture is from the same person, if 0, then shows two face pictures from different people.
Claims (2)
1. the improvement SIFT face feature extraction method based on key point, it is characterised in that: the implementing procedure of this method is as follows,
Step 1 reads face picture, and utilizes five crucial pixels on three layer depth convolutional networks cascade locating human face's picture
Point;Five pixels are respectively the pixel in left eye middle position, the pixel in right eye middle position, supratip pixel
Point, the pixel of the pixel of the left corners of the mouth and the right corners of the mouth;
The cascade of depth convolutional network used in the step includes three layers;First layer is using depth convolutional network to five crucial pictures
Vegetarian refreshments is accurately positioned, and is reaffirmed using positioning result of the convolutional network to first layer for other two layers;In order to guarantee
The accuracy of positioning, in every layer the positioning result of each depth convolutional network be merged together take it is average as final positioning
As a result;Deep layer convolutional network includes four convolutional layers, pond layer and two full articulamentums, and initial layers obtain the overall situation of face picture
Contextual information, since convolutional network is to five crucial pixels while to predict, so between each key pixel
Relative position convolutional network training while be also carried out coding, and then weaken expression shape change, illumination variation and its
It is influenced caused by his environmental factor;
Step 2, five crucial pixels for extracting in step 1, carry out feature description using SIFT method, extract face figure
The feature of piece;
SIFT feature extracting method is a kind of method for detecting local feature, not only has scale invariability, while having rotation
Invariance;This method generally comprises Four processes: (1) constructing scale space, detect key point;(2) unstable key is rejected
Point;It (3) is key point assignment directioin parameter;(4) description of key point is generated;In general, SIFT feature extracting method is suitable
For the identification of the general object with higher contrast, and face picture has lower contrast and skirt response, due to
SIFT feature extracting method cannot be accurately positioned the key point in face picture, therefore with positioning crucial pixel in step 1
Method replaces the step (1) and step (2) in SIFT feature extracting method;
Then, for five obtained in step 1 crucial pixels, some pixels in each crucial neighborhood of pixel points are taken,
Calculate gradient modulus value and the direction of each pixel;The coordinate for defining some pixel is P (x, y):
θ (x, y)=tan-1((P(x,y+1)-P(x,y-1))/(P(x+1,y)-P(x-1,y)))
Wherein, m (x, y) is the gradient modulus value of the pixel, and θ (x, y) is the gradient direction of the pixel;
According to the calculated result of above formula, the gradient direction of pixel in statistics with histogram neighborhood is utilized;It influences, needs to reduce mutation
Histogram is carried out smoothly with Gaussian function;So, the peak value of histogram represents the ladder of crucial neighborhood of pixel points pixel
Spend principal direction, namely the direction of crucial pixel;
In order to keep rotational invariance, reference axis is rotated to be to the direction of crucial pixel, then centered on key point, takes 16
The neighborhood window of × 16 sizes calculates the histogram of 8 gradient directions in the grid of every 4 × 4 size, ultimately form 4 × 4 ×
The SIFT face feature vector of 8=128 dimension;
Step 3, by maps feature vectors obtained in step 2 into the subspace intra-personal;
In this step, in order to weaken the influence of noise, it is necessary first to which feature vector obtained in step 2 is utilized PCA method
(principal component analysis) carries out dimensionality reduction, forms eigenface;Its covariance matrix expression formula is as follows:
Wherein, n is face sample size, xiIndicate that face vector, m are the mean value of n face vector;Since covariance matrix is retouched
The correlation between vector is stated, therefore the feature vector of above-mentioned covariance matrix forms mapping matrix, according to mapping matrix pair
Face image data is mapped, and eigenface can be formed;Then, in order to guarantee between the different face picture of the same person
Eigenface is mapped in the subspace intra-personal by invariance in class, and covariance matrix expression formula is as follows:
Wherein, S indicates the face picture set of the same person, xiAnd xjIndicate different in the face picture set of the same person
The face vector of two face pictures;∧={ λ1..., λkAnd V=(v1..., vk) numerical value respectively indicate above formula covariance
The preceding k characteristic value and feature vector of matrix;Similarly, preceding k feature vector forms mapping matrix, by features described above face data
It is mapped by the mapping matrix, to guarantee invariance in the class between the face picture of the same person;If CsBeing can
Inverse, then, eigenface is mapped to the subspace intra-personal with following formula to express:
Wherein, V is the mapping matrix formed by above-mentioned k feature vector,It is by above-mentioned k characteristic value
Composed diagonal matrix,I.e. finally formed eigenmatrix;
Step 4, the similarity calculated between two face feature vectors using bilinearity similarity function and mahalanobis distance are obtained
Score value;
Mahalanobis distance has been widely used for recognition of face neighborhood, but its recognition effect is not especially good, and research table in recent years
Bright, bilinearity similarity function has obtained good effect in picture similarity search neighborhood;Therefore, in this method, in conjunction with
Bilinearity similarity function and mahalanobis distance calculate the similarity score value between two face feature vectors, and expression formula is such as
Under:
Wherein,It indicatesWithBilinearity similarity function between eigenmatrix,It indicatesWith
Mahalanobis distance between eigenmatrix;G and M is the matrix of k × k size, needs that suitable M and G is trained to guarantee class as far as possible
Guarantee the maximum distinguishability between class while interior invariance;Therefore, the subspace intra-personal similarity measure is learnt
Expression formula be defined as following form:
Wherein, S and D denotes like face to two face pictures of the i.e. same person and dissimilar face to i.e. different people
Two face pictures label;||·||FIt is the F norm of matrix, i.e. the quadratic sum of matrix element absolute value evolution again;It is made
With 2 norms for being similar to vector, therefore, expression formulaWhile guaranteeing invariance in class effectively
Prevent over-fitting;ξtIt is the loss function that experience differentiates, minimizing the parameter can be enhanced distinguishability between class;Thus
As it can be seen that ξtIt ensure that the maximum distinguishability between class, andIt ensure that invariance in class, and positive number γ
For coordinating influence brought by the two expression formulas;For the inequality in above-mentioned expression formula, when a pair of of face picture comes from
When the same person, yij=1, and ξijValue it is smaller, thenValue can be as big as possible;And when a pair of of face picture
When from different people, yij=-1, and ξijValue it is smaller, thenValue can be as small as possible;Therefore, whenWhen being worth larger, then show a pair of of face picture from the same person, conversely, then show a pair of of face from
Different people;
Whether step 5 judges two pictures from the same person using KELM classifier;
ELM is the neural network for only including a hidden layer and an output layer;Its most significant feature is its hidden layer ginseng
Number does not need to be debugged, but set at random, there is stronger generalization ability;Assuming that the hidden layer of ELM includes L section
Point, then its output function is as follows:
Wherein, x ∈ Rd, y ∈ Rc, β indicates the weight between the L node and output layer of hidden layer, h (x) indicate L node and
The relationship between x is inputted, it is a nonlinear excitation function, in fact, its effect is that d dimension data is mapped to L dimension
In data space, wiIndicate the connection weight of hidden layer i-th of node and input layer, biIndicate the inclined of i-th of node of hidden layer
Difference;
On the basis of ELM, a kind of ELM method based on kernel function, i.e. KELM method are had also been proposed;This method is to use kernel function
The original excitation function H of ELM is hidden, to preferably improve the generalization ability of algorithm;For a certain sample xi, output
Function expression is as follows:
Y=[y1;...;yn]∈Rn×c
Wherein, C is a regression coefficient;
Using similarity score value obtained in step 4 as the input of KELM classifier, obtained output then shows if 1
Two face pictures are from the same person, if 0, then show two face pictures from different people.
2. the improvement SIFT face feature extraction method according to claim 1 based on key point, it is characterised in that: we
Method specifically includes following steps,
1) data in face database are divided into 10 groups of carry out cross-validation experiments, wherein 9 groups of data as training data,
Remaining 1 group of data are used as test data, and face picture in every group comprising 300 pairs from the same person and 300 pairs are from different
The face picture of people;For every face picture, the seat of positioning five crucial pixels of three layer depth convolutional networks cascade is utilized
Cursor position, the pixel in left eye middle position, the pixel in right eye middle position, supratip pixel, the pixel of the left corners of the mouth
The pixel of point and the right corners of the mouth;
It include three depth convolutional networks in the first layer of three layer depth convolutional network cascade structures;First depth convolution
The input of network is whole face, exports the position of five crucial pixels;The input of second depth convolutional network is face
Top half, export eyes and nose shape key pixel position;The input of third depth convolutional network is face
Lower half portion, export nose and mouth position key pixel position;Finally, by the output of these three depth convolutional networks
As a result the final output of first layer is averagely obtained;The cascade second layer of deep layer convolutional network and third layer take preceding layer defeated
The neighborhood in result around crucial pixel reaffirms the coordinate position of crucial pixel, is pair as input out
The supplement of first layer output result;
2) 14 × 14 neighborhood window is taken centered on each crucial pixel for five crucial pixels in step 1)
Size calculates gradient modulus value and the direction of each pixel in the window;Then the histogram in statistical gradient direction;In order to reduce
The influence of mutation is smoothed histogram with the Gaussian function that parameter σ is 1.5 × 14;Finally with each crucial pixel
The gradient direction distribution characteristic of vertex neighborhood pixel determines the direction of each crucial pixel;
Behind the direction that crucial pixel has been determined, reference axis is rotated to be to the direction of crucial pixel, to ensure rotational invariance;
Then 16 × 16 neighborhood window size is taken centered on crucial pixel, and 8 are then calculated on the fritter of every 4 × 4 size
The histogram of gradients in direction, finally, each feature vector form description of 4 × 4 × 8=128 dimension;Due to every in the present invention
The crucial pixel number of face picture is 5, and therefore, the feature vector of every face picture is 5 × 128=640 dimension, this is not only
Influence caused by greatly reducing intrinsic dimensionality, and weaken expression shape change, illumination variation and other environmental factors etc. increases
The strong robustness of face characteristic;
3) by feature vector obtained in step 2) using PCA (principal component analysis) method carry out dimensionality reduction, obtain it is preceding 400 it is main at
Variation per minute forms the eigenface of 400 dimensions;Then, in order to guarantee invariance in the class between the different face picture of the same person,
Eigenface is mapped in the subspace intra-personal, wherein the intrinsic dimensionality in the space intra-personal takes 300;
4) the M parameter and G parameter in bilinearity similarity function and mahalanobis distance are obtained by 9 groups of training data training;It utilizes
The obtained M parameter of training process and G parameter calculate the similarity in training data and test data between each pair of face picture
Score value;
5) judge two pictures whether from the same person using KELM classifier;In KELM classifier, present invention selection
Radial basis kernel function (RBF) is used as kernel function, regression coefficient C=1024;It will be between two face pictures obtained in step 4)
Input of the similarity score value as KELM classifier, obtained classifier output then shows two face pictures if 1
From the same person, if 0, then show two face pictures from different people.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510977092.3A CN105550657B (en) | 2015-12-23 | 2015-12-23 | Improvement SIFT face feature extraction method based on key point |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510977092.3A CN105550657B (en) | 2015-12-23 | 2015-12-23 | Improvement SIFT face feature extraction method based on key point |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105550657A CN105550657A (en) | 2016-05-04 |
CN105550657B true CN105550657B (en) | 2019-01-29 |
Family
ID=55829840
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510977092.3A Active CN105550657B (en) | 2015-12-23 | 2015-12-23 | Improvement SIFT face feature extraction method based on key point |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105550657B (en) |
Families Citing this family (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106127103B (en) | 2016-06-12 | 2019-06-25 | 广州广电运通金融电子股份有限公司 | A kind of offline identity authentication method and device |
CN106528662A (en) * | 2016-10-20 | 2017-03-22 | 中山大学 | Quick retrieval method and system of vehicle image on the basis of feature geometric constraint |
CN106845331B (en) * | 2016-11-18 | 2018-03-13 | 深圳云天励飞技术有限公司 | A kind of image processing method and terminal |
CN106778531A (en) * | 2016-11-25 | 2017-05-31 | 北京小米移动软件有限公司 | Face detection method and device |
CN106778775B (en) * | 2016-11-29 | 2020-08-18 | 南京邮电大学 | Image classification method based on SIFT feature soft matching |
CN106534576A (en) * | 2016-12-07 | 2017-03-22 | 深圳市传奇数码有限公司 | Side face unlocking method applied to mobile phone |
CN107169420B (en) * | 2017-04-19 | 2020-04-21 | 南京邮电大学 | Eyeball center positioning method based on image gradient |
CN107945219B (en) * | 2017-11-23 | 2019-12-03 | 翔创科技(北京)有限公司 | Face image alignment schemes, computer program, storage medium and electronic equipment |
CN108038476B (en) * | 2018-01-03 | 2019-10-11 | 东北大学 | A kind of facial expression recognition feature extracting method based on edge detection and SIFT |
CN110751005B (en) * | 2018-07-23 | 2022-09-30 | 合肥工业大学 | Pedestrian detection method integrating depth perception features and kernel extreme learning machine |
CN108921140A (en) * | 2018-08-07 | 2018-11-30 | 安徽云森物联网科技有限公司 | Pedestrian's recognition methods again |
CN109344760A (en) * | 2018-09-26 | 2019-02-15 | 江西师范大学 | A kind of construction method of natural scene human face expression data collection |
CN111008589B (en) * | 2019-12-02 | 2024-04-09 | 杭州网易云音乐科技有限公司 | Face key point detection method, medium, device and computing equipment |
CN114936597B (en) * | 2022-05-20 | 2023-04-07 | 电子科技大学 | Method for extracting space true and false target characteristics of local information enhancer |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20100073749A (en) * | 2008-12-23 | 2010-07-01 | 고려대학교 산학협력단 | Apparatus and method for extracting feature point based on sift, and face recognition system using thereof |
CN103729625A (en) * | 2013-12-31 | 2014-04-16 | 青岛高校信息产业有限公司 | Face identification method |
CN104036524A (en) * | 2014-06-18 | 2014-09-10 | 哈尔滨工程大学 | Fast target tracking method with improved SIFT algorithm |
CN104794441A (en) * | 2015-04-15 | 2015-07-22 | 重庆邮电大学 | Human face feature extracting method based on active shape model and POEM (patterns of oriented edge magnituedes) texture model in complicated background |
CN105095857A (en) * | 2015-06-26 | 2015-11-25 | 上海交通大学 | Face data enhancement method based on key point disturbance technology |
-
2015
- 2015-12-23 CN CN201510977092.3A patent/CN105550657B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20100073749A (en) * | 2008-12-23 | 2010-07-01 | 고려대학교 산학협력단 | Apparatus and method for extracting feature point based on sift, and face recognition system using thereof |
CN103729625A (en) * | 2013-12-31 | 2014-04-16 | 青岛高校信息产业有限公司 | Face identification method |
CN104036524A (en) * | 2014-06-18 | 2014-09-10 | 哈尔滨工程大学 | Fast target tracking method with improved SIFT algorithm |
CN104794441A (en) * | 2015-04-15 | 2015-07-22 | 重庆邮电大学 | Human face feature extracting method based on active shape model and POEM (patterns of oriented edge magnituedes) texture model in complicated background |
CN105095857A (en) * | 2015-06-26 | 2015-11-25 | 上海交通大学 | Face data enhancement method based on key point disturbance technology |
Also Published As
Publication number | Publication date |
---|---|
CN105550657A (en) | 2016-05-04 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105550657B (en) | Improvement SIFT face feature extraction method based on key point | |
Jin et al. | Face detection using improved LBP under Bayesian framework | |
CN102938065B (en) | Face feature extraction method and face identification method based on large-scale image data | |
Barnouti et al. | Face recognition: A literature review | |
CN110532920A (en) | Smallest number data set face identification method based on FaceNet method | |
Setiowati et al. | A review of optimization method in face recognition: Comparison deep learning and non-deep learning methods | |
JP5153434B2 (en) | Information processing apparatus and information processing method | |
CN107392105B (en) | Expression recognition method based on reverse collaborative salient region features | |
CN110826408A (en) | Face recognition method by regional feature extraction | |
Sikder et al. | Intelligent face detection and recognition system | |
Essa et al. | Face recognition based on modular histogram of oriented directional features | |
Zhang et al. | 3d face recognition using multi-level multi-feature fusion | |
Lamsal et al. | Effects of the Unscented Kalman filter process for high performance face detector | |
Ghalleb et al. | Face recognition improvement using soft biometrics | |
Wang et al. | Robust face recognition from 2D and 3D images using structural Hausdorff distance | |
Reddy et al. | A novel face recognition system by the combination of multiple feature descriptors. | |
Bindu et al. | Kernel-based scale-invariant feature transform and spherical SVM classifier for face recognition | |
Vairavan et al. | An Efficient Age Estimation System with Facial Makeover Images Based on Key Points Selection. | |
Kumar et al. | Skin based occlusion detection and face recognition using machine learning techniques | |
Venkatramaphanikumar et al. | Face Recognition with Modular Two Dimensional PCA under Uncontrolled Illumination Variations | |
Dehghani et al. | Face detection speed improvement using bitmap-based Histogram of Oriented gradien | |
Hafez et al. | 3D face recognition based on normal map features using selected Gabor filters and linear discriminant analysis | |
Al-Taie et al. | The effect of distance similarity measures on the performance of face, ear and palm biometric systems | |
Qadir et al. | Black hole algorithm along edge detector and circular hough transform based iris projection with biometric identification systems | |
Upraity et al. | Efficient face recognition using morphological operations |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |