CN103632149A - Face recognition method based on image feature analysis - Google Patents

Abstract

The invention discloses a face recognition method based on image feature analysis. The method comprises the steps of 1, processing the image of a face to be recognized, extracting scale-invariant feature transform (SIFT) feature point vector, taking the SIFT feature point vector as basis of recognition and judgment, and carrying out rotating optimization on the extracted SIFT feature point vector; 2, reducing the dimensionality of the extracted SIFT feature point vector; 3, selecting the SIFT feature points of the image, and matching the feature points. After the method based on feature analysis and matching is adopted, the dimensionality of the feature vector can be reduced, the image matching effect is not influenced when the number of the feature points is reduced, the speed and the accuracy of face recognition are improved, the robustness of recognition can be improved, and the real-time online running speed and the matching accuracy of the whole method are improved.

Description

A kind of face identification method based on image characteristic analysis

Technical field

The present invention relates to a kind of face identification method, particularly relate to a kind of face identification method based on image characteristic analysis.

Background technology

In display model structure and knowledge method for distinguishing, wherein important class methods are mainly divided into overall outward appearance and two kinds of methods of local appearance identification, and its core embodies the extraction of the global characteristics and the local feature that are called optical imaging.In recognition of face, in the identification fields such as handwritten Kanji recognition, feature extraction is one of most important link.The eighties in last century, Movarac, these two scientists of Hannah have proposed an algorithm about angle point, then according to this algorithm Stephens, Harris has been converted into this initial idea the speciality problem of two character numerical values of structure tensor after 8 years, also can be called second-order matrix problem.To two of second-order matrix character numerical values, give the different manifestations producing after various combination specially afterwards, the scientists such as Rohr, Kanade, Tomasi, Triggs and Shi have launched to study and sum up the various classification of Corner detector according to this mode of thinking.Wherein Kenney and Triggs etc. have mentioned the tolerance of broad sense angle point.Kenney has announced the method for angle point axiomatization in 2005.Until nineteen ninety-five Zhang Zhengyou has proposed to utilize near the image area information of point of interest to mate this saying, the initial form of ability You Liao local feature description.Schmid in 2002 and the Mikolajczyk further affine unchangeability of research in the theoretical foundation of Lindeberg, and the method for this problem that is resolved in theory.2004, Lowe, according to the graphical rule Choice Theory of Lindeberg, formally proposed SIFT descriptor for the first time, and this Feature Descriptor performance is good, and the abundant outstanding property that it has makes it have unprecedented Research Prospects.

At present, common face identification method mainly contains global characteristics method, the recognition methods based on algebraic characteristic and the recognition methods based on geometric properties.Yet global characteristics generally can only carry out rough coupling, and local feature can provide meticulousr confirmation; Recognition methods based on algebraic characteristic becomes proper vector face representation, form a vector subspace that is called eigenface, in space, people's face is carried out to projection, then obtain each corresponding characteristic coordinates coefficient, this group coordinate coefficient has reflected the position of people's face, thereby the position of finding people's face to exist that can be highly stable reaches identifying purpose, but because eigenface is easily subject to the impact of extraneous factor, as: illumination, shooting angle etc., and cause the effect of identification to decline; Recognition methods based on geometric properties becomes a geometric properties vector by face representation, and the classifier design thought of binding hierarchy cluster is identified.Based on geometric properties vector, be that certain that people's face shape and geometry produce is before related to that basis produces, its composition generally includes the Euclidean distance of specifying two points in face, angle, curvature etc., this method intuitively, more adept to searching of face organ's profile and unique point, but tend to take a part for the whole, by the characteristic phenomenon of minority, go to replace the picture feature of whole face, ignore facial detail problem, cause the not high of identification accuracy.

Summary of the invention

The deficiency existing for overcoming above-mentioned prior art, the present invention's object is to provide a kind of face identification method based on image characteristic analysis, it is by adopting based on signature analysis and the method for mating, reduced the dimension of proper vector, reduce the quantity of unique point and do not affected image matching effect, reinforcement, to the speed of recognition of face and accuracy, improves the robustness of identification, has improved the real-time online travelling speed of whole method and the accuracy of coupling.

For reaching above-mentioned and other object, the present invention proposes a kind of face identification method based on image characteristic analysis, comprises the steps:

Step 1, knows others face to needs and carries out image processing, extracts SIFT unique point vector as the foundation of identification judgement, the SIFT unique point vector extracting is rotated to optimization simultaneously;

Step 2, the dimension of the SIFT unique point vector that reduction is extracted;

Step 3, chooses the SIFT unique point of image and carries out Feature Points Matching.

Further, a kind of face identification method based on image characteristic analysis as claimed in claim 1, is characterized in that, step 1 also comprises the steps:

Step 1.1, generates difference of Gaussian metric space;

Step 1.2, detects the extreme point of this metric space;

Step 1.3 accurately to determine position and the yardstick of key point, is removed key point and the unsettled skirt response point of low contrast by the three-dimensional quadratic function of matching simultaneously;

Step 1.4, utilizes the gradient direction distribution characteristic of key point neighborhood territory pixel, the direction parameter that the gradient direction of selecting amplitude maximum is each key point, and by this direction rotation the forward to Y-axis, make operator possess rotational invariance;

Step 1.5, generates SIFT unique point vector.

Further, in step 1.1, according to the Gaussian difference pyrene of different scale and image convolution institute this difference of Gaussian metric space of varied configurations.

Further, in step 1.2, whole sampled points are adjacent a little and are compared, obtain the magnitude relationship of the consecutive point in its yardstick, two regions of image, middle check point and its compare with 8 consecutive point on metric space and 18 points on neighbouring yardstick, in the difference of Gaussian image of image pyramid different layers, detect different scale extreme point.

Further, in step 2, by choosing the 2x2 sub regions centered by unique point, every sub regions is got 4 gradient directions, makes the dimension of this SIFT unique point vector be reduced to 16 dimensions by 128 dimensions.

Further, in step 3, the coupling of unique point adopt to calculate that two unique point Euclidean distances decide, and when this distance is less than setting threshold, thinks that these two unique points mate.

Further, this setting threshold is the lowest distance value of coupling.

Further, in step 3, while choosing the SIFT unique point of image, preferentially choose the SIFT unique point with obvious characteristic.

Further, in step 3, by calculating the amplitude size of comparative feature point gradient, from gradient direction ratio concentrate rather than homodisperse unique point, choose N the SIFT unique point of putting as this image.

Further, N choose by total feature, count 25%～30% carry out value.

Compared with prior art, a kind of face identification method based on image characteristic analysis of the present invention by taking based on feature interest points matching method in face recognition algorithms, reduce significantly the dimension of SIFT proper vector simultaneously, make SIFT proper vector from 128 dimensions/become, 16 dimensions/point, significantly reduce the SIFT unique point quantity of image, choose the good a part of unique point of Grad, save storage resources, improve computing time, for real-time online, calculating recognition of face etc. requires strict task, the present invention to have larger advantage.Gradient vector is after over-rotation is processed in addition, and the impact that image is rotated has produced good robustness, has improved arithmetic speed and the accuracy of real-time online.

Accompanying drawing explanation

Fig. 1 is the flow chart of steps of a kind of face identification method based on image characteristic analysis of the present invention;

Fig. 2 is the SIFT feature point extraction schematic diagram based on people's face in preferred embodiment of the present invention

Fig. 3 is the schematic diagram that in the present invention, adjacent this image subtraction of yardstick two panel heights obtains DOG metric space image;

Fig. 4 is key point proper vector schematic diagram in the present invention;

Fig. 5 is the recognition result figure based on features of human face images in preferred embodiment of the present invention.

Embodiment

Below, by specific instantiation accompanying drawings embodiments of the present invention, those skilled in the art can understand other advantage of the present invention and effect easily by content disclosed in the present specification.The present invention also can be implemented or be applied by other different instantiation, and the every details in this instructions also can be based on different viewpoints and application, carries out various modifications and change not deviating under spirit of the present invention.

Fig. 1 is the flow chart of steps of a kind of face identification method based on image characteristic analysis of the present invention.As shown in Figure 1, a kind of face identification method based on image characteristic analysis of the present invention, comprises the steps:

Step 101, knows others face to needs and carries out image processing, extracts SIFT unique point vector, and the foundation as identification judgement is rotated optimization to SIFT unique point vector, to improve the robustness of coupling simultaneously.

Fig. 2 is the SIFT feature point extraction schematic diagram based on people's face in preferred embodiment of the present invention.In the present invention, SIFT feature operator is that Lowe proposes according to the graphical rule Choice Theory of Lindeberg, the local feature of its Description Image, rotation, yardstick convergent-divergent, brightness are changed and maintained the invariance, visual angle change, affined transformation, noise are also kept to stability to a certain degree, its uniqueness, volume, high speed and extensibility are all fine simultaneously, the abundant outstanding property that it has makes it have unprecedented Research Prospects, and SIFT proper vector is the mathematical description to SIFT operator.

Particularly, in step 101, the generation of SIFT unique point vector also comprises the steps:

(1) generate difference of Gaussian metric space.

Utilize Gaussian convolution to examine the Analysis On Multi-scale Features that existing change of scale carrys out simulated image data, the metric space that defines a secondary two dimensional image is:

L(x,y,σ)＝G(x,y,σ)*I(x,y)

Wherein G (x, y, σ) is changeable scale Gaussian function,

(x, y) is volume coordinate, and σ is yardstick coordinate.The size of σ determines the level and smooth degree of image, the general picture feature of large scale correspondence image, and the minutia of small scale correspondence image, the large corresponding coarse scale of σ value (low resolution), otherwise, corresponding fine dimension (high resolving power).

According to the varied configurations difference of Gaussian metric space (DOG scale-space) of the Gaussian difference pyrene of different scale and image convolution, to improve the ability of the invariant feature point that image detected in metric space:

D(x,y,σ)＝(G(x,y,kσ)-G(x,y,σ))*I(x,y)＝L(x,y,kσ)-L(x,y,σ)。

(2) detect yardstick spatial extrema point.

In order to find the extreme point of metric space, whole sampled points are adjacent a little and are compared, obtain the magnitude relationship of the consecutive point in its yardstick, two regions of image, middle check point and its compare totally with 8 consecutive point on metric space and 9+9=18 point on neighbouring yardstick (26 points), in the difference of Gaussian image of image pyramid different layers, detect different scale extreme point, as shown in Figure 3.

(3) by the three-dimensional quadratic function of matching accurately to determine position and the yardstick of key point, remove key point and the unsettled skirt response point (because DoG operator can produce stronger skirt response) of low contrast, to strengthen coupling stability, to improve noise resisting ability simultaneously.

First to space scale function:

$D(x,y,\mathrm{\σ})=D(x,y,\mathrm{\σ})+\frac{{\∂D}^T}{\∂x}x+\frac{1}{2}{x}^T\frac{{\∂}^{2}D}{{\∂x}^2}x---\left(1\right)$

Differentiate, and to make it be 0, obtains accurate position

Secondly, in the unique point having detected, remove unique point and the unsettled skirt response point of low contrast.Remove the point of low contrast: formula (2) substitution formula (1), only getting first two can obtain:

$D\left(\hat{x}\right)=D(x,y,\mathrm{\σ})+\frac{1}{2}\frac{{\∂D}^T}{\∂x}\hat{x}$

If this unique point just remains, otherwise abandons.

The removal of skirt response a: extreme value that defines bad difference of Gaussian has larger principal curvatures in the place across edge, and has less principal curvatures in the direction of vertical edge.Principal curvatures is obtained by the Hessian matrix H of 2 * 2:

$H=\left[\begin{array}{cc}{D}_{\mathrm{xx}}& D_{\mathrm{xy}}\\ D_{\mathrm{xy}}& D_{\mathrm{yy}}\end{array}\right]$

Derivative is obtained by the adjacent poor estimation of sampled point.

The privilege value of the principal curvatures of D and H is directly proportional, and making α is eigenvalue of maximum, and β is minimal eigenvalue,

Tr(H)＝D _xx+D _yy＝α+β,

Det(H)＝D _xxD _yy-(D _xy) ²＝αβ

Make α=r β,

$\frac{\mathrm{Tr}{\left(H\right)}^2}{\mathrm{Det}\left(H\right)}=\frac{{(\mathrm{\α}+\mathrm{\β})}^2}{\mathrm{\α\β}}=\frac{{(\mathrm{r\β}+\mathrm{\β})}^2}{r{\mathrm{\β}}^2}=\frac{{(r+1)}^2}{r}$

(4) utilize the gradient direction distribution characteristic of key point neighborhood territory pixel, the direction parameter that the gradient direction of selecting amplitude maximum is each key point, and by this direction rotation the forward to Y-axis, make operator possess rotational invariance, the impact that image rotated produces robustness.

(5) generate SIFT unique point vector.

Step 102, the dimension of reduction unique point vector:

Because SIFT unique point vector dimension is 128,4x4x8 namely, the gradient direction of every sub regions forms (at rectangular coordinate, fastening interval 45 degree) by 8 directions, and such subregion centered by unique point total 4x4=16.We only get the 2x2 sub regions centered by unique point, and same every sub regions is got 4 gradient directions, and the unique point vector drawing so thus has just become 2x2x4=16 dimension.Reduce vectorial dimension, but do not reduced the quantity of unique point, therefore there will not be the phenomenon that can mate without unique point above-mentioned.Owing to having reduced the quantity of dimension, there will be some false match points, the probability of correct coupling declines to some extent, but does not affect the matching effect of image, in the experimental result of this point below, can see.Reduce vectorial dimension just the discriminating function aspect unusual trickle distinguishing reduced.

What at this, need special proposition is, 4 gradient directions getting in the every sub regions proposing in the present invention, be not that all gradient directions of this subregion finally merge formation, the gradient direction of this subregion is still divided into 8 directions, the amplitude of gradient is through standardization processing, as divided by a certain constant or the amplitude of greatest gradient, obtain the normalized one group of gradient vector of amplitude, to possess the robustness that image light and shade is changed.

In fact in the SIFT of standard algorithm, when calculating the gradient vector of subregion, constraint (referring to Fig. 4) due to the dimensional Gaussian weighting function centered by unique point, make originally to reduce away from the subregion gradient magnitude of unique point (center), its weights that participate in unique point gradient calculation have also reduced accordingly, thereby decentering is put nearest subregion leading role in the calculating of SIFT proper vector.

Step 103, the quantity of minimizing unique point, carries out Feature Points Matching.

The optimization of proper vector dimension is to realize in the SIFT of image production process, next in the process of Feature Points Matching, is optimized.Because a lot of images have a large amount of unique points that can detect, especially all the more so when the noise pollution of image is larger, and these unique points have a big chunk can not correctly to mate in matching process, even there are some false match points, correct coupling is caused to interference.It is necessary suitably reducing some unnecessary unique points.

Two characteristics of correspondence are put vectorial matching process and are adopted its Euclid of calculating (Euclidean) distance to decide, and when this distance is in certain thresholding, just can think that mate at 2.We are made as threshold value this coupling lowest distance value, when this threshold value more hour, the point that may mate is just fewer, but the accuracy of coupling is higher.In fact, have in time domain and have 3～5 correct match points just can complete the task of identification, realize so the needed unique point quantity of correct coupling just less.Therefore we are when calculating the unique point of piece image, preferentially choose those points with obvious characteristic, amplitude (Magnitude) size that compares its gradient by calculating, gradient direction is relatively concentrated (mono-Orientation) rather than dispersed, choose the point of some (N) as the SIFT unique point of this width image, what the choosing of N also can be counted by total feature 25%～30% carrys out value, generally for 50 left and right that are chosen at of the image N of 320x240.Choose few unique point, improve the requirement (reduction threshold value) of Feature Points Matching, do not reduce images match rate as far as possible.Fig. 5 is the identification schematic diagram based on features of human face images in preferred embodiment of the present invention.

Visible, by above-mentioned optimization, it is original 1/8th that dimension based on SIFT proper vector reduces to, and the data of unique point reduce to 25% original left and right, for the storage resources of unique point, requires, or reduces coupling and all played vital role computing time.

In sum, a kind of face identification method based on image characteristic analysis of the present invention by taking based on feature interest points matching method in face recognition algorithms, reduce significantly the dimension of SIFT proper vector simultaneously, make SIFT proper vector from 128 dimensions/become, 16 dimensions/point, significantly reduce the SIFT unique point quantity of image, choose the good a part of unique point of Grad, save storage resources, improve computing time, for real-time online, calculating recognition of face etc. requires strict task, the present invention to have larger advantage.Gradient vector is after over-rotation is processed in addition, and the impact that image is rotated has produced good robustness, has improved arithmetic speed and the accuracy of real-time online.

Above-described embodiment is illustrative principle of the present invention and effect thereof only, but not for limiting the present invention.Any those skilled in the art all can, under spirit of the present invention and category, modify and change above-described embodiment.Therefore, the scope of the present invention, should be as listed in claims.

Claims (10)

1. the face identification method based on image characteristic analysis, comprises the steps:

Step 1, knows others face to needs and carries out image processing, extracts SIFT unique point vector as the foundation of identification judgement, the SIFT unique point vector extracting is rotated to optimization simultaneously;

Step 2, the dimension of the SIFT unique point vector that reduction is extracted;

Step 3, chooses the SIFT unique point of image and carries out Feature Points Matching.

2. a kind of face identification method based on image characteristic analysis as claimed in claim 1, is characterized in that, step 1 also comprises the steps:

Step 1.1, generates difference of Gaussian metric space;

Step 1.2, detects the extreme point of this metric space;

Step 1.3 accurately to determine position and the yardstick of key point, is removed key point and the unsettled skirt response point of low contrast by the three-dimensional quadratic function of matching simultaneously;

Step 1.4, utilizes the gradient direction distribution characteristic of key point neighborhood territory pixel, the direction parameter that the gradient direction of selecting amplitude maximum is each key point, and by this direction rotation the forward to Y-axis, make operator possess rotational invariance;

Step 1.5, generates SIFT unique point vector.

3. a kind of face identification method based on image characteristic analysis as claimed in claim 2, is characterized in that: in step 1.1, according to this difference of Gaussian metric space of the varied configurations of the Gaussian difference pyrene of different scale and image convolution.

4. a kind of face identification method based on image characteristic analysis as claimed in claim 3, it is characterized in that: in step 1.2, whole sampled points are adjacent a little and are compared, obtain the magnitude relationship of the consecutive point in its yardstick, two regions of image, middle check point and its compare with 8 consecutive point on metric space and 18 points on neighbouring yardstick, in the difference of Gaussian image of image pyramid different layers, detect different scale extreme point.

5. a kind of face identification method based on image characteristic analysis as claimed in claim 4, it is characterized in that: in step 2, by choosing the 2x2 sub regions centered by unique point, every sub regions is got 4 gradient directions, makes the dimension of this SIFT unique point vector be reduced to 16 dimensions by 128 dimensions.

6. a kind of face identification method based on image characteristic analysis as claimed in claim 5, it is characterized in that: in step 3, the coupling of unique point adopts calculates that two unique point Euclidean distances decide, and when this distance is less than setting threshold, thinks that these two unique points mate.

7. a kind of face identification method based on image characteristic analysis as claimed in claim 6, is characterized in that: this setting threshold is the lowest distance value of coupling.

8. a kind of face identification method based on image characteristic analysis as claimed in claim 6, is characterized in that: in step 3, while choosing the SIFT unique point of image, preferentially choose the SIFT unique point with obvious characteristic.

9. a kind of face identification method based on image characteristic analysis as claimed in claim 8, it is characterized in that: in step 3, by calculating the amplitude size of comparative feature point gradient, from gradient direction ratio concentrate rather than homodisperse unique point, choose N the SIFT unique point of putting as this image.

10. a kind of face identification method based on image characteristic analysis as claimed in claim 8, is characterized in that: what the choosing of N counted by total feature 25%～30% carrys out value.

Images