CN104268584A - Human face detection method based on hierarchical filtration - Google Patents

Abstract

The invention provides a human face detection method based on hierarchical filtration. A traditional human face detection method can not consider the detection speed and the accuracy at the same time. The human face detection method includes the following steps that (1) feature values of different scales are generated through an image pyramid; (2) Haar-like matrix features are trained through a dual-threshold FloatBoost, so that a strong classifier with the low false drop rate and the low level is generated and used for rapidly and accurately extracting human face candidate regions of multiple scales; (3) LBP textural features and LGP textural features are fused; (4) the fused features are trained through the dual-threshold FloatBoost, so that a cascade classifier with the high detection rate is generated and used for accurately locating the human faces of different scales in the candidate regions of multiple scales; (5) detection results of the human faces of the different scales are displayed. According to the method, the detection rate is increased, meanwhile, the false drop rate in target regions is reduced, and human face detection time is shortened.

Description

A kind of method for detecting human face based on layered filtration

Technical field

The invention belongs to field of image detection, particularly a kind of method for detecting human face.

Background technology

Method for detecting human face is different, substantially can be divided into: based on the method for color characteristic, Knowledge based engineering method, Statistics-Based Method and the method based on probability.Because face is nonrigid, have the change of high-freedom degree, explicit description face characteristic is more difficult, and therefore, Statistics-Based Method is more paid attention to, as SVM, Adaboost and neural network detection algorithm etc.Traditional Adaboost detection algorithm can not take into account speed and accuracy rate: the Face datection based on matrix character can detect face fast, but its verification and measurement ratio is not high; Face datection based on complex characteristic can accurate locating human face, but detection time is very long.And traditional detection method all progressively detects based on single pixel, the plenty of time is placed on non-face region, causes detection time long, often also can not get testing result accurately.Accurately and fast method for detecting human face is more and more subject to the attention of researcher.

Fusion Features is merged multiple feature, generates a new feature.New feature has possessed the advantage of multiple merged feature, but its complexity does not increase, and just the number of eigenwert adds, and is the summation of merged Characteristic Number.Tradition improves the method for eigenwert, mainly through improving single feature or directly using multiple feature.The method that tradition improves eigenwert can obtain good accuracy rate, but extends detection time.

The invention provides a kind of method for detecting human face of layered filtration, first with simply, fast algorithm extract the candidate region of face, carry out the face in accurate locating candidate region again with algorithm that is complicated, pin-point accuracy, thus make Face datection process fast, accurately.

Summary of the invention

The object of the present invention is to provide a kind of method for detecting human face based on layered filtration, detection speed and accuracy rate can be improved simultaneously.Choose face candidate region with haar-like matrix character, carry out accurate locating human face with the fusion feature of LBP and LGP, and whole process all uses dual threshold FloatBoost to carry out training characteristics value.

The technical solution used in the present invention is as follows:

Based on a method for detecting human face for layered filtration, it comprises following steps:

Step 1: read image to be detected, utilizes image pyramid to generate the training sample set of different scale;

Step 2: the haar-like matrix character extracting different scale, on each yardstick, generates strong classifier A by dual threshold FloatBoost;

Step 3: the multiple dimensioned human face region carrying out to extract fast and accurately candidate in original image by the strong classifier A of different scale;

Step 4: LBP and the LGP textural characteristics extracting different scale, on each yardstick, both merged, the new feature after fusion is LBP and LGP feature summation, and train generation strong classifier B by dual threshold FloatBoost, the form of this strong classifier is H (C);

Step 5: strong classifier B is accurate locating human face in candidate region, and shows.

Described method for detecting human face, in step 2, dual threshold FloatBoost is the improvement to FloatBoost, using dual threshold function as classifier functions.

The training sample set of 12 kinds of different scales is generated in above-mentioned steps 1.

Said method step 2 extracts the haar-like matrix character of 6 kinds of different scales, need not 12 kinds of yardsticks be in order to rapid extraction face candidate region.6 kinds of yardsticks are respectively 16*16,25*25,38*38,60*60,94*94,146*146.

In said method step 2, dual threshold FloatBoost is the improvement to FloatBoost, using dual threshold function as classifier functions.

FloatBoost improves on Adaboost basis, this algorithm combines floating search with Adaboost algorithm, introduce backtrack mechanism when sweep forward, the Weak Classifier declined causing overall classification performance in the strong classifier selected is deleted from cascade Weak Classifier.Solve the problem that nonmonotonicity brings, and characteristic sum can be selected efficiently to generate strong classifier.

What FloatBoost algorithm Weak Classifier function adopted is single threshold function, and for eigenwert M, should have bound, i.e. M ₀≤ M≤M ₁if eigenwert to be detected falls into region [M ₀, M ₁] in, then represent face, otherwise be non-face.Dual threshold FloatBoost adopts dual threshold function as Weak Classifier, can extract better Weak Classifier, just can improve the performance of FloatBoost further.

Said method step 2 generates the strong classifier A of these yardsticks by dual threshold FloatBoost, and number is low and false drop rate is low.

In said method step 3 in the process of extraction candidate region, first can extract candidate region with the matrix character of large scale, then extract next candidate region with the matrix character of small scale.Preferred method is once the matrix character of large scale has extracted candidate region further, and the matrix character of small scale just need not carry out having judged in this candidate region again, so just can extract the candidate region of multiple dimensioned face fast and accurately.

Train LBP and LGP fusion feature with dual threshold FloatBoost in said method step 4, the strong classifier H (C) of formation is:

$H\left(C\right)=\underset{X\∈S_T^{\mathrm{LBP}}}{\mathrm{\Σ}}{h}_x\left(B\left(X\right)\right)+\underset{X\∈S_T^{\mathrm{LGP}}}{\mathrm{\Σ}}{h}_x\left(G\left(X\right)\right)$

In formula, C is LBP and LGP characteristic image.B is LBP textural characteristics, and G is LGP textural characteristics.And be respectively the set of LBP and LGP textural characteristics.X=(type, x, y), type represent LBP feature or LGP feature, and x, y are the coordinate figure of pixel.

Beneficial effect:

The invention discloses the method for detecting human face based on dual threshold FloatBoost layered filtration and Fusion Features.Wherein layered filtration is to accelerate detection speed, and ground floor is with simply removing most non-face region based on haar-like matrix character cascade classifier, and in order to accurately extract face candidate region, the false drop rate of this cascade classifier is low.The second layer complicated based on fusion feature cascade classifier candidate region locating human face, in order to locating human face accurately, the verification and measurement ratio of this cascade classifier wants high.

In the process of layered filtration, dual threshold FloatBoost is all adopted to carry out training characteristics value, dual threshold FloatBoost not only can extract better Weak Classifier, and these Weak Classifiers are best combinations, it trains the progression of the cascade classifier that obtains not only cascade little, and in every one-level, Weak Classifier number is few.So not only can accelerate detection speed, can also verification and measurement ratio be improved.

In order to locating human face accurate in candidate region, adopt fusion feature, merged by LBP and LGP textural characteristics, it is considered that LBP textural characteristics can remove the impact of global illumination, and LGP textural characteristics can remove local noise.And the feature after merging is relative to independently LBP and LGP feature, does not improve the computation complexity of detection algorithm, just when training, best Weak Classifier can be selected from LBP and LGP fusion feature.So just can improve verification and measurement ratio when not increasing detection time.

Accompanying drawing explanation

Fig. 1 is the present inventor's face detecting method process flow diagram;

Fig. 2 is 7 kinds of forms in haar-like matrix character;

Fig. 3 is LBP textural characteristics;

Fig. 4 is LGP textural characteristics;

Fig. 5 is the sample set of 12 groups of different scales;

Fig. 6 is the candidate region of multiple dimensioned face;

Fig. 7 is the testing result of multiple dimensioned face.

Embodiment

Below in conjunction with the drawings and specific embodiments, the present invention is further described.

One, embodiment

As shown in Figure 1, the method for detecting human face specific implementation step of layered filtration of the present invention is as follows:

1. initialization

In face database obtain 1500 faces and 8000 non-face, for training classifier.

By image pyramid, training sample is divided into the sample set that 12 scale size differ.Wherein the size of base sample set is 60*60.And then by basic pattern, this constantly amplifies 1.25 times, generates 6 corresponding large samples; Basic pattern is originally constantly reduced 1.25, generates 5 corresponding small samples.Concrete sample size size as shown in Figure 5.

2. the training stage

As shown in Figure 2, structure haar-like matrix character, for reflecting the local feature of face.Comprising the intensity variation of the intensity variation in traditional edge feature, linear feature, central feature and the gradient direction of face that newly increases and nose, eye areas.This feature is simple, and computing velocity is fast, can accelerate choosing of face candidate region.

The local feature of 6 kinds of yardstick training samples is extracted, by the haar-like feature of generation 6 different scales with haar-like matrix character.Feature for each yardstick can generate 4 grades of corresponding strong classifiers, and 4 grades of strong classifiers of these 6 different scales are for choosing the candidate region of multiple dimensioned face.Rudimentary strong classifier is adopted to be to accelerate choosing of candidate region.

Strong classifier what be formed by is dual threshold FloatBoost algorithm.What in this algorithm, Weak Classifier function adopted is dual threshold function, and when choosing Weak Classifier, employing be FloatBoost algorithm.And in the process of training, being set to of classifier parameters: require that the verification and measurement ratio of every one-level is not less than 0.85, but false drop rate is not higher than 0.05.Relatively low verification and measurement ratio is in order to rapid extraction face candidate region, and low false drop rate is in order to accurate locating human face candidate region.

As shown in Figure 3,4, with the local feature of LBP and LGP texture feature extraction different scale training sample, the textural characteristics of 12 different scales will be generated respectively.Both textural characteristics of same yardstick are merged, and fusion feature is trained, generate 15 grades of strong classifier H (C).And the verification and measurement ratio of every one-level otherwise lower than 0.99, false drop rate is not higher than 0.1.So just can generate 15 grades of strong classifiers of 12 different scales.These strong classifiers are used for accurate locating human face in candidate region.The progression of this strong classifier is high, and verification and measurement ratio is high and false drop rate is low, and this is the precision in order to improve Face datection.

The concrete training process of H (C) is as follows:

Calculate classification error rate by whole LBP and LGP features, then choose the feature that error rate is minimum.Namely this feature may be LBP, also may be LGP.By this feature as best Weak Classifier, and this Weak Classifier is added the optimal characteristics matrix H of dual threshold FloatBoost algorithm _min, H _monce selected, just H can be judged by the type value in X _min Weak Classifier be LBP feature, or LGP feature, so just can form strong classifier H (C).Just fusion feature perfectly can be combined with dual threshold FloatBoost by above-mentioned steps.

By above-mentioned combination, adopt dual threshold FloatBoost can generate 15 grades of strong classifier H (C).Because this strong classifier have employed the textural characteristics of LBP and LGP fusion, so noise resisting ability is stronger.Wherein LBP can eliminate global illumination, and LGP can eliminate local noise.This fusion can select quantity few and the Weak Classifier that performance is good, not only can accelerate detection speed, can also improve verification and measurement ratio.

3. detection-phase

First use 4 grades of strong classifiers of 146*146 yardstick (yardstick maximum in 6 yardsticks) to extract the candidate region of face, detection window is moved forward with 29 (146/5) individual pixels.Once find face, just expanded by human face region and be twice, as the candidate region of face, and this candidate region is once determine, judges with regard to not needing again.Repeat process above, namely continue at the remainder of detected image the candidate region choosing face with 4 grades of strong classifiers of next yardstick (16/25 of last scale size), wherein moving pixel value is 1/5 scale size.So just can extract the candidate region of multiple dimensioned face fast.As shown in Figure 6.

With 15 grades of strong classifiers accurate locating human face in multiple dimensioned candidate region of different scale.In testing process, first carry out Face datection with 15 grades of strong classifiers of large scale, once face be detected, just mark this face, and this human face region is deleted from candidate region.Then in candidate region, Face datection is carried out with 15 grades of strong classifiers that yardstick is relatively little gradually.Wherein detection window moves forward with a pixel.So just can locate different scale face fast and accurately.As shown in Figure 7.Two, the Some principles of summary involved by the inventive method:

The step of dual threshold FloatBoost Face datection:

1. training sample is (x ₁, y ₁) ..., (x _n, y _n), wherein N=a+b; y _i=(-1 ,+1), be used for respectively marking negative sample and positive sample, and the quantity of positive sample is being a, the quantity of negative sample is b.

(1) the maximum quantity M of Weak Classifier _max.

(2) error rate function J (H is asked for _m) and maximum acceptable threshold value J ^*.

2. initialization

(1) initial weight: ${\mathrm{\ω}}_i^{\left(0\right)}=\left\{\begin{array}{cc}\frac{1}{2a}& y_i=+1\\ \frac{1}{2b}& y_i=-1\end{array}\right.$

(2) current iteration number of times M=0 is made, minimal characteristic set H ₀={ }; The minimal error rate of the strong classifier be made up of current M optimum Weak Classifier wherein M=0,1 ..., M _max.

3. Weak Classifier is selected

(1) loop iteration M=M+1;

(2) Weak Classifier h _mselection:

$h_M(x,f,p,\mathrm{\&theta;},{\mathrm{\&theta;}}^{\&prime;})=\left\{\begin{array}{ccc}1& \mathrm{if}& \mathrm{pf}\left(x\right)<p{\mathrm{\&theta;}}^{\&prime;}\mathrm{or\; pf}\left(x\right)\&GreaterEqual;\mathrm{p\&theta;},\mathrm{when\; p}=1;\mathrm{or}\\ & \mathrm{if}& \mathrm{pf}\left(x\right)<p{\mathrm{\&theta;}}^{\&prime;}\mathrm{and\; pf}\left(x\right)\&GreaterEqual;\mathrm{p\&theta;},\mathrm{when\; p}=-1;\\ 0& & \mathrm{others}\end{array}\right.$

(3) weight is upgraded: ${\mathrm{\&omega;}}_i^{\left(M\right)}\&LeftArrow;{\mathrm{\&omega;}}_i^{(M-1)}\exp [-y_i{h}_M\left(x_i\right)],$ And normalization ${\mathrm{\&Sigma;}}_i{\mathrm{\&omega;}}_i^{\left(M\right)}=1.$

(4) optimal characteristics matrix H _m=H _m-1∪ { h _m; If then

4. delete Weak Classifier according to condition

(1) H is obtained _mthe Weak Classifier h ' that middle performance is the poorest:

(2) if H _mnot the combination of a front M optimal characteristics, namely then

A () deletes the poorest Weak Classifier h ' of performance, namely

$H_{M-1}=H_M-h^{\&prime;};J_{M-1}^{\min }=J(H_M-h^{\&prime;});M=M-1;$

If (b) h '=h _{m '}then recalculate and h _j(j=m ' ..., M);

C () jumps to 4. (1).

(3) if H _mthe combination of a front M optimal characteristics, namely then

A () is if M=M _maxor J (H _m) < J ^*, then 5 are jumped to;

B () jumps to 3. (1).

5. export $H\left(x\right)=\mathrm{sign}\left[{\mathrm{\&Sigma;}}_{h\left(x\right)\&Element;H_M}h\left(x\right)\right].$

Claims (8)

1., based on a method for detecting human face for layered filtration, it is characterized in that comprising following steps:

Step 1: read image to be detected, utilizes image pyramid to generate the training sample set of different scale;

Step 2: the haar-like matrix character extracting different scale, on each yardstick, generates strong classifier A by dual threshold FloatBoost;

Step 3: the multiple dimensioned human face region carrying out to extract fast and accurately candidate in original image by the strong classifier A of different scale;

Step 4: LBP and the LGP textural characteristics extracting different scale, on each yardstick, both merged, the new feature after fusion is LBP and LGP feature summation, and train generation strong classifier B by dual threshold FloatBoost, the form of strong classifier B is H (C);

Step 5: strong classifier B is accurate locating human face in candidate region, and shows.

2. method for detecting human face as claimed in claim 1, it is characterized in that, in described step 2, dual threshold FloatBoost is the improvement to FloatBoost, using dual threshold function as classifier functions.

3. method for detecting human face as claimed in claim 1, is characterized in that, when choosing human face region in described step 3, first judge by large scale, then use small scale.

4. method for detecting human face as claimed in claim 3, is characterized in that, when choosing human face region in described step 3, the candidate region that large scale confirms, small scale does not need to judge again.

5. method for detecting human face as claimed in claim 1, is characterized in that, the form of strong classifier B in described step 4 $H\left(C\right)=\underset{X\&Element;S_T^{\mathrm{LBP}}}{\mathrm{\&Sigma;}}{h}_x\left(B\left(X\right)\right)+\underset{X\&Element;S_T^{\mathrm{LGP}}}{\mathrm{\&Sigma;}}{h}_x\left(G\left(X\right)\right).$

6. method for detecting human face as claimed in claim 1, it is characterized in that, described strong classifier A progression is little and false drop rate is low.

7. method for detecting human face as claimed in claim 1, it is characterized in that, described strong classifier B progression high detection rate is high.

8. method for detecting human face as claimed in claim 1, is characterized in that, generates the training sample set of 12 kinds of different scales in described step 1; The haar-like matrix character of 6 kinds of different scales is extracted in step 2; Be the strong classifier A of 6 kinds of different scales in step 2; LBP and the LGP textural characteristics of 12 kinds of different scales is extracted in step 4.