KR20160037424A - A Novel Multi-view Face Detection Method Based on Improved Real Adaboost Algorithm - Google Patents

Abstract

The present invention relates to a face detection method, and more particularly, to a novel multi-view face detection (MVFD) method based on an improved real adaptive boosting (AdaBoost) algorithm. The improved real AdaBoost algorithm for trained Haar features devises a dynamic weight and a previously divided sample. As compared with an existing real AdaBoost algorithm, a time complexity of a weak classifier is 0(M*N), a time complexity of a strong classifier is 0(T*M*N), and a training speed is reinforced by 0(N). Only twelve category classifiers for the multi-view face detection (MVFD) based on our Haar feature processing are required for generating rotation-out-of-image (ROP) and changing an angle of rotation-in-plane (RIP). The devised algorithm significantly reduces training complexity. In order to use the real AdaBoost algorithm, an LUT weak classifier is used to enhance the Haar features. Instead of each trained pose measuring device, the confidence of the strong classifier defined by the AdaBoost algorithm is used. The first four layers among 16 layers of a cascade classifier are used for pose estimation, and extra calculation for pose estimation is unnecessary. An algorithm test is performed based on CMU and MIT face databases. As a result of the testing, the convergence performance of the improved real AdaBoost algorithm is more enhanced than that of a previous real AdaBoost algorithm. Also, the multi-view face detection system according to the present invention shows a high sensing rate and has an excellent timeliness and a high detection rate.

Description

A Novel Multi-view Face Detection Method Based On Improved Real Adaboost Algorithm Based on Improved Real AdaBoost Algorithm [

The present invention relates to a face detection method, and more particularly, to a new multi-view face detection method based on an improved real AdaBoost algorithm.

Over the last decade, facial recognition has emerged as an active research area for computer vision and a number of potential applications, including biometrics, surveillance, human-computer interaction, video-mediated communication, and content-based access to image and video databases.

The first step of face recognition is the automatic face recognition step. The reliability has a great influence on the performance and usability of the whole face recognition system. A given image, video, or ideal face detector should be able to find and distinguish all current faces, regardless of the location, size, orientation, age, or facial expression of the image or video.

In addition, the face must be detected regardless of the external lighting conditions, image and video content. Face detection is performed based on various signals such as skin color, motion, face or head shape, face shape, and the like.

Most successful face recognition algorithms do not use any other signal, they are based on appearance, the input image is scanned at all possible positions and scales by the sub-window, the face detection is classified into one face or sub- A face, or a non-face classification is processed using a statistical learning method.

Statistical methods have been widely adopted in conventional face detection methods. Moghaddam and Pentland introduced an eigenface recognition method in which the probability of facial patterns was created by distance-in-feature-space (DIFS) and distance-from-feature-space (DFFS). Osuna proposed a face detection method based on SVM to recognize facial face.

Unlike the eigenface method, which only estimates the density of the positive, this method attempts to learn the boundaries between face and non-face patterns. After learning, an 'important' sample at the boundary is chosen to build a decision function.

Soulie introduced a system that uses neural networks (NNs) to detect faces. They have implemented a multi-mode architecture in which various rejection criteria are used for false recognition of trade-offs against false rejection.

Sung and Poggio also introduced an NN-based face detection system. They devised six prototypes (faces) and six prototypes (not faces) in the hidden layer.

Recently, Viola and Jones proposed a face detection method that can be efficiently computed using a simple rectangle function called so-called integral image. This system has been developed as a more general system than rotational invariant face detection and multi view face detection (MVFD). The AdaBoost and cascade methods are also used to detect faces according to similar functions to those described above.

Li devised a similar but more general function (a function that can be calculated from the difference of blocks), and Floatboost proposed a way to overcome the simplicity of sequential AdaBoost algorithm learning.

Previous research on multi-view face detection (MVFD) has focused on multi-view face detection (MVFD) work including studies by Schneiderman et al. Based on Bayesian decision rules and the first proposed real- Li pyramid structure detectors.

In Hongming Zhang's study, they emphasized effective classification schemes in which learning of beneficial functions was achieved through minimizing the error rate of the ensemble ECOC multi-class classification.

The ultimate goal of real-time rotational invariant multiview face detection (MVFD) systems is to meet the needs of various applications.

Although the conventional MVFD method can be applied to overcome the problem by methods such as iterative procedure and image rotation, there is a need to research and develop a face detection method for a large time-consuming and high failure rate.

KR 10-2011-0097794

The present invention has been made to solve the above-mentioned problems of the prior art,

That is, the object of the present invention is to propose a new improved real AdaBoost algorithm-based face detection method based on the improved AdaBoost classifier proposed by Schapire and Singer.

The LUT weak classifier is used to train Haar functions (mirrors and rotations) processed by the AdaBoost algorithm. In addition, the pose estimating method using the first four layers to estimate the face pose based on the trust of the strong classifier is proposed in the present invention. The present invention tested a multi-view face detection (MVFD) scheme designed in the present invention through the CMU and MIT face database, and this experimental result shows that the detector system described above has better performance than discrete and real AdaBoost algorithms . The multi-view face detection (MVFD) method aims at accelerating convergence of a strong classifier, reducing the number of weak classifiers, and reducing detection time.

In a new multi-view face detection method (MVFD) based on an improved real AdaBoost algorithm, the multi-view face detection (MVFD) method is composed of 16 layers and extracts facial expressions from a plurality of sample face images A cascade classifier; A sub-window formed by dividing the extracted sample space evenly by a weak classifier; A pause estimation unit based on the reliability of the strong classifier; And a multi-view face recognition (MVFD) unit designed for real AdaBoost algorithm testing.

As described above, the new multi-view face detection method (MVFD) according to the present invention has been tested with multi view face detection (MVFD) through CMU and MIT face database. As a result, It is proved that the algorithm is superior in terms of performance, such as convergence speed of the strong classifier faster than the algorithm, reduction of the number of weak classifiers, and reduction of the detection time.

Detailed results are shown in Figure 1, Figure 2, Figure 3, Figure 4, Table 1, and Table 2.

Figure 1. Face detection ratio comparison

Table 1

Figure 2. Comparison of ROC curves detected in the MIT frontal face test set

Figure 3. Front face detection results from CMU and MIT test set

Table 2. Multiview face detection results with CMU profile face test set

Figure 4. Multiview face detection results with CMU profile face test set

The detection time of the cascaded classifier is shorter than that of the conventional real AdaBoost algorithm by 8 ms, and the number of training layers is 16, which is 9 times more than the previous real AdaBoost algorithm and 62 discrete AdaBoost algorithms. Better performance. The number of weak classifiers in the cascade experiment using the real Adaboost algorithm compared with the previous real Adaboost algorithm was 7492, the discrete Adaboost algorithm 4297, and the previous real Adabus algorithm 1784 Good results.

The first layer of the cascade was used for pose detection, and as a result, the failure rate was lower with a pose detector (PE) than without a pose detector (PE). The results of the experiment are shown in Fig.

Figure 5. ROC curve comparison with CMU profile face data set

In the present invention, a multi-view face detection system is devised based on VS2008 and public CV library for real-time face detection. Three persons with different poses, facial expressions, and other elements (glasses, hair, etc.) were selected as experimental specimens and the results of the experiment are shown in Fig.

Figure 6. Real-time multi-view face detection result

Experiments have shown that the convergence performance of the improved real AdaBoost algorithm is better than the previous real AdaBoost algorithm, as a result of the algorithm tests based on the CMU and MIT face data bases.

In addition, the multi-view face detection system designed has a high detection rate for face detection, can be applied in real time, and has a high detection rate even when detecting a pose of a target image.

FIG. 1 illustrates a face detection process using an improved real AdaBoost algorithm according to an embodiment of the present invention.
Figure 2 shows a Haar feature pattern selected by the real AdaBoost algorithm of an embodiment of the present invention.

In a new multi-view face detection method (MVFD) based on an improved real AdaBoost algorithm, the multi-view face detection (MVFD) method is composed of 16 layers and extracts facial expressions from a plurality of sample face images A cascade classifier; A sub-window formed by dividing the extracted sample space evenly by a weak classifier; A pose estimation unit based on the reliability of the strong classifier, and a multi-view face recognition (MVFD) unit designed for real AdaBoost algorithm testing.

The sub window processes the Haar feature corresponding to the weak classifier, and the weak classifier is formed by evenly distributing the upper space into n sub categories, and each weak classifier corresponds to each sub category.

It is an object of the present invention to detect multi-view face detection (MVFD0, i.e. to detect vertical face images (rotaion-out-of-plane change and rotation invariance) -plane) means to detect a face having a pose change.

According to the ROP angle, 180 degrees can be set to [-90 °, -75 °], [-75 °, -45 °], [-45 °, -15 °] + 45 °], [+ 45 °, + 75 °], [+ 75 °, + 90 °]. According to the RIP angle, the face is divided into 12 categories, each covering a range of 30 degrees. Thus, a total of 7 * 12 view categories correspond to 84 detectors. Based on the AdaBoost algorithm, 84 category face detectors are designed for 84 categories of multi-view faces. Since the Haar feature can be flipped horizontally and rotated 90 degrees, anything else can be created with the original image distortion.

The above table is a table summarizing the geometric transformation of the first image using Haar features.

The first detectors in the front view are 60 degrees and 90 degrees. Half of the profile views of the original detector are half the remaining profiles, and half of the profile views of the original detector are 60 degrees, 90 degrees, 120 degrees. The overall profile element is equal to half of the profile. In the previous study, the first sensor was formed as 7 * 12 = 84 category sorter, but the category sorter required the first sensor in the present invention is 4 * 3 = 12 categories.

&Quot; 4 " means 4 out of 7 horizontally inverted angles, and " 3 " means at least 0, 30 and 60 degrees in 12 categories along the RIP angle. Other elements are created by modifying the Haar feature.

The figure above illustrates the rotation and reflection of the image.

The AdaBoost algorithm is a learning algorithm, in which a set of weak classifiers is selected from a virtual large space to form a strong classifier.

The performance of the strong classifier depends on the nature of the virtual space of the weak classifier. The weak classifier of the threshold is that the virtual space is the input to accelerate the procedure, and the main disadvantage of the threshold model is that it is too simple to fit into the complex distribution.

Therefore, in the present invention, the LUT weak classifier is used to use the real Adaboost algorithm.

Therefore, in the present invention, an actually evaluated LUT weak classifier is used.

Assuming that f _haar is a Haar feature and f _haar is normalized to [0,1], if the parent category is equally divided into subcategories, then the jth LUT item corresponds to the subrange:

의 분할에 대응되고, 약분류기는 아래의 식으로 산출된다.The divided categories are

, And the weak classifier is calculated by the following equation.

일때,

when,

이다.here

to be.

는

이고, 여기서

이다.Given exponential function

The

, Where

to be.

로 나타낼 수 있다. LUT분류기는 모든 종류의 확률분산을 가능하게 한다. 부-윈도우의 샘플 사이즈가 확정되면, 모든 Haar 특징요소가 정해지고, 각각의 Haar 특징은 하나의 LUT 약분류기에 대응되고, 이것은 약분류기 공간 H를 구성한다.The LUT weak classifier

. The LUT classifier enables all kinds of probability distributions. Once the sample size of the sub-window is determined, all Haar feature elements are determined, and each Haar feature corresponds to one LUT weak classifier, which constitutes the weak classifier space H.

In the present invention, each layer of the cascade is a strong classifier trained by an improved real AdaBoost algorithm. In the present invention, the threshold value of each layer is set to b,

로 로 b가 산출되고, 얼굴의 99.9%가 통과하면 반례를 처리하기 위해 노력한다. 후자의 공간에서 분류기는 더욱 복잡하고, 이것은 더 많은 LUT에 기반한 분류기를 포함한다. 따라서 그것들은 강분류기의 성능을 가진다.

B is calculated, and when 99.9% of the face passes, we try to deal with the counterexample. In the latter space, the classifier is more complex, including a classifier based on more LUTs. Therefore, they have the performance of a strong classifier.

The improved real AdaBoost algorithm can be said to be better than the previous algorithm because the real AdaBoost algorithm produces a more continuous confidence value when compared to the discrete AdaBoost algorithm. In addition, the LUT weak classifier of the present invention works well with the real AdaBoost algorithm in particular. The algorithm learning procedure is as follows.

1) Enter the maximum false positive rate for each layer, enter d for the minimum pass rate, and F _target for the total false positive rate. A given amount of training set is Pos, and a negative training set is Neg.

2) Initialization phase:

3)

A. Train the i-th layer using Pos and Neg; f, the false positive rate f_iAnd the threshold value b is set to set the pass ratio when it is higher than d.

B.

A. 이면, 캐스케이드 분류기를 사용하여 비 얼굴이미지를 스캔하고, 수집된 모든 false positive를 Neg로 둔다.C.

If A., scan the non-facial image using a cascaded classifier and place all false positives collected in Neg.

Wherein said weak classifier does not redistribute any further training sample space if an optimal weak classifier is selected at each stage of face detection, wherein said weak classifier comprises a sub window size determination step; Haar feature candidate is confirmed and Haar feature determination step; A corresponding step wherein each Haar feature corresponds to a respective weak classifier; , And the LUT weak classifier can realize all kinds of distributability.

A strong classifier trained by the improved real AdaBoost algorithm constitutes each layer of the cascade.

Learning of the improved real AdaBoost algorithm involves constructing a data set; An initialization step; obtaining t; And a strong classifier calculation step.

1) a data set construction step;

로 둔다.Data set S

.

, 약 분류기 풀 H, 약 분류기의 수를 T, 분류기를 위한 특징의 수를 s 일 때,

에 가중치를 준다.

, The number of features of the weak classifier pool H, the classifier of the weak classifier, T, s,

.

1) an initialization step;

(1) Initialize sample distribution.

(2) Initialize the gray value distribution difference of the Haar feature.

에 대응된다.(3) Process each Haar feature corresponding to the weak classifier, and evenly divide into n subcategories. The division of categories

.

)를 구하는 단계 3) t (

)

(1) To obtain H 'approximate classifier h

A. Find the distribution D _t . The equation is as follows.

이다.here

to be.

Substitute the result of B. h into each X _j .

Where e is a positive constant.

Calculate the normalization factor.

,

의 식으로 구한다.(1) Z 'and h' _t

,

.

의 식으로 샘플분배를 업데이트하고.(2)

≪ / RTI > to update the sample distribution.

으로 훈련가중치

_t를 구한다.

Training weighted by

_{Find t} .

를 정규화한다.(3) To find the conditional probability of the distribution function

Is normalized.

4) Step of calculating the strong classifier

이고, 여기서 b는 임계값 디폴트로 0이다.The last river classifier H

, Where b is a threshold default of zero.

로 산출한다.The reliability of H

.

In addition, the real AdaBoost algorithm described above is improved in terms of weight adjustment and data set partitioning compared to the conventional AdaBoost algorithm, and shows improved performance in the cascade classifier designed in the present invention. Improved performance is demonstrated by weighted control.

는

에 근거하고,

는 상기의 식에서 도출된 값과 같다.

The

Based on this,

Is the same as the value derived from the above equation.

로 산출되고,

이다.The stochastic leading variable

Lt; / RTI >

to be.

The strong classifier of the real AdaBoost algorithm is equal to the sum of all weak classifiers.

으로 산출되고,

_t는

로 산출된다.The strong classifier

Lt; / RTI >

_t is

.

,

으로 구할 수 있다.The mean value and variance of this

,

.

일때, m는 평균값이고, s_t ²은 분산이다.

, M is the mean value, and s _t ² is the variance.

In the present invention, the pose estimating unit uses the reliability of the strong classifier without being separated and trained, so that real-time face detection is deviated from the previous method which was difficult.

으로 산출된다. 멀티 뷰에 근거한 d 감지기와 각각의 감지기는 n개의 레이어를 가진다고 가정할 때, i번째 감지기의 j번째 레이어 신뢰도는

로 산출된다.The reliability of the strong classifier

. Assuming that the multi-view based detector and each detector have n layers, the jth layer reliability of the ith detector is

.

로,The reliability of the first k layers is

in,

로 산출된다.The first m-layer pose estimator

.

Through this process, if x is estimated to be not a face, it is discarded in the first m layer, and the pose estimation process is not performed again.

Unlike the previous Rowley study, pose estimation and face detection are not performed separately, so no additional computation is required. The pose estimation procedure is shown in the figure below.

In the present invention, there are four view base detectors, the cascade classifier includes sixteen layers, and four layers are used for face pose detection.

Claims (8)

In a new multi-view face detection scheme (MVFD) based on the improved real AdaBoost algorithm,
The multi-view face detection (MVFD)
A cascade classifier configured to extract a facial expression from a plurality of sample facial images, the facial expressions being composed of 16 layers;
A sub-window formed by dividing the extracted sample space evenly by a weak classifier;
A pause estimator based on the reliability of the strong classifier;
Multi-view face detection method that includes multi-view face recognition (MVFD) part designed for real AdaBoost algorithm test. The method according to claim 1,
The sub window may be formed by processing the Haar feature corresponding to each weak classifier so that the weak classifier is formed by evenly distributing the upper space into n subcategories such that each weak classifier corresponds to each subcategory Features multi-view face detection. 3. The method of claim 2,
Wherein said weak classifier does not redistribute further training sample space if an optimal weak classifier is selected at each stage of face detection. 3. The method of claim 2,
The weak classifier is an LUT weak classifier in which a Haar feature candidate is determined when the size of a sub window sample is determined and each Haar feature is obtained through a step corresponding to each weak classifier, And the distribution possibility of the multi-view face detection method can be realized. The method according to claim 1,
In the above cascade classifier,
A multi-view face detection system characterized in that a strong classifier trained by an improved real AdaBoost algorithm constitutes each layer of a cascade. The method according to claim 1,
The real AdaBoost algorithm described above is improved in terms of weight adjustment and data set partitioning compared to the existing AdaBoost algorithm, thereby showing improved performance in the cascade classifier devised in the present invention. The method according to claim 1,
The improved real AdaBoost algorithm, which operates the cascade classifier described above,
The multi-view face detection method is characterized by better performance than existing algorithms by improving the control of segmentation and weight adjustment of data sets. The method according to claim 1,
The pose estimation unit does not train the face detection and the pose estimation separately by using the reliability of the strong classifier for the pose estimation out of the conventional method in which real time face detection is difficult and does not cause any additional calculation problem Multi-view face detection method.

Images