KR20070118830A - Face and facial feature detection method using evolutionary algorithms - Google Patents

Abstract

A method for detecting a facial image through an evolutionary algorithm is provided to enable facial image detection and facial image feature detection even in a system, such as a desktop computer, or a cellular phone, having a restricted operation capacity. Rectangular features are extracted from a moving picture or a still image by using a rectangular feature sample. The sets of the rectangular features(201-206) are made based on the rectangular features through an AdaBoost scheme. The weight values of the sets of the rectangular features are reconstructed by using an evolutionary algorithm, so that the number of the rectangular features is reduced, and then the sets of rectangular features(400) are reconstructed.

Description

 Face and facial feature detection method using evolutionary algorithms}

1 is a schematic diagram illustrating an image system configuration for detecting a face and facial features.

2 is a diagram showing the types of rectangular features used in the present invention.

3 is a diagram illustrating a classifier learning method using an AdaBoost method and an evolutionary algorithm.

4 is a diagram illustrating a structure having a plurality of stages by sequentially arranging a plurality of learned classifiers.

를 진화 알고리즘을 이용하여 재구성하기 위한 인코딩을 보여주는 도면이다.5 is a weight

Is a diagram showing the encoding for reconstruction using the evolutionary algorithm.

를 재구성하는 방법을 설명한 도면이다.6 using an evolutionary algorithm

A diagram illustrating a method of reconfiguring.

The present invention relates to a method of extracting a face region of a person from moving image or still image information and detecting facial feature components such as eyes, nose and mouth in the face region.

As a conventional face detection method, a method using a template or a method using a color information of a face has been frequently used. The method using the template is far from real-time detection because it takes too much time to perform while changing the possible position and size of the input image without knowing the size and position of the face. The method using color information guarantees fast detection time, but it is not effective in terms of detection rate and false detection rate because of the skin color of people who are sensitive to lighting environment.

The present invention has been made to solve the above problems, and an object of the present invention is to provide a method of learning a classifier that accurately classifies two classes of face-to-face learning data. If you learn the classifier that separates the two types of learning data, you can use this to classify the face and the non-face image and perform face detection. Furthermore, the facial features such as eyes, nose, and mouth are also detected through learning. I can make it. The purpose of this paper is to improve the detection rate and computation time by introducing an evolutionary algorithm to the existing AdaBoost learning method.

을 재구성하는 방법을 이용하여 줄어든 개수의 사각형 특징 집합을 만드는 제 3단계를 포함하여 이루어진 발명이 제시된다.In order to achieve the object of the present invention, in the present invention, in the method for detecting a face region of a person from moving image or still image information, the method for extracting a rectangular feature using a rectangular feature sample from the moving image or still image information; Step 1, the second step of creating a rectangular feature set using the AdaBoost method from the extracted rectangular feature, and the second feature after the step 2 using the evolution algorithm

An invention is provided that includes a third step of creating a reduced number of rectangular feature sets using a reconstruction method.

Hereinafter, with reference to the accompanying drawings, the configuration and operation of the invention for the embodiment of the present invention will be described in detail.

The face detection and face feature detection method according to the present invention includes a face detector for detecting a face from an input image, a face feature extractor for extracting eyes, nose, mouth, etc. from the detected face, AdaBoost and an evolution algorithm for each detector configuration. It includes a classifier learning method using.

In the present invention, a method for face detection and face feature detection using only grayscale still image information will be described. Since most color images and videos basically include still image information of grayscale, the method using the grayscale still image information can operate under various image acquisition conditions.

를 구성할 수 있다. 즉, 하나의 사각형 특징은 하나의 약분류기를 구성한다. 수학식 1은 입력 영상 벡터

에 대해서 각각의 사각형 특징값

와 문턱값

비교를 통해서 간단히 분류를 하는 약분류기를 나타낸다. 이때 부등호의 방향을 결정하기 위해서 부호값

이 곱해진다.

은

값을 가진다.Rectangular features are used to detect facial features such as face, eyes, nose and mouth. Six types are used for the rectangular shape as shown in FIG. 2. The five features of 201 to 205 are the pixel values (brightness values) corresponding to the rectangular areas of the bright areas minus the pixel values of the dark areas, and the features shown in 206 are the variances of the pixel values of the rectangular areas. It has a value as a feature value. These six types of rectangular features may exist about 180,000 when the reference window size 24 pixels * 24 pixels. Among the large number of rectangular features, the AdaBoost learning method is used to reduce the number of rectangular features that can distinguish the face and non-face image sets to hundreds or thousands. Weak classifier that extracts valid rectangular features from a set of 180,000 levels of rectangular features and allows each rectangular feature to make one decision.

Can be configured. That is, one rectangular feature constitutes one weak classifier. Equation 1 is the input image vector

Each rectangular feature value for

And threshold

We show a weak classifier that simply classifies through comparison. At this time, to determine the direction of inequality sign

Is multiplied.

silver

Has a value.

가 얼굴일 경우이고, 0은 얼굴이 아닐 경우이다. 이러한 약분류기 들은 그 분류 성능이 부정확하므로 여러 가지 약분류기들의 결정을 종합하여 최종 결정을 하는 강분류기(strong classifier)를 AdaBoost 방법에 의해서 수학식2 와 같이 얻을 수 있다.The weak classifier output with a simple form of decision is 1 or 0. 1 is input video

Is a face and 0 is not a face. Since the classifiers are inaccurate in their classification performance, a strong classifier that combines the decisions of various classifiers and makes a final decision can be obtained by Equation 2 by AdaBoost method.

는 강분류기를 구성하는 T개의 약분류기 각각의 출력값에 가중치

가 곱해진 형태이다. Output value

Is weighted to the output values of each of the T weak classifiers constituting the strong classifier.

Is multiplied.

AdaBoost learning obtains a set consisting of several classifiers by sequentially selecting the useful features one by one from the rectangular features (Fig. 2), and constructs the strong classifier (Equation 2) by combining the decisions of the selected classifiers. . The operation of the AdaBoost learning method from the prepared training data follows the following procedure.

개의 얼굴 영상과

개의 얼굴이 아닌 영상의 두 가지 집합으로 구성된 학습 데이터를 설정한다. 전체 데이터의 개수는 K=a+b이다. 얼굴영상에 대해서는 레이블 값

를 1으로 설정하고 아닐 경우에는 0으로 설정한다.first,

Face images

Set training data consisting of two sets of images instead of two faces. The total number of data is K = a + b. Label value for face image

Is set to 1, otherwise it is set to 0.

Second, when the training data is prepared, the weights are initialized as in Equation 4 for each image.

Third, AdaBoost learning is conducted by the number T of all the weak classifiers.

t = 1,… Repeat steps a through d for T.

a. The weight is normalized as shown in Equation 5.

를 구한다. b. For all features, as shown in equation (6)

Obtain

가 최소가 되는

,

를 선택한다. c.

Is the minimum

,

Select.

d. The weight is updated as shown in equation (7).

이고,

이다.here,

ego,

to be.

Fourth, after the T iterations, the strong classifier is finally determined.

Equation 8 is the same as Equation 2, and in this case, the detection rate and the false detection rate may be changed by adjusting TH, which is a threshold value for discrimination. In general, when constructing a detector, a desired detection rate and a false detection rate are selected to find a suitable TH so as to satisfy a predetermined detection rate or more and a predetermined error detection rate or less. If no TH can be found, increase the number of weak classifiers and proceed with learning AdaBoost until the desired detection rate and false detection rate are met.

와 강분류기를 위한 문턱값 TH정보를 모두 포함하는 것을 말한다.In this way, if the final decision is made by combining T weak classifier decisions, a classifier that satisfies high detection rate and low false detection can be designed. At this time, one strong classifier learns by increasing the number of weak classifiers one by one until it satisfies a desired detection rate and a false detection rate in previously prepared training data. Rectangle feature set means the rectangle feature and its weight

And the threshold TH information for the classifier.

The evolutionary algorithm is applied to the rectangular feature set selected by the AdaBoost method to find a smaller number of rectangular feature sets while satisfying the desired detection rate and false detection rate and store them as the final rectangular feature set.

이다. 즉

를 진화 알고리즘에서 사용하는 개체로 인코딩한다. 인코딩을 하는 것은 0과 1을 이용하여 2진수로 표현하는 방법을 사용하거나 실수 형태로 표현할 수 있다. 하나의 개체는 도 5와 같이

들로 구성되어 있는데 이러한 개체들이 1개 이상 모여서 개체들의 집단

를 구성한다. 여러 가지 다양한

들의 조합으로 구성된 개체들이 모여 있기 때문에 이중에는 우수한 개체가 있는 반면 그렇지 못한 개체들도 있다. 진화 알고리즘을 이용하여

를 재구성하는 방법은 도 6과 같다. The set of rectangular features (a set of weak classifiers) already found by the AdaBoost method satisfies the desired detection rate and false detection rate, but an evolutionary algorithm is used to find a smaller number of rectangular feature sets. Evolutionary algorithms are a probabilistic optimization method that has been proven to perform well in many optimization problems. It has the advantage that it works well even when the search range is wide and local optimal exists. The object of optimization is

to be. In other words

Is encoded into the object used by the evolutionary algorithm. The encoding can be expressed in binary form using 0 and 1 or in real form. One object is as shown in FIG.

Consisting of one or more of these entities

Configure Various

Because of the grouping of individuals, some of them are superior, while others are not. Using evolutionary algorithms

The method of reconstructing is as shown in FIG.

를 재구성하여 원하는 검출률과 오검출률을 만족하면서 더 높은 검출률과 더 낮은 오검출률을 얻을 수 있도록 동작한다. 진화 알고리즘에서는 매 세대마다 우수한 평가를 받은 개체들이 살아 남아가기 때문에 세대가 진행될수록 우수한 해를 찾을 가능성이 높아진다. 이를 위해서는 적절한 평가 척도를 설정해야 한다. 평가는 학습 데이터에 적용하여 검출률과 오검출률을 얻은 다음 이 두 가지 값의 조합으로써 평가 함수값을 얻을 수 있다. 어떠한 형태로 평가 함수를 구성하더라도, 평가 함수의 값이 검출률이 높을수록 오검출률이 낮을수록 높은 값을 가지도록 설정하고 진화 알고리즘은 평가 함수를 최대화 하는 방향으로 동작시킨다면 무방하다. 즉, 평가 함수의 값이 클수록 우수한 개체로 평가하는 것이다. 혹은 검출률이 높을수록 오검출률이 낮을수록 평가 함수값이 더 낮은 값을 가지도록 설정하고 진화 알고리즘은 평가 함수를 최소화 하도록 동작시켜도 같은 효과를 볼 수 있다. 좋은 평가를 받은 우수한 개체들은 낮은 평가를 받은 개체들보다 높은 확률로 자식을 형성하는데, 즉 P(t-1)로부터

를 생성한다. 새로 생성된 개체들에 크로스 오버, 돌연변이 등의 일반적인 진화 알고리즘에서 사용하는 여러 가지 변형 테크닉을 적용하여 다양한 탐색을 수행한다.Evolutionary algorithm

By reconstructing, it operates to obtain higher detection rate and lower false detection rate while satisfying the desired detection rate and false detection rate. In evolutionary algorithms, the most well-established individuals survive every generation, so the chances of finding a better solution increase as generation progresses. To this end, appropriate evaluation measures should be established. The evaluation is applied to the training data to obtain a detection rate and a false detection rate, and then a combination of the two values can be used to obtain an evaluation function value. Whatever the evaluation function is, the value of the evaluation function should be set to have a higher value as the detection rate is higher and the false detection rate is lower, and the evolution algorithm may operate in the direction of maximizing the evaluation function. In other words, the larger the value of the evaluation function, the better the individual. Alternatively, the higher the detection rate, the lower the false detection rate, the lower the evaluation function value, and the evolutionary algorithm can be operated to minimize the evaluation function. Good individuals with good evaluation have a higher probability of forming offspring than low rated individuals, ie from P (t-1).

Create The newly created entities are subjected to various exploration techniques by applying various transformation techniques used in general evolutionary algorithms such as crossovers and mutations.

를 제거하고 다시 원하는 조건을 만족하는 해를 찾는다. As evolution progresses, if the desired conditions (detection rate and false detection rate) are satisfied, the solution is analyzed and the smallest

Remove it and again find the solution that satisfies the desired condition.

가 사라졌기 때문에 개체의 길이는 하나 줄어들게 되고 하나 줄어든 길이의 개체들 집단

을 구성하고 다시 진화 알고리즘을 적용한다. 이러한 작업을 반복함으로써

개보다 적은 수의 약분류기 집합을 찾아낼 수 있다. 또한 더 적은 수의 약분류기 집합을 찾을 수 없을 경우에도 AdaBoost 방법에서 얻어진 약분류기 집합이 가지는 검출률과 오검출률 보다 더 높은 검출률과 더 낮은 오검출률을 얻을 수 있는

집합을 구할 수 있다. With the smallest value

Disappears, the length of the object is reduced by one,

Construct and apply evolutionary algorithms again. By repeating this operation

You can find fewer classifier sets. In addition, even if a smaller set of weak classifiers cannot be found, higher detection rates and lower false detection rates can be obtained than those of the classifier sets obtained by the AdaBoost method.

You can get a set.

집합보다 더 높은 검출률과 더 낮은 오검출률을 만족하는

집합을 얻게 되면 이는 더 빠른 연산 시간을 가능하게 한다.Since the number of rectangular features is directly proportional to the computation time, a set of fewer rectangular features ensures faster computation time. In addition, even if the same number of weak classifier sets are used, the evolutionary algorithm is applied to find the AdaBoost method.

Satisfying higher detection rate and lower false detection rate than aggregation

Once you have the set, this allows for faster computation time.

The series of learning processes is performed as shown in FIG. 3. Experimentally confirmed that the rectangular feature set 400 found using the evolutionary algorithm is composed of a smaller number of rectangular features of about 60% than the rectangular feature set 300 obtained by the conventional AdaBoost method.

Using the finally obtained rectangular features to construct a strong classifier as shown in Equation 2, and to connect the strong classifier in several stages (stage) to construct an efficient detector. As shown in FIG. 4, the process of FIG. 3 is required to learn one stage classifier, and for the next stage of learning, new process data is constructed to repeat the process of FIG. 3. This process is repeated to implement a multistage detector. In addition, the overall efficiency of the classifier is increased through the structure in which the whole classifier in the front stage is used as one weak classifier in the rear stage. If each stage determines that the classifier is a face, the next stage operation is performed. If it is not a face, the classifier discards the current input window and performs classification on the next input window.

, 오검출률

를 기대할 수 있다.In order to learn the classifier of each stage, the desired detection rate and false detection rate should be selected. Mainly, the detection rate can be selected from about 0.9 to 1 (90% to 100%), and the false detection rate can be selected from about 0.05 to 0.5. These values ultimately determine the detection rate and false detection rate of a multi-stage detector. For example, a total of 10 detectors configured such that each stage satisfies a detection rate of 0.999 and a false detection rate of 0.4 may finally be detected.

, False detection rate

You can expect.

In this way, the facial region is detected and face feature elements are detected again based on the detection result. In order to detect facial elements such as eyes, nose, and mouth, only the training data is replaced with a corresponding image such as eyes, nose, mouth, and the like, and then the learning method of FIG. Thereafter, a multi-stage detector of FIG. 4 is configured to perform facial feature detection.

According to the present invention, face detection and facial features that are fast and robust can be detected freely under the constraints of the image capturing device and the computing device. Efficient computation enables face detection and facial feature detection not only on the desktop but also on systems with limited computing power, such as mobile phones. The present invention will enable effective implementation of facial recognition systems, human-robot interaction, security and surveillance systems.

Claims (14)

In the method for detecting a face area of a human from moving image or still image information, A first step of extracting a rectangular feature using a rectangular feature sample from the video or still image information; A second step of creating a rectangular feature set using the AdaBoost method from the extracted rectangular features; And a third step of creating a reduced number of rectangular feature sets by using a method of reconstructing the weights of the rectangular feature sets by using the evolutionary algorithm of the rectangular feature sets after the second step.  The method according to claim 1, A face detection method using an evolutionary algorithm, characterized in that a face is detected or a face feature element is detected using the rectangular feature set of the third step. The method according to claim 2, The face detection is a face detection method using an evolutionary algorithm, characterized in that for detecting the face of a plurality of angles by forming a learning image of the front face, the left face, the right face. The method according to claim 2, Face detection method using an evolutionary algorithm, characterized in that the first, second, third step learning results of claim 1 arranged in multiple stages for the face detection. The method according to claim 1, And a rectangular feature set comprising weak classifiers satisfying a detection rate of 90% to 100% and a false detection rate of 0.05% to 0.5% using the evolutionary algorithm. The method according to claim 2, A face detection method using an evolutionary algorithm, characterized by detecting each feature in a detected face region by constructing a learning image using the face feature elements of eyes, nose, and mouth. The method according to claim 2, The face detection method using the evolutionary algorithm, characterized in that the first, second, third stage learning results of claim 1 arranged in multiple stages for detecting the facial feature elements. In the method of detecting a human face area from moving picture or still image information, A first step of extracting a rectangular feature using a rectangular feature sample from the video or still image information; And a second step of creating a rectangular feature set by using a method of reconstructing weights of the rectangular feature set by using an evolution algorithm in the first step rectangular feature.  The method according to claim 8, The face detection method using the evolutionary algorithm, characterized in that for detecting the face or the facial feature element using the rectangular feature set of the two steps. The method according to claim 8, The face detection is a face detection method using an evolutionary algorithm, characterized in that for detecting the face of a plurality of angles by forming a learning image of the front face, the left face, the right face. The method according to claim 9, Face detection method using an evolutionary algorithm, characterized in that the first and second stage learning results of claim 1 arranged in multiple stages for the face detection. The method according to claim 8, And a rectangular feature set comprising weak classifiers satisfying a detection rate of 90% to 100% and a false detection rate of 0.05% to 0.5% using the evolutionary algorithm. The method according to claim 9, A face detection method using an evolutionary algorithm, characterized by detecting each feature in a detected face region by constructing a learning image using the face feature elements of eyes, nose, and mouth. The method according to claim 9, A face detection method using an evolutionary algorithm, characterized in that the first and second stage learning results of claim 1 are arranged in multiple stages to detect the facial feature elements.

Images