KR20060057110A - Apparatus and method for detecting and authenticating facial data - Google Patents

Abstract

The present invention relates to a face detection and authentication apparatus and a method thereof, and more particularly, to store face data in a server and to extract face features and extract facial features from a mobile wearable terminal, and to detect face data detected by the server. And a face detection and authentication device and a method for comparing the stored face data and outputting the presence or absence of face authentication to the personal wearable terminal. The present invention has a detection unit for detecting a face and a storage unit for storing a detected face image and an detection algorithm in a wearable personal station (WPS) terminal, which is a personal wearable terminal, and stores a face image and a face authentication algorithm in a server. And a processing unit for comparing and detecting the detected image and the stored image to detect a face and extract a feature of a video, and display only authentication, a login function, and a face for the individual in the server by placing a lot of individuals or more users. It provides an efficient solution for implementing video conferencing and identification of the other party.

Description

Apparatus and method for detecting and authenticating facial data

1 shows a prototype of a Haar-like feature applied to the present invention.

FIG. 2 is a view for explaining each step grouping using the AdaBoost algorithm applied to the present invention.

3 is a flowchart of face detection using a Halle feature for face detection according to the present invention.

4 is a diagram illustrating a generation method and probability calculation principle of face data using the HMM algorithm for face authentication according to the present invention.

5 is a view for explaining the training process and the database of the face data for face detection and authentication according to the present invention.

6 is a block diagram of a face detection and face authentication process according to the present invention.

7 is a view for explaining the face authentication results according to the present invention.

8 is a conceptual diagram illustrating a face detection and authentication process according to a preferred embodiment of the present invention.

<Description of main parts of drawing>

60 Videos 62 Face Detection

64 ... block extractor 66 ... feature extractor

71, 72 Similar model calculator 76 Maximum probability selector

78.Face authentication 80 ... Personal mobile terminal

82 Camera 86 Face Detection Wireless Transmission and Authentication Output

90 ... server

The present invention relates to a face detection and authentication apparatus and a method thereof, and more particularly, to store face data in a server, detect a face for a video in a personal wearable terminal, and detect the face data and stored face data in the server. The present invention relates to a face detection and authentication device for outputting the presence or absence of face authentication to the personal portable wearable terminal and a method thereof.

In image processing and interpretation, the human face is an important factor for visual discrimination and identification. Since the early 1990s, analyzes of face recognition and facial expression interpretation have been extensively developed. Recently, MPEG-7 face descriptors have been proposed for face searching and identification in the flow of video. In contrast to conventional face recognition algorithms, the main function of the face descriptor is to retrieve face images that are identical to the query image as quickly and effectively as possible.

Face recognition technology refers to a technology for verifying the identity of one or more faces present in a still image or video using a given face database. Unlike other biometrics, such as fingerprint recognition, face recognition technology does not require direct contact with a part of its body to the recognition device, and the method of obtaining biometric information is less mandatory. The recognition rate is lower than other biometric systems because the change is severe and very sensitive to the surrounding environment.

Conventional technologies for detecting and recognizing faces include Daewoo Electronics Co., Ltd.'s "Personal Recognition Method Using Image Processing of Human Face" (Publication No. 10-1995-29985), "Face Component Extraction System" of Korea Electronics and Telecommunications Research Institute And its extraction method (Publication No. 10-1998-20738), "Real-time face tracking method using the color model of the face and the ellipsoid approximation model and its recording medium" (Publication No. 10-2000-60745), Samsung Electronics Co., Ltd. "Face Recognition Method and Apparatus Using Dimensional Spatial Transformation of Gabor Filtering Response Values" (Publication No. 10-2001-87487), but all of them face limitations in face recognition rate and are not practical.

The technical problem to be achieved by the present invention is to detect a face using a fast face detection algorithm applicable to a video stored on a personal wearable terminal, the detection using the face database and facial authentication algorithm stored in the server in the personal wearable terminal It is an object of the present invention to provide a face detection and authentication device for comparing and authenticating the face data stored in the server with the face data stored in real time.                         

Another technical problem to be solved by the present invention is to detect a face using a fast face detection algorithm applicable to a video stored in a personal wearable terminal, and detected by the personal wearable terminal using a face database and a face authentication algorithm on a server. Another object is to provide a face detection and authentication method for comparing and authenticating face data and face data stored in the server in real time.

According to an aspect of the present invention, in the face detection and authentication apparatus, a storage unit for storing face data and a face detection algorithm and a detection unit for detecting a face for an input video and outputting parameters of feature values Personal portable wear terminal comprising; And a server for storing a face database and a face authentication algorithm, and a server for comparing the detected face data with face data stored in the face database according to a parameter of the feature value and outputting an authentication status. And an authentication device.

In order to achieve the above technical problem, the present invention provides a face detection and authentication method comprising: storing a face database and a face authentication algorithm in a server; Detecting a face for a video input by a face detection algorithm stored in a personal wearable terminal and outputting a parameter of a feature value; And comparing the face data detected from the personal portable terminal with the face data stored in the face database to the server and outputting the presence or absence of the authentication according to the parameter of the feature value.

Preferably, the face detection uses a Haar-like feature prototype and an AdaBoost learning algorithm to cover the face area with a window, and then calculates the Halle feature for the area. The face is detected by grouping and storing the values.

Preferably, the face authentication is characterized by using the HMM algorithm (Hidden Markov Model algorithm) modeled using the features of the face and by comparing the modeled face image to the image of the stored face sequentially from top to bottom to authenticate.

More preferably, the HMM algorithm consists of an extension of a block of a face, each of which is divided into a state for each super state, calculates each probability, and then calculates a probability for the whole. Further characterized in that it further comprises a secondary HMM is passed to.

Hereinafter, with reference to the accompanying drawings will be described in detail the configuration and operation of the preferred embodiment of the apparatus and method for face detection and authentication of the present invention.

In the field of face recognition technology, face detection algorithm using haar-like feature has advantages of very fast training time and processing speed among many related algorithms. Therefore, it can be particularly useful for face detection in moving pictures. The facial data detected in this way is compared with the face database already trained using the Hidden Markov Model (HMM) algorithm to implement an identification system that authenticates the person with the highest probability of facial recognition as his face. Can be.

In the identification system, both the face detection rate and the success rate of identification are determined by the learning process, so the performance depends on how much learning is done efficiently. These systems can be identified by simply showing their face to the camera, providing a very convenient feature compared to other system architectures that require a cumbersome identification assumption.

In the present invention, the apparatus and method for face detection and authentication are implemented by using the Halle feature and the HMM algorithm. In the present invention, face authentication based on the Hidden Markov Model (HMM) is performed as one of methods for increasing the recognition rate. In the HMM technique, a face image is processed by dividing it into parts, and in the present invention, a face extraction method for constructing input image data is extracted using a haar-like feature (see paper: P). Viola and MJ Jones, "Robust real-time object detection", Technical Report Series, Compaq Cambridge research Laboratory, CRL 2001/01, Feb. 2001).

In other words, the present invention compares a face data extracted using a haar-like feature with a face database that has already been trained using an HMM algorithm, and outputs the person with the most probable face as an output. It uses authentication, and it raises such a face detection rate and recognition rate economically.

First, face detection applied to the present invention will be described.

Haar-like features and AdaBoost learning algorithms are one of the face detection algorithms applied to the present invention. Haar-like features are a common technique used for face searching, and there are many prototypes. These prototypes use more efficient faces because they use only the faces that are well represented by the AdaBoost learning algorithm. Detection is possible (Reference papers: Sunghoon Park, Jaeho Lee, Hoeul Kim, "Face Recognition Using Haar-like Feature / LDA", 2004 16th Workshop on the Processing and Understanding, January 2004).

Haar-like features are viola's first recognizer for face detection, and the method is simple and uses only simple concatenation in the calculation method. It is easy to detect the face in video by calculating the sum of the area by adding the values of the pixels in each area using the feature, not the operation using the pixel to detect the face. Can be used. 1 is a prototype of a haar-like feature, and since the position is changed according to the feature in the window, it is possible to display many feature values according to an image to be recognized. 1 are prototypes of a Halle feature where 1 represents an edge feature, 2 represents a line feature, and 3 represents a center feature.

In the present invention, the recognition rate is improved by using an AdaBoost learning algorithm when detecting a face. In the present invention, a simple and efficient AdaBoost learning algorithm is used among boost algorithms in order to distinguish between a face and a non-face area and increase a probability of judging by a face. This algorithm has the advantage that the position of prototypes of each Haar-like feature can be as fine as the grouping stage. The feature values obtained by the AdaBoost learning algorithm are stage grouped as shown in FIG. 2. In FIG. 2, the entire sub-window 10 determines the authenticity at node 1 and proceeds to node 2 if true (T), and rejects the secondary window (20). Similarly, it proceeds to nodes 2, 3, 4, and so on (30).

This grouping is for implementing a more powerful recognition algorithm. As the number of steps is repeated, more groups of feature values are created and grouped. In the present invention, the grouping is performed up to 25 levels, and the number of feature values is 200. .

Hereinafter, a face detection method using a haar-like feature will be described with reference to FIG. 3.

Firstly, image data of the face image is received and a window is put on the face part, and then a haar-like feature value is obtained for the area. The selected Haar-like features are then grouped and stored using the AdaBoost learning algorithm. At this point, the first nine groups contain the least nine Haar-like features. Each stage is grouped into a total of 25 stages, with 200 haar-like features. The grouped data is saved as a text file and used as a hidden cascade for face recognition. In the present invention, the size of the window is 24 * 24 size, and the number of haar-like features increases as the level goes up to 25 levels, and each prototype is located in the detail. : Alexander Kuranov, Rainer Lienhart, and Vadim Pisarevsky, "An Empirical Analysis of Boosting Algorithms for Rapid Objects With an Extended Set of Haar-like Features", Intel Technical Report MRL-TR-July 02-01, 2002).

The principle of face detection using such a hake feature will be described.

Figure 3 is a simplified representation of the principle of face detection using the Halle feature. That is, by using the group of faces 300 trained in one frame of image, the image is reduced in the form of a pyramid structure to determine an area that is a face, and in the process, a plurality of candidate areas are used to restore an image whose face is reduced. Is outputted, so that an area obtained by averaging the area is outputted (steps 1 to 5). In this case, the trained face is classified into a face and a non-face using a face classifier to detect the face.

In the present invention, the face thus detected is compared with the face database stored using the HMM to perform face authentication.

The HMM algorithm uses the Markov property that the future depends only on the present regardless of the past in the past and present, so that the Markov state is hidden and the process is observed only through other probabilities. Papers: F. Samaria and S. Young, "HMM based architecture for face identification", Image and Computer Vision, vol. 12, pp. 537-543, October 1994).

In face recognition, HMM algorithm recognizes face features. The biggest feature of these faces is head, forehead, eyes, nose, and mouth. These features are modeled using the primary HMM (1D-HMM). Each state is dependent on the process of the HMM and the order is applied sequentially from top to bottom of the face image. In addition, the interval between dividing the face is set in a fixed position in the image of the face only. This primary HMM has a recognition rate of about 85% (Ref .: AV Nefian and MH Hayes, "A Hidden Markov Model for face recognition", in ICASSP 98, vol. 5, pp. 2721-2724, 1988; AV Nefian and MH Hayes, "Face detection and recognition using Hidden Markov Models", in International Conference on Image Processing, 1988).

Extensions to primary HMMs can be extended to pseudo 2D-HMMs (see S. Kuo and O. Agazzi, "Keyword spotting in poorly printed documents using pseudo 2-D Hidden Markov Models", IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 16, pp. 842-848, August 1994). Each of these 2D-HMMs consists of an extension of the block of the face, which is divided into a predetermined state 44 for each super state 42, and the respective probabilities are calculated again. After calculating the probabilities, we move on to the next step. In Figure 4 shows the progression from the face to the forehead, eyes, nose, mouth, chin sequentially.

5 is a view for explaining the training process and the database of the face data for face detection and authentication according to the present invention.

In FIG. 5, the block extractor 510 extracts a block by receiving the training image 500 for the N-th person, and the feature extractor 520 receives the extracted block by using features O ₁ ^N , O ₂ ^N , and the like. ..., O _J ^N ) is output to the initial parameter evaluator 540 and the Baum-welch parameter reevaluator 550. Meanwhile, the initial parameter evaluator 540 receives the HMM model λ ₀ from the prototype HMM model unit 530 and evaluates the feature sent from the feature extractor 520 to calculate the parameter evaluation result λ _init . The parameter is sent to the parameter reevaluation unit 550. The Baum-Welch parameter reevaluator 550 transmits the features O ₁ ^N , O ₂ ^N ,..., O _J ^N sent from the feature extractor 520 and the initial parameter evaluator 540. After receiving the parameter evaluation result (λ _init ), the trained face image (λ ^N ) is output and stored in the face database.

6 is a block diagram of a face detection and face authentication process according to the present invention. In FIG. 6, when the authentication is desired, a face area is automatically detected in a video using a haar-like feature and stored in a BMP format. At this time, one face and 6 images per face are used as data, and a face database to be authenticated is added to the face database by combining the database of previously collected faces. In FIG. 6, it is intended to automatically extract and store a face area of a person to be added from a video rather than a picture. The detected face image is outputted through the block extractor 64 and the feature extractor 66, and the parameter is compared with the parameters stored in the face database by the similar model calculator 71,..., 72. By selecting the most similar image in the maximum probability selection unit 80 to output the authentication through the output unit (90).

Table 1 shows the trained location information and thresholds for prototypes of Haar-like features at 24 * 24 window size. As a result of verifying the facial authentication system, the result shown in Table 2 was obtained.

extern const char * FaceCascade [] = {/ * Stage 0 * / "9 1 2 6 4 12 9 0 -1 6 7 12 3 0 3 haar_y3 -0.031512 0 -1 2.087538 -2.217210 1 2 6 4 12 7 0 -1 \ n "" 10 4 4 7 0 3 haar_x3 0.012396 0--1.863394 1.327205 1 2 3 9 18 9 0 -1 3 12 18 3 0 3 \ n "" haar_y3 0.021928 0 -1 -1.510525 1.062573 1 2 8 18 9 6 0 -1 8 20 9 2 0 3 haar_y3 \ n "                                                  ... "1 2 5 8 12 6 0 -1 5 11 12 3 0 2 haar_y2 0.002734 0 -1 -1.691193 0.440097 1 2 11 14 4 10 0 -1 \ n" "11 19 4 5 0 2 haar-y2- 0.018859 0 -1 -1.476954 0.443501 1 2 4 0 7 6 0 -1 4 3 7 3 0 2 \ n "" haar_y2 0.005974 0 -1 -0.859092 0.852556 -5.042550 \ n ", / * Stage 1 * /" 16 1 2 6 6 12 6 0 -1 6 8 12 2 0 3 haar_y3 -0.021110 0 -1 1.243565-1.571301 1 2 \ n "" 6 4 12 7 0 -1 10 4 4 7 0 3 haar_x3 0.020356 0 -1 -1.620478 1.181776 1 2 1 8 19 12 0 -1 \ n "                                                  ... "9 5 2 10 0 3 haar_x3 -0.007507 0 -1 0.927821 -0.663499 1 2 10 8 6 10 0 -1 12 8 2 10 0 3 \ n" "haar_x3 -0.009877 0 -1 1.157747 -0.297748 1 2 2 5 4 9 0 -1 4 5 2 9 0 2 haar_x2 0.015814 0 -1 \ n "                                                  ...                                             

 Number of experiments  Total experimental population  Certification Success  Certification failure  Certification success probability  Certification Failure Probability  10th  48 people  46 people  2 people  95.65%  4.35%

7 is a view for explaining the face authentication results according to the present invention. The data consisted of 8 faces and 48 images of 6 each, and a database of 48 faces was constructed by combining the database of previously collected faces (40 faces and 240 images of 6 each). . As shown in FIG. 7, the features of the wrongly verified faces can be seen that the forehead part of the face is wider than the other person's face because the forehead part is wider in common. This is because face segmentation for training with 2D-HMM. In the process, the eye position is out of the area.

The face detection and authentication process of the present invention has been described so far, but with reference to a wearable personal station (WPS), a personal wearable terminal of FIG. 8 to which the present invention is applied, according to a preferred embodiment of the present invention Describes the face detection and authentication process.

A wearable personal station (WPS) as a personal wearable terminal of FIG. 8 includes a detection unit for detecting a face and a storage unit for storing a face image and a face detection algorithm in the terminal, and includes a face database and a face authentication algorithm in a server. And a processing unit for storing and comparing the image detected by the terminal with the image stored in the server. The WPS 80 receives a face image input from a built-in camera 82 and detects a face by a haar-like feature (84). The detected face image is transmitted to the server 90 via Bluetooth or wireless Internet. In this case, as described above, a number of face images are trained and classified using a Haar-like feature and an AdaBoost learning algorithm, which are face detection algorithms stored in the WPS 80, and each trained. With 25 groups of steps, the higher the level, the finer the face area mask is generated, the image to be processed is received, and the image is reduced to detect the face that matches the face area mask. The detected face compares facial parameters by using the HMM algorithm, which is a face authentication algorithm stored in the server 90, and the face database of a plurality of people, and completes authentication (86).

As described above, the present invention has described a system for authenticating a face using a haar-like feature and an HMM. Such a system is very useful for extending the face authentication method for still images to perform face detection and authentication in real time video. By using the Haar-like feature, it is easy to apply to a video due to the fast face detection time, and by using the detected face region data directly to the HMM, it is possible to quickly detect and detect a face in a video image requiring real time processing. Authentication is now possible. The improvement of the authentication rate can be sufficiently improved by applying the HMM after the face detection in the future to further refine the position of the eyes and mouth, and to use the skin color detection.

As described above, the apparatus and method for detecting and authenticating a face of the present invention automatically detects only a face area in a terminal, wirelessly transmits a parameter of a feature value to a server, and then outputs the presence or absence of authentication after comparing with a face database in the server. Face detection and facial feature extraction for video can be performed on the terminal, and multiple data can be stored on the server, which can be used by individuals or multiple users for authentication, login function, video conferencing with only faces, and identification of the other party. Provide an efficient solution to implement

Claims (8)

In the face detection and authentication device, A personal portable wear terminal comprising a storage unit for storing face data and a face detection algorithm and a detection unit for detecting a face of an input video and outputting a parameter of a feature value; And And a server including a storage unit for storing a face database and a face authentication algorithm and a processing unit for comparing the detected face data with face data stored in the face database according to the parameter of the feature value and outputting an authentication status. Authentication device. The method of claim 1, wherein the face detection comprises using a Haar-like feature prototype and an AdaBoost learning algorithm to cover the area of the face, and then obtain a Halle feature value for the area. And detecting the face by grouping and storing the Halle feature values. According to claim 1 or 2, wherein the face authentication is modeled using the facial features through the HMM algorithm (Hidden Markov Model algorithm) by comparing the modeled face image to the image of the stored face sequentially from top to bottom Face detection and authentication apparatus characterized in that the authentication. The method of claim 3, wherein the HMM algorithm consists of an extension of a block of a face. The face detection and authentication device further comprises a second HMM to be passed to the next process. In the face detection and authentication method, Storing a face database and a face authentication algorithm in a server; Detecting a face for a video input by a face detection algorithm stored in a personal wearable terminal and outputting a parameter of a feature value; And Comparing the face data detected from the personal wearable terminal with the face data stored in the face database in the server and outputting the presence or absence of authentication to the personal wearable terminal according to the feature value parameter. And authentication method. The method of claim 5, wherein the face detection is performed using a Haar-like feature prototype and an AdaBoost learning algorithm to cover the area of the face, and then obtain a Halle feature value for the area. And detecting the face by grouping and storing the Halle feature values. According to claim 5 or 6, The face authentication is modeled using the facial features through the HMM algorithm (Hidden Markov Model algorithm) by comparing the modeled face image to the image of the stored face sequentially from top to bottom Face detection and authentication method characterized in that the authentication. The method of claim 7, wherein the HMM algorithm consists of an extension of a block of a face. The method further comprises a second HMM to be passed to the next step.

Images