KR100852630B1 - Biometric method using probabillistic access in video image - Google Patents

Abstract

The present invention relates to a biometric method using a probabilistic approach in a video image, and more particularly, to accumulate a feature acquired in a video image in a stochastic manner, thereby improving performance for proper biometric recognition even when the feature is not accurately acquired. And a biometric method using a probabilistic approach in video images based on the same.

To this end, the present invention comprises the steps of designing a feature extractor for feature extraction; Extracting a feature into an image using the feature extractor; Obtaining each probability from successive images using a stochastic approach and accumulating and combining the respective probabilities; There is provided a biometric method using a probabilistic approach in a video image comprising the step of learning a classifier using the probability obtained through the feature extraction.

SVM, MLP, Probabilistic Approach, Biometrics

Description

Biometric method using probabillistic access in video image}

1 is a schematic representation of a representative SVM classifier in two dimensions;

2 is a structural diagram of an MLP classifier;

3 is a schematic diagram illustrating a method of accumulating and combining information of consecutive images according to a preferred embodiment of the present invention.

4 is a schematic diagram illustrating a biometric system using an integrated feature extractor and a binary classifier to learn conventional classifiers by extracting features.

5 is a schematic diagram showing a biometric system using a conventional integrated feature extractor and a multiple classifier,

6 is a schematic diagram illustrating a biometric system using an individual feature extractor and a binary classifier in accordance with a preferred embodiment of the present invention.

7 is a graph comparing the results of FIGS. 4, 5, and 6;

8 is a chart comparing the results of FIGS. 4, 5, and 6;

9 is a structural diagram showing a biometric system using the side profile,

10 is a structural diagram showing a biometric system using the gait,

11 is a graph showing the results of a general biometric system,

12 is a graph showing the results of a biometric system using a probabilistic approach in accordance with a preferred embodiment of the present invention.

FIG. 13 is a diagram illustrating a result of comparing a biometric system with a general biometric system using a probabilistic approach according to a preferred embodiment of the present invention.

The present invention relates to a biometric method using a probabilistic approach in a video image, and more particularly, to accumulate a feature acquired in a video image in a stochastic manner, thereby improving performance for proper biometric recognition even when the feature is not accurately acquired. And a biometric method using a probabilistic approach in video images based on the same.

Biometric systems developed to date are classified according to methods of obtaining data of a part of the body.

To date, the most commonly used method is to obtain images of body parts and extract and compare features. Fingerprint recognition, face recognition, retinal recognition, iris recognition, vein recognition, palmprint recognition, etc. Recognition uses the method of using human voice.

The biometric system includes an input unit for acquiring an image or voice, and a processing unit for extracting features from the input data and comparing the features with the features stored in the database to identify an identity.

For example, Japanese Patent Laid-Open No. 10-2001-0042659 (face recognition from a video image) includes a method for recognizing an object in an image frame, wherein the object is detected in the image frame, and the image frame is associated with the object. Delimiting a portion of; Generating a converted image by transforming a boundary portion of the image frame using a wavelet transform; LOCATING nodes (NODEs) related to the distinguishing feature of the object defined by the wavelet jet of the collective graph BUNCH GRAPH generated from a plurality of representative object images. step; And identifying the object based on a similarity between the wavelet jet associated with the object image in the gallery of object images and the wavelet jet at the node NODE of the transformed image. And an image portion not included in the boundary portion of the image frame does not perform a group graph analysis.

In addition, Korean Patent Laid-Open Publication No. 10-1998-0058349 (Person Identification Method Using Image Information) discloses a person identification method using image information in an image recognition system, wherein the person identification method includes facial feature information for identification. A facial feature information registration mode and a facial feature information recognition mode for identifying a person using the registered facial feature information rate, the facial feature information registration mode comprising: a first step of separating a face part from a face image; Dividing the separated face part into a plurality of cells to create a first template including a face shape in each cell; A third extracting a portion having the highest similarity with the first set of templates as the face region for the input image by comparing the first template with the similarity with the first set of templates classified according to the K-means clustering; step; A fourth step of creating a second template including a face shape in each cell by dividing the extracted face region into a plurality of cells; A fifth step of classifying and storing the second template into a second set of templates according to K-means clustering; And a sixth step of extracting facial feature data from the extracted face region and registering it in a database.

In general, in a pattern recognition system, it is desirable to reduce the size of data for efficient storage of input data and real-time transmission to a pattern classification stage. In this case, only information necessary for classification is extracted from the original data.

In other words, feature extraction can be seen as converting high-dimensional data into low-dimensional data.

In general, classification using a small number of features other than the original data in a pattern classification problem results in poor classification accuracy, but using a large number of features in a real pattern recognition system is a classification cost. It is very inefficient in terms of cost.

Therefore, in order to implement an efficient pattern recognition system, the minimum features should be used, and the loss of information necessary for classification should be minimized.

1 is a diagram showing a support vector machine (SVM) as an example of a representative classifier, which was originally developed for binary classification, and now bioinformatics, character recognition, handwriting recognition, face and object. It has been successfully applied in various fields such as recognition.

In addition, there is a classifier using a neutral network such as MLP (Multi Layer Perceptron) of FIG. 2.

However, since the conventional biometric system using statistical characteristics only uses the current single frame information, a problem that adversely affects the monitoring result may not be obtained because the perfect feature cannot be extracted according to the lighting and noise image changes during the image acquisition process of the current frame. do.

The present invention has been made in view of the above, and by employing a probabilistic method of accumulating image information, even if some of the images obtained from the video image are wrongly extracted, it is possible to accurately recognize them without being affected. The purpose is to provide a biometric method using a stochastic approach in video images.

In the present invention for achieving the above object in the biometric method of extracting features from the image data to perform biometrics,

Designing a feature extractor for feature extraction; Extracting a feature into an image using the feature extractor; Obtaining each probability from successive images using a stochastic approach and accumulating and combining the respective probabilities; And learning a classifier using the probability obtained through the feature extraction.

In a preferred embodiment, using the probabilistic approach, each probability is obtained from successive images, and the cumulative result is obtained.

이고,

ego,

I _t is the t-th image, p (w = n ⁱ | I _t ) is the probability that the t-th image belongs to the n ⁱ classification, I _{1: t} is the set of 1st to tth images, p (w = n ⁱ I _{1: t} ) is a cumulative probability obtained from the first to tth images.

In a more preferred embodiment, the accumulated probability is

의 log-odds를 이용하여 구해지는 것을 특징으로 한다.

It is characterized by using the log-odds.

In particular, the step of learning the classifier using the probability obtained through the feature extraction is characterized in that it comprises the step of making and using a separate feature extractor for each classification.

In addition, the step of learning the classifier using the probability obtained through the feature extraction is characterized in that it comprises using a binary classifier as the classifier, learning a binary classifier for each classification.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.

The present invention is a method of fusing accumulated image information by a probabilistic approach to develop a suitable biometric system in a video image, and extracts a suitable feature representing a given situation into an image and obtains a prior probability with a pre-trained classifier. Obtaining a post probability.

The method of fusing the accumulated image information by the probabilistic approach is a technique of combining the prior probabilities by combining the prior probabilities by Bayes' theorem as shown in Equation 1 and then combining the information.

That is, when a video image is observed, it is not a single image but a probability is obtained from successive images and accumulated by combining them.

In order to use the probabilistic approach, the prior probabilities of each successive image must be determined. In other words, the results of each image from the classifier must be transformed probabilisticly.

Equation 2 is used to determine the probability of outputs from classifiers such as SVM and MLP.

Where I _t is t th represents the image, this is a t-th image (w = n ⁱ │I _t) represents the probability p n belongs to category ^i, f _classifier (I _t) is the output of the SVM and MLP.

When the prior probability of each successive image is obtained, it is represented as a set as shown in Equation (3).

Where I _{1: t} represents the set of images from the 1st to the tth.

The probability of belonging to the n ⁱ classification through the continuous image information is expressed as in Equation 4.

In Equation 4, since I _t is only affected by w regardless of the value of I _{1: t-1} , it can be replaced with Equation 5.

Applying to Bayes' theorem, it is as shown in Equation 6.

Equation 6 is applied to Equation 5 to derive Equation 7.

In the opposite case, Equation 8 is derived through the same derivation process.

In Equations 7 and 8, p (I _t ) and p (I _t │I _t-1 ) are difficult problems to calculate, so dividing Equations 7 and 8 by p (I _t ) and p Eliminate (I _t | I _{t -1} ). In this way, it can be represented by Equation (9).

Equation 9 is again developed to derive Equation 10.

Taking logarithm in Equation 10, log-odds are derived as in Equation 11.

When the equation 11 is expanded again, it is derived as shown in equation 12.

이다. 수학식 12에서 보면 이전 log-odds 값만을 사용하여 l_t(w=nⁱ)를 도출할 수 있게 된다. 이런 방법을 통해 nⁱ분류에 속할 확률을 구하게 되고, 결국 nⁱ분류에 속할 확률이 사후확률이 된다. Where l ₀ (w = n ⁱ ) is the initial log-odds value,

to be. In Equation 12, it is possible to derive l _t (w = n ⁱ ) using only previous log-odds values. In this way the chance of being rescued belong to classify n ^i, n ⁱ finally a chance to belong to this classification is posterior.

Hereinafter, the present invention will be described in more detail based on the following examples, but the present invention is not limited by the following examples.

Example

The preferred embodiment according to the present invention applies information fusion using the proposed probabilistic approach to the development of biometric systems. The data used in the experiment is a video image of the person's side and gait, and a given database consists of side profile and gait information.

Then, the feature information of side profile and gait was extracted and used as learning data and experimental data of classifier.

At this time, the method of extracting feature information and learning the classifier uses the new learning method as shown in FIG. 6 along with the general method as shown in FIGS. 4 and 5.

In other words, the existing method uses the same feature extractor for several classifications, but the newly used method creates a feature extractor for each classification, and uses this to extract the features to learn the classifier.

The feature extractor can be used for both commonly used independent component analysis (ICA) and principal component analysis (PCA).

In this case, when the binary classifier is used as the classifier, the binary classifier must be learned for each classification, and when the multiple classifier is used, only one classifier is used. From the result of the trained classifier, the biometric recognition is performed through Equation 13.

7 and 8 are graphs and charts comparing the results of experiments of the conventional biometric methods of the present invention.

Of the four methods shown in FIGS. 7 and 8, Standard ICA + SVM and Standard ICA + MLP show the results of existing feature extractors and classifiers, and Individual ICA + SVM and Individual ICA + MLP are the probability of the present invention. The results of the feature extractor and classifier are shown when the adaptive approach is applied.

The ICA is a feature extractor, and a standard ICA means a feature extractor that learns using several individuals, and means that a feature is extracted by learning for each individual ICA. MLP is also a multiple classifier.

7 and 8, the method using the individual ICA than the standard ICA + SVM and Standard ICA + MLP shows a better result. In FIG. 8, the cumulative number of frames is fixed to three frames, and the recognition rates for each individual object in four situations are shown.

On the other hand, the case where the feature extraction is not performed correctly is shown in FIG. 11, FIG. 12, and FIG.

That is, although the feature extraction is not performed in the third and fifth frames of FIGS. 11 and 12, it is not properly recognized in FIG. 11 without using a probabilistic method, but is properly recognized in FIG. 12. Can be.

Therefore, although the method of performance is different for each individual, the recognition rate is on average, but the biometric system in which the cumulative image information is converged using the probabilistic approach according to the present invention does not have better feature extraction than the conventional system. It also shows superior performance.

This is because the cumulative information is used, and even if a feature that is incorrectly extracted comes in temporarily, a serious error can be avoided by fusion with existing information.

9 and 10 show the structure of the biometric system using the gait and the side view, and FIG. 9 extracts the side view feature information (Side View Face Detection) and recognizes it through a previously learned classifier (Side View Face Recognition). 10 is a method of extracting gait feature information (Gait sihouette detection) and recognizing it through a previously learned classifier (Gait sihouette recognition).

Although the present invention has been illustrated and described with respect to specific preferred embodiments, the invention is not limited to these embodiments, and the invention is claimed in the claims by one of ordinary skill in the art to which the invention pertains. It includes all embodiments of the various forms that can be carried out without departing from the spirit of the invention.

As described above, according to the biometric method using the probabilistic approach in the video image according to the present invention, by developing a biometric system that is most suitable for the video image by accumulating information using the stochastic approach, user identification There is an effect that can be done efficiently.

Claims (5)

delete Designing a feature extractor for feature extraction; Extracting a feature into an image using the feature extractor; Obtaining each probability from successive images using a stochastic approach and accumulating and combining the respective probabilities; Learning a classifier using the probability obtained through the feature extraction; In the biometric method using a probabilistic approach in a video image comprising a, Using the probabilistic approach, each probability is obtained from successive images and accumulated

ego, I _t is the t-th image, p (w = n ⁱ | I _t ) is the probability that the t-th image belongs to the n ⁱ classification, I _{1: t} is the set of 1st to tth images, p (w = n ⁱ I _{1: t} ) is a biometric method using a probabilistic approach in a video image, characterized in that the cumulative probability obtained from the first to t-th image. The method according to claim 2, The cumulative probability is

Biometric method using a stochastic approach in a video image, characterized in that is obtained using the log-odds of. The method according to claim 3, Learning the classifier using the probability obtained through the feature extraction comprises the step of making and using a separate feature extractor for each classification, biometric method using a probabilistic approach in a video image. The method according to claim 4, Learning a classifier using the probabilities obtained through the feature extraction includes using a binary classifier as the classifier and learning a binary classifier for each classification, using a probabilistic approach in a video image. Recognition method.

Images