KR100981832B1 - Device and method for face features selection using boosting algorithm and feature vector segmentation - Google Patents

Abstract

Disclosed are a face feature selection method and apparatus using a boosting algorithm and a feature vector division technique. A face feature selection apparatus using a boosting algorithm and a feature vector division technique, the method comprising: receiving and storing a feature vector group, assigning a learning weight to each feature vector of the feature vector group, and at least one feature vector group Dividing into partial vector groups, learning a facial feature for each partial vector group, extracting a group having a minimum learning error, selecting a facial feature from the extracted partial vector group, and adding feature weights to the selected facial feature Giving and storing.

Face recognition, feature selection, boosting, feature vector segmentation, weight

Description

DEVICE AND METHOD FOR FACE FEATURES SELECTION USING BOOSTING ALGORITHM AND FEATURE VECTOR SEGMENTATION}

The present invention relates to a method and apparatus for selecting a facial feature using a BOOSTING algorithm and a feature vector segmentation technique. In particular, a facial feature is selected by performing a learning while repeatedly dividing a facial feature vector into a predetermined number of partial vectors. A selection apparatus and method are disclosed.

The face recognition system uses a pre-processing device that makes the system suitable for processing the input face image, a facial feature extraction device that extracts features from the preprocessed face image, and a face feature extraction process. A facial feature selection device that selects only useful features from the extracted features again, a learning device that learns a registered face with the selected features, and an input of a face image based on the learned content. It consists of a recognition device that finds a matching face by comparing the feature with the feature of the learned face image.

A number of facial features extracted from one face image by the face feature extraction device are expressed in a standardized vector form called feature vector, and each vector component constituting the feature vector corresponds to each feature in the face image. do. Therefore, facial feature vectors obtained from multiple face images have the same format but different vector component values. On the other hand, the face feature vector extracted from the face image at the beginning is generally very large in dimension so that the amount of computation is increased when it is used as it is, and unnecessary features act as a noise component, causing performance degradation of the entire face recognition system. do. Therefore, in order to prevent this, by selecting only some useful features of the extracted facial features, the dimension of the facial feature vectors can be reduced and used for face learning and recognition to improve the performance of the entire face recognition system.

Among the facial feature selection methods, the facial feature selection method using the Boosting algorithm learns each feature on the feature vector one by one and compares the learning errors. Therefore, there is a limit in selecting a face feature combination that is more optimal for face recognition in selecting a predetermined number of face features. Also, in the time required for face feature selection, all the features on the feature vector must be learned one by one in order to select one face feature. Do. Therefore, it takes quite a long time to select a feature, and this time is increased in proportion to the dimension of the feature vector.

Therefore, there is an urgent need for a facial feature selection device and method for screening an optimal facial feature combination more quickly and effectively.

The present invention is more useful for face recognition from a plurality of face features extracted from a face image by using a boosting algorithm and a feature vector division technique in the face feature selection device and method constituting a part of the overall face recognition system. The present invention provides a facial feature selection device and method for more effectively extracting a plurality of predetermined facial features to increase face recognition rate.

In the present invention, in the face feature selection apparatus and method for forming a part of the entire face recognition system, by using a boosting algorithm and a feature vector division technique, the processing time of the face recognition system is drastically reduced by reducing the time required for face feature selection. It provides a facial feature selection apparatus and method for improving the.

According to an aspect of the present invention, there is provided a method for selecting a facial feature, comprising: receiving and storing a feature vector group, assigning a learning weight to each feature vector of the feature vector group, and at least one partial vector of the feature vector group. Dividing into groups, learning facial features for each partial vector group, extracting a group having a minimum learning error, selecting a facial feature from the extracted partial vector group, and assigning feature weights to the selected facial features Storing.

According to an aspect of the present invention, the step of selecting the facial feature, determining whether each vector dimension of the group having the minimum learning error is less than or equal to a predetermined size, and if the learning dimension is greater than the predetermined dimension, the learning error is minimum And repeating the step of extracting the phosphorus group.

According to an aspect of the present invention, the method may further include determining whether the number of stored facial features is smaller than a preset number, and when the number of stored facial features is smaller than a preset number, assigning a learning weight to each feature vector. And repeating the step of assigning and storing a feature weight to the selected face feature.

In accordance with an aspect of the present invention, there is provided a face feature selection apparatus comprising: a weight adjusting unit configured to receive a feature vector group and assign a weight to each feature vector, a vector divider dividing the feature vector group into at least one partial vector group, A feature learning unit learning face features for each of the partial vector groups and generating a learning result, and extracting a group having a minimum learning error among the partial vector groups based on the learning results, and extracting a facial feature from the extracted partial vector group. A feature selection unit for selecting a feature and a feature storage unit for storing the selected face feature by weighting the selected feature;

According to an embodiment of the present invention, in the face feature selection apparatus and method for constituting a part of the entire face recognition system, the face is extracted from a plurality of face features extracted from the face image by using a boosting algorithm and a feature vector division technique. A facial feature selection apparatus and method for more effectively extracting a plurality of predetermined facial features that are more useful for recognition to increase face recognition rate are provided.

According to an embodiment of the present invention, in the face feature selection device and method constituting a part of the entire face recognition system, by using a boosting algorithm and a feature vector division technique, the time required for face feature selection is remarkably reduced. Provided are a facial feature selection device and method for reducing and improving the processing speed of the facial recognition system.

Hereinafter, with reference to the contents described in the accompanying drawings will be described in detail an embodiment according to the present invention. However, the present invention is not limited to or limited by the embodiments. Like reference numerals in the drawings denote like elements.

1 is a flowchart illustrating a process of selecting a facial feature using a boosting algorithm and a feature vector division technique according to an embodiment of the present invention.

Referring to FIG. 1, in step S110, a feature vector group consisting of a plurality of face feature vectors is received and stored. In this case, one facial feature vector may be extracted from one facial image, and N facial feature vectors may be extracted from N facial images to form a facial feature vector group.

In operation S120, a learning weight (first weight) is assigned to each vector in the facial feature vector group stored in operation S110. In this case, when the learning weight is initially assigned, the same learning weight is assigned to each feature vector, and if more than one facial feature is selected, the learning weight is assigned according to the learning error. In this case, the larger the learning error, the larger the learning weight may be given to allow more learning. Here, the learning is the same as the concept used in the existing boosting algorithm (ex. AdaBoost algorithm), and may mean a process of extracting prominent facial features. Likewise, 'learning error' is the same concept used in the existing AdaBoost algorithm, and may mean an error generated by testing a learned vector.

In step S130, the vector group is divided to generate a predetermined number of partial vector groups. That is, a partial vector group can be generated by dividing all the vectors in the vector group equally into a predetermined number of partial vectors and classifying them into partial vectors having the same vector component.

In step S140, the facial features are learned for each partial vector group and a learning error is calculated. That is, the learning error can be calculated by performing the learning independently for each subvector group.

In step S150, the least error group is selected by comparing the calculated learning errors of the partial vector groups. That is, the minimum error group may be selected by comparing the learning errors of each partial vector group calculated in step S140.

In step S160, it is determined whether the vector dimension of the selected group is smaller than or equal to a preset size. At this time, when the vector size of the selected group exceeds the predetermined size, steps S130 to S150 are performed again.

In operation S170, when the vector dimension of the selected group is equal to or less than a predetermined size, the characteristics corresponding to the vector component are selected, and in operation S180, a feature weight (second weight) is assigned to the selected feature and stored. That is, the learning error may be calculated based on the selected feature and the learning result, and the feature weight may be assigned to the selected feature and stored. In this case, the feature weight may be regarded as a more important feature and may be given a relatively larger feature weight as the learning error of the selected feature is smaller.

In step S190, it is determined whether the number of stored features is smaller than the predetermined number. If the number of stored features is greater than or equal to the predetermined number, the selection of the features is terminated.

That is, by selecting the feature feature by applying the feature vector division technique to the boosting algorithm as described above, the face feature is extracted more effectively, the face recognition rate is increased, and the processing speed of the face recognition system is generated.

FIG. 2 is a diagram illustrating a configuration of a face feature selection device using a boosting algorithm and a feature vector division technique according to an embodiment of the present invention.

Referring to FIG. 2, a facial feature selection apparatus using a boosting algorithm and a feature vector division technique includes a weight controller 220, a vector divider 230, a feature learner 240, a feature selector 250, and a feature store. The unit 260 is included.

The weight adjusting unit 220 receives and stores a feature vector group consisting of a plurality of face feature vectors extracted from a plurality of face images, and outputs a learning weight to each feature vector in the group. At this time, the learning weight is initially assigned the same weight to all the vectors, and each time the face feature is selected once, the learning result is received from the feature storage unit 260 to calculate the learning error of each vector, Relatively larger learning weights can be assigned to allow more learning.

The vector divider 230 receives one vector group from the weight adjuster 230 or the feature selector 250, divides all the vectors in the group equally into a predetermined number of partial vectors, and has the same vector component. By classifying by vector, a predetermined number of partial vector groups are generated and transmitted to the feature learner 240.

The feature learner 240 independently learns a predetermined number of partial vector groups received from the vector divider 230 and outputs the partial vector groups to the feature selector 250 together with the respective learning results. .

The feature selector 250 receives the partial vector groups and the respective learning results from the feature learning unit 240, calculates a learning error based on the received learning results, and compares only the partial vector groups having the minimum learning error. Take and exclude the rest of the group. When the vector size of the partial vector group thus taken is larger than the preset size, the vector size is transferred to the vector divider 230 so that the vector is divided again. The result is output to the feature storage unit 260 together with the result.

The feature storage unit 260 receives the selected features and learning results from the feature selection unit 250, calculates a learning error based on the learning results, assigns and stores a feature weight value to the selected features, and stores the number of stored features. The learning result is transmitted to the weight adjusting unit 220 until the number is reached. At this time, the feature weight value may be regarded as a more important feature and given a relatively larger weight as the learning error of the selected feature is smaller.

FIG. 3 is a diagram for comparing and comparing a conventional technology with a vehicle space recognition success rate according to an embodiment of the present invention.

Referring to FIG. 3, a recognition rate for the same human car space and another car space is shown according to the number of features selected from the feature vectors. Recognition rate of the same human car space is significantly higher than that of using only boosting algorithm (312) when the feature vector segmentation technique is applied to the boosting algorithm (312), while the recognition rate of car space between others is boosting algorithm. When the feature vector segmentation technique is applied together (321), which is slightly lower than the case of using only the boosting algorithm (322), the recognition rate is higher on average when the feature vector segmentation technique is applied to the boosting algorithm. Seems.

4 is a view for comparing and comparing a conventional technology with a vehicle space recognition failure rate according to another embodiment of the present invention.

Referring to FIG. 4, the recognition failure rates for the same human car space and each other car space are shown according to the number of selected features. In general, biometric fields such as face recognition are related to security in terms of their characteristics and use false-positive rate and false-negative rate as one of the indicators. A positive error occurs when the same person is not recognized as the same person. On the contrary, a voice error is an error caused by mistaken different people as the same person. Therefore, the recognition error rate of the same human car space as indicated by reference numeral 410 may be viewed as a positive error rate, and the recognition error rate of the human car space as indicated by reference numeral 420 may be regarded as a voice error rate.

As shown by reference numeral 420, when the feature vector segmentation technique is applied to the boosting algorithm (421) and when only the boosting algorithm is used (422), the voice error rate is close to 0% and is higher than the positive error rate indicated by reference numeral 410. It is low. When the feature vector segmentation technique is applied, the speech error rate is slightly higher, but the positive error rate is significantly lower than when the feature vector segmentation technique is applied to the boosting algorithm (411) than when only the boosting algorithm is used (412). Recognition failure rate is lower.

FIG. 5 is a diagram illustrating a comparison between a conventional technology and a feature selection time in one embodiment of the present invention.

Referring to FIG. 5, the time required for each feature selection when the feature vector division technique is used in the boosting algorithm and when it is not used is illustrated. Here, since feature selection is a process that is performed simultaneously without distinguishing between the same human car space and another car space, the time required for feature selection is also expressed as a single chart without distinction according to the car space. ) Is significantly shorter than the case of applying only the boosting algorithm (502). At this time, as the number of selected features increases, the difference in the time required becomes larger, and when selecting 200 features, the speed is improved by about four times.

As described above, when the feature vector division technique is used, the recognition rate of the same person is significantly higher than that of the boosting algorithm alone, and the positive error rate is significantly lowered. In addition, although the voice error rate was slightly increased, the average face recognition rate was reached, which means that more useful facial features could be selected. In addition, although the average recognition rate improved, the time required for feature selection was rather reduced. In other words, applying the feature vector segmentation technique to the face feature selection process together with the boosting algorithm improves not only the face recognition rate but also the processing speed for feature selection.

As described above, the present invention has been described by way of limited embodiments and drawings, but the present invention is not limited to the above-described embodiments, which can be variously modified and modified by those skilled in the art to which the present invention pertains. Modifications are possible. Accordingly, the spirit of the present invention should be understood only by the claims set forth below, and all equivalent or equivalent modifications thereof will belong to the scope of the present invention.

1 is a flowchart illustrating a process of selecting a facial feature using a boosting algorithm and a feature vector division technique according to an embodiment of the present invention.

FIG. 2 is a diagram illustrating a configuration of a face feature selection device using a boosting algorithm and a feature vector division technique according to an embodiment of the present invention.

FIG. 3 is a diagram for comparing and comparing a conventional technology with a vehicle space recognition success rate according to an embodiment of the present invention.

4 is a view for comparing and comparing a conventional technology with a vehicle space recognition failure rate according to another embodiment of the present invention.

FIG. 5 is a diagram illustrating a comparison between a conventional technology and a feature selection time in one embodiment of the present invention.

<Explanation of symbols for the main parts of the drawings>

210: feature vector group 220: weight control unit

230: vector division unit 240: feature learning unit

250: feature selection unit 260: feature storage unit

Claims (5)

In the facial feature selection method using a boosting algorithm, Receiving a feature vector group including one or more facial feature vectors; Assigning a first weight to each facial feature vector of the feature vector group; Generating one or more partial vector groups by dividing the facial feature vector to which the first weight is assigned into a plurality of partial vectors and classifying the plurality of partial vectors into partial vectors having the same vector component; Performing learning, which is a process of extracting prominent facial features for each partial vector group; Extracting a partial vector group having a minimum learning error generated by testing a learned vector which is a prominent facial feature extracted for each partial vector group according to the performance of the learning; When the vector dimension of the extracted partial vector group exceeds 1, the step of generating the at least one partial vector group or extracting the partial vector group having the minimum learning error is performed again, and the vector dimension is 1 Selecting one facial feature from the vectors in the extracted partial vector group; Assigning and storing a second weight which is a value determined by using the first weight to the selected facial feature; And Determining whether the number of stored facial features is smaller than a preset number; Including, The determining step, If the number of stored facial features is less than a preset number, repeating the step of assigning the first weight to the storing of the second weight until the stored facial features is equal to or greater than the preset number. Facial feature selection method. delete delete The method of claim 1, The step of giving a first weight to each facial feature vector, Equal weight is given to each of the facial feature vectors, The step of storing by giving a second weight, Face feature selection method characterized in that the weight is given in consideration of the magnitude of the learning error. In the facial feature selection apparatus using a boosting algorithm, A weight adjusting unit configured to receive a feature vector group including one or more face feature vectors and to assign a first weight to each face feature vector; A vector dividing unit for dividing the first feature weighted face feature vector into a plurality of partial vectors and classifying the plurality of partial vectors into partial vectors having the same vector component to generate at least one partial vector group; A feature learning unit that performs learning, which is a process of extracting a prominent face feature for each partial vector group; Feature selection for extracting a group having a minimum learning error among the partial vector groups based on a learning result which is a prominent face feature extracted for each partial vector group according to the performance of the learning, and selecting a facial feature from the extracted partial vector group part; And A feature storage unit for storing the selected face feature by giving a second weight value which is a value determined by using the first weight value. Facial feature selection device comprising a.

Images