KR20160110741A - Device and method for human age estimation - Google Patents

Abstract

The present invention relates to an apparatus and a method for age estimation and, more specifically, to an apparatus and a method for age estimation, which increase age estimation performance by considering the quality of an inputted face image. The apparatus for age estimation includes: a face area detection unit detecting a face area in the inputted face image; an image quality measurement unit measuring the image quality of the detected face area; and an age estimation unit calculating an age characteristic of the face area, and estimating an age by using support vector regression (SVR) data trained by groups according to the image quality.

Description

≪ Desc / Clms Page number 1 > DEVICE AND METHOD FOR HUMAN AGE ESTIMATION &

The present invention relates to an age estimation apparatus and method, and more particularly, to an age estimation apparatus and method for improving the age estimation performance in consideration of the quality of an input facial image.

People change their faces as they get older. Based on these characteristics, many computer vision systems have been developed to perform a variety of specific tasks. For example, age information has been used in e-commerce systems to acquire information such as customer's taste, preference, and expectation, and has also been used in security control and surveillance systems. In addition, the age estimation technique of a person can be applied to various fields such as a method of estimating the age of each person by detecting the face region in the input face image, and the provision of age-based advertisement,

On the other hand, the age estimation technique can obtain a good performance estimation result when a clear and focused face image is used. However, when the user uses a system with actual age estimation technology, the age estimation performance may be degraded due to the degradation of the image quality due to the shooting environment and the user's motion.

Background Art [0002] The background art of the present invention is disclosed in Korean Patent Publication No. 2008-0012488 (published on Feb. 12, 2008, apparatus and method for visual age measurement).

The present invention provides an age estimation apparatus and method for accurately estimating the age even when the focus of a face image is blurred.

The objects of the present invention are not limited to the above-mentioned objects, and other objects not mentioned can be clearly understood from the following description.

According to an aspect of the present invention, an age estimating apparatus is provided.

According to an aspect of the present invention, there is provided an age estimating apparatus including a face region detecting unit for detecting a face region in an input face image, an image quality measuring unit for measuring an image quality of the detected face region, And an age estimation unit for estimating an age using SVR (Support Vector Regression) data trained in groups according to image quality.

According to another aspect of the present invention, there is provided an age estimation method and a computer program for executing the method.

A method for estimating an age and a computer program for executing the same according to an embodiment of the present invention includes detecting a face region in an inputted face image, measuring an image quality of the detected face region, and calculating an age characteristic of the face region , And estimating the age using Support Vector Regression (SVR) data trained in groups according to image quality.

The present invention extracts age characteristics using Multi-level Local Binary Pattern (MLBP) and Gabor filtering and applies the dimensionality of age characteristics by applying Principal Component Analysis (PCA) The accuracy of the age estimation is improved.

The present invention classifies the face image into one of the groups classified according to the image quality based on the measured face score, and uses it for the grouped data of the group to further improve the accuracy of the age estimation.

The present invention estimates two ages when the measured image quality corresponds to a portion overlapping two groups and averages the two age values by mixing, so that the classification error of the group is alleviated and the accuracy of the age estimation is further improved .

1 is a view for explaining an age estimation system according to an embodiment of the present invention;
2 is a block diagram of an age estimating apparatus according to an embodiment of the present invention.
3 to 12 are views for explaining a method of estimating an age according to an embodiment of the present invention.
13 to 17 are views for explaining the effect of the age estimation method according to an embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily carry out the present invention. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Also, when a part is referred to as "including " an element, it does not exclude other elements unless specifically stated otherwise.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Reference will now be made in detail to embodiments of the present invention, examples of which are illustrated in the accompanying drawings.

1 is a view for explaining an age estimation system according to an embodiment of the present invention.

Referring to FIG. 1, an age estimation system according to the present invention includes an age estimation apparatus 100 and a facial image input apparatus 200.

The age estimation apparatus 100 measures the image quality (Focus Score) in the face region of the input image and estimates the age using Support Vector Regression (SVR) data that is trained for each group classified according to the image quality. The age estimating apparatus 100 extracts age characteristics through Multi-level Local Binary Pattern (MLBP) and Gabor filtering, processes PCA (Principal Component Analysis) of extracted age characteristics, do. This will be described in more detail in Fig.

The age estimation apparatus 100 includes a processor for performing image processing, an input / output interface for input / output of image data, and a memory for storing data and programs. The age estimation apparatus 100 may be, for example, a server computer, a personal computer, an electronic device including a processor, an input / output interface, and a memory, such as a notebook, a tablet, a pad and a smart phone.

The face image input apparatus 200 inputs an image of an object including a face region. The facial image input apparatus 200 may include, for example, a camera.

2 is a configuration diagram of an age estimating apparatus according to an embodiment of the present invention.

2, the age estimation apparatus 100 includes an image input unit 110, a face region detection unit 120, a face image correction unit 130, an image quality measurement unit 140, and an age estimation unit 150 do.

The image input unit 110 inputs an image including a face area.

The face region detection unit 120 detects a face region from the input face image. The face area detecting unit 120 may use the AdaBoost method for detecting the face area.

The face image correcting unit 130 corrects the face image and redefines the face region. The face image correction unit 130 performs in-plane rotation correction by measuring a slope of a straight line connecting the centers of the detected eyes. In addition, the face image correcting unit 130 redefines the face area in order to minimize the unnecessary background factor for the age estimation.

The image quality measurement unit 140 measures the image quality (FS) of the face region. The image quality measurement unit 140 extracts high frequency component values extracted from the face region by applying a convolution mask to the images including the face region and calculates the image quality using the nonlinear normalization function for the extracted high frequency component values. The image quality measuring unit 140 classifies the calculated image quality into at least one of a group of blurred images of a preset image.

The age estimating unit 150 calculates the age characteristics of some regions using the age characteristic and Gabor filtering using the MLBP of the face region, and combines the calculated characteristics. The age estimating unit 150 normalizes the combined characteristics, reduces the dimension through the PCA, and estimates the age using SVR data separately grouped according to the image quality (FS)

FIGS. 3 to 12 are views for explaining an age estimation method according to an embodiment of the present invention.

Referring to FIG. 3, in step S310, the age estimation apparatus 100 inputs a face image.

In step S320, the age estimation apparatus 100 detects a face region and an eye region in the input face image. The input face image generally includes a face area as well as a background area. Accordingly, the age estimating apparatus 100 first detects and defines the face region. The age estimation apparatus 100 may use an AdaBoost method for detecting a face region. The Adsboost method has good detection results and is suitable for real-time image processing. The age estimation apparatus 100 performs eye detection together with face detection. The OpenCV package can be applied to perform these functions. The age estimating apparatus 100 may further use at least one of a subblock-based template matching method and an adaptive template matching method other than the AdaBoost method, for example .

In step S330, the age estimation apparatus 100 corrects the face image and redefines the face area.

The age estimation apparatus 100 corrects the rotated face region due to the human's natural face posture. This is because if the face is not aligned, the result of age estimation may be bad. The age estimating apparatus 100 measures an inclination of a straight line connecting the centers of the detected eyes and performs an in-plane rotation correction by the angle of Equation 1 below.

Equation 1

? = tan ^-1 ((Ry-Ly) / (Rx-Lx))

Where (Rx, Ry) is the position coordinate of the right eye, (Lx, Ly) is the position coordinate of the left eye.

Thereafter, the age estimation apparatus 100 redefines the face area to minimize unnecessary background elements for age estimation. 4, the age estimating apparatus 100 redefines the horizontal and vertical sizes of the face ROI region 440 based on the d value obtained by measuring the distance between the eyes 420 and 430 in the detection region 410 do. In FIG. 4, the values k1 , k2, and k3 can be experimentally set to define a face region.

In step S340, the age estimation apparatus 100 measures a focus score on a face area.

Referring to FIG. 5, in step S342, the age estimation apparatus 100 extracts a high frequency component value by applying a convolution mask to a face region.

First, the age estimating apparatus 100 generates an observation image by convoluting a PSF to an image in which a face area is defined as shown in Equation 2 below.

Equation 2

g (x, y) = (f * h) (x, y)

Here, gIx, y represents an observation image, f (x, y) represents an image in which a face region is defined, and h (x, y) represents a point-spread function of blurring.

The age estimation apparatus 100 extracts high frequency component values extracted from the face region by applying a convolution mask in the observed image. Here, the convolution mask may be, for example, three separate element masks 610 of FIG. 6 or a load mask of 620, or the like.

The age estimation apparatus 100 calculates the image quality using the nonlinear normalization function of Equation (3) below for the extracted high frequency component values.

Equation 3

FS = 100 x (x ² / (x ² + C ² ))

Where FS is the image quality (Focus Score), x is the high frequency component value, and c is used to cancel the normalization with a normalization offset value representing the slope of the nonlinear normalization function, which is determined experimentally

Referring to FIG. 7, an example of the image quality of the face region image according to the degree of blur of the image can be seen. As shown in Fig. 7, a blurred image is calculated to be lower than that of the other image quality (FS).

In step S346, the age estimating apparatus 100 classifies the blur of a preset image into at least one of the groups according to the calculated image quality. Referring to FIG. 12, the age estimation apparatus 100 can be divided into, for example, a blurred image group 1210, a slightly blurred image group 1220, and a grouped image 1230, which are image quality evaluation criteria. The age estimation apparatus 100 also includes a first overlapping group 1240 and a slightly blurry image group 1220 overlapping between the blurred image group 1210 and the slightly blurred image group 1220, A second overlapping group 1250 that overlaps the first overlapping group 1250. [

In step S350, the age estimation apparatus 100 estimates the age according to the measured image quality. This will be described in detail below with reference to FIG.

Referring to FIG. 8, in step S351, the age estimation apparatus 100 divides the detected face region into a plurality of blocks, and extracts image pixel values for each block as histogram characteristics.

9, the age estimation apparatus 100 divides a redefined face ROI region into subblocks, such as 910, and extracts characteristics using a local binary pattern (LBP) do. Here, 920 corresponds to the first block when the image of 910 is divided into 4 × 4, and if the histogram is generated according to the pixel value of the corresponding block, the histogram characteristic can be expressed as 930. The age estimation apparatus 100 combines the histogram characteristics of the number of subblocks in this manner to generate one histogram characteristic 940. [

Thereafter, in step S352, the age estimating apparatus 100 fuses a plurality of single-level LBP characteristics having different numbers of subblocks divided, and calculates a sum of the number of subblocks for the entire area by the MLBP (Multi-level Local Binary Pattern) Creates texture properties. Referring to FIG. 10, the MLBP characteristic generation of multi-level has a plurality of single-level LBP characteristics having different radial parameters R, the number of peripheral pixels P, and the number of sub-blocks M Respectively. That is, the multi-level local binary pattern (MLBP) property can collect age characteristics at resolutions of other scales R and LBP.

In step S353, the age estimation apparatus 100 In parallel with step S352, texture characteristics of a partial area such as a corrugation of the face area are extracted through Gabor filtering. Here, the Gabor filter uses eight kinds of directions and four kinds of scales. The age estimating apparatus 100 defines a plurality of corrugated regions based on the eye position to extract corrugation characteristics of a certain region and extracts high-frequency components by culling Gabor filtering on the corrugated regions defined. Then, the age estimation apparatus 100 uses the average and standard deviation of the filtered image as extracted wrinkle characteristics. Referring to FIG. 11, the age estimation apparatus 100 may define five corrugated regions based on the eye position.

In step S353, the age estimation apparatus 100 The characteristics of the MLBP extracted in step S352 and the characteristics of the Gabor filtering extracted in step S353 are normalized through Equation 4 and combined through Equation 5 to obtain final characteristics .

Equation 4

fi * = (fi-mu) / sigmai

Here, fi represents feature values extracted through MLBP or Gabor filtering, and μi and σi mean the mean and standard deviation of feature values, respectively.

Equation 5

_{f = [f 1 *, f} 2 *]

In step S355, the age estimation apparatus 100 Collaborative properties are reduced by using principal component analysis (PCA). This is because the LBP feature increases the dimension of the dimension as well as the continuous LBP feature dimension is included in the MLBP feature, because the face region is divided into several regions when generating the LBP feature. When the dimension of the data is high, the age estimating apparatus 100 can take a long time to process the data and can estimate the age with low accuracy.

In step S356, the age estimating apparatus 100 The age is estimated from the dimensionally reduced characteristics using the image quality regression method.

The age estimating apparatus 100 generates training data by training data to form a relationship between age characteristics and verification age. The age estimation apparatus 100 can select a kind of a kernel, a parameter thereof, a support vector suitable for an input value and a verification age by using the generated training data, store training data and estimate an age for a new input value thereafter When used. The age estimation apparatus 100 can use, for example, the LibSVM software package for application of the regression method.

Referring to FIG. 12, the age estimation apparatus 100 can be divided into, for example, a blurred image group 1210, a slightly blurred image group 1220, and a grouped image 1230, which are image quality evaluation criteria. There are three SVR (Support Vector Regression) training data according to each image quality evaluation group.

The age estimating apparatus 100 estimates the age by using the SVR training data corresponding to 1210 and the SVR training data corresponding to 1220 when the image quality corresponds to the first overlapping group 1240 region, The average value can be used as the result value. Similarly, when the image quality corresponds to the second overlapping group 1250, the age estimating apparatus 100 estimates the age using the SVR training data corresponding to 1220 and the SVR training data corresponding to 1230, Can be used as a result value.

13 to 17 are diagrams for explaining the effect of the age estimation method according to an embodiment of the present invention.

Referring to FIG. 13, the age estimation apparatus 100 may use a PAL database to evaluate the performance of age estimation. The PAL database contains approximately 580 face images including each gender, different facial expressions and other racial faces, and includes facial images from 18 to 93 years old.

The age estimation apparatus 100 generates three kinds of experiment databases by applying a Gaussian blurring corresponding to sigma values of 0 to 5 to each face image of the PAL database as shown in FIG. Here, the larger the sigma value, the more blurred the face image. The experimental database includes a focused database 1310 with sigma values of 0 and 1, a slightly blurred database 1320 with a system value of 2, and a blurred database with a sigma value of 3 to 5 ) 1330. < / RTI >

Referring to FIG. 14, the age estimating apparatus 100 measures the age estimation performance of the age estimation method according to the present invention, which includes the MLBP method and the Gabor filtering method, by using the MAE (Mean Absolute Error) Respectively. Here, the MAE measures the mean error of the estimated age compared to the verification data, and the lower the value, the better the age estimation performance. The age estimation method 1430 according to the present invention calculates the age of the user by performing only the MLBP and the MAE is calculated to be lower than the method 1410 of extracting the age characteristics or 1420 of extracting the age characteristics by performing only Gabor filtering, The estimated performance is considered to be better.

Referring to FIG. 15, the age estimating apparatus 100 measures the age estimation performance of the age estimation method according to the present invention in which the data dimension is reduced through PCA processing on an experimental database using MAE (Mean Absolute Error) . The age estimation method 1510 according to the present invention is considered to have lower MAE and better age estimation performance than the age estimation method 1520 without PCA processing.

Referring to FIG. 16, the age estimating apparatus 100 calculates the age of the age estimation method according to the present invention including the first overlapping region 1610 and the second overlapping region 1620, The estimated performance was measured using Mean Absolute Error (MAE). The age estimation method 1610 according to the present invention calculates MAE lower than the age estimation method (FIG. 14) that does not include the first superposition group area 1610 and the second superposition group area 1620, It is judged to be better.

Referring to FIG. 17, the estimated age of the age estimation method (Estimated age 1) according to the present invention and the estimated age of the method (Estimated age 2) not including the technical features of the present invention of FIGS. Compared with age. Here, 1710 is a clear image, 1720 and 1730 are slightly blurry images, and 1740 to 1760 are blurry images. The estimation method according to the present invention (Estimated age_1) is closer to the actual age than the other estimation methods (Estimated age_2). Therefore, it can be seen that the estimation method of age according to the present invention has a better estimation performance than the estimation method (Estimated age_2) which does not include the technical features of FIGS. 14 to 16 of the present invention.

The method of estimating the age according to various embodiments of the present invention may be implemented in the form of program instructions that can be executed through various means such as servers. Further, a program for executing the multiple input image correction method according to the present invention may be installed in a computer means and recorded in a computer-readable medium. The computer readable medium may include program instructions, data files, data structures, and the like, alone or in combination.

Program instructions to be recorded on a computer-readable medium may be those specially designed and constructed for the present invention or may be available to those skilled in the computer software arts. Examples of computer-readable media include magnetic media such as hard disks, floppy disks, and magnetic tape; optical media such as CD-ROMs and DVDs; magnetic media such as floppy disks; Includes hardware devices specifically configured to store and execute program instructions such as magneto-optical media and ROM, RAM, flash memory, and the like.

The embodiments of the present invention have been described above. It will be understood by those skilled in the art that the foregoing description of the present invention has been presented for illustrative purposes and that those skilled in the art will readily understand that various changes in form and details may be made therein without departing from the spirit and scope of the invention. It will be possible. It is therefore to be understood that the above-described embodiments are illustrative in all aspects and not restrictive. The scope of the present invention is defined by the appended claims rather than the detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalents should be construed as being included in the scope of the present invention.

100: Age estimation device
200: facial image input device

Claims (16)

In the age estimation apparatus,
A face region detection unit for detecting a face region in the input face image;
An image quality measuring unit for measuring an image quality of the detected face area; And
And an age estimator for calculating an age characteristic of the face region and estimating an age using Support Vector Regression (SVR) data trained for each group according to the image quality.

The method according to claim 1,
And a face image correcting unit for correcting the face image and redefining the face area.
The method according to claim 1,
Wherein the image quality measuring unit extracts a high frequency component value by applying a convolution mask in an image including the face region and calculates the image quality using a nonlinear normalization function for the high frequency component value.
The method of claim 3,
Wherein the image including the face region is an image generated by convoluting a point-spread function of blurring (PSF) on an image in which the face region is defined.
The method according to claim 1,
Wherein the age estimating unit calculates an age characteristic of a partial region using overall age characteristics and Gabor filtering using a multi-level local binary pattern (MLBP) of a face region, and calculates an age characteristic by combining the calculated characteristics.
6. The method of claim 5,
Wherein the age estimating unit normalizes the combined characteristics, reduces the dimension through PCA (Principal Component Analysis), and estimates the age using SVR data separately trained according to the image quality.
The method according to claim 1,
The group classified according to the image quality may include a blurred image group, a slightly blurred image group, and a group with a clear image according to a preset reference,
A first region in which the blurred image group and the slightly blurred image group overlap, a second region in which the slightly blurred image group and the blurred image group overlap each other,
Estimating the age of the image quality corresponding to the first region as an average value of the estimated age of the blurred image group and the slightly blurred image group, A blurred image group, and an estimated value of the age of each of the blurred image group and the blurred image group.
In the age estimation method,
Detecting a face region in the input face image;
Measuring an image quality of the detected face region; And
Calculating an age characteristic of the face region, and estimating an age using Support Vector Regression (SVR) data trained in each group according to the image quality.
9. The method of claim 8,
And correcting the facial image and redefining the facial region.

9. The method of claim 8,
The step of measuring the image quality of the detected face region
Extracting a high frequency component value by applying a convolution mask to the face region;
Calculating an image quality using the nonlinear normalization function on the extracted high frequency component values; And
And classifying the blur of the image into at least one of the groups according to the calculated image quality.

11. The method of claim 10,
The step of calculating the image quality using the nonlinear normalization function for the extracted high frequency component value
A method for estimating an age using Equation (3).
Equation 3
FS = 100 x (x ² / (x ² + C ² ))
Here, FS is the image quality (Focus Score), x is the high frequency component value, and c is the normalization offset value indicating the slope of the nonlinear normalization function.
9. The method of claim 8,
The step of calculating the age characteristics of the face region and estimating the age using Support Vector Regression (SVR) data trained for each group according to the image quality
Dividing the detected face region into a plurality of blocks and extracting histogram characteristics of image pixel values per block;
Extracting a total age characteristic by MLBP (Multi-level Local Binary Pattern) generated by combining the histogram characteristics; And
Extracting an age characteristic of a partial region of the face region through Gabor filtering;
And calculating an age characteristic of the face region by normalizing and combining the overall age characteristic and the partial region age characteristic.

13. The method of claim 12,
Wherein the age characteristic of the partial region of the face region is a wrinkle characteristic, and a region is selected based on the position of the eye.
13. The method of claim 12,
And performing PCA (Principal Component Analysis) on the age characteristics of the face region.
9. The method of claim 8,
The step of estimating the age using SVR (Support Vector Regression) data trained for each group according to the image quality
Setting a blurred image group, a slightly blurred image group, and an image into a clear group according to a preset reference;
Setting a first region in which the blurry image group and the slightly blurry image group are overlapped with a second region in which the blurry image group and the image are clearly overlapped; And
Estimating the age of the image quality corresponding to the first region as an average value of the estimated age of the blurred image group and the slightly blurred image group, Estimating an average value of a blurred image group and an estimated age of each of the blurred image group and the blurred image group.
17. A computer program for implementing the method of estimating age according to any one of claims 8 to 15.

Images