JPH06168318A - Face image recognizing device - Google Patents

Abstract

PURPOSE:To provide a face image recognizing device appropriate for the recognition of a face image. CONSTITUTION:A binarization part 1 binarizes multi-level face image data to be recognized obtained by photographing the face of a person to be recognized and inputted from a face image input part 8 by means of binarization OF ternary coding. A 1st calculation part 2 calculates a pair of the number of '1's continued in the prescribed determined direction of the generated binarization image data and the number of continued '0's adjacent to the '1'. Then the 2nd calculation part 3 calculates the previously determined number of feature data such as the average value or the like of the ratio of the number of continued '1's to the number of continued '0's adjacent to the '1's from the calculated pair. A judgement analyzing part 7 executes the judgement analysis of the feature data calculated by the 2nd calculating part 3 based upon reference data for judgement analysis which are previously registered in a dictionary 6 and identifies a person corresponding to the inputted face image data.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a face image recognition apparatus for performing image recognition on multi-gradation recognition target face image data.

[0002]

2. Description of the Related Art Television ratings mainly include household ratings and individual ratings, the former simply showing the viewing status of televisions installed in the households surveyed, while the latter It includes information about who is actually watching the households surveyed, and the demand for them is increasing in recent years.

By the way, methods for measuring the personal audience rating are roughly classified into active type and passive type.
The active type is a method in which the surveyee operates the pushbuttons assigned to him / herself at the start and end of viewing. Burden tends to be heavy.

On the other hand, the passive type is a system for automatically judging the person watching the television, and the burden on the person to be surveyed is remarkably reduced as compared with the active type.

[0005]

As described above, the passive type has an excellent merit that the subject to be surveyed has a small load in measuring the personal audience rating, but there are some problems to be solved in realizing it. One of them is a method of identifying a person who is watching TV.

[0006] That is, in general, as this type of identification method,
It is considered that a method of performing image recognition processing on the image data of the person watching the television with a camera and performing the image recognition processing is suitable. Image recognition processing is required, but unlike popular fields such as character recognition, little research has been done on face image recognition. In particular, a person's face cannot be identified unless it is observed in detail unlike characters, so it is difficult to binarize the face image and set the feature data.
The reality is that a recognition device effective for face images has not yet been put to practical use, which is a bottleneck in the practical use of passive personal audience rating measurement.

The present invention has been made in view of the above circumstances, and its purpose is to apply it to passive personal audience rating measurement, etc., which is devised in binarizing a face image and setting characteristic data. It is to provide a possible face image recognition device.

[0008]

A face image recognition apparatus according to the present invention includes a binarizing unit for binarizing image data by performing nOFm binarization on multi-gradation target face image data. Two
A first calculation unit that calculates a set of a continuous number of 1s and a continuous number of 0s adjacent to each other in a predetermined direction for the binarized image data generated by the binarizing unit; From the set of the number of consecutive 1's calculated by the calculator and the number of consecutive 0's,
A second calculator that calculates a predetermined number of feature data, a dictionary that holds reference data for discriminant analysis, and feature data calculated by the second calculator based on the reference data in the dictionary. And a discriminant analysis section for performing discriminant analysis of.

Further, as one of the characteristic data calculated by the second calculating section, an average value of the ratio between the number of consecutive 1s and the number of consecutive 0s is used.

Here, nOFm binarization means that all pixels of the face image data to be recognized are sorted in ascending order by gradation level, and the pixel having the smallest gradation level is arranged in order from the top (total number of pixels / m).
Each of the m pixel groups is extracted, and the first x pixel group and the last y pixel group (however, x, y ≧ 1, x + y = n)
It means that the levels of the pixels belonging to the above are unified to 0, and the levels of the pixels belonging to the remaining pixel groups are unified to 1, respectively.

As a direction for extracting a set of the number of consecutive 1's and the number of consecutive 0's adjacent to each other, the vertical and horizontal directions of the binarized image data, the method of diagonally lower right 45 degrees, and the diagonally lower left 45 degree. Any one or more of the directions of

[0012]

In the face image recognition apparatus of the present invention, the binarization unit performs nOFm binarization on the multi-gradation recognition target face image data of the face of the person to be recognized to obtain the binarized image. Data is generated, and then the first calculation unit calculates, for the generated binarized image data, a set of a consecutive number of 1 and an adjacent consecutive number of 0 in a predetermined direction, Second
The calculation unit calculates a predetermined number of characteristic data from the set of the calculated continuous number of 1s and the adjacent continuous number of 0s, and the discriminant analysis unit calculates the reference data stored in the dictionary. Based on the above, the discriminant analysis of the characteristic data calculated by the second calculation unit is performed.

[0013]

Embodiments of the present invention will now be described in detail with reference to the drawings.

Referring to FIG. 1, one embodiment of the face image recognition apparatus of the present invention is a face image input unit 8, a binarization unit 1, a first calculation unit 2, and a second calculation unit 3. A selector 4, a dictionary registration unit 5, a dictionary 6, and a discriminant analysis unit 7.

The face image input section 8 is composed of a video camera and a processing section for cutting out an image of a face portion from the output image thereof. In addition, you may make it perform the process which further removes a hair part from the cut-out face image.

The binarizing unit 1 performs 2OF3 binarization on the multi-gradation face image data input from the face image input unit 8, and the binary face image data in which the level of each pixel is 1 or 0. Is the part that generates. In the 2OF3 binarization, all pixels of the face image data are sorted in ascending order by gradation level, and the pixel having the smallest gradation level is extracted at the beginning (total number of pixels / 3) of three pixel groups. The first 1 pixel group and the last 1
This is performed by unifying the levels of the pixels belonging to each pixel group to 0 and the levels of the pixels belonging to one intermediate pixel group to 1.

Here, 2 is applied to multi-gradation face image data.
The OF3 binarization is performed in order to suppress the relative gradation level change of the image even if the brightness of the illumination changes when the face image to be recognized is input from a video camera. (2) Since simple binarization with a certain threshold tends to result in an image in which only the features of facial components such as eyebrows, eyes, under the nose, and mouth tend to be emphasized. In order to emphasize other components, it is possible to extract features of other parts such as cheeks and unevenness of the forehead. Is.

Other than 2OF3 binarization, other nOFm binarization such as 4OF5 binarization may be adopted.

Next, the first calculating unit 2 sets a predetermined 4 for the binarized image data generated by the binarizing unit 1.
With respect to the direction, a set of a continuous number of 1 and a continuous number of adjacent 0 is calculated.

For example, the set of the number of consecutive 1's in the vertical direction and the number of consecutive 0's for the binarized image data composed of 10 × 10 pixels as shown in FIG. 2 is calculated as follows. Is done in this way.

Assuming that the vertical columns of the binarized image data of FIG. 2 are line 1 to line 10 as shown in FIG.
The number of identical pixels from the first pixel to the last pixel of line1 is calculated. In the case of the figure, the result is, as in the lower part of FIG. 2, 7 consecutive 0s at the beginning, 2 consecutive 1s at the beginning, and 1 1 at the end. Similarly, the remaining line2 to line2
The results calculated for 10 are shown in the lower part of FIG.

Next, for each line, a set of a continuous number of 1s and a continuous number of adjacent 0s is extracted. The result is shown in FIG. In the case of this example, a total of 20 sets R _{1 to} R ₂₀ are extracted. It should be noted that there is no pair in the line 9 having only 0.

In the same manner as described above, regarding the horizontal columns of the binarized image data of FIG. 2, for the arrangement in the direction of the lower right diagonal 45 degrees, and for the arrangement in the direction of the lower left diagonal 45 degrees, the number of consecutive ones is 1 respectively. A pair of adjacent 0s is calculated.

The set of sets in each direction extracted as described above well reflects the characteristics of the binarized image data, that is, the original face image. Incidentally, a scatter diagram similar to the scatter diagram with frequency as shown in FIG. 4 in which each set R _{1 to} R ₂₀ in FIG. 3 is plotted by taking the continuous number of 0 on the vertical axis and the continuous number of 1 on the horizontal axis, As a result of creating several face images, it was confirmed that the scatter diagram of each person was different and the features of each face were well captured.

Since the set of groups shown in FIG. 2 captures the features of the face image of each person in this way, it is possible to use the set as feature data, but the number of sets is large. And the difference in the number of sets between different face images, etc.
In the present invention, the characteristic data is further extracted from the set of the above sets.

That is, the second calculating unit 3 in FIG. 1 uses the following combinations of the number of consecutive 1's in each direction calculated by the first calculating unit 2 and the number of consecutive 0's as follows. 9 pieces of feature data are calculated.

X1; average of the number of consecutive 1's X2; average of the number of consecutive 0's X3; ratio of the number of consecutive 1's to the number of consecutive 0's in all sets (number of consecutive 0's / number of consecutive 1's) Average X4; total number of pairs X5; total ratio of the number of consecutive 1's to the number of consecutive 0's in all pairs X6; even when the same consecutive number appears multiple times, only one appears Average of the number of consecutive 1's X7; Average of the number of consecutive 0's when it is considered that only one appears even when the same number of consecutive appears more than once X8; Only one when multiple similar sets appear Average of the ratio of the number of consecutive 1's to the number of consecutive 0's in all pairs when it is considered not to appear X9; Even when multiple similar pairs appear, only one is considered to appear Total number of cases

The above nine pieces of feature data X1 to X9
Is created for each direction, so in the case of 4 directions, a total of 3
Six pieces of characteristic data will be created. Hereinafter, each characteristic data will be described with reference to FIGS. 3 and 4.

The characteristic data X1 corresponds to each set R _{1 to} R _{20 in} FIG.
It is a value obtained by dividing the total sum of the number of 1s in 1 by the number of sets 20.

The characteristic data X2 is the sets R _{1 to} R _{20 of} FIG.
It is a value obtained by dividing the total sum of the numbers of 0 in the table by 20.

The characteristic data X3 is an average value of the ratio between the number of consecutive 1's in all the sets R _{1 to} R ₂₀ and the number of consecutive 0's.

The characteristic data X4 is the number of sets R _{1 to} R ₂₀ . That is, 20.

The characteristic data X5 is a value obtained by summing the ratios of the number of consecutive 1's of all the sets R _{1 to} R ₂₀ and the number of consecutive 0's. Therefore, there is a relationship of X5 / X4 = X3.

The characteristic data X6, X7, and X8 all correspond to X1, X2, and X3 in the scatter diagram without frequency (the figure in which the values of frequency 2 or more on the scatter diagram in FIG. 4 are all 1), and the characteristic data X9 corresponds to the total number of plots in the scatter plot without frequency.

In the above feature data, especially the feature data X3 and X8, since the size of the face portion in the recognition target face image data, in other words, the pixel density is almost invariable, The feature data is not affected by the reduction or expansion of image data. Therefore, if discriminant analysis is performed based on such feature data, image recognition that is strong against reduction and enlargement becomes possible.

Next, the selector 4 of FIG. 1 is a switching unit that outputs the feature data calculated by the second calculation unit 3 to the dictionary registration unit 5 at the time of so-called learning and outputs it to the discrimination analysis unit 7 at the time of actual recognition. is there.

Further, the dictionary registration unit 5 creates data as a reference of the discriminant analysis of the discriminant analysis unit 7 based on the characteristic data calculated by the second calculation unit 3 at the time of learning, and the dictionary 6 is created.
The discriminant analysis unit 7 performs the discriminant analysis of the characteristic data calculated by the second calculator 3 based on the reference data registered in the dictionary 6 at the time of actual recognition, and the discriminant analysis result is obtained. Is a means for outputting. This discrimination analysis unit 7
Various known discriminant analysis methods can be applied to the discriminant analysis method.

In the face image recognition apparatus of the present embodiment having such a configuration, first, the selector 4 is switched to the dictionary registration unit 5 side during learning, and the face image data of the person to be recognized is input to the face image input unit 8 It will be given sequentially from. This way
Each processing is performed by the binarization unit 1, the first calculation unit 2, the second calculation unit 3, and the dictionary registration unit 5, and the reference data necessary for the discriminant analysis of the face image of the person is registered in the dictionary 6. To be done.

When the registration of the reference data of all the necessary persons in the dictionary 6 is completed in this way, the preparation for actual recognition is completed, and the selector 4 is switched to the discriminant analysis section 7 for recognition. . At the time of actual recognition, when face image data of a certain person is input from the face image input unit 8, the binarization unit 1, the first calculation unit 2, and the second calculation unit 3 perform the respective processes. The discriminant analysis unit 7 performs discriminant analysis of the characteristic data calculated by the second calculation unit 3 based on the reference data of each person in the dictionary 6, and outputs the discrimination result indicating which person.

[0040]

According to the face image recognition apparatus of the present invention described above, the following effects can be obtained.

NO is applied to multi-gradation target face image data.
Since the Fm value conversion is performed, even if the brightness of the illumination changes when the face image to be recognized is input from a video camera, the relative change in the gradation level of the image is suppressed, and the influence of the illumination is reduced. You can In addition, simple binarization with a certain threshold tends to result in an image in which only the features of facial components such as eyebrows, eyes, under the nose, and mouth are emphasized. Since the features of other parts such as cheeks and foreheads are further extracted while emphasizing various constituent elements, the features of the individual can be more accurately captured and the recognition accuracy can be improved.

For the binarized image data, a set of a continuous number of 1s and a continuous number of 0s adjacent to each other in a predetermined direction is calculated, and then the calculated continuous number of 1s and adjacent 0s are calculated.
Since a predetermined number of feature data is calculated from the set of the continuous number of, the amount of calculation involved in the feature data calculation is relatively small, and the discriminant analysis is performed based on the limited number of feature data. , The overall calculation amount is also reduced. Further, since the set of the number of consecutive 1's and the number of consecutive 0's does not change even when the binarized image data is moved in parallel, the feature data obtained from them is not affected by the parallel movement. Face image recognition that is strong against movement is possible.

Particularly, the average value of the ratio of the number of consecutive 1's, which is one of the feature data, and the number of consecutive 0's, is the feature data that is not affected by the reduction or enlargement of the face image data. Strong face image recognition can be performed.

[Brief description of drawings]

FIG. 1 is a configuration diagram of an embodiment of the present invention.

FIG. 2 is an explanatory diagram of a method of calculating a set of a continuous number of 1s and a continuous number of 0s adjacent to each other from binary image data binarized by 2OF3 binarization.

FIG. 3 is a diagram showing an example of a set of a calculated continuous number of 1s and a continuous number of adjacent 0s.

FIG. 4 is a scatter diagram with frequency in which each set of FIG. 3 is plotted with the number of consecutive 0s on the vertical axis and the number of consecutive 1s on the horizontal axis.

[Explanation of symbols]

 1 ... Binarization unit 2 ... First calculation unit 3 ... Second calculation unit 4 ... Selector 5 ... Dictionary registration unit 6 ... Dictionary 7 ... Discrimination analysis unit 8 ... Face image input unit

Claims (2)

[Claims] 1. n for multi-gradation recognition target face image data
A binarization unit that performs OFm binarization to create binarized image data, and the binarized image data generated by the binarization unit is adjacent to a continuous number of 1 in a predetermined direction. Of a predetermined number from a pair of a first calculation unit that calculates the number of consecutive 0s and a number of consecutive 1s that are calculated by the first calculation unit A second calculation unit that calculates the characteristic data, a dictionary that holds the reference data of the discriminant analysis, and a judgment that performs the discriminant analysis of the characteristic data calculated by the second calculation unit based on the reference data in the dictionary. A face image recognition apparatus comprising: an analysis unit. 2. The face image recognition apparatus according to claim 1, wherein one of the feature data calculated by the second calculation unit is an average value of a ratio between the number of consecutive 1s and the number of consecutive 0s.