JP2002342760A - Facial picture processor and facial picture processing method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a facial picture processor for which a large capacity storage device is not required. SOLUTION: A facial picture processor is provided with a face detecting part 13 for detecting the position of one's face from a picture read by picture inputting equipment 11, a facial picture normalizing part 14 for normalizing the picture on the basis of the position of one's face, a feature value calculating part 15 for calculating a feature value from the normalized picture, a dictionary preparing part 16 for preparing a facial picture dictionary from the feature value, and a recognizing part 20 for recognizing the facial picture by comparing the feature value calculated by the feature value calculating part 15 with the facial picture dictionary.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a face image processing apparatus and method, and more particularly, to a face image processing apparatus and method for automatically creating a face image dictionary from face images and recognizing the face images.

[0002]

2. Description of the Related Art With the spread of computers in recent years, many methods for performing personal authentication using face images have been studied.

[0003] For example, "Kazuhiro Hotta, Takio Kurita, Takenori Mishima," A Recognition Method Insensitive to Changes in Object Size Using Spectral Features Extracted from Log-Polar Images ", IEICE Technical Committee. Report PRMU99-
110 (1999-11) "(hereinafter referred to as" prior art 1 "), the input image is Log-Polar converted,
A method of using a spectral feature extracted from a Log-Polar image for face image recognition has been reported.

[0004] Also, "Sei Kuriyama, Kazuhiko Yamamoto, Hitoshi Hongo, Kunito Kato," Proposal of face recognition method using higher-order surface features ", IEICE Technical Report, PRMU 99-115 (199)
9-11) "(hereinafter referred to as" prior art 2 ").
Face image recognition is performed using higher-order local autocorrelation features, four-way face features, higher-order face features, and the like.

Further, "Masato Kawade, Seiji Hosoi," Robust Face Recognition Technology in Real Environment and Its Application ", Japan Opportunity Society I
In the IP2000 Information / Intelligence / Precision Instruments Division Lecture Paper [2000.3.27, 28] (hereinafter referred to as “Prior Art 3”), a feature amount is calculated using Gabor Wavelet transform, and graph matching is performed. To recognize a face image.

[0006]

However, in prior arts 1 and 2, since the number of dimensions of the feature vector is affected by the size of the image, the number of dimensions generally tends to increase. Therefore, the amount of calculation at the time of recognition and dictionary generation increases, and the size of the dictionary becomes huge. Further, the prior art 3 has a problem that the calculation amount is large because the calculation of the feature amount is a convolution operation. Therefore, these existing methods are applied to PDAs (Personal Di-Os) that cannot be equipped with a high-speed arithmetic device, a large-capacity external storage device, or a large-capacity memory.
It is difficult to use it in a portable information terminal device such as a digital assistant or a mobile phone.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problems, and an object of the present invention is to provide a face image processing apparatus and method which do not require a large-capacity storage device.

Another object of the present invention is to provide a face image processing apparatus and method capable of high-speed processing.

[0009]

A face image processing apparatus according to an aspect of the present invention includes image input means for inputting a face image, and is connected to the image input means, and detects a position of the face from the input face image. Face detection means for detecting, face image normalization means connected to the face detection means for normalizing the face image based on the position of the face detected by the face detection means, and face image normalization means And Z from the normalized face image
a feature amount calculating means for calculating an ernik moment and calculating a feature amount based on the Zernike moment; and a dictionary is created from a plurality of feature amounts obtained from a plurality of face images, connected to the feature amount calculating means. And a dictionary creating means for the user.

[0010] A dictionary is created from feature quantities based on Zernike moments. Therefore, a low-dimensional feature amount vector can be created, and the dictionary size can be reduced. Therefore, a PDA in which the face image processing apparatus cannot be mounted with a large-capacity external storage device or memory.
(Personal Digital Assista
nt) or a portable information terminal such as a mobile phone.

Preferably, the face image processing apparatus further comprises:
A face image recognition unit connected to the feature amount calculation unit and the dictionary creation unit, for comparing the feature amount obtained by the feature amount calculation unit with the dictionary created by the dictionary creation unit, and performing face image recognition; Including.

A face image can be recognized using a low-dimensional feature vector. For this reason, face images can be recognized at high speed.

[0013] More preferably, the feature value calculating means is configured to obtain Ze determined for each annular region when a reference circle provided in the normalized face image is divided by a plurality of concentric circles.
The feature amount is calculated based on the rnike moment.

A Zernike moment is determined for each annular region. Therefore, a plurality of feature amounts can be calculated from one image, and highly accurate face image recognition can be performed.

[0015] More preferably, the characteristic amount calculating means obtains a Zernike moment only for a predetermined annular region.

A circular region having low discrimination ability is defined as Zernike.
It can be excluded from the calculation of the moment. Therefore, the calculation of the feature amount can be performed at high speed, and the recognition rate is improved.

A face image processing method according to another aspect of the present invention includes a step of inputting a face image, a step of detecting a position of the face from the input face image, and a step of detecting a face based on the detected position of the face. Normalizing the image; calculating a Zernike moment from the normalized face image;
The method includes a step of calculating a feature amount based on the like moment and a step of creating a dictionary from a plurality of feature amounts obtained from a plurality of face images.

A dictionary is created from feature quantities based on Zernike moments. Therefore, a low-dimensional feature amount vector can be created, and the dictionary size can be reduced.

Preferably, the face image processing method further comprises:
The method includes a step of comparing a feature amount obtained for the input image with a dictionary created in advance and performing face image recognition.

A face image can be recognized using a low-dimensional feature vector. For this reason, face images can be recognized at high speed.

[0021]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS [First Embodiment] Referring to FIG. 1, a face image processing apparatus according to an embodiment of the present invention includes an image input device 11 for inputting an image, and an image input device 1 for inputting an image.
1 and the image input by the image input device 11
A / D conversion unit 12 for performing / D (Analog to Digital) conversion
And an image memory 18 connected to the A / D converter 12 and storing digital images output from the A / D converter 12.
A face detection unit 13 connected to the image memory 18, reads an image digitized by the A / D conversion unit 12 from the image memory 18, and detects a face position;
, A memory 17 for holding the position of the face detected by the face detection unit 13, a memory 17 and an image memory 1
And a face image normalizing unit 14 that reads the position of the face detected by the face detecting unit 13 from the memory 17, normalizes the image stored in the image memory 18, and writes the image into the image memory 18. Including.

The face image processing apparatus is further connected to a memory 17 and an image memory 18, reads out the image normalized by the face image normalizing section 14 from the image memory 18, calculates a feature amount, and writes the calculated feature amount into the memory 17. Feature calculating unit 15
And a dictionary creating unit 16 connected to the memory 17 and creating a face image dictionary from the feature amounts stored in the memory 17, and connected to the dictionary creating unit 16 and storing the face image dictionary created by the dictionary creating unit 16. The external storage device 19 which is connected to the memory 17 and the external storage device 19, reads out the feature amount calculated by the feature amount calculation unit 15 from the memory 17, and compares it with the face image dictionary stored in the external storage device 19. And a recognition unit 20 for performing face image recognition.

The image input device 11 may be any device for inputting an image, and examples thereof include a video camera, a digital camera, and a scanner.

Referring to FIG. 2, a process for creating a face image dictionary will be described. The A / D converter 12 digitizes the image input from the image input device 11 (S2).
1). The face detection unit 13 detects the position of the face from the digitized image (S22). The face image normalizing unit 14 calculates S
The face image is normalized based on the position of the face detected in step S22 (S23). The feature amount calculation unit 15 calculates a feature amount from the face image normalized in S23 (S24). It is checked whether all the images for creating the face image dictionary have been input (S25), and if there is an image for which input has not been completed (No in S25), the process returns to S21. If all images have been input (Yes in S25), the dictionary creation unit 1
In step S6, a face image dictionary is created from a plurality of feature amounts obtained by repeating the processes from S21 to S24.
26).

The end check of the image input in S25 may be performed by the user by explicitly telling the end when an arbitrary number of images are read, or by checking all the images when a predetermined number of images are input. May be determined to have been input. The termination may be determined by another method.

The detection of the face position (S22), the normalization of the face image (S23), the calculation of the feature amount (S24), and the creation of the dictionary (S26) will be described later.

The processing for recognizing a face image will be described with reference to FIG. The A / D converter 12 is a part of the image input device 1
The image input from step 1 is digitized (S31). The face detection unit 13 detects the position of the face from the digitized image (S32). The face image normalizing unit 14 normalizes the face image based on the position of the face detected in S32 (S3).
3). The feature amount calculation unit 15 calculates a feature amount from the face image normalized in S33 (S34). Recognition unit 20
Is the face image dictionary stored in the external storage device 19 and S
The facial image is recognized by comparing the feature amount calculated in step S34 (S35).

The face position detection (S32), face image normalization (S33), feature value calculation (S34), and face image recognition (S35) will be described later.

Referring to FIG. 4, the processing for detecting the face position (S22 in FIG. 2 and S32 in FIG. 3) will be described.

The present method focuses on the fact that when the frequency of the pixel luminance on the circumference centered on the eyebrows is obtained, the value ideally becomes 2.

The maximum intensity MaxP (x, y) of the Fourier transform at each coordinate position of the image and the radius Ma at that time
xR (x, y) is initialized (S41). N in advance
r radii r [i] (i = 0, 1,..., Nr) are set (S42). Since the radius r [i] corresponds to half the distance between the pupils of the face to be detected, it is set according to the size of the face to be detected. The radius index i is initialized to 0.

The radius R currently focused on is set to r [i] (S43). The coordinate values x and y on the image of interest are initialized to 0 (S44 and S45). Consider a circle having a radius R centered on coordinates (x, y), and the luminance value f of a pixel on this circumference
Is subjected to a Fourier transform of frequency 2 defined by the following equation (1) (S46).

[0033]

(Equation 1)

Here, N is the number of points on the circumference, and j is the imaginary unit. Based on the calculated F ₂ ,
The intensity P (x, y, R) at the coordinates (x, y) is defined by the following equation (2) (S47).

[0035]

(Equation 2)

The intensity P (x, y, R) calculated in S47
Is compared with the maximum value (maximum intensity MaxP (x, y)) of the intensity P (x, y, R) determined so far, and the maximum intensity MaxP (x, y) and the radius M giving the maximum intensity are compared.
axR (x, y) is updated (S48 and S49).
When this process is repeated for all coordinate positions and radii (S410 to S415), the intensity Ma at each coordinate is obtained.
xP (x, y) and radius MaxR giving the maximum intensity
(X, y) is obtained. In FIG. 4, width and he
"light" indicates the width and height of the image, respectively.

The maximum intensity MaxP at each coordinate position
The maximum value max (MaxP (x, y) of (x, y)
And arg, which is the x and y coordinates at that time
{Max (MaxP (x, y))} is converted to the eyebrow coordinate (x
o, yo) (S416).

(Xo, yo) = arg {max (MaxP (x, y))} (3) Referring to FIG. 5, processing for normalizing a face image (S2 in FIG. 2)
3 and S33) of FIG. 3 will be described.

Here, the reduction processing at the time of normalization is performed by thinning out the luminance values of the pixels, but the luminance values may be averaged or interpolated. However, the size of the image after normalization (hereinafter, referred to as “normalized image”) is 2N × 2N.

Face detection step (S22 in FIG. 2 or FIG. 3)
Using the eyebrow coordinates (xo, yo) and the radius MaxR (xo, yo) obtained in S32), the following equations (4) to
Based on (7), the range (sx,
sy)-(ex, ey) is determined (S51).

Sx = xo−2MaxR (xo, yo) (4) ex = xo + 2MaxR (xo, yo) (5) sy = yo−MaxR (xo, yo) (6) ey = yo + 3MaxR (xo, yo) (7) The reduction ratio xratio in the x direction and the reduction ratio yratio in the y direction of the image are determined by the following equations (S52).

Xratio = (ex−sx + 1) / 2N (8) yratio = (ey−sy + 1) / 2N (9) Variable i indicating the y coordinate and the x coordinate of the normalized image, respectively.
And j are initialized to 0 (S53 and S54). The coordinates (x, y) of the original image corresponding to the coordinates (j, i) of the normalized image are obtained according to the following equations (10) and (11) (S55).

X = sx + xratio * j (10) y = sy + yratio * i (11) The pixel value of the coordinates (x, y) of the original image is read out and set as the pixel value of the coordinates (j, i) of the normalized image. Writing to the image memory 18 (S56). It is checked whether pixel values are set for all coordinates of the normalized image (S57 to S57).
S510). For coordinates for which a pixel value has not been set, a pixel value is set according to the method described above. If pixel values have been set for all coordinates of the normalized image (NO in S510), the process ends.

With reference to FIG. 6, a description will be given of a process of calculating a feature amount from a normalized image (S24 in FIG. 2 and S34 in FIG. 3).

A preset Zernike, which will be described later,
According to the maximum order n _max of the polynomial, the order n and the number of iterations m
Is set (S61). The number of elements in the table at this time is #table. The elements of the actual table are n, m satisfying the relationship of the following equations (12) and (13).
Are all combinations of However, if n and m match at the time of dictionary creation and at the time of recognition, only an arbitrary combination may be used. FIG. 7 shows an example in which the maximum order n _max is 8, and m is limited to non-negative.

N− | m |: even number (12) | m | <n (13) The index number i of the table is initialized to 0 (S6)
2). The Zernike moment is calculated from the normalized image f (x ′, y ′) created by the face image normalization unit 14 according to the definition formula of the Zernike moment shown in Expression (14) (S63).

[0047]

(Equation 3)

Here, n and m represent n and m in the i-th element of the table, and V _{n, m} (x, y) is expressed by the following equation (1).
This is a polynomial called a Zernike polynomial defined in 5). * Is a symbol representing complex conjugate.

V _{n, m} (x, y) = r _{n, m} (ρ) e ^jm θ (15) Further, r _{n, m} (ρ) is called a radius polynomial defined by the following equation (16). Is a polynomial, and ρ = (x ² +
y ² ) ^1/2 and θ = tan ⁻¹ (y / x).

[0050]

(Equation 4)

In the calculation of the Zernike moment shown in equation (14), coordinates (x, y) are defined as in equations (17) and (18), and all (x, y) have a radius of 1
Is converted to be within the unit circle of.

X = x ′ / N−N (17) y = y ′ / N−N (18) The index number i is incremented (S64),
The index number i is compared with the table element number #table (S65). If there is an unprocessed table element (YES in S65), the process returns to S63.

Zerni for all table elements
If the ke moment is determined (NO in S65),
Zerni for each table element obtained in S63
The ke moment A _{n, m} is converted to a vector (S6
6). In this embodiment, the Zer shown in the following equation (19)
The vectorization is performed so that the power of the naked moment A _{n, m} becomes a vector element. Of course, vectorization may be performed using another conversion method. When the above vectorization is performed and the table of FIG. 7 is used, the number of dimensions of the feature vector is 25.

[0054]

(Equation 5)

Referring to FIG. 9, the process of creating a face image dictionary from the feature vector (S26 in FIG. 2) will be described.

In the present embodiment, face image recognition is performed using the subspace method as described later. For this reason,
Here, the dictionary creation based on the subspace method will be described. However, the recognition and dictionary creation may be performed using a kNN (k Nearest Neighbor) method or the like.

The variance-covariance matrix S of the feature vectorized by the feature calculator 15 is calculated by the following equation (20) (S71). At this time, by repeating the processing from S21 to S25 in FIG.
It is assumed that the number of feature vectors have been calculated. The dimension of the vectorized feature is d.
Here, μ is an average vector of the feature vectors, and is represented by Expression (21).

[0058]

(Equation 6)

As shown in equations (22) and (23),
The variance-covariance matrix S calculated in S71 is unitarily similarized using the eigenvector u _i and eigenvalue σ _i of the matrix S, and is diagonalized as follows. Note that matrix U is a matrix composed of eigenvectors u _i , and matrix Σ is a diagonal matrix composed of eigenvalues σ _i .

[0060]

(Equation 7)

Σ _i has the relationship of the following equation (24). _{σ 1 ≧ σ 2 ≧ ... ≧} σ d ... (24) the dimension r of the smallest r subspaces satisfying the following equation (25) (S73).

[0062]

(Equation 8)

Where dep is a predetermined cumulative contribution ratio. The eigenvectors up to the dimension r obtained in S73 are registered as a dictionary in the external storage device 19 (S7).
4).

Referring to FIG. 10, by comparing the feature amount calculated by feature amount calculation section 15 with the face image dictionary stored in external storage device 19, a process of recognizing a face image ( Step S35 in FIG. 3 will be described. As described above, in this embodiment, recognition using the subspace method will be described, but another recognition method may be used.

The plurality of dictionaries registered in the processing of S74 in FIG. 9 are read from the external storage device 19 (S81). At this time, the number of read dictionaries is #dic. S66 in FIG.
Is read out from the memory 17 (step S8).
2). The dictionary number i representing the dictionary to be compared is initialized to 0, the variable i _min representing the dictionary number of the dictionary that minimizes the distance between the input vector and the dictionary is set to 0, and the variable L _min representing the distance at that time. Is initialized to -1 (S
83). The input vector x and the dictionary u _j ⁽ⁱ⁾ , (j
= 0, ..., to calculate the distance L _i of the ^{r (i)) (S84)} .

[0066]

(Equation 9)

Here, r ⁽ⁱ⁾ is the number of dimensions of the dictionary of the dictionary number i obtained in the process of S73 in FIG. The sign of the current shortest distance L _min is negative or L _i is L
If it is smaller than _min (YES in S85), the following equation (2)
7) and updates the i _min and L _min by (28) (S85A).

I _min = i (27) L _min = L _i (28) Whether the condition of S85 is not satisfied (NO in S85) or after the process of S85A, the dictionary number i is incremented by one (S86). . It is checked whether or not the distance has been calculated for all dictionaries (S87). If there is an unprocessed dictionary (YES in S87), the process returns to S84. If the distance calculation has been completed for all dictionaries (N in S87
O), and a predetermined threshold TH and the shortest distance L _min are compared (S88). If L _min is greater than TH (NO in S88), it is determined that the face image of the person shown in the input image is not registered in the face image dictionary (S810). If L _min is smaller than TH (YES in S88), dictionary number i
Is recognized as a person in the input image.

FIG. 11 shows the result of comparison between the number of dimensions, the dictionary size, and the amount of calculation of the feature vector according to the method described in this embodiment and those of the conventional method.

The dictionary size is a dictionary size necessary when all the eigenvalues and eigenvectors are held when a recognition dictionary is created from the feature vectors according to the flowchart shown in FIG. Here, each element is 8 By
It is calculated as te.

The calculation amount is the number of product-sum operations when recognition is performed according to the flowchart shown in FIG.

The Zernike moment in the present invention has a maximum order of 8, uses {m, n} shown in FIG. 7, and sets a value obtained by squaring each Zernike moment as a feature vector. Therefore, the number of dimensions is 25.

Log Polar FSP (Fourier power spectrum) is a method described in the prior art 1, and the number of dimensions of the feature vector is proportional to the size of the image. If the size of the image is 32 × 32, the number of dimensions is 1024. The four-direction surface feature is a technique described in the related art 2, and the dimension number with the highest recognition rate of this technique is 256 (= 4 × 8 × 8). Therefore, it is apparent that the feature vector obtained by the present invention has the lowest dimension, the dictionary size is small, and the amount of calculation is small. Therefore, by using the face image processing method according to the present invention, a face image recognition process can be realized even in a portable information terminal such as a PDA or a mobile phone which cannot be equipped with a high-speed arithmetic device, a large-capacity external storage device or a memory. Become like

[Modification 1 of Zernike Moment Calculation] The calculation method of the Zernike moment in S63 of FIG. 6 may be modified as shown in FIG. Referring to FIG. 12, ring number i is 0, Zernike moment A
_{n and m} are each initialized with 0 (S91). i-th circular Zernike moment A ⁱ _n, the _m is calculated in accordance with the following equation (29) (S92).

[0075]

(Equation 10)

[0076] Here, the circular Zernike moment, the time of dividing the unit circle as shown in FIG. 8 in a plurality of concentric circles, a Zernike moment of each annular area S _i. In FIG. 8, the number of rings is four. Circular calculated by S92 Zernike moment A ⁱ _n, Zerni the _m
It is added to the ke moment _{An, m} (S93). The ring number i is updated (S94), the ring number i and the ring number #ring
Are compared (S95). If there is an unprocessed ring (S9
(YES at 5) returns to S92. Otherwise, the process ends.

A Zernike moment is determined for each annular region. Therefore, a plurality of feature amounts can be calculated from one image, and highly accurate image recognition can be performed.

[Modification 2 of Zernike Moment Calculation] The Zernike moment calculation method of S63 in FIG. 6 may be modified as shown in FIG. Referring to FIG. 13, ring number i is set to 0, Zernike moment A
_{n and m} are each initialized with 0 (S111). If the ring number i is a preset non-target area described later (YES in S112), the flow advances to S115 to update the ring number. If the ring number i is not the non-target area (S112
NO), and proceeds to S113. The non-target region is a ring that has been previously determined to have low recognition ability in an experiment. For example, FIG.
In the example, S ₀ corresponds to the periphery of the nose, and the characteristics of the individual are not so remarkable. For this reason, it is set as a non-target area. In S113, i-th circular Zernike moment A ⁱ _n, the _m calculated by the above equation (29). S11
Ring Zernike moment A ⁱ calculated by Eq.
_{n, m} is added to the Zernike moment A _{n, m} (S1
14). The ring number i is updated (S115), and the ring number i is updated.
And the number of rings #ring (S116). If there is an unprocessed ring (YES in S116), the process returns to S112. Otherwise, the process ends.

A circular region having low discrimination ability is defined as Zernike.
It can be excluded from the calculation of the moment. Therefore, the calculation of the feature amount can be performed at high speed, and the recognition rate is improved.

The embodiments disclosed this time are to be considered in all respects as illustrative and not restrictive. The scope of the present invention is defined by the terms of the claims, rather than the description above, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

[0081]

According to the present invention, a face image can be quickly recognized without requiring a large-capacity storage device.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a hardware configuration of a face image processing apparatus according to an embodiment of the present invention.

FIG. 2 is a flowchart of a face image dictionary creation process.

FIG. 3 is a flowchart of a face image recognition process.

FIG. 4 is a flowchart of a face position detection process.

FIG. 5 is a flowchart of a face image normalization process.

FIG. 6 is a flowchart of a process of calculating a feature amount from a normalized image.

FIG. 7 is a diagram for explaining a table of an order n and an iteration number m.

FIG. 8 is a diagram for explaining an annular Zernike moment.

FIG. 9 is a flowchart of a process of creating a face image dictionary from a feature vector.

FIG. 10 is a flowchart of a face image recognition process.

FIG. 11 is a diagram showing a comparison result between the present invention and a conventional method.

FIG. 12 is a flowchart of Zernike moment processing.

FIG. 13 is a flowchart of Zernike moment processing.

[Explanation of symbols]

Reference Signs List 11 image input device, 12 A / D converter, 13 face detector, 14 face image normalizer, 15 feature amount calculator, 1
6 Dictionary creator, 17 memories, 18 image memories, 1
9 external storage device, 20 recognition unit.

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁷ Identification symbol FI Theme coat ゛ (Reference) G06T 7/60 150 G06T 7/60 150A F term (Reference) 5B043 AA09 BA04 CA00 DA04 EA15 FA00 GA02 5B057 AA20 BA24 CA12 CA16 CB12 CB16 CC01 CE20 CG05 CH01 CH11 CH12 DA12 DB02 DC02 DC33 5B075 ND06 NK07 NK13 NK37 UU08 5L096 BA18 CA02 CA24 DA02 EA14 FA23 FA55 GA59 HA09 JA11 KA01 LA05

Claims (6)

[Claims] 1. An image input means for inputting a face image, a face detection means connected to the image input means for detecting a position of a face from the input face image, and a connection to the face detection means A face image normalizing means for normalizing the face image based on the position of the face detected by the face detecting means; and a Zernike connected to the face image normalizing means and performing Zernike Calculate the moment, and
a feature value calculation unit for calculating a feature value based on the ike moment; a dictionary creation unit connected to the feature value calculation unit for creating a dictionary from a plurality of feature values obtained from a plurality of face images And a face image processing device. 2. The method according to claim 1, further comprising the step of: connecting the feature quantity calculated by the feature quantity calculation means to the dictionary created by the dictionary creation means, connected to the feature quantity calculation means and the dictionary creation means. The face image processing apparatus according to claim 1, further comprising a face image recognition unit for performing face recognition. 3. The Zernik calculated for each annular region when a reference circle provided in the normalized face image is divided by a plurality of concentric circles.
3. The face image processing device according to claim 1, wherein the feature amount is calculated based on the e-moment. 4. The face image processing apparatus according to claim 3, wherein said feature amount calculating means obtains a Zernike moment only for a predetermined annular area. 5. A step of inputting a face image; a step of detecting a position of the face from the input face image; a step of normalizing the face image based on the detected position of the face; Face image processing, comprising: calculating a Zernike moment from the obtained face image, calculating a feature amount based on the Zernike moment; and creating a dictionary from a plurality of feature amounts obtained from the plurality of face images. Method. 6. The face image processing method according to claim 5, further comprising the step of comparing a feature amount obtained for the input image with a dictionary created in advance and performing face image recognition.