JPH08315133A - Person's face image collation device - Google Patents

Abstract

PURPOSE: To automatically extract the structure of a face with ease and obtain a discrimination result which is stable against the change in lighting conditions and the change with time, etc. CONSTITUTION: An input image signal 31 showing the face is inputted to 1st and 2nd V-shaped edge extracting devices 33 and 34 in the person's face image collation device 32, and V-shaped edges which have minimum luminance value in mutually different directions are extracted. The obtained signal 35 and 36 are ANDed by an AND part 37 and structure points which are viewed as a V shape in both the directions are found. The structure point signal 38 is inputted to a cost calculation part 39 to obtain data regarding the structure points of the images of person's faces as candidates from a data base storage part 41, and the distances between the structure points are calculated by those faces. The obtained distance signal 44 is inputted to a minimum cost selecting means 45 to select the person whose distance between the structure points is shortest.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a person face image collating apparatus for identifying a person from a face image, and more specifically, a person face image capable of obtaining a stable identification result even with changes in illumination and aging. Regarding a collating device.

[0002]

2. Description of the Related Art Identifying a person using a face image is
This is useful for various uses for identifying a person as well as a method for identifying a person using a personal identification number. Therefore, a human face image matching device for this purpose has been conventionally proposed. Prior art such devices generally used one of two approaches:

(1) A method of using the gray value of the image itself of the face portion cut out from the image based on the reference points such as the eyes. (2) A method that uses facial structure points and line information represented by the corners of the eyes, nose, end points of lips, contour lines, and the like.

Next, these two representative examples will be described. Here, the document “Study on Automatic Identification Method of Frontal Face Image (Tsutomu Sasaki, Shigeru Akamatsu, Eio Fukamachi, Yasuhito Suenaga, IEICE Transactions on Image Engineering IE-91-50pp.1-8
(1991)) ”, a conventional method for identifying a person using a face image will be described.

FIG. 6 shows a state in which a collation pattern is cut out using three reference points.
To extract the representative points of the left and right eye regions and the mouth region from the face image and cut out the face matching pattern image, first, the face input image 1
From 1, the representative point 12 of the right eye region and the representative point 13 of the left eye region
Also, the representative point 14 of the lip region is extracted. This extraction process itself is not directly related to the present invention, and therefore its explanation is omitted.
Using the representative points of the left and right eye areas and lip area as a reference,
An affine transformation 15 is applied to the face input image 11 to cut out a face matching pattern 16 to have a size of 128 × 128 pixels. At this time, the parameters C ₁ , C ₂ , C ₃ , C ₄
And D are converted to have values given in advance.

The representative points 12, 13, and 14 of the left and right eye regions and the lip region, which are the reference for cutting out, may be displaced when the illumination condition is deteriorated. Therefore, in order to absorb the fluctuations of the representative points 12, 13, and 14 due to this positional shift, the face matching pattern is subjected to a two-dimensional Fourier transform to be converted into a frequency feature pattern.

Each frequency matching pattern is generated from the face image group for a plurality of names as a database for matching by the method already described. After calculating the covariance matrix from these frequency matching patterns, KL (Kah
unen Lo'eve) expansion is performed to extract L eigenvectors in descending order of eigenvalue. After all, these L
The person is collated by measuring the distance between the database and the unknown input pattern by the Euclidean distance in the L-dimensional space defined by the individual eigenvectors.

As the technique (1), there is also a technique (Japanese Patent Laid-Open No. 3-34075) in which an image of a face is input and the primary image is reduced to make the features of the face stand out. There is.

Next, another example of identifying a person using a face image will be described in the document "Analyse und Klassifikation von".
Gesichtsbildern (R.Buhr, ntzArchiv Bd.8 pp.245-156
(1986) ”.

First, face structural lines such as eyebrow contours, upper eyelid lines, lip contour lines, face contour lines, and facial structure points such as inner corner points, outer corner points of the eyes, nasal wing points, and lower nose points are extracted from the face image. Extract. Regarding the extraction of these structural points, an edge image adapted to a unique shape such as eyes and nose is adopted, and an image processing method such as edge tracking is applied to obtain them.

The method of extracting the inner and outer canthus points according to this proposal will be described with reference to FIG. First, it is assumed that the image 21 of the eye part and the center position of the pupil are extracted by an appropriate method as shown in FIG. A Laplacian-Gaussian obvlator is applied to the image of this eye portion and binarized by "0" to generate an eye edge image 22 as shown in FIG. The following processing is performed from this eye edge image 22 to generate an eye edge image 23 shown in FIG.

Eliminate edges caused by bright areas of illumination reflected from the pupil surface. The x-axis 24 and the y-axis 25 shown in FIG. 7B are set as shown in FIG. 7 with the pupil as the center. These x-axis 24 and y
The axis 25 divides the eye edge image into four quadrants. The following processing is performed in each quadrant.

(A) In the upper left quadrant, an edge with a change in brightness from bright to dark in the x-axis direction and an edge with a change in brightness from bright to dark in the y-axis direction are obtained. Erase all edges. (B) In the lower left quadrant, find the edge with brightness change from bright to dark in the x-axis direction and the edge with brightness change from dark to light in the y-axis direction, and erase all other edges. To do. (C) In the upper right quadrant, find the edge with a change in brightness from dark to bright in the x-axis direction and the edge with a change in brightness from bright to dark in the y-axis direction, and erase all other edges. To do. (D) In the lower right quadrant, find an edge with a change in brightness from dark to bright in the x-axis direction and an edge with a change in brightness from dark to bright in the y-axis direction, and find the other edges. Erase everything.

When the edge curve on the pupil is first extracted from the edge image 23 of the eye shown in FIG. 3C, the edge lines are traced to the left and right. The end points are determined under the condition that the edge curvature as the curvature of the edge line becomes a certain value or more, and these are set as the inner corner point and the outer corner point of the eye. The face structure points or structure lines thus obtained are normalized by the left and right pupil center positions. Then, it is converted into a plurality of feature amounts such as the distance between the left and right outer corners of the eye for use in matching. The person matching process is similarly performed based on the distances in these feature amount spaces.

As another technique shown in (2), Japanese Patent Laid-Open No.
There is one disclosed in Japanese Patent Publication No. 1-100872. Here, a predetermined fixed point included in the face of the individual to be discriminated is extracted, and the relative position data is collated with prestored data to discriminate the individual. This technique focuses on the fact that, for example, when focusing on the eyes of the human eye, the positions of the outer corners and the inner corners of the eyes do not change even if the facial expression is changed.

In this technique, an image is stored in a frame memory by using a CCD camera and a light emitter, and a fixed point such as the corner of the eye is extracted by a general method using a computer. Then, the distance between these fixed points is measured to discriminate the individual. The light emitter is for obtaining a still image with good contrast.

[0017]

Among the conventional human face image matching processes described above, in the method that uses the gray value as a feature, it is easy to automatically extract this feature, and the illumination condition is stable. It has been found that in such cases, very good matching performance is exhibited. However, there is a problem that the collation cannot be performed at all when the illumination condition changes.

On the other hand, if the face structure points or lines described above are used, the original structure itself is stable regardless of the illumination conditions. However, it is very difficult to stably and automatically extract these structures of the face. Further, in the technique disclosed in Japanese Patent Laid-Open No. 61-100872,
A light emitting device is required to obtain a still image with good contrast, and if such conditions are not satisfied, it is not possible to perform good collation of the human face image.

Therefore, an object of the present invention is to perform human face image collation in which the structure of a face can be easily and automatically extracted and a stable identification result can be obtained even with respect to fluctuations in illumination conditions and changes over time. To provide a device.

[0020]

According to a first aspect of the present invention, (a) V of which the luminance is minimized along two or more different directions in the face image of the person to be collated.
V-shaped edge extracting means for individually extracting the character edges,
(B) V extracted by each of the V-shaped edge extracting means
The logical product of the character edges, and the logical product means for obtaining the structural points or the continuation of the structural points that look like V-shaped edges in any direction, and (c) the data of the structural points for each face of the candidate person. Structure point data output means for outputting, and (d) Face distance between the structure points of the face obtained by the AND means and the structure points of the face output from the structure point data output means A distance calculation means between the structure points for each
(E) A person selecting means for selecting, as a person to be collated, a person having the smallest distance between the structure points calculated by the distance between structure points calculation means out of the candidates who are stored in the structure point database storage means. Is provided in the human face image collation device.

That is, according to the first aspect of the present invention, the V-shaped edges having the minimum brightness are individually extracted along two or more different directions in the image of the face of the person to be collated. The logical product is taken to determine the structure points or the continuation of the structure points that appear as V-shaped edges in any direction. This causes stable discrimination of a person's face image against variations in lighting and aging. That is, a structure that is resistant to illumination fluctuations and secular changes has the following two characteristics. (A) The surface reflectance changes significantly. (B) The shape of the surface changes significantly.

The face structure points used in the above-described conventional method or the face structure line as a set of these face structure points includes a portion having such a feature, but a person such as a mole or a scratch. Some did not include a featured structure. In addition, these face structure points or face structure lines are those that humans discriminate and name, such as "eye corner points" and "eye corner points", and do not consider the ease of extraction by image processing. . According to the first aspect of the present invention, in order to obtain a structure that is resistant to illumination variation and secular change and can be easily extracted by image processing, use is made of a structure point in which the brightness value change is locally minimum in two or more directions. I chose Such structure points can be extracted under the condition that the first derivative of the image is 0 and the second derivative is positive.

A person is collated by comparing the detected structure point distribution with a structure point distribution prepared for a candidate person's face in advance. At this time, the respective points between the structural points are associated so that the correspondences do not overlap and the cost represented by the following formula (1) is minimized. Cost = (sum of distances between corresponding structural point coordinates) (1)

According to the second aspect of the invention, the structure point data output means in the first aspect of the invention stores the image of the face of the candidate person, and from this, the structure point data for each face is created. It is characterized by outputting as. That is, the face structure point distribution of each person may be provided as a database, or the face image may be stored and the structure point data for each face may be created if necessary.

According to the third aspect of the invention, (a) a first V-shaped edge extracting means for extracting a V-shaped edge having a minimum brightness along the first direction in the image of the face of the person to be collated. And (b) second V-shaped edge extraction means for extracting a V-shaped edge having a minimum brightness along a second direction different from the first direction in the image of the face of the person to be collated. (C) A logical product means for obtaining a logical product of the first and second V-shaped edge extraction means to obtain a structural point or a series of structural points that appear as a V-shaped edge in the first and second directions. (D) Structure point database storage means that stores a database of structure points for each face of a candidate person,
(E) Calculation of inter-structure-point distance for each face of candidate person by calculating the inter-structure-point distance between the facial structure points obtained by the logical product means and the facial structure points stored in the structural point database storage means And (f) a person who selects the person having the smallest inter-structure-point distance calculated by the inter-structure-point distance calculating means from among the candidate persons stored in the structure-point database storage means as the person to be collated. The person face image collation device is provided with a selection means.

That is, according to the third aspect of the present invention, the first and second V-shaped edge extracting means are prepared and the V-shaped edge is extracted from two different directions. In addition, a structure point database storage means for storing a database of structure points for each face of the candidate person is prepared, and the face structure points obtained by the logical product means and the face structure stored in the structure point database storage means are prepared. The distance between the points and the structure is calculated for each face of the candidate person. The person selection means is
Among the candidates who are stored in the structure point database storage means, the person having the smallest distance between the structure points calculated by the distance between structure points calculation means is selected as the person to be collated.

According to the fourth aspect of the invention, (a) first V-shaped edge extracting means for extracting a V-shaped edge having a minimum brightness along the first direction in the image of the face of the person to be collated. And (b) second V-shaped edge extraction means for extracting a V-shaped edge having a minimum brightness along a second direction different from the first direction in the image of the face of the person to be collated. (C) A logical product means for obtaining a logical product of the first and second V-shaped edge extraction means to obtain a structural point or a series of structural points that appear as a V-shaped edge in the first and second directions. (D) Feature extracting means for extracting texture features corresponding to each pixel from the image of the face of the person to be collated, and (e) Structure point database storage storing a database of structure points for each face of the candidate person. Means and
(F) a texture feature database storing means that stores a database of texture features for each face of a candidate person; and (g) face structure points obtained by the ANDing means and face structure points existing in the database. The inter-structure-point distance calculating means for calculating the inter-structure-point distance for each face of the candidate person, and the face stored in the texture feature database storing means corresponding to the face structural points obtained by the (h) ANDing means. An inter-texture-feature distance calculation means for calculating the inter-texture-feature distance with respect to the structural point for each face of a candidate person,
(I) Cost calculation means for obtaining the sum of the calculation results of the structure point distance calculation means and the texture feature distance calculation means for each face of a candidate person; The person face image matching device is provided with a person selecting means for selecting a person as

That is, according to the invention described in claim 4, the texture feature is newly used for collation of the human face image. The reason for this will be explained. The structure point has texture information such as the second partial differential value at that point. Document "Image recognition and image understanding for artificial intelligence and computer vision (R. Nevatia, translated by Satoshi Minami),
pp. 161-175, Keigaku Shuppan (1986) ”,
Regarding the texture feature, luminance average, variance, and histogram distribution are given as primary measures, convolution outputs of various filter templates are given as texture energy measures, and frequency feature / secondary measure etc. are given as Fourier measures. The texture information at the structure points is used as the texture feature in the partial image including the structure points, and in order to ensure the correspondence between the structure points, the cost represented by the following equation (2) is minimized. It is designed to be attached. Cost = (sum of distances between corresponding structure point coordinates) + (sum of distances between corresponding structure point texture information) (2) In the invention according to claim 4, the input face image and the database having the smallest cost It is a method to perform person verification by selecting both images in the face.

The invention according to claim 5 is characterized in that known texture information such as a histogram distribution in the partial image is used as the texture feature.

[0030]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The present invention will be described in detail below with reference to embodiments.

First embodiment

FIG. 1 shows the configuration of a human face image collating apparatus according to the first embodiment of the present invention. An input image signal 31 obtained from an image input device (not shown) is input to the first and second V-shaped edge extraction devices 33 and 34 in the human face image matching device 32, and different first or second V-shaped edge extraction devices 33 and 34, respectively. The V-shaped edge having the minimum luminance value in the direction is extracted. Input image signal 3 here
In No. 1, the face is already cut out and its position, size, and angle are corrected, as shown as the figure on the right side of FIG. The first V-shaped edge signal 35 output from the first V-shaped edge extraction device 33 and the first V-shaped edge signal 3 output from the second V-shaped edge extraction device 34.
6 is input to the logical product unit 37, the logical product of both is taken,
Structural points or a series of structural points that appear as V-edges in the first and second directions are determined. The structure point signal 38 representing these is input to the cost calculation unit 39.

The cost calculation unit 39 is connected to a database storage unit 41 that stores a database of structural points of the face image of the person who is a candidate for collation. The database storage unit 41 supplies the database structure point signal 42 to the cost calculation unit 39, and the distance between structure points is calculated for each face of the candidate person. The distance signal 44 as a result of this is input to the minimum cost selecting means 45, and the person having the minimum distance between structural points is selected. This result is output as the person collation signal 46 from the person face image collation device 32.

An input image signal 31, which is an image of a face of a person to be collated, is input to the first V-shaped edge extraction device 33, and V having a minimum luminance value in a certain direction.
Character edges are detected. In this embodiment, x described in FIG.
A V-shaped edge having a minimum brightness value in the direction of the axis 24 is detected. This is output as the first V-shaped edge signal 35. Further, the second V-shaped edge extracting device 34 detects a V-shaped edge having a minimum brightness value in the other direction. In this embodiment, the V-shaped edge having the minimum brightness value in the y-axis 25 direction described in FIG. 7 is detected. This is output as the second V-shaped edge signal 36.

FIG. 2 shows a circuit configuration of the first V-shaped edge extraction device. The circuit configuration of the second V-shaped edge extraction device 34 is the same as that of the first V-shaped edge extraction device 33.
Since it is exactly the same as that of the above, and only the direction of obtaining the minimum of the luminance value is different, the description of this circuit is omitted.
Note that the same devices as the V-shaped edge extracting devices 33 and 34 of this embodiment are disclosed in, for example, Japanese Patent Application No. 5-189566.

The first V-shaped edge extracting device 33 inputs the above-mentioned input image signal 31 to both the primary partial differential calculating section 51 and the secondary partial differential calculating section 52. First partial differential calculator 51
The primary differential signal 55 output from is input to the zero point detection unit 56, and the point where the value of the image subjected to the primary partial differential is 0 is detected. The extreme value edge signal 57 thus obtained is input to one of the logical product calculation sections 58. Further, the secondary differential signal 59 output from the secondary partial differential calculating section 52 is input to the positive area detecting section 61, and the area (positive area) in which the value of the second partial differential image is positive is detected. Be seen.
The positive region signal 62 resulting from this becomes the other input of the logical product calculation unit 58. The logical product calculation unit 58 calculates the logical product of the extreme value edge signal 57 and the positive region signal 62 and outputs the first V-shaped edge signal 35.

As described above, in the human face image collating apparatus according to the first embodiment, an image of a human face is photographed by using the first V-shaped edge extracting device 33 and the second V-shaped edge extracting device 34. The change in the luminance of is obtained from each of the two directions, and the first V-shaped edge signal 35 and the second V-shaped edge signal 36 are obtained. Since the logical product unit 37 takes these logical products, V in either direction
The structure point signal 38 that looks like a V-shaped edge is obtained, and a V-shaped edge that does not depend on the illumination condition is obtained. Therefore, the cost calculation unit 39, the database storage unit 41, and the minimum cost selection unit 45 thereafter can select the corresponding person with reliability from the candidate persons.

In the first embodiment, the two V-shaped edge extracting devices 33 and 34 are combined to extract the structural points of the face, but it goes without saying that three or more may be combined. In addition, in the first embodiment, the database storage unit 41
Decided to keep the structure points of the face.However, the images of the face of the candidate person are directly held, and the structure points of the face are extracted from these images at the required time, and the work for comparison is performed. It goes without saying that you may do so.

Second embodiment

FIG. 3 shows the configuration of a human face image collating apparatus according to the second embodiment of the present invention. In this figure, the same parts as those in FIG. 1 are designated by the same reference numerals. This second
In the human face image collation device 71 of the embodiment, the input image signal 31 obtained from an image input device (not shown) is input to the first and second V-shaped edge extraction devices 33 and 34.
The luminance distribution feature extraction unit 72 is also input. First and second V-shaped edge extraction devices 33, 34
Extracts the V-shaped edges having the minimum brightness value of the image in different directions based on the input image signal 31. The first V-shaped edge signal 35 output from the first V-shaped edge extraction device 33 and the first V-shaped edge signal 36 output from the second V-shaped edge extraction device 34 are supplied to the logical product section 37. It is input, the logical product of both is taken, and the structural points or the series of structural points that appear as V-shaped edges in the first and second directions are obtained. The structure point signal 38 representing these is input to the cost calculation unit 73.

On the other hand, the brightness distribution feature extraction unit 72 is adapted to calculate the brightness distribution in the partial image including each pixel point from the human face image and extract the brightness distribution texture feature. The feature signal 74 output from the brightness distribution feature extraction unit 72 is also input to the cost calculation unit 73.

The cost calculation section 73 is a database storage section 7.
5 is connected. The database storage unit 75 stores the brightness distribution texture feature corresponding to the face structure point to be matched. The database structure point signal 76 representing the face structure points for each face stored and the database feature signal 77 which is the texture feature corresponding to the structure point are
It is sent from the database storage unit 75 to the cost calculation unit 73. The cost calculation unit 73 performs the above-mentioned expression (2) between the structure point signal 38 and the feature signal 74, the database structure point signal 76 that is the structure point of a certain face as a candidate, and the corresponding database feature signal 77. Based on, the calculation is performed in association with each other so that the distance becomes smallest. And
The minimum distance is output as the distance signal 78. This distance signal 78 is input to the minimum cost selection unit 79, and the person with the smallest distance is selected from the faces in the database. The selection result is output from the human face image matching apparatus 71 as a person matching signal 81.

In the second embodiment described above, the database storage unit 75 holds face structure points and corresponding texture features. Also in this case, the face image itself may be held as described in the first embodiment, and the face structure points may be calculated at any time. Further, in the second embodiment as well, the two V-shaped edge extraction devices 33 and 34 are combined to extract the facial structure points, but it is natural that three or more may be combined. Further, in the second embodiment, the brightness distribution value in the partial image is used as the texture feature in each pixel, but it is also possible to use a known texture feature.

Modification of the second embodiment

FIG. 4 shows, as a modification of the second embodiment, a circuit configuration of a portion of a texture feature extraction device using known texture features. The texture feature extraction unit 90 receives the input image signal 31 and creates a cumulative histogram of brightness values.
1 and a brightness average calculator 92 that calculates the average value of the brightness of the input partial image, which is also supplied with the input image signal 31. The cumulative distribution signal 93 output from the cumulative distribution creating unit 91 and the brightness average signal 94 output from the brightness average calculating unit 92 are input to the cumulative distribution normalizing unit 95. The cumulative distribution normalization unit 95 is adapted to parallelize and normalize the cumulative distribution so that the brightness average value becomes a certain value. The cumulative distribution normalization unit 95 outputs a normalized cumulative distribution signal 96, and the cost calculation unit 73 shown in FIG.
Will be input together with the structure point signal 38.

Texture feature extraction apparatus 90 of this modification
Is a document "NUMERICAL RECIPES in C (Japanese version) (Will
iam H. Press, Saul A. Teukolsky, William T. Vetterl
ing, Brian P. Flannery, Katsuichi Dankei, Haruhiko Okumura, Toshiro Sato, Makoto Kobayashi) Technical Review, PP.460-463 (1993) "
Based on the Kolmogonov-Smirnov test. This texture feature extraction device 9
At 0, the texture in the partial image on the input image is calculated.

The texture feature extraction device 90 will be described as a clock enemy. The input image signal 31 is input to the cumulative distribution creating unit 91, converted into a cumulative distribution of luminance values in a partial image on the input image, and a cumulative distribution signal 93 is obtained. The cumulative distribution is a value P (x) at a certain brightness value x.
Is given by the following equation (3). P (x) = (number of pixels in partial image whose brightness value is −∞ to x) (3) The input image signal 31 is also input to the brightness average calculation unit 92,
The brightness average signal 94 is obtained by converting the brightness value in the partial image on the input image into the average value.

FIG. 5 is for explaining the cumulative distribution. In this figure, the horizontal axis represents the luminance value and the vertical axis represents the number of pixels. If the minimum value K _{MIN of} brightness that can be taken is the origin of the coordinates and the maximum value that can be taken is K _MAX , the cumulative distribution represented by the curve 101 is obtained therebetween.

The cumulative distribution signal 93 and the luminance average signal 94 in FIG. 4 are input to the cumulative distribution normalization unit 95, and the distribution signal is averaged so that the average luminance value becomes "0", and the luminance value changes. Is performed to obtain a normalized cumulative distribution signal 96.

This parallel movement operation will be described with reference to FIG. By using the average value K _{AVE of} luminance as a reference, the parallel movement 103 is performed so that the value becomes 0, and the normalized cumulative distribution 104 that does not depend on the average luminance is obtained in the partial image. The present invention can be realized by using the normalized cumulative distribution signal 96 as a texture feature. In addition, reference "Image Recognition and Image Understanding for Artificial Intelligence / Computer Vision (R. Nevatia, Translated by Minami Satoshi), pp.161-17
The present invention can also be realized by using known texture information introduced in 5, Keigaku Shuppan (1986), for example, histogram distribution in partial images, various filter template outputs, frequency components, and the like.

[0051]

As described above, claims 1 to 5 are as follows.
According to the invention described above, since the structure point in which the change in the brightness value is commonly minimized in two or more directions is used, a human face image collation device that is resistant to illumination variation and aging variation can be provided by simple image processing. Can be realized.

According to the second aspect of the invention, the structure point data output means decides to create the structure point data of the face from the image itself of the face of the candidate person. By adding, the person to be collated can be easily added.

Further, according to the invention described in claim 4, since the feature extracting means for extracting the texture feature corresponding to each pixel from the image of the face of the person to be collated is further used,
There is an advantage that the person to be collated can be identified more accurately.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a human face image matching apparatus according to a first exemplary embodiment of the present invention.

FIG. 2 is a block diagram showing a circuit configuration of a first V-shaped edge extraction device in the first embodiment.

FIG. 3 is a block diagram showing a configuration of a human face image matching apparatus according to a second embodiment of the present invention.

FIG. 4 is a block diagram specifically showing a luminance dispersion feature extraction unit in the second embodiment.

FIG. 5 is a characteristic diagram for explaining a cumulative distribution in a modification of the second embodiment.

FIG. 6 is an explanatory diagram showing a first conventional method for identifying a person using a face image.

FIG. 7 is an explanatory diagram showing a second conventional method for identifying a person using a face image.

[Explanation of symbols]

 31 input image signal 32, 71 human face image collation device 33 first V-shaped edge extraction device 34 second V-shaped edge extraction device 37 logical AND section 39, 73 cost calculation section 41, 75 database storage section 45, 79 minimum Cost selection means 51 First partial differential calculation unit 52 Second partial differential calculation unit 56 0 Point detection unit 61 Positive area detection unit 58 Logical product calculation unit 72 Luminance variance feature extraction unit 90 Texture feature extraction unit 91 Cumulative distribution creation unit 92 Luminance average Calculation unit 95 Cumulative distribution normalization unit

Claims (5)

[Claims] 1. A V-shaped edge extracting means for individually extracting V-shaped edges having minimum brightness along two or more different directions of a face image of a person to be collated, and the V-shaped edge extraction. The logical product of the V-shaped edges extracted by each of the means, and the logical product means for obtaining the structural points or the continuation of the structural points that look like the V-shaped edges in any direction, and the structure for each face of the candidate person Structure point data output means for outputting point data, and the inter-structure point distance between the face structure points obtained by the AND means and the face structure points output from the structure point data output means are the candidates. An inter-structure-point distance calculation means for calculating each face of a person, and a person who has the smallest inter-structure-point distance calculated by the inter-structure-point distance calculation means among candidate persons stored in the structure point database storage means In front Human Face image collation apparatus characterized by comprising a person selecting unit configured to select as a person being a target for verification. 2. The structure point data output means stores an image of a face of a candidate person, and creates and outputs data of a structure point for each face from the image.
The described human face image matching device. 3. A first V-shaped edge extracting means for extracting a V-shaped edge having a minimum brightness along a first direction in an image of a face of a person to be collated, and a person to be collated. For the face image, the brightness V has a minimum value along the second direction different from the first direction.
A structure point that looks like a V-shaped edge in the first and second directions by taking the logical product of the second V-shaped edge extraction means for extracting the V-shaped edge and the first and second V-shaped edge extraction means. Alternatively, a logical product means for obtaining a sequence of structural points, a structural point database storage means for storing a database of structural points for each face of a candidate person, and a structural point for a face and a structural point database obtained by the logical product means An inter-structure-point distance calculating means for calculating the inter-structure-point distance from the structural points of the face stored in the storage means for each face of the candidate person, and a candidate person stored in the structure-point database storage means A person face image collation device, comprising: person selecting means for selecting a person having a minimum inter-structure point distance calculated by the inter-structure point distance calculating means as the person to be collated. 4. A first V-shaped edge extracting means for extracting a V-shaped edge having a minimum brightness along a first direction in an image of a face of a person to be collated, and a person to be collated. For the face image, the brightness V has a minimum value along the second direction different from the first direction.
A structure point that looks like a V-shaped edge in the first and second directions by taking the logical product of the second V-shaped edge extraction means for extracting the V-shaped edge and the first and second V-shaped edge extraction means. Alternatively, a logical product means for obtaining the continuity of structure points, a feature extraction means for extracting texture features corresponding to each pixel from the image of the face of the person to be collated, and a database of structure points for each face of the candidate person Existing in the database, the texture point database storage means that stores the database of texture features for each face of the candidate person, and the face structure points obtained by the logical product means. A structure point distance calculating means for calculating a distance between a structure point of a face and a structure point for each face of the candidate person; and a face structure point obtained by the ANDing means An inter-texture-feature distance calculating means for calculating, for each face of the candidate person, an inter-texture-feature distance with respect to a face structure point stored in the corresponding texture feature database storage means; Cost calculation means for obtaining the sum of the calculation results of the inter-texture feature distance calculation means for each face of the candidate person, and person selection means for selecting the person who has become the minimum by the cost calculation means as the person to be collated. A person's face image collation device comprising. 5. The human face image matching device according to claim 4, wherein known texture information such as a histogram distribution in a partial image is used as the texture feature.