JP2000194849A - Individual identification device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an individual identification device high in robustness to the change of the direction and expression of a face and capable of enhancing an identification ratio. SOLUTION: This device is provided with an affine transformation part 230 for detecting the positions of eyes and a mouth from the face area of an input image and then affine-transforming the positions of the eyes and the mouth so as to bring them to preset positions, a sample set 250 composed by registering the samples of a face image vector to be the reference of respective persons beforehand, a nonlinear main component analysis part 240 for performing the dimensional compression of the face image vector based on the learning of 5-layer perceptron using the sample set and outputting a compressed face image vector, a reference face image database part 260 for registering a set transformed to the compressed face image vector dimensionally compressed by using 3-layer perceptron in the nonlinear main component analysis part and a face image identification part 270 for outputting an identified result at the time of inputting the compressed face image vector to be identified based on the learning using the compressed face image vector included in the reference face image database part consisting of the three-layer perceptron.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a personal identification device for identifying an individual using a face image, and more particularly, to a personal identification device applicable to automatic login to a computer.

[0002]

2. Description of the Related Art With the spread of the Internet, computers have been used as information terminals, and people who are unfamiliar with keyboard operations have been treated. Therefore, the need for a more natural interface between a computer and a human has been increasing, and research on a man-machine interface modeled on the interface between a human and a human has been conducted.

FIG. 5 is a block diagram showing a configuration of a personal identification device according to a conventional example. The personal identification device illustrated in FIG. 5 includes an image input unit 100 that captures an image including a face by a video camera or the like, and parameters such as size, inclination, and brightness for the face image captured by the image input unit 100. A normalization unit 110 for making constant, a principal component analysis for a large number of normalized face images, a feature extraction by dimensional compression to obtain a base, and a principal component analysis unit 120 for obtaining a basis thereof. (Karhunen-Loeve) conversion unit 130 that expresses the face image as a vector using the basis obtained by the principal component analysis unit 120, and a database in which a face image vector serving as a reference of a person to be identified in advance is registered. From the reference face image dictionary 140 as a reference and the vectors registered in the reference face image dictionary 140, the vector having the shortest distance from the face image vector obtained by the KL conversion unit 130 is detected. Minimum distance discriminating section 15, the identification result of a person of that vector with
0 is provided.

[0004]

However, in the conventional personal identification device having the above-described configuration, the principal component analyzing section 1
The processing by the method 20 has a problem that the feature extraction is limited due to the dimensional compression by the linear processing with respect to the image having the change of the expression, the direction of the face, and the like, so that the recognition rate is reduced.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems of the prior art, and has a personal identification device which has high robustness to changes in the face direction and facial expressions and can improve the identification rate. It is intended to provide.

[0006]

In order to achieve the above object, a personal identification device according to the present invention comprises an image input means for inputting an image including a face image, and an image input by the image input means. A face area extracting means for extracting a face area by template matching with a standard face image from
The eyes and mouth detecting means for detecting the position of the eyes and the mouth from the face area extracted by the face area extracting unit, and the positions of the eyes and the mouth detected by the eyes and the mouth detecting means are set to the preset positions. Affine transformation means, and affine transformation means for constructing a face image vector having a dimension equal to the number of pixels from the affine transformed image data, and a sample of a face image vector created from a face image used as a reference for each person are registered in advance. A compressed face image comprising a sample set and a five-layer perceptron identity mapper, performing dimension compression of the face image vector from the affine transformation unit based on learning of the five-layer perceptron using the sample set. Non-linear principal component analysis means for outputting a vector, and dimensional compression of a face image vector included in the sample set using the three-layer perceptron of the non-linear principal component analysis means A reference face image database unit for registering a set converted into the compressed face image vector, and the three-layer perceptron, and the nonlinear face based on learning using a compressed face image vector included in the reference face image database unit. Face image identification means for outputting an identification result when a compressed face image vector to be identified is input from the principal component analysis means.

[0007]

FIG. 1 is a block diagram showing a configuration of a personal identification device according to an embodiment of the present invention. The personal identification device having the illustrated configuration includes an image input unit 200 and an image input unit 20 that capture an input image including a face image from a video camera or the like.
0, a face area extraction unit 210 that extracts a face area from a face image captured from a video camera or the like by template matching with a standard face image, contour extraction in the face area extracted by the face area extraction unit 210, and binarization processing The affine transformation is performed on the face image such that the center position of the eye and the mouth is detected by the eye, the mouth detection unit 220, the eye, the eye detected by the mouth detection unit 220, and the center position of the mouth come to a preset position. , An affine transformation unit 23 forming a face image vector having a dimension equal to the number of pixels from the affine-transformed image data
It is composed of an identity mapper composed of a 0-layer and 5-layer perceptron. Based on the learning of the 5-layer perceptron using the sample set 250, the face image vector from the affine transformation unit 230 is dimensionally compressed and a compressed face image vector is output. A non-linear principal component analysis unit 240, a sample set 250 in which samples of face image vectors created from face images used as references for each person are registered in advance, and a face image vector included in the sample set 250 is converted to a non-linear principal component analysis unit. Reference face image database units 260 and 3 for registering sets converted into compressed face image vectors that have been dimensionally compressed using a three-layer perceptron 240
The identification result when the compressed face image vector is input from the non-linear principal component analysis unit 240, which is composed of a layer perceptron and is desired to be identified based on learning using the compressed face image vector included in the reference face image database unit A face image identification unit 270 is provided.

Next, the operation of the personal identification device having the above-described configuration will be described with reference to FIGS. Here, FIG. 2 shows the image input unit 200 and the eye / mouth detection unit 22 of FIG.
FIG. 7 is a diagram for explaining processing of 0 and an affine transformation unit 230;
(A) is an example of an image input by the image input unit 200 and an example of a face image area extracted by the face area extraction unit 210 in FIG. 1, (b)
Shows an example of the result of detecting the center position of the eye and the mouth detected by the eye and mouth detecting unit 220, and FIG. FIG. 3 shows the non-linear principal component analysis unit 240 and the face image identification unit 270 of FIG.
FIGS. 7A and 7B are explanatory diagrams of the processing in FIG. 7A, wherein FIG. 7A is a five-layer perceptron in the nonlinear principal component analysis unit 240, FIG. Face image identification unit 270
Describes the three-layer perceptron.
Further, FIG. 4 shows an example of teacher data of an output at the time of learning of the three-layer perceptron in the face image identification unit 270 of FIG.

In the configuration shown in FIG.
0 captures an input image including a face image from a video camera or the like as shown in FIG. The face image captured from the video camera or the like by the image input unit 200 is a face area extraction unit 2
10, a face area is extracted from the face image by template matching with a standard face image as shown in FIG. The eye / mouth detection unit 220 detects the center position of the eyes and mouth as shown in FIG. 2B by performing contour extraction, binarization processing, and the like on the face region extracted by the face region extraction unit 210.

[0010] The eye, the eye detected by the mouth detection unit 220,
The center position of the mouth is provided to the affine transformation unit 230. The affine transformation unit 230 affine transforms the face image such that the detected center position of the eyes and the mouth is at a preset position as shown in FIG. Perform the conversion. Then, a face image vector having a dimension equal to the number of pixels is formed from the affine-transformed image data.

Next, the nonlinear principal component analysis unit 240
As shown in FIG. 3A, it is composed of an identity mapper composed of a five-layer perceptron.
The number of neurons in the layer is equal, and the number of neurons in the third layer is set smaller than the number of neurons in the first layer.
In the sample set 250, a face image vector created from a face image used as a reference for each person is registered in advance. The nonlinear principal component analysis unit 240 performs learning of the 5-layer perceptron using each element of the sample set 250. Learning is performed such that when an image vector is input to the first layer using the error back propagation method, the same value is output to the fifth layer output.
After learning is completed, 3 of the first to third layers out of the five layers
Using a layer perceptron, the face image vector is dimensionally compressed by using the three-layer output to create a compressed face image vector (see FIG. 3B).

On the other hand, in the reference face image database section 260, a set obtained by converting face image vectors included in the sample set 250 into compressed face image vectors obtained by dimensionally compressing the face image vectors using the three-layer perceptron of the non-linear principal component analysis section 240. Is stored.

Further, the face image identification unit 270 is provided with a
As shown in FIG. 9C, the number of neurons in the first layer is determined by the nonlinear principal component analyzer 240.
The number of neurons in the third layer is equal to the number of registered neurons. The number of neurons in the second layer is set to a value larger than the first and third layers. In the learning, the reference face image database 260
Using the compressed face image vector included in
One neuron of the third layer is associated with each person, and the output of the corresponding neuron is 1 or a number close to 1 (for example, 0.9), and the other neurons are 0 or a number close to 0 (for example, 0.9). The learning is performed using the error back propagation method so as to be 0.1) (see FIG. 4). After the learning is completed, a person corresponding to the neuron having the largest value among the outputs of the neurons of the third layer when the compressed face image vector to be identified is input is output as the identification result.

Therefore, according to the present embodiment, when performing personal identification using a face image, a nonlinear principal component analysis unit 240 that performs dimension reduction using an identity mapper composed of a five-layer perceptron, It has a face image identification unit 270 consisting of a three-layer perceptron that has been trained using a compressed data group of a reference image of a person to be identified in advance, and performs non-linear processing. The robustness to the change in is increased.

[0015]

As described above, according to the present invention, when performing personal identification using a face image, the face image is subjected to dimensional compression using an identity mapper composed of a five-layer perceptron, and Non-linear processing is performed by identifying face images using a three-layer perceptron that has been trained using the compressed data group of the reference image of the person to be identified, so that it is robust against changes in face direction and facial expressions. Performance can be increased, and the identification rate can be improved.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a personal identification device according to an embodiment of the present invention.

FIG. 2 shows an example of an image input by the image input unit 200 and an example of a face image area extracted by the face area extraction unit 210, an example of an eye / mouth center position detection result detected by the eye / mouth detection unit 220 in FIG. ,
FIG. 9 is an explanatory diagram of each unit process showing an example of a face image after affine transformation by the affine transformation unit 230;

FIG. 3 illustrates a five-layer perceptron in the nonlinear principal component analyzer 240, a three-layer perceptron for generating a compressed image vector in the nonlinear principal component analyzer 240, and a three-layer perceptron in the face image identification unit 270, respectively. FIG.

FIG. 4 is an explanatory diagram of an example of teacher data serving as an output when learning a three-layer perceptron in the face image identification unit 270 of FIG. 1;

FIG. 5 is a block diagram showing a configuration of a personal identification device according to a conventional example.

[Explanation of symbols]

 200 image input unit (image input unit) 210 face region extraction unit (face region extraction unit) 220 eyes, mouth detection unit (eye and mouth detection unit) 230 affine transformation unit (affine transformation unit) 240 nonlinear principal component analysis unit ( Non-linear principal component analysis means) 250 sample set 260 face image identification unit (face image identification means) 270 reference face image database unit

Claims (1)

[Claims] An image input unit for inputting an image including a face image; a face area extracting unit for extracting a face region from the image input by the image input unit by template matching with a standard face image; An eye and mouth detecting means for detecting an eye and a mouth position from the face area extracted by the face area extracting unit; and an eye and a mouth position detected by the eye and the mouth detecting means come to a preset position. Affine transformation means for performing a affine transformation in such a manner as to form a face image vector having a dimension equal to the number of pixels from the affine transformed image data, and a face image vector sample created from the face image used as a reference for each person. A registered sample set and a five-layer perceptron identity mapper. Based on learning of the five-layer perceptron using the sample set, the affine transformation unit Non-linear principal component analysis means for performing dimensional compression of the face image vector and outputting a compressed face image vector, and dimensional compression of the face image vector included in the sample set using the three-layer perceptron of the non-linear principal component analysis means A reference face image database unit for registering a set converted into the compressed face image vector, and the three-layer perceptron, wherein the nonlinear face is formed based on learning using the compressed face image vector included in the reference face image database unit. A face image identification unit that outputs an identification result when a compressed face image vector to be identified is input from the principal component analysis unit.