JP2019096146A - Image identification device, image identification method, computer program, and storage medium - Google Patents

Abstract

To provide an image identification device which achieves high accuracy face authentication.SOLUTION: In an image identification device, a face image for registration included in the image acquired by an image acquisition part 110 is registered in a face authentication dictionary storage part 150, and a pair dictionary of a low resolution dictionary and a high resolution dictionary which are created by learning the face image for registration is stored in a high resolution dictionary storage part 190. In the image identification device, a high resolution image is created by high resolution enhancement of the face image for registration included in the image acquired by the image acquisition part 110 using the pair dictionary, and the face image for authentication is identified on the basis of the created high resolution image and the face image for registration.SELECTED DRAWING: Figure 2

Description

  The present invention relates to image identification technology.

  2. Description of the Related Art Face authentication technology for extracting a face image of a person from an image captured by an imaging device such as a camera and identifying the person from the face image has been widely used. Non-Patent Document 1 discloses a technique for extracting a feature quantity called an LBP (Local Binary Pattern) feature from an input luminance face image, and identifying a person according to the LBP feature quantity. In this technique, a person is identified from the input face by comparing the LBP feature amount with the feature amount extracted from the luminance face image registered in advance. Such face recognition technology has been used under relatively good imaging conditions such as a short distance from an imaging device to an object as seen in the automatic shutter and entry / exit management of digital cameras.

  In recent years, research has been conducted on face recognition for low resolution images such as small faces and blurs in distant places such as obtained by surveillance cameras. One of the means for face recognition using low resolution images is to increase the resolution of images. For example, Non-Patent Document 2 discloses a technique called Face Halucination which approximates a face image of one person by a linear sum of face images of another person. Non-Patent Document 3 discloses a technology using a hallucination technology that performs face recognition using a high resolution image.

・ Description of hallucination technology
The hallucination technology is a technology for generating a high resolution face image from a low resolution face image. The outline of the principle is that an input low resolution face image is approximated by another person's high resolution face image. Details of the hallucination technology will be described with reference to the illustration of the hallucination technology in FIG.
As shown in FIG. 16, a high resolution dictionary (Formula 1) in which a high resolution dictionary and a low resolution dictionary are paired is previously prepared by learning using face images of various persons. A plurality of pairs of high resolution dictionary and low resolution dictionary are stored as shown in Equation 2. Further, each of the first pair constituting the high resolution dictionary, the low resolution dictionary constituting the first pair, and the high resolution dictionary constituting the first pair is described as Expression 3. . The relationship between the high resolution dictionary, the low resolution dictionary, and the high resolution dictionary is as shown in Formula 4.

When a luminance image is used as the high resolution dictionary D, a high resolution luminance image and a low resolution luminance image are stored as a pair. For example, a high resolution luminance image is an image as clear as possible enough to identify who the face in the video is. Low-resolution luminance images are images in which it is difficult to identify who is the face in the video because the face is too small or the face is blurred.
Next, the low-resolution partial images are cut out from the input low-resolution face image I _L. This low resolution partial image is approximated by the linear sum of the low resolution dictionary among the high resolution dictionary and the low resolution dictionary (Dual Equation 2) stored in the high resolution dictionary D. Equation 5 is the result of the approximation. Thereby, a low resolution dictionary approximating a low resolution partial image and coupling coefficients α (α 1, α 2, α 3,...) Are obtained. Then, a high resolution partial image is generated using the high resolution dictionary corresponding to the low resolution dictionary and the coupling coefficient α (α1, α2, α3,...). Equation 6 represents the generated high resolution partial image.

  The high-resolution dictionary D may use not a luminance image but a base image common to face images such as edges. As an example of the base image, there is an eigenface by principal component analysis as in Non-Patent Document 4.

T. Pajdla, and J. Matas, “Face Recognition with Local Binary Patterns”, ECCV, pp. 469-481, 2004 K. Huang, R. Hu, “Facehallucination via K-selection mean constrained sparse representation,” ICIP, pp. 882-885, 2012 B. Li, H. Chang, “Hallucinating Facial Images and Features”, ICPR, pp. 1-4, 2008 M. Turk, A. Pentland, Eigenfaces for Recognition, Journal of cognitive neurosicence, Vol. 3, No. 1, 1991, pp. 71-86

  As in Non-Patent Document 2, in the Face Hallucination technology that approximates with the linear sum of another person's face, the high resolution dictionary D prepared in advance is important for high resolution of the image. In addition to small information such as small faces and blurs, the high-resolution dictionary D needs an expression ability to sufficiently approximate the faces of various persons even in a state in which noise is added. Otherwise, the face image after high resolution will be a face image that is not similar to the person, and it may be determined to be a different person than the person even if it is input to face recognition technology There is However, it is very difficult to prepare a high-resolution dictionary D that can cope with noise as well as sufficiently represent the faces of various persons from small information such as small faces and blurs.

  The present invention has been made in view of the above problems, and provides an image identification apparatus that realizes face recognition with high accuracy even for a face image including distant faces such as small faces and blurs, and noise. The purpose is to

  The image identification apparatus according to the present invention comprises an image acquisition unit for acquiring an image, a registration unit for registering a face image for registration included in the image acquired by the image acquisition unit, and learning the face image for registration. A learning unit that generates a pair dictionary of a low resolution dictionary and a high resolution dictionary, a storage unit that stores the generated pair dictionary, and an authentication face image included in an image acquired by the image acquisition unit High resolution image generation means for generating a high resolution image by using the pair dictionary for high resolution, and the high resolution image generated with the high resolution image generation means and the registration And face identification means for identifying the face image for authentication based on the face image for registration registered in the means.

  According to the present invention, it is possible to realize face recognition with high accuracy even for a face image including a small face or blur that is far away and noise.

Explanatory drawing of the kind of face at the time of learning a high resolution dictionary. The functional block diagram of an image identification device. 6 is a flowchart illustrating face identification processing. 6 is a flowchart illustrating face identification processing. The detailed block diagram of a high-resolution dictionary learning data generation part. 6 is a flowchart showing high-resolution dictionary learning data generation processing. Explanatory drawing of a cutout image. Explanatory drawing of a cutout image. Explanatory drawing of a low resolution image. Explanatory drawing of noise addition in the case of supposing a video compression noise. The example figure when dividing learning data into a plurality of blocks. Explanatory drawing of a pair dictionary. Explanatory drawing of a pair dictionary. Explanatory drawing of the effect which learns a high resolution dictionary using only the registered face image. Explanatory drawing of the effect by a high resolution dictionary. Illustration of hallucination technology.

  Hereinafter, embodiments will be described in detail with reference to the drawings.

  Focusing on face recognition technology, the types of faces to be approximated are limited. That is, since the face recognition technology is a technology for determining that the input face image is the face of the same person as the face image registered in advance, the high resolution dictionary D capable of sufficiently expressing any face is learned. There is no need. Therefore, the face recognition technology may prepare the high resolution dictionary D capable of sufficiently expressing only the face registered in advance. In the present embodiment, by using the registered face image registered in the face authentication as the high resolution dictionary D, the high resolution dictionary D that sufficiently approximates only the registered face image is generated. In addition, a moving image is generated from the registered face image so that noise caused by moving image compression that can not be reproduced only by image processing can be coped with, and a high resolution dictionary is generated using the face image extracted from the generated moving image. Learn D

  FIG. 1 is an explanatory view of the type of face when learning the high-resolution dictionary D used in the hallucination technology. In the prior art, the high resolution dictionary D is learned using face images (for example, Mr. X, Mr. Y, Mr. Z) different from the person registered in the face authentication, for example, collected from the Internet in advance. . Then, in face recognition, the high resolution dictionary D is used to perform high resolution on the input face image.

  On the other hand, in the present embodiment, one common high-resolution dictionary D is learned from registered face images (for example, Mr. A, Mr. B) (first method) or from registered face images The high resolution dictionary D for each individual is learned (second method). This is because, as described above, it is only necessary to determine whether or not a person is a registered person in face recognition technology, so it is intended to improve face recognition accuracy by learning a high resolution dictionary D that approximates only a registered face image. It is.

(Constitution)
FIG. 2 is a functional block diagram of the image identification apparatus of the present embodiment. The image identification apparatus is realized by an information processing apparatus including a central processing unit (CPU), a read only memory (ROM), and a random access memory (RAM) (not shown). The information processing apparatus includes an imaging device such as a camera, or is connected to the imaging device. The CPU causes the information processing apparatus to function as an image identification apparatus by executing the computer program stored in the ROM using the RAM as a work area. The image identification apparatus functions as an image acquisition unit 110, a face detection unit 120, a face organ detection unit 130, a feature extraction unit 140, a face authentication dictionary storage unit 150, and a face identification unit 160 for face authentication. The image identification apparatus also includes a high resolution dictionary learning data generation unit 170, a high resolution dictionary learning unit 180, a high resolution dictionary storage unit 190, and a high resolution image generation for high resolution. It functions as the unit 200. Each functional block may be realized at least in part by hardware.

(processing)
The image identification apparatus includes a first operation mode for registering a face image and learning of a high resolution dictionary by each functional block as described above, and a second operation mode for performing high resolution of an image and face recognition. Works with FIG. 3 and FIG. 4 are flowcharts showing face identification processing in the first and second operation modes.

  The image identification apparatus first determines whether the operation mode is the first mode (S1000). The operation mode is set by the user using, for example, an input device (not shown). In the case of the first mode, the image identification device starts the operation in the first operation mode (S1000: Y). In the case of the second operation mode, the image identification apparatus starts the operation in the second operation mode shown in FIG. 4 (S1000: N).

(First operation mode: Image registration and high-resolution dictionary learning mode)
The first operation mode is an operation mode for registering a face image used for face authentication and learning a dictionary for enhancing the resolution of the image. Here, the registration image used to register the face image includes a face image of a person to be registered, and is assumed to be a high resolution image photographed under a good illumination environment. The high resolution image is an image having a resolution that can identify a person from a face image.

  The image acquisition unit 110 acquires a registration image from the imaging device (S1100). The imaging device includes a focusing element such as a lens, an imaging element that converts light into an electrical signal, and an AD converter that converts an analog signal into a digital signal, and generates digital image data representing a captured image. The imaging device is, for example, a complementary metal oxide semiconductor (CMOS) image sensor or a charge coupled device (CCD) image sensor. The image acquisition unit 110 acquires a registration image as digital image data. The image acquisition unit 110 acquires, for example, a registration image converted to VGA (640 × 480 [pixel]) or QVGA (320 × 240 [pixel]) by performing thinning processing or the like on the registration image. It is also possible. The image acquisition unit 110 may also acquire an image for registration via a storage medium such as a flash memory, in addition to the imaging device. In any case, the image acquisition unit 110 acquires an image from an external device. The image acquisition unit 110 transmits the acquired registration image to the face detection unit 120 and the high-resolution dictionary learning data generation unit 170.

The face detection unit 120 detects the position of the center of gravity of the face, the left and right eyes, and the mouth from the registration image (S1200). For example, “P. Viola, M. Jones,“ Rapid Object Detection using a Boosted Cascade of Simple Features ”, in Proc. Of CVPR, vol. 1, pp. 511-518, December, 2001. This process is performed according to the disclosed technology.
The face detection unit 120 is a face image in which the size of the face is a predetermined size and the orientation of the face is erected from the detected center of gravity of the face, the left and right eyes, and the mouth using affine transformation. A first normalized image is generated (S1201). The size of the face may be defined as Euclidean distance between the left and right eyes. In the first operation mode, the face detection unit 120 transmits the generated first normalized image to the face organ detection unit 130.

The facial organ detection unit 130 detects the center-of-gravity position of finer feature points such as the corner of the eye and the corner of the eye from the first normalized image (S1300). The feature extraction unit 140 is, for example, “TF Cootes, CJ Taylor, DH Cooper, and J. Graham,“ Active Shape Models-The Training and Application ”, Computer Vision and Image Understanding, Vol. 61, No. 1, January, pp. This process is performed according to the technology disclosed in “38-59, 1995”.
The face organ detection unit 130 generates a second normalized image which is a face image having a face size of a predetermined size and an orientation of the face erected using the barycentric position of the detected feature point. (S1301). In the first operation mode, the face organ detection unit 130 transmits the generated second normalized image to the feature extraction unit 140 and the high-resolution dictionary learning data generation unit 170.

  The feature extraction unit 140 sets a feature extraction region for the second normalized image, and extracts LBP features as shown in Non-Patent Document 1 for the set region (S1400). The feature extraction unit 140 causes the face authentication dictionary storage unit 150 to store a dictionary used for face authentication in which the extracted LBP feature is linked to a personal ID specifying a person of the face (S1500).

  The processes of S1000 to S1500 described above are registration processes generally performed in face authentication. The face authentication dictionary storage unit 150 is a registration unit in which a face image for registration is registered. In the following process, the image identification apparatus learns the high resolution dictionary D necessary for high resolution using the face image for registration.

  The high-resolution dictionary learning data generation unit 170 learns the high-resolution dictionary D based on the registration image received from the image acquisition unit 110 and the second normalized image received from the facial organ detection unit 130. Data (high-resolution dictionary learning data) necessary for the image processing is generated (S1700). Specifically, the high-resolution dictionary learning data generation unit 170 generates, from the registration image, a pair image in which the high-resolution image and the low-resolution image are in one-to-one correspondence. FIG. 5 is a detailed block diagram of the high-resolution dictionary learning data generation unit 170. As shown in FIG. The high-resolution dictionary learning data generation unit 170 includes an image cutout unit 171, a reduced image generation unit 172, a blur imparting unit 173, and a noise imparting unit 174. FIG. 6 is a flowchart showing the high-resolution dictionary learning data generation process of S1700.

  The image cutout unit 171 generates at least one cutout image from the second normalized image generated in the process of S1301. 7 and 8 are explanatory diagrams of the cutout image. In the example of FIG. 7, the image clipping unit 171 generates a clipped image 17101 in which the face area center coincides with the clipped image center and a clipped image 17102 in which the face area center is on the left side of the clipped image center. In the process of S1200 (face detection) or the process of S1300 (face organ detection), a detection error occurs to some extent. In order to create a high resolution dictionary D capable of coping with this detection error, a cutout image 17102 is generated. In the example of FIG. 8, the image cropping unit 171 uses the 3D model so as to be able to cope with the orientation of the face as well as the position, and uses the 3D model to cut out the cut-out image 17101, which is a face image facing front, Generate 17103 The processing of the 3D model is disclosed, for example, in "I. Kemelmacher-Shlizerman," 3D Face Reconstruction from a Single Image Using a Single Reference Face Shape ", PAMI, pp. 394-405, 2011.

As described above, the high-resolution dictionary learning data generation unit 170 learns the high-resolution dictionary D by clipping from the normalized registration image (second normalized image) and using the 3D model. To generate the high resolution image required for
Next, the high-resolution dictionary learning data generation unit 170 generates a low-resolution image necessary for learning the high-resolution dictionary D. In the present embodiment, a low resolution image is generated by image reduction, blurring, and noise. Whether blurring and noise are to be applied by the user using, for example, an input device (not shown) is set. Therefore, a low resolution image may be generated only by image reduction, or at least one of blur and noise may be added to the reduced image to generate a low resolution image. FIG. 9 is an explanatory view of a low resolution image.

  The reduced image generation unit 172 generates at least one reduced image from the second normalized image (S1720). This processing is to enable high resolution for a small face located far away. There are various image enlargement / reduction methods such as bicubic and bilinear. In this embodiment, bilinear is used. In the example of FIG. 9, a reduced image 17201 is generated from the cutout image 17101. In addition, a reduced image may be generated from the face image 17103 facing diagonally.

  The blur imparting unit 173 determines whether to apply blur to the reduced image 17201 generated by the reduced image generation unit 172 (S1730). When blur is added (S1730: Y), the blur applying unit 173 generates an image 17301 (see FIG. 9) in which the reduced image 17201 is added with blur (S1731). There are various methods for blurring, but the blurring unit 173 of the present embodiment applies blurring using a Geisian filter according to a predetermined kernel size and standard deviation. This is to make it possible to achieve high resolution even for blurring caused by a focal length or the like.

  If blurring is not applied (S1730: N), or after blurring is applied, the noise applying unit 174 determines whether noise is applied to the reduced image 17201 (S1740). When applying noise (S1740: Y), the noise applying unit 174 generates an image 17401 (see FIG. 9) obtained by applying noise to the reduced image 17201 (S1741). When sensor noise is assumed, the noise giving unit 174 gives Geithian noise according to a predetermined average and standard deviation. In addition to sensor noise, noise addition in the case of assuming moving image compression noise over a frame such as H264 will be described with reference to FIG. As shown in FIG. 10, the noise giving unit 174 generates a plurality of converted face images with shifted face positions from the reduced image, and uses a codec such as H264 for the generated plurality of converted face images. Generate moving pictures. The noise giving unit 174 gives a moving image compression noise by acquiring a still image from the generated moving image again.

If noise is not added (S1740: N) or noise is added, the high-resolution dictionary learning data generation unit 170 ends the low-resolution image generation processing. By this processing, the high-resolution dictionary learning data generation unit 170 generates various low-resolution reduced images 17201 and images 17301 and 17401 as shown in FIG. 9 from the registration image.
As described above, the high-resolution dictionary learning data generation unit 170 performs the process of S1700 to use a pair image of the high-resolution image and the low-resolution image as learning data to be used for learning the high-resolution dictionary D. Generate

When the process of generating learning data for learning the high resolution dictionary D is completed, the high resolution dictionary learning unit 180 performs learning of the high resolution dictionary D using the learning data (S1800). . Since resolution enhancement is performed for each local region as shown in FIG. 16, the resolution enhancement dictionary learning unit 180 divides the learning data generated in the process of S1700 into a plurality of blocks. FIG. 11 is an exemplary view when learning data is divided into a plurality of blocks. The p-th block high-resolution dictionary D ^p is expressed by Equation 1.

For example, as disclosed in Non-Patent Document 2, the learning of the p-th block high-resolution dictionary D ^p satisfies the following equation 7, that is, the p-th block image I ^p and D ^p α ^p It is carried out by finding D ^p and α ^p which minimize the L 2 norm of The second term on the right side of Equation 7 is a penalty term for preventing overfitting, and is generally also called an L1 regularization term.

The p-th block image I ^p is a p-th block high resolution image obtained by dividing the cutout image 17101 from the registration image into blocks, and a p-th block low resolution image generated from the cutout image 17101; Is a connected image. The low resolution images are the reduced image 17201 and further images 17301 and 17401 to which noise and blur are added.
The high resolution dictionary D ^p is a pair dictionary in which the high resolution dictionary and the low resolution dictionary of the p th block are connected. The high-resolution dictionary D ^p is set to an initial value such as a random number at the first stage of learning. However, higher resolution dictionary D ^p is in advance by learning a high resolution dictionary by using a face image of a person that is not registered in advance, which may be set as the initial value.
W1 and W2 are sizes of one divided image respectively. α ^p is a linear combination coefficient.

Although there are various methods for satisfying the formula 7 in this embodiment, the K-SVD method is used as in Non-Patent Document 2 in the present embodiment. That is, the high-resolution dictionary learning unit 180 approximates the learning image I ^p with the linear combination D ^p α ^p of the current high-resolution dictionary D ^p , and then performs singular decomposition on the difference to obtain an eigenvalue The coefficients of the high-resolution dictionary are updated in the direction in which the difference is reduced by using the size of. Therefore, the high resolution dictionary DI ^p created here is a base image such as an eigenface. However, the processing of the high-resolution dictionary learning unit 180 is not limited to this.

  FIG. 12 is an explanatory diagram of a pair dictionary generated by the learning process of S1800. When the learning process of S1800 ends, as shown in FIG. 12, it is possible to acquire a pair dictionary (high resolution dictionary) of the high resolution dictionary and the low resolution dictionary for each block position.

The high resolution dictionary generated by the first method described at the beginning learns one common high resolution dictionary from the registered face image. The image I ^p is a divided image of the p-th block of all registered persons. Therefore, in the first method, one learning dictionary (pair dictionary) is generated after learning as shown in FIG.

The high resolution dictionary generated by the second method learns the individual high resolution dictionary from the registered face image. The image I ^p is a divided image of the p-th block of the same person among the registered face images. Therefore, in the second method, after learning, a high resolution dictionary (pair dictionary) for each person is generated as shown in FIG. FIG. 13 is also an explanatory diagram of a pair dictionary generated by the learning process of S1800.

  The high-resolution dictionary learning unit 180 that has generated the high-resolution dictionary stores the generated high-resolution dictionary together with the block position p in the high-resolution dictionary storage unit 190, which is a memory (S1900). As described above, the processing of the first operation mode (image registration and high-resolution dictionary learning mode) ends.

(Second operation mode: high resolution and face recognition mode)
The second operation mode is an operation mode in which the resolution is increased for the face image for authentication (image for authentication) to be input, and the face authentication is performed using the high resolution face image for authentication It is.

  The processes of S1101 to S1203 illustrated in FIG. 4 are the same as the processes of S1101 to S1201 of the first operation mode. However, the image acquired by the image acquisition unit 110 is an authentication image, and the first normalized image generated by the face detection unit 120 is generated from the authentication image. In the second operation mode, the face detection unit 120 transmits the face detection result and the first normalized image to the high resolution image generation unit 200.

The high resolution image generation unit 200 acquires the high resolution dictionary D stored in the process of S1900 from the high resolution dictionary storage unit 190 (S2000). The high resolution image generation unit 200 performs high resolution on the first normalized image using the high resolution dictionary D (S2001). First, the high resolution image generation unit 200 determines whether to perform high resolution based on the face detection result in S1201. For example, when the Euclidean distance between the left and right eyes is equal to or less than a predetermined threshold, the high-resolution image generation unit 200 determines that high resolution is to be performed.
When high resolution is to be performed, the high resolution image generation unit 200 divides the input face image j _L (first normalized image) normalized as shown in FIG. 11 into a plurality of blocks. Next, the high resolution image generation unit 200 generates a linear sum using the low resolution dictionary of the high resolution dictionary D ^p as the image of the p th block of the input face image j _L is expressed by Equation 8. And approximate linear combination coefficients.

  Next, the high resolution image generation unit 200 generates a high resolution image of the p-th block from the linear sum of the linear combination coefficient and the high resolution dictionary corresponding to the low resolution dictionary as shown in Formula 9.

The high-resolution image generation unit 200 performs the above-described processing on all blocks. In the case of the first method (learning one common high-resolution dictionary from registered face images), the high-resolution dictionary is common to all registered face images. Therefore, one high resolution image is generated from the p-th block image of the input face image. In the case of the second method (learning a high resolution dictionary for each individual from a registered face image), the high resolution dictionaries exist as many as the number of registered persons. Therefore, n high resolution images are generated from the p-th block image of the input face image.
As described above, the high-resolution image generation unit 200 performs processing for generating a high-resolution image from the p-th block image of the input face image for all blocks, and connects these results to obtain a high-resolution solution. Generate an image. In the second method described above, it goes without saying that n high resolution images can be obtained from one authentication face image.

The face organ detection unit 130 generates a second normalized image from the high resolution image generated by the high resolution image generation unit 200 by the processing similar to the processing of S1300 and S1301 of FIG. 3 (S1302, S1303). The feature extraction unit 140 extracts the feature amount from the second normalized image by the same process as the process of S1400 in FIG. 3 (S1401). The face identification unit 160 acquires a dictionary used for face authentication from the face authentication dictionary storage unit 150 (S1600). As exemplified in Non-Patent Document 1, the face identification unit 160 performs face authentication based on the acquired dictionary used for face authentication and the feature value extracted in the process of S1401 (S1601).
Face recognition based on the face image registered as described above is performed.

  An effect of learning the high resolution dictionary D using only the registered face image will be described. FIG. 14 is an explanatory diagram of an effect of learning the high resolution dictionary D using only the registered face image. If the input low-resolution face image is Mr. A, then raising the resolution of the low-resolution face image using the high-resolution dictionary generated from Mr. A's face image resembles Mr. A High resolution face image is generated. When raising the resolution of the low resolution face image of Mr. A using the high resolution dictionary generated from Mr. B's face image, a high resolution face image resembling Mr. B is generated. That is, when the input low-resolution face image matches the person in the high-resolution dictionary, a high-resolution face image like the individual is generated, but the input low-resolution face image and the high-resolution face image are high. When the person with the resolving dictionary does not match, a high resolution face image that is not like the person is generated.

  FIG. 15 is an explanatory diagram of the effect of the high resolution dictionary D. As shown in FIG. As shown in the figure, learning is performed using only the face image of Mr. A rather than the similarity between the face image of Mr. A and the registered face image that have been enhanced with the high-resolution dictionary D learned with face images of various persons. The degree of similarity between the face image of Mr. A who has been made high resolution by the high resolution dictionary D and the registered face image is higher. As described above, learning the high resolution dictionary D using only the registration image has an effect of improving the face authentication accuracy.

  The image identification apparatus according to the present embodiment as described above learns and uses a high resolution dictionary D that can express only a registered image used for face authentication, thereby achieving high resolution for low resolution face images. Accurate face recognition can be realized. In addition, the image identification apparatus generates a moving image from the registered face image, and uses the face image cut out from the generated moving image for learning of the high resolution dictionary D to perform only image processing such as Gaussian noise. It also copes with the noise caused by moving picture compression that can not be reproduced with. Therefore, the image identification apparatus can realize highly accurate face recognition.

  The present invention supplies a program that implements one or more functions of the above-described embodiments to a system or apparatus via a network or a storage medium, and one or more processors in a computer of the system or apparatus read the program. It can also be realized by the process to be executed. It can also be implemented by a circuit (eg, an ASIC) that implements one or more functions.

Claims (12)

Image acquisition means for acquiring an image;
Registration means for registering a face image for registration included in the image acquired by the image acquisition means;
Learning means for learning the face image for registration to generate a pair dictionary of a low resolution dictionary and a high resolution dictionary;
Storage means in which the generated pair dictionary is stored;
A high resolution image generation unit configured to generate a high resolution image by performing high resolution on a face image for authentication included in the image acquired by the image acquisition unit using the pair dictionary;
Face identification means for identifying the face image for authentication based on the high resolution image generated by the high resolution image generation means and the face image for registration registered in the registration means; Characterized in that,
Image identification device. The learning means is characterized in that it generates the pair dictionary which is a dictionary common to all persons registered in the registration means.
The image identification device according to claim 1. The learning means generates the pair dictionary for each person registered in the registration means.
The image identification device according to claim 1. The learning means is characterized in that a dictionary obtained by learning a face image not registered in the registration means is set as an initial value, and the face image for registration is learned to generate the pair dictionary.
The image identification device according to any one of claims 1 to 3. The high resolution image generation unit generates a plurality of high resolution images for each face image acquired by the image acquisition unit using a pair dictionary for each person stored in the storage unit. Characterized by
The image identification device according to claim 4. It further comprises learning data generation means for generating a low resolution face image and a high resolution face image from the image acquired by the image acquisition means,
The learning means learns the low resolution face image and the high resolution face image to generate the low resolution dictionary and the high resolution dictionary.
The image identification device according to any one of claims 1 to 5. The learning data generation unit generates a moving image composed of a plurality of converted face images having different positions and sizes of faces from the image acquired by the image acquiring unit, and extracts a face image extracted from the moving image. Generate
The learning means is characterized in that the pair dictionary is generated by learning the extracted face image.
The image identification device according to claim 6. The learning data generation means generates the high resolution face image by cutting out a face image from the image acquired by the image acquisition means, and reduces the cut out face image to reduce the low resolution face Characterized by generating an image,
The image identification device according to claim 6. The learning data generation means generates the low resolution face image by adding at least one of blur and noise to the reduced face image.
The image identification device according to claim 8. The information processing apparatus
A low resolution dictionary and a high resolution dictionary generated by extracting a face image for registration of a person from an image acquired from an external device and registering the face image in a predetermined registration unit and learning the face image for registration Store the pair dictionary of in the predetermined storage means,
A high resolution image is generated by increasing the resolution of the face image for authentication acquired from the external device using the pair dictionary, and the high resolution image and the registration unit registered in the registration unit Performing face authentication of the face image for authentication based on the face image;
Image identification method. Computer,
Image acquisition means for acquiring an image,
Registration means for registering a face image for registration included in the image acquired by the image acquisition means;
Learning means for learning the face image for registration to generate a pair dictionary of a low resolution dictionary and a high resolution dictionary;
Storage means for storing the generated pair dictionary;
A high resolution image generation unit that generates a high resolution image by raising the resolution of the face image for authentication included in the image acquired by the image acquisition unit using the pair dictionary.
Face identification means for identifying the face image for authentication based on the high resolution image generated by the high resolution image generation means and the face image for registration registered in the registration means;
Computer program to function as.   A computer readable storage medium storing the computer program according to claim 11.