JP2008217688A - Collation device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To shorten collation processing time by solving the problem that very long response time is required in collation of one-to-multiple authentication. <P>SOLUTION: An image coincident with a collation image is specified from registration images by a block similarity calculation part 105 which calculates block similarities for each block for a collation image and registration images each of which are divided into predetermined blocks, a first similarity calculation part 106 which selects a registered image in which all block similarities exceed a predetermined threshold, and calculates a first similarity between the collation image and the registration image based on the block similarities, and a second similarity calculation part 107 which calculates, when the registration image with the similarities exceeding the threshold is selected by the calculation part 106, a second similarity between the collation image and the registration image. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a matching technique, and more particularly, to a matching technique used in a biometric authentication system based on images such as fingerprints, finger veins, palm veins, and irises.

  Biometric authentication technology is used to determine whether a person accessing the system is qualified when accessing important information or areas such as bank ATM authentication, office entry / exit management, and computer system login procedures. Is done. The basis of biometric authentication technology is performed based on collation of two images or the like. The collation is performed by determining whether or not they are similar by calculating the similarity between the two images. A typical method of matching is normalized correlation matching.

と照合画像

And matching image

との正規化相関

Normalized correlation with

は次の式で与えられる。

Is given by:

However,

  The normalized correlation takes a value from −1 to 1, and the similarity is higher as the value is larger. Therefore, by obtaining the maximum value of the normalized correlation, an image most similar to an image given from a set including a plurality of images can be extracted. Normalized correlation has the advantage that it is resistant to uniform density changes throughout the image and can accommodate minor geometrical changes. However, there is a problem that the calculation amount is large due to the necessity of calculating the square root.

  A technique has been proposed in which biometric authentication data is divided into arbitrary sizes and recorded, and authentication results are obtained in divided units to obtain authentication results in divided units to improve performance (see, for example, Patent Document 1). .

  This technology divides and registers biometric authentication data related to biometric authentication into an arbitrary size, and authenticates the registered biometric authentication data with biometric authentication data captured via a sensor in units of division. Is to do. Further, by setting and defining an authentication level for each division unit, it is possible to perform partial authentication or highly flexible authentication by designating a portion to be focused on.

JP 2006-85249 A

  As described above, the image collation process tends to be computationally intensive, and there is a demand for speeding up the calculation process in consideration of practicality. In particular, in the verification of one-to-many authentication, that is, the processing for extracting the registered image that most closely matches the verification target image from N registered images registered in advance, the processing speed is required to be increased. If one-to-many authentication collation is many, that is, if N is a large number, if an accurate collation result is required, collation processing with a large amount of calculation such as normalized correlation matching must be executed many times.

  By the way, generally the response time regarding the collation processing permitted is several seconds or less. However, in the verification of one-to-many authentication, there is a problem that this response time is often exceeded.

For example, in the case of one-to-many authentication in a large-scale biometric authentication system used in the entire company, the number N of registered images corresponds to the number of employees, for example. Therefore, N is in the order of hundreds to tens of thousands, and it has been difficult to achieve a required response time of several seconds or less using a general one-to-many authentication method.
An object of the present invention is to shorten the verification processing time.

  According to one aspect of the present invention, there is provided a collation apparatus that performs one-to-many collation that collates a collation target image and a plurality of registered images, and divides the collation target image and the registered image into a plurality of blocks, and And an inter-block similarity calculating unit that calculates an inter-block similarity, and excluding registered images whose inter-block similarity is less than a first threshold, and only the registered images that are greater than or equal to the first threshold A first similarity calculating unit that calculates a first similarity between the collation image and the registered image based on the inter-block similarity calculated by the inter-similarity calculating unit, and the first similarity is a second threshold value And a second similarity calculating unit that calculates a second similarity between the matching image and the registered image for only registered images that are equal to or greater than the second threshold, Matches the matching image in the image Collating apparatus is provided, wherein the identifying things.

The advantages of the authentication method of the present invention are as follows.
1) The comparison between the similarity between the blocks with a relatively small amount of calculation and the first similarity eliminates many registered images, and the amount of the second similarity with a relatively large amount of calculation can be minimized. Is achieved.
2) The present invention is applicable even when the registered image and the collation image are binary images or images with a low density level.
3) Collation of similarities between blocks handles information in units of blocks, and is not easily affected by fine noise.

  Hereinafter, a verification technique according to an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a functional block diagram showing a configuration example of a collation device according to the present embodiment. The collation device shown in FIG. 1 includes a processing device 104 that performs similarity calculation processing, a storage device 108, and input devices a1 to an (n is an integer). Furthermore, you may provide the display 111 for confirming collation. The input device a1... Is connected to an image input device (image acquisition device) 101 such as a digital camera, an input interface such as a keyboard 102 and / or a mouse 103, and the processing device 104. The image input device 101 includes, for example, an imaging device that acquires grayscale images such as fingerprints, finger veins, and palm veins. The keyboard 102 and / or the mouse 103 are used for starting commands and inputting parameters. In these input systems, many interfaces from a1 to an are connected to the processing device 104 in many cases. A display 111 used as an output device is also coupled to the system. The display 111 displays a system status, an input prompt, and the like.

  The processing device 104 includes an inter-block similarity calculation unit 105, a first similarity calculation unit 106, and a second similarity calculation unit 107. The storage device 108 includes a registered image storage unit 109 and a collation image storage unit 110, and the collation image storage unit 110 stores at least an image to be collated temporarily at the time of collation processing. The three similarity calculation units 105, 106, and 107 of the processing device 104 compare the registered image 109 and the collation image 110, and calculate the similarity between them.

  FIG. 2 is a flowchart showing the flow of collation processing according to this embodiment. As described in Japanese Patent Application Laid-Open No. H10-228688, the entire image corresponding to each of the collation image and the registered image is divided into a plurality of block areas for the areas corresponding to the collation image and the registration image. The number of blocks to be divided is not limited, but may be about 3 to 100, for example.

  Next, symbols used in FIG. 2 will be described. Variable i is the index of the registered image. The constant I is the total number of registered images.

  The verification image and the registered image are divided into a plurality of corresponding blocks as described above, and the variable j is an index of the block. The constant J is the total number of blocks. The variable R indicates the similarity between the block j of the collation image and the block j of the registered image i. The constant RT is a threshold for similarity between blocks. Similarity / dissimilarity is determined using this threshold value. The variable R1 is a value indicating the first similarity between the collation image and the registered image i. The constant R1T is a threshold value for the first similarity. The variable R2 is a variable indicating the second similarity between the collation image and the registered image i. The constant RT2 is a threshold regarding the second similarity. The variable R2Max and the variable iMax are variables used when calculating the maximum value of the second similarity R2.

  In FIG. 2, first, in step 201, an initial value is set for the variable R2Max. After step 202, a loop for variable i is formed. In step 202, an initial value 0 is set in the variable i, and in step 203, it is determined whether or not processing has been completed for all registered images. If the process has not been completed (YES), a loop for variable j after step 206 is entered.

  In step 205, an initial value 0 is set in the variable j, and in step 206, it is determined whether or not processing has been completed for all blocks. If not completed (YES), in step 208, the similarity R between blocks is calculated. In step 209, it is determined whether the similarity R between the blocks is greater than or equal to a threshold RT. If so (YES), the process proceeds to step 207, the value of variable j is incremented by 1, and the loop for variable j is repeated. Otherwise (NO), since it is determined that the collation image and the registered image i are not similar as a whole, the value of the variable i is increased by one in step 204, and the loop for the variable i is repeated. If it is determined in step 206 that all blocks have been processed (NO), the process exits the loop for variable j and proceeds to step 210.

  In step 210, a first similarity R1 between the collation image and the registered image I is calculated. In step 211, it is determined whether or not the first similarity R1 is greater than or equal to a threshold value R1T. If the first similarity R1 is not greater than or equal to the threshold value R1T (NO), the process proceeds to step 204, the value of the variable i is increased by 1, and the loop for the variable i is repeated. If the first similarity R1 is equal to or greater than the threshold value R1T (YES), that is, if all the corresponding blocks of the images to be compared are similar, the similarity based on the first similarity is recognized, so that The similarity based on the following second similarity is determined only for a simple image. That is, the process proceeds to step 212, and the second similarity R2 between the collation image and the registered image i is calculated. In step 213, it is determined whether the second similarity R2 is greater than or equal to a variable R2Max. If not (NO), the process proceeds to step 204, the value of the variable i is incremented by 1, and the loop for the variable i is repeated. If it is above (YES), the process proceeds to step 214, and the value of the second similarity is substituted for the variable R2Max. In step 215, the value of the index i of the registered image at this time is stored in the variable iMax. Next, the value of variable i is incremented by one (step 204), and the loop for variable i is repeated.

  If all the registered images have been processed in step 203 (NO), the process exits the loop for variable i, proceeds to step 216, and determines whether variable R2Max is greater than or equal to threshold value R2T. If it is above (YES), the process proceeds to step 217, the collation is successful, and all the processes are completed. If not (NO), the process proceeds to step 218, the collation fails and the entire process is terminated.

  Hereinafter, an example in which the present invention is applied to finger vein authentication will be described more specifically. The similarity R between blocks can be defined by the histogram intersection as follows (see Reference 1 below). A density histogram of a block in an image

とする。

And

の値はヒストグラムの性質上、そのブロックの画素数となる。正規化ヒストグラム

The value of is the number of pixels in the block due to the nature of the histogram. Normalized histogram

とは濃度ヒストグラムの各要素を画素数で割ったものである。すなわち、

Is obtained by dividing each element of the density histogram by the number of pixels. That is,

は１に正規化されている。登録画像のブロックの正規化ヒストグラムを

Is normalized to 1. Registered image block normalization histogram

とし、対応する照合画像のブロックの正規化ヒストグラムを

And the normalized histogram of the corresponding matching image block

とすると、ヒストグラムインタセクション

Then the histogram intersection

は次の式で定義される。

Is defined as:

  FIG. 3 is a diagram showing the concept of a histogram intersection for one block. FIG. 3A shows a normalized histogram of a block of a registered image, and FIG. 3B shows a normalized histogram of a corresponding block of a collation image. In both FIG. 3A and FIG. 3B, the horizontal axis represents pixel density (brightness level: luminance), and the vertical axis represents the number of pixels. FIG. 3 (c) is a diagram in which the two normalized histograms of FIG. 3 (a) and FIG. 3 (b) are overlapped, and the area of the histogram intersection which is a common part thereof is represented in gray. Since the similarity R between blocks is normalized, it takes values from “0” to “1”, and it can be determined that the two blocks are more similar as the value increases.

  For simplicity, the registered image and the collation image are images of the same size. If they are not the same, either one of the images is enlarged or reduced so as to be the same.

  The registered image and the collation image are divided into, for example, 15 blocks, and a normalized histogram is generated for each block. A normalized histogram of a block of registered images can be generated by preprocessing before authentication. Therefore, the time required for this generation process does not need to be included in the authentication response time, and does not limit the response speed.

  Next, the following calculation processing is performed in the same manner as the processing shown in FIG. That is, first, the similarity of the entire image is calculated only for images having a sufficiently high similarity R between all the blocks.

As the first similarity R1 between the registered image and the verification image, for example, the following two definitions can be used.
1) Average value of similarity between blocks for the entire image.
2) An average value for the entire image of values binarized by applying threshold processing to the similarity between blocks.

  In both cases 1) and 2), the first similarity R1 clearly takes a value from 0 to 1, and it can be determined that the larger the value, the more similar the two images are.

  Hereinafter, an example using the above definition 1) will be described. Note that the above two definitions of the first similarity R1 are directly linked to the similarity R between blocks, and therefore have the advantage of requiring less calculation amount.

  The second similarity R2 is calculated for the registered image and the collated image only for those having a sufficiently high first similarity R1 between the registered image and the collated image. Prior to the calculation process of the second similarity R2, pruning (exclusion) related to two-stage collation: exclusion from the target image based on the similarity R between blocks and exclusion from the target image regarding the first similarity R1. Is running.

  Therefore, the number of times of calculating the second similarity R2 is reduced. For the amount corresponding to the second similarity R2, it is preferable to use collation based on the mismatch rate, focusing on the advantage of relatively high accuracy. Non-patent document 2 describes collation based on the mismatch rate.

  As described above, in the embodiment of the present invention, in order to realize high-speed authentication in advance by minimizing the application of a matching process with a high accuracy but a large amount of calculation required. Pruning is performed based on the similarity between blocks and the first similarity. More precisely, the mismatch rate M takes a value from 0 to 1, and the smaller the value, the more similar the two images. Therefore, it is necessary to redefine 1-M as the similarity. The mismatch rate is defined as follows.

  Taking vein authentication as an example, an image is divided into 1) a finger vein region, 2) a background region, and 3) an ambiguous region. The number of mismatches is the number when the pixel of the registered image belongs to the finger vein region and the corresponding pixel of the verification image belongs to the background region, or vice versa. The mismatch rate is obtained by dividing the number of mismatches by the total number of pixels in the finger vein region between the registered image and the collation image.

  The result of collation between the registered image and the collation image is that the collation was successful (success), the collation was unsuccessful due to an oversight (rejected person), and other images were collated incorrectly. There are three ways: “Accept others”. In the verification method in the verification technique according to the present embodiment, the ratio of rejecting the person is slightly increased, but the processing time can be reduced to a fraction.

As described above, the matching technique according to the present embodiment has the following advantages.
1) The target image of the second similarity matching process with a relatively large calculation amount by eliminating many registered images by the matching process based on the similarity between blocks and the first similarity, which has a relatively small calculation amount Since the number can be minimized, the collating process can be speeded up.
2) The present invention is applicable even when the registered image and the collation image are binary images or images with a low density level.
3) About the similarity collation process between blocks, since the information of a block unit is handled, it is hard to be influenced also to a fine noise.

  Note that the present invention can also be applied to a program using the above collation processing technique and a computer-readable recording medium storing the program.

(References)
1) Reference 1
Murase Hiroshi, VV Vinode: Fast object search using local color information --- Active search method--: IEICE Transactions, D-II, Vol. J81-D-II, No. 9, pp. 2035-2042, 19
1998 2) Reference 2
Naoto Miura, Goro Nagasaka, Takafumi Miyatake: Extraction of finger vein pattern based on repeated trials of line tracking and application to personal authentication: IEICE Transactions, D-II, Vol. J86-D-II, No. 5 , Pp.678
−687, 2003

  The present invention can be used for an image verification device, an authentication device based on biometric information, and the like.

It is a functional block diagram which shows the structural example of the image collation apparatus by one embodiment of this invention. It is a flowchart figure which shows the flow of the collation process by this Embodiment. (A)-(c) is a figure which shows the concept of a histogram intersection.

Explanation of symbols

DESCRIPTION OF SYMBOLS 101 ... Image input device, 102 ... Keyboard, 103 ... Mouse, 104 ... Processing device, 105 ... Block similarity module, 106 ... First similarity module, 107 ... Second similarity module, 108 ... Storage device, 109 ... Registration Image, 110 ... collation image, 111 ... display.

Claims (1)

A collation device that performs one-to-many collation for collating a target image and a plurality of registered images,
An inter-block similarity calculating unit that divides the verification target image and the registered image into a plurality of blocks and calculates an inter-block similarity in units corresponding to each other;
Based on the inter-block similarity calculated by the inter-block similarity calculation unit for only registered images that exclude the registered image whose inter-block similarity is less than the first threshold and is equal to or greater than the first threshold. A first similarity calculation unit that calculates a first similarity between the verification image and the registered image;
A second similarity that excludes registered images whose first similarity is less than the second threshold and calculates a second similarity between the matching image and the registered image for only registered images that are equal to or greater than the second threshold A collation apparatus comprising: a degree calculation unit, wherein a matching image is identified from registered images.

Images