Personal authentication method and device

BACKGROUND OF THE INVENTION

0001The present invention relates to a technology of personal authentication in which features are obtained from biological information by frequency analysis and personal authentication is performed based on the similarity between the features.

0002As individual identification techniques using human biological information (biometrics), there have been proposed many methods using information of fingerprints, irises, blood-vessel arrangements in the retinas, faces, and the like. Among these, individual identification using an iris, in particular, is expected to be mainstream of biometrics authentication in the future for the reasons that: <ol id="ol0001" compact="compact"><li>(1) an iris pattern can be acquired with a camera in a noncontact manner,</li><li>(2) the false acceptance rate (FAR) is significantly low due to complexity of the iris pattern, and</li><li>(3) the iris pattern remains unchanged substantially through the life of the owner.</li></ol>

0003Techniques for extracting iris features from iris images and identifying individuals are disclosed in USP 5,291,560, Japanese National Phase PCT Laid-Open Publication No. 8-504979, and "High Confidence Visual Recognition of Persons by a Test of Statistical Independence", IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 15, No. 11, Nov. 1993 (these three disclosures are roughly the same in contents).

0004In the above techniques, an iris image is analyzed at multiple resolutions using multi-scale self-similarity type two-dimensional quadrature band-pass filters (Gabor filters, for example) to generate a feature (iris code). To state as a specific procedure, a digitized image of a human eye is captured with a video camera, and the boundary between the iris and the sclera and the boundary between the iris and the pupil are determined to separate an iris region from others. A polar coordinate system is applied to the separated iris image, and a plurality of ring analysis bands are determined. Analysis and coding are then performed for the analysis bands using a signal processor comprised of multi-scale quadrature band-pass filters. The thus-generated iris codes are compared with each other by calculating a hamming distance between the codes as shown in FIG. <b>24</b>, to determine whether or not the compared iris codes originate from an identical person.

-Prior art -

0005Cooke M. et al. in " Missing data techniques for robust speech recognition" (1997 IEEE International Conference on Acoustics, Speech, and Signal Processing) elaborates the missing feature theory in the context of speech recognition, the principle of theory is to divide spectral features in two sets of reliable and unreliable and apply modifications to the reference templates in order to ignore the effect of unreliable features. Bourlard H. et al. in "A new ASR approach based on independent processing and recombination of partial frequency bands" (1996 Fourth International Conference on Spoken Language Processing), again in the context of speech recognition, outlines a process which comprises the step of splitting the whole frequency band into sub-bands on which different recognizer are independently applied and then recombined to yield a global score.

-Problems to be solved-

0006The above technique is based on the premise that multi-scale frequency analysis is performed for predetermined fixed frequency bands using an image capture device providing a predetermined fixed resolution in both cases of iris registration and comparison. To fulfill this premise, dedicated registration and authentication devices are necessary.

0007In view of the recent sophistication in function of cellular phones and personal digital assistants (PDAs), increase in capacity of communication bands, and the like, the following use of personal authentication is considered possible in the near future. That is, a cellular phone or PDA equipped with an image capture device (camera) may be used for taking an iris image of a person and authenticating the person. And this capability may be utilized in occasions of access control, such as logging in to a cellular phone or PDA, authentication in electronic commerce (EC), control of entrance/exit into/from a place requiring physical security, and alternative use to a key of a house. When the above occasions are to be realized, an image capture device incorporated in or mounted externally on a cellular phone or PDA will possibly be comparatively low in resolution at the beginning. In addition, specifications for image capture devices will possibly be different among the types of the devices. Moreover, it is considered that authentication will be effected via a variety of apparatuses such as a terminal mounted on the door, not only a cellular phone and a PDA.

0008As described above, an iris image may be taken with a variety of apparatuses providing lower to higher resolutions during authentication. Under this situation, if frequency analysis is performed at fixed frequency bands by the conventional method described above, the following problem will arise. That is, when a low-resolution image is input, a part of a feature obtained by analysis at a high frequency (specifically, a frequency component equal to or higher than Fs/2 where Fs is a sampling frequency) is no more useful as the feature. Therefore, if this part obtained by high-frequency analysis is counted as part of the feature, the entire correlation value decreases and thus authentication precision possibly degrades.

SUMMARY OF THE INVENTION

0009The object of the present invention is defined in the claims.

BRIEF DESCRIPTION OF THE DRAWINGS

0010<ul id="ul0001" list-style="none" compact="compact"><li>FIG. <b>1</b> is a conceptual illustration of technical features of a personal authentication method of Embodiment 1 of the present invention.</li><li>FIG. <b>2</b> is a schematic block diagram of a personal authentication system of the embodiments of the present invention.</li><li>FIG. <b>3</b> is an illustration of an appearance of a camera-equipped cellular phone as an iris authentication terminal.</li><li>FIG. <b>4</b> is a block diagram of an internal configuration of the iris authentication terminal.</li><li>FIG. <b>5</b> is a flowchart of processing during registration in the personal authentication method of Embodiment 1 of the present invention.</li><li>FIG. <b>6</b> is a flowchart of processing during authentication in the personal authentication method of Embodiment 1 of the present invention.</li><li>FIG. <b>7</b> is an illustration of a structure of an iris.</li><li>FIG. <b>8</b> is a view representing an iris pattern in terms of the frequency range using Fourier transform.</li><li>FIG. <b>9</b> is a view illustrating frequency analysis using four Gabor filters.</li><li>FIG. <b>10</b> represents Gaussian function used for the Gabor filter.</li><li>FIGS. <b>11A, 11B</b> and <b>11C</b> illustrate an example of an iris code.</li><li>FIG. <b>12</b> is an illustration of generation of a registration iris code.</li><li>FIG. <b>13</b> is an illustration of generation of an identification iris code.</li><li>FIG. <b>14</b> is an illustration of comparison between the registration iris code and the identification iris code.</li><li>FIG. <b>15</b> is an illustration of calculation of a hamming distance.</li><li>FIG. <b>16</b> is a conceptual illustration of distributions of authentication distances between identical persons and between different persons.</li><li>FIG. <b>17</b> is a flowchart of processing in a personal authentication method in Embodiment 2 of the present invention.</li><li>FIGS. <b>18A</b> and <b>18B</b> are conceptual illustrations of distributions of authentication distances between identical persons and between different persons.</li><li>FIG. <b>19</b> is an illustration of a method for generating a plurality of iris images with different resolutions using low-pass filters.</li><li><b>FIG. 20</b> is an illustration of selection of Gabor filters.</li><li>FIGS. <b>21A</b> to <b>21D</b> are views representing the images generated using low-pass filters in terms of the frequency range using Fourier transform.</li><li>FIG. <b>22</b> is an illustration of selection of Gabor filters.</li><li>FIG. <b>23</b> is an illustration of iris DBs in which extracted features are stored separately by combination of analysis frequency bands.</li><li>FIG. <b>24</b> is an illustration of conventional personal authentication.</li></ul>

DESCRIPTION OF THE PREFERRED EMBODIMENTS

0011Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings. In the following description, iris authentication, that is, personal authentication using an image of an iris of an eye, is exemplified. It should however be noted that the present invention is also applicable to other types of personal authentication using other biological information such as a fingerprint and a voiceprint.

Embodiment 1

0012FIG. <b>1</b> is a conceptual illustration of technical features of a personal authentication method of Embodiment <b>1</b> of the present invention. As shown in FIG. <b>1,</b> in this embodiment, during registration, frequency analysis at a plurality of frequencies <b>f1</b> to <b>fn</b> is performed for an iris image <b>1</b> taken, and features are generated for the respective frequencies <b>f1</b> to <b>fn.</b> During authentication, the frequencies <b>f2</b> to <b>fn</b> among the plurality of frequencies <b>f1</b> to <b>fn</b> used during the registration are selected for use in frequency analysis for authentication based on the resolution of an iris image <b>2</b> taken. Frequency analysis at the selected frequencies <b>f2</b> to fn is performed for the iris image <b>2</b> of a person to be authenticated, to generate features for the respective frequencies <b>f2</b> to <b>fn.</b> The generated respective features are compared with those for the same frequencies obtained during the registration, to perform personal authentication.

0013By adopting the method described above, even when an iris image of a person to be authenticated is taken with an image capture device providing a resolution lower than that used during the registration, personal authentication can be performed using this lower-resolution iris image. In addition, since a feature for a non-selected frequency, which is useless as a feature, is eliminated, degradation in authentication precision can be suppressed in the personal authentication.

0014FIG. <b>2</b> is a schematic diagram of a personal authentication system of this embodiment. Referring to FIG. <b>2,</b> an iris authentication server <b>11</b> has an iris database (DB) <b>12</b> storing iris data of a plurality of persons and is connected to a network <b>13</b> such as the Internet, private lines and public lines. At least one Iris registration device <b>14</b> and at least one iris authentication terminal <b>15,</b> which are provided with an image capture device, are connected to the network <b>13.</b>

0015FIG. <b>3</b> is an illustration of a camera-equipped cellular phone <b>21</b> applied as the iris authentication terminal <b>15</b> of the iris authentication system shown in FIG. <b>2.</b> FIG. <b>4</b> is a block diagram of an inner configuration of the iris authentication terminal <b>15.</b>

0016A plurality of iris authentication servers 11 may be provided, for placement in each area or organization using the network and/or for placement of mirror servers for dispersing the load. The iris DB <b>12</b> may be connected to the iris authentication server via the network.

0017The function of the iris registration device <b>14</b> may be incorporated in the iris authentication server <b>11,</b> or the iris authentication terminal <b>15</b> may have both functions of registration and authentication. When a plurality of iris authentication terminals <b>15</b> are provided, the image capture specifications for the terminals are not necessarily the same.

0018FIGS. <b>5</b> and <b>6</b> are flowcharts of processing during registration and that during authentication, respectively, in the personal authentication method of this embodiment. Hereinafter, the processing during registration and that during authentication in the personal authentication method of this embodiment will be described separately.

<Registration>

0019During registration, in step <b>SA01</b> for image capture, an iris image is taken with the iris registration device <b>14</b> at a "resolution" adequate to capture fine details of the iris pattern. The "resolution" as used herein refers to "how finely the iris was sampled", which may also be called the resolving power. An image capture device/lens system incorporated in the iris registration device <b>14</b> should be the one capable of taking an iris image at an adequate resolution.

0020FIG. <b>7</b> illustrates a structure of an iris. Referring to FIG. <b>7,</b> an iris <b>41</b> mainly includes folds <b>42</b> (a pattern extending radially from around the boundary with a pupil <b>45</b>), iris pits <b>43,</b> an iris frill <b>44,</b> and the like. The pattern of the iris <b>41</b> differs among individuals. Personal authentication is realized using this difference in pattern, in which a feature is extracted from the pattern of the iris <b>41</b> and the extracted feature is compared with others.

0021FIG. <b>8</b> is a view representing the iris pattern in terms of the frequency range using Fourier transform. Although 2D Fourier transform is actually used since the actual iris pattern is a two-dimensional image, one-dimensional representation is shown in FIG. <b>8</b> for simplification of description. The one dimension in this case is a dimension in the circumferential direction of a two-dimensional iris pattern obtained by separating from other portions in a manner described later and applying polar coordinates. That is, to state differently, a radial pattern of the iris such as the pattern of the folds <b>42</b> is frequency-analyzed in the circumferential direction. The reason why the frequency analysis is performed in the circumferential direction is that the radial pattern of the iris is known important for personal authentication.

0022In this embodiment, the frequency of a sine wave of which one period covers 360 degrees in a circumferential direction (one cycle) is expressed as frequency 1 (the unit of the frequency is Hz, which is also applicable to the rest of the description).

0023In FIG. <b>8,</b> Fm represents the upper-limit frequency of a frequency band effective in use of the iris pattern for authentication. According to the sampling theorem, in order to use the upper-limit frequency Fm for analysis of the iris pattern, an iris image should be captured at a sampling frequency Fs = Fm × 2 or higher. By this capture, the "adequate resolution" described above is obtained. Alternatively, the upper-limit frequency Fm may be determined by capturing iris images at various sampling frequencies Fs in preliminary experiments and selecting a sampling frequency Fs at which the highest recognition performance is exhibited. Otherwise, the upper-limit frequency Fm may be estimated from gaps between the finest wrinkles observable obtained by observing iris patterns of various persons.

0024The iris image taken during the registration is sent to the iris authentication server <b>11</b> via the network <b>13,</b> together with the ID of the person concerned obtained separately. Note that if there is available an iris authentication terminal <b>15</b> provided with the ability of taking an iris image at an adequate resolution, registration of an iris image can also be performed using such a terminal <b>15.</b>

0025In steps <b>SA02</b> to <b>SA07,</b> a feature (iris code) used for authentication is extracted from the iris image taken. Note that although the iris authentication server <b>11</b> performs the extraction of a feature in this embodiment, the iris registration device <b>14</b> may perform a series of processing from capturing of an iris image through extraction of a feature, and send the generated iris code via the network <b>13</b> to be stored in the iris DB <b>12.</b>

0026Steps <b>SA02</b> to <b>SA05</b> can be implemented by any method as long as the iris region can be extracted stably. In this embodiment, the method described in Japanese National Phase PCT Laid-Open Publication No. 8-504979 is employed. Details of this method are omitted here, and only the outline is described as follows.

0027In step <b>SA02,</b> the inner boundary of the iris is first determined for extraction of the iris region. More specifically, utilizing the fact that the brightness is different between the pupil and the iris, a circle in which the integral of the brightness abruptly changes on the circumference is sought among circles having regularly increasing radii, and the center (x0, y0) and the radius r0 of the circle are obtained.

0028Likewise, in step <b>SA03,</b> the outer boundary of the iris is determined. In this step, also, the fact that the brightness is different between the iris and the sclera is utilized. In this case, since the boundary between the iris and the sclera is often occluded by the upper eyelid and the lower eyelid, calculation of the integral of the brightness is made for only the right and left arc portions (called "pie wedges" in the above publication) excluding the upper and lower portions of the circle. In other words, a circle in which the integral of the brightness at the pie wedges abruptly changes is sought among circles having regularly increasing radii, and the center (x1, y1) and the radius r1 of the circle are obtained.

0029Thus, from steps <b>SA02</b> and <b>SA03,</b> the inner and outer boundaries of the iris are determined, and thus the iris region is extracted.

0030In step <b>SA04,</b> a polar coordinate system is applied to the extracted iris region. The origin of the polar coordinate system is set at the center (x0, y0) of the pupil. In the dimension in the radial direction, the circumference of the pupil (that is, the innermost radius of the iris) is determined as radius 0 while the circumference of the iris (that is, the outermost radius of the iris) is determined as radius 1. In-between values of the radius are linearly interpolated between 0 and 1 depending on the distance from the pupil circumference to the iris circumference. In the dimension in the angular direction, values between 0 and 360 degrees are determined. Thus, although the size of the iris image varies with the difference in iris size among individuals, the zoom value of a camera, the distance between the camera and the iris, and the like, extraction of a feature from the iris image can be performed without being influenced by the iris size by setting the polar coordinate system described above.

0031In step <b>SA05,</b> the iris region is divided into a plurality of concentric ring regions (for example, eight regions).

0032In step <b>SA06,</b> as disclosed in the above publication, the polar coordinates-applied iris image is subjected to 2D Gabor filters, which are multi-scale band-pass filters. The Gabor filter is represented by the following expression. <maths id="math0001" num="(1)"><math display="block"><mrow><mi mathvariant="normal">G</mi><mrow><mo>(</mo><mrow><mi mathvariant="normal">r</mi><mo>,</mo><mi mathvariant="normal">θ</mi></mrow><mo>)</mo></mrow><mo>=</mo><msup><mi mathvariant="normal">e</mi><mrow><mn>2</mn><mi mathvariant="normal">π</mi><mi mathvariant="normal">j</mi><mi mathvariant="normal">ω</mi><mrow><mo>(</mo><mrow><mi mathvariant="normal">θ</mi><mo>−</mo><msub><mi mathvariant="normal">θ</mi><mn>0</mn></msub></mrow><mo>)</mo></mrow></mrow></msup><msup><mi mathvariant="normal">e</mi><mrow><mo>−</mo><mrow><mrow><msup><mrow><mrow><mo>(</mo><mrow><mi mathvariant="normal">r</mi><mo>−</mo><msub><mi mathvariant="normal">r</mi><mn>0</mn></msub></mrow><mo>)</mo></mrow></mrow><mn>2</mn></msup></mrow><mo>/</mo><mrow><msup><mi mathvariant="normal">α</mi><mn>2</mn></msup></mrow></mrow></mrow></msup><msup><mi mathvariant="normal">e</mi><mrow><mo>−</mo><mrow><mrow><mrow><mo>(</mo><mrow><mi mathvariant="normal">θ</mi><mo>−</mo><msub><mi mathvariant="normal">θ</mi><mn>0</mn></msub></mrow><mo>)</mo></mrow></mrow><mo>/</mo><mrow><msup><mi mathvariant="normal">β</mi><mn>2</mn></msup></mrow></mrow></mrow></msup></mrow></math><img file="EP1237117B1_D0001.tif" /></maths> where r is the radius, θ is a radial angle, ω is the angular frequency, and α and β are constants. The parameters α and β vary inversely with the angular frequency ω. Herein, a plurality of angular frequencies ω are prepared, and analysis of the iris image is performed using a plurality of Gabor filters corresponding to the respective angular frequencies ω.

0033FIG. <b>9</b> illustrates analysis using four frequencies. In FIG. <b>9</b>, the x-axis represents the frequency, showing how the frequency band from frequency 0 (DC component) to the upper-limit frequency Fm is analyzed with four Gabor filters having different passbands. F1 to F4 denote the center frequencies of the passbands of the four Gabor filters. When expressed in terms of the angular frequencies described above, the following are established. ω1 = 2πF1, ω2 = 2πF2, ω3 = 2πF3, ω4 = 2πF4 Note that the upper-limit frequency Fm and the upper limit F4U of the passband of the Gabor filter are not necessarily identical to each other as shown in FIG. <b>9</b>, but may establish the relationship F4U ≦ Fm.

0034In FIG. <b>9</b>, the passband of each Gabor filter is defined as between points at which the peak value of the Gaussian function used in the Gabor filter is halved, as shown in FIG. <b>10</b>. To state differently, if the fringes of the Gaussian function (portions apart from the center) are taken into consideration, the passbands of the Gabor filters overlap each other. The points at which the passbands of the Gabor filters meet may be set at points other than those at which the Gaussian function is halved, such as points of σ, 2σ or 3 σ where σ is the standard deviation of the Gaussian function.

0035A feature is then extracted from a Gabor-filtered signal as disclosed in the above publication. That is, the signal is changed to a binary value determined by whether the output of the Gabor filter is positive or negative and coded, as represented by the following expressions. <maths id="math0002" num="(2)"><math display="block"><mtable columnalign="left"><mtr><mtd><msub><mi mathvariant="normal">MSB</mi><mrow><mi mathvariant="normal">Re</mi></mrow></msub><mrow><mo>(</mo><mrow><mi mathvariant="normal">r</mi><mo>,</mo><mi mathvariant="normal">θ</mi></mrow><mo>)</mo></mrow><mo>=</mo><mn>1</mn><mi mathvariant="normal"> </mi><mi mathvariant="normal">if</mi><mi mathvariant="normal"> </mi><mi mathvariant="normal">Re</mi><mstyle displaystyle="true"><mrow><msub><mo>∫</mo><mi mathvariant="normal">ρ</mi></msub><mrow><mstyle displaystyle="true"><mrow><msub><mo>∫</mo><mi mathvariant="normal">φ</mi></msub><mrow><msup><mi mathvariant="normal">e</mi><mrow><mn>2</mn><mi mathvariant="normal">π</mi><mi mathvariant="normal">j</mi><mi mathvariant="normal">ω</mi><mrow><mo>(</mo><mrow><mi mathvariant="normal">θ</mi><mo>−</mo><mi mathvariant="normal">φ</mi></mrow><mo>)</mo></mrow></mrow></msup><msup><mi mathvariant="normal">e</mi><mrow><mo>−</mo><mrow><mrow><msup><mrow><mrow><mo>(</mo><mrow><mi mathvariant="normal">r</mi><mo>−</mo><mi mathvariant="normal">ρ</mi></mrow><mo>)</mo></mrow></mrow><mn>2</mn></msup></mrow><mo>/</mo><mrow><msup><mi mathvariant="normal">α</mi><mn>2</mn></msup></mrow></mrow></mrow></msup><msup><mi mathvariant="normal">e</mi><mrow><mo>−</mo><mrow><mrow><mrow><mo>(</mo><mrow><mi mathvariant="normal">θ</mi><mo>−</mo><mi mathvariant="normal">φ</mi></mrow><mo>)</mo></mrow></mrow><mo>/</mo><mrow><msup><mi mathvariant="normal">β</mi><mn>2</mn></msup></mrow></mrow></mrow></msup><mi mathvariant="normal">I</mi><mrow><mo>(</mo><mrow><mi mathvariant="normal">ρ</mi><mo>,</mo><mi mathvariant="normal">φ</mi></mrow><mo>)</mo></mrow><mi mathvariant="normal">ρ</mi><mi mathvariant="normal">d</mi><mi mathvariant="normal">ρ</mi></mrow></mrow></mstyle><mi mathvariant="normal">d</mi><mi mathvariant="normal">φ</mi></mrow></mrow></mstyle><mo>></mo><mn>0</mn></mtd></mtr><mtr><mtd><msub><mi mathvariant="normal">MSB</mi><mrow><mi mathvariant="normal">Re</mi></mrow></msub><mrow><mo>(</mo><mrow><mi mathvariant="normal">r</mi><mo>,</mo><mi mathvariant="normal">θ</mi></mrow><mo>)</mo></mrow><mo>=</mo><mn>0</mn><mi mathvariant="normal"> </mi><mi mathvariant="normal">if</mi><mi mathvariant="normal"> </mi><mi mathvariant="normal">Re</mi><mstyle displaystyle="true"><mrow><msub><mo>∫</mo><mi mathvariant="normal">ρ</mi></msub><mrow><mstyle displaystyle="true"><mrow><msub><mo>∫</mo><mi mathvariant="normal">φ</mi></msub><mrow><msup><mi mathvariant="normal">e</mi><mrow><mn>2</mn><mi mathvariant="normal">π</mi><mi mathvariant="normal">j</mi><mi mathvariant="normal">ω</mi><mrow><mo>(</mo><mrow><mi mathvariant="normal">θ</mi><mo>−</mo><mi mathvariant="normal">φ</mi></mrow><mo>)</mo></mrow></mrow></msup><msup><mi mathvariant="normal">e</mi><mrow><mo>−</mo><mrow><mrow><msup><mrow><mrow><mo>(</mo><mrow><mi mathvariant="normal">r</mi><mo>−</mo><mi mathvariant="normal">ρ</mi></mrow><mo>)</mo></mrow></mrow><mn>2</mn></msup></mrow><mo>/</mo><mrow><msup><mi mathvariant="normal">α</mi><mn>2</mn></msup></mrow></mrow></mrow></msup><msup><mi mathvariant="normal">e</mi><mrow><mo>−</mo><mrow><mrow><mrow><mo>(</mo><mrow><mi mathvariant="normal">θ</mi><mo>−</mo><mi mathvariant="normal">φ</mi></mrow><mo>)</mo></mrow></mrow><mo>/</mo><mrow><msup><mi mathvariant="normal">β</mi><mn>2</mn></msup></mrow></mrow></mrow></msup><mi mathvariant="normal">I</mi><mrow><mo>(</mo><mrow><mi mathvariant="normal">ρ</mi><mo>,</mo><mi mathvariant="normal">φ</mi></mrow><mo>)</mo></mrow><mi mathvariant="normal">ρ</mi><mi mathvariant="normal">d</mi><mi mathvariant="normal">ρ</mi></mrow></mrow></mstyle><mi mathvariant="normal">d</mi><mi mathvariant="normal">φ</mi></mrow></mrow></mstyle><mo>≤</mo><mn>0</mn></mtd></mtr><mtr><mtd><msub><mi mathvariant="normal">MSB</mi><mrow><mi mathvariant="normal">Im</mi></mrow></msub><mrow><mo>(</mo><mrow><mi mathvariant="normal">r</mi><mo>,</mo><mi mathvariant="normal">θ</mi></mrow><mo>)</mo></mrow><mo>=</mo><mn>1</mn><mi mathvariant="normal"> </mi><mi mathvariant="normal">if</mi><mi mathvariant="normal"> </mi><mi mathvariant="normal">Im</mi><mstyle displaystyle="true"><mrow><msub><mo>∫</mo><mi mathvariant="normal">ρ</mi></msub><mrow><mstyle displaystyle="true"><mrow><msub><mo>∫</mo><mi mathvariant="normal">φ</mi></msub><mrow><msup><mi mathvariant="normal">e</mi><mrow><mn>2</mn><mi mathvariant="normal">π</mi><mi mathvariant="normal">j</mi><mi mathvariant="normal">ω</mi><mrow><mo>(</mo><mrow><mi mathvariant="normal">θ</mi><mo>−</mo><mi mathvariant="normal">φ</mi></mrow><mo>)</mo></mrow></mrow></msup><msup><mi mathvariant="normal">e</mi><mrow><mo>−</mo><mrow><mrow><msup><mrow><mrow><mo>(</mo><mrow><mi mathvariant="normal">r</mi><mo>−</mo><mi mathvariant="normal">ρ</mi></mrow><mo>)</mo></mrow></mrow><mn>2</mn></msup></mrow><mo>/</mo><mrow><msup><mi mathvariant="normal">α</mi><mn>2</mn></msup></mrow></mrow></mrow></msup><msup><mi mathvariant="normal">e</mi><mrow><mo>−</mo><mrow><mrow><mrow><mo>(</mo><mrow><mi mathvariant="normal">θ</mi><mo>−</mo><mi mathvariant="normal">φ</mi></mrow><mo>)</mo></mrow></mrow><mo>/</mo><mrow><msup><mi mathvariant="normal">β</mi><mn>2</mn></msup></mrow></mrow></mrow></msup><mi mathvariant="normal">I</mi><mrow><mo>(</mo><mrow><mi mathvariant="normal">ρ</mi><mo>,</mo><mi mathvariant="normal">φ</mi></mrow><mo>)</mo></mrow><mi mathvariant="normal">ρ</mi><mi mathvariant="normal">d</mi><mi mathvariant="normal">ρ</mi></mrow></mrow></mstyle><mi mathvariant="normal">d</mi><mi mathvariant="normal">φ</mi></mrow></mrow></mstyle><mo>></mo><mn>0</mn></mtd></mtr><mtr><mtd><msub><mi mathvariant="normal">MSB</mi><mrow><mi mathvariant="normal">Im</mi></mrow></msub><mrow><mo>(</mo><mrow><mi mathvariant="normal">r</mi><mo>,</mo><mi mathvariant="normal">θ</mi></mrow><mo>)</mo></mrow><mo>=</mo><mn>0</mn><mi mathvariant="normal"> </mi><mi mathvariant="normal">if</mi><mi mathvariant="normal"> </mi><mi mathvariant="normal">Im</mi><mstyle displaystyle="true"><mrow><msub><mo>∫</mo><mi mathvariant="normal">ρ</mi></msub><mrow><mstyle displaystyle="true"><mrow><msub><mo>∫</mo><mi mathvariant="normal">φ</mi></msub><mrow><msup><mi mathvariant="normal">e</mi><mrow><mn>2</mn><mi mathvariant="normal">π</mi><mi mathvariant="normal">j</mi><mi mathvariant="normal">ω</mi><mrow><mo>(</mo><mrow><mi mathvariant="normal">θ</mi><mo>−</mo><mi mathvariant="normal">φ</mi></mrow><mo>)</mo></mrow></mrow></msup><msup><mi mathvariant="normal">e</mi><mrow><mo>−</mo><mrow><mrow><msup><mrow><mrow><mo>(</mo><mrow><mi mathvariant="normal">r</mi><mo>−</mo><mi mathvariant="normal">ρ</mi></mrow><mo>)</mo></mrow></mrow><mn>2</mn></msup></mrow><mo>/</mo><mrow><msup><mi mathvariant="normal">α</mi><mn>2</mn></msup></mrow></mrow></mrow></msup><msup><mi mathvariant="normal">e</mi><mrow><mo>−</mo><mrow><mrow><mrow><mo>(</mo><mrow><mi mathvariant="normal">θ</mi><mo>−</mo><mi mathvariant="normal">φ</mi></mrow><mo>)</mo></mrow></mrow><mo>/</mo><mrow><msup><mi mathvariant="normal">β</mi><mn>2</mn></msup></mrow></mrow></mrow></msup><mi mathvariant="normal">I</mi><mrow><mo>(</mo><mrow><mi mathvariant="normal">ρ</mi><mo>,</mo><mi mathvariant="normal">φ</mi></mrow><mo>)</mo></mrow><mi mathvariant="normal">ρ</mi><mi mathvariant="normal">d</mi><mi mathvariant="normal">ρ</mi></mrow></mrow></mstyle><mi mathvariant="normal">d</mi><mi mathvariant="normal">φ</mi></mrow></mrow></mstyle><mo>≤</mo><mn>0</mn></mtd></mtr></mtable></math><img file="EP1237117B1_D0002.tif" /></maths> where I(ρ,ϕ) is an input iris image represented by the polar coordinate system.

0036FIGS. <b>11A</b> to <b>11C</b> illustrate an example of extraction of an iris code. An iris image is actually a two-dimensional signal (in the radial direction and the angular direction (circumferential direction) on the polar coordinate system). However, in these figures, the iris image is shown one-dimensionally in the angular direction for simplification of description. FIG. <b>11A</b> illustrates an original signal representing the brightness along the circumference of a circle having a certain radius of the iris image, FIG. <b>11B</b> illustrates a signal obtained by Gabor-filtering the original signal of FIG. <b>11A,</b> and FIG. <b>11C</b> illustrates a signal obtained by changing the signal of FIG. <b>11B</b> to a binary value, that is, an iris code. The iris code shown in FIG. <b>11C</b> is generated for each frequency every ring region obtained by the division in step <b>SA05</b>.

0037As shown in FIG. <b>12</b>, sub-features Fsi obtained from analysis at respective frequencies Fi are put together to obtain a registration iris code FT. In step <b>SA07</b>, the registration iris code FT and analysis frequency information are stored in the iris DB <b>12</b> in association with the ID of the person concerned.

0038In this embodiment, the polar coordinate system was adopted for the iris image. Alternatively, a normal coordinate system may be adopted. Multi-resolution frequency analysis using Gabor filters, a type of multi-scale self-similarity type two-dimensional band-pass filters, was performed. Alternatively, another method utilizing multi-resolution analysis, such as wavelet analysis, may be employed.

0039In this embodiment, the feature was calculated using the real part and imaginary part of the Gabor filter separately. Alternatively, the power obtained by calculating the square root of sum of squares of the two parts may be used as the feature. The output of the Gabor filter may not be changed to a binary value, but the multilevel output may be used as it is as the feature.

<Authentication>

0040During authentication, the following processing is performed.

0041First, in step <b>SB01</b> for image capture in the flowchart of FIG. <b>6</b>, the user intending to receive authentication, that is, the person to be authenticated inputs an iris image using the cellular phone <b>21</b> shown in FIGS. <b>3</b> and <b>4</b>, for example. In more detail, the user examines an image currently captured by a camera <b>22</b> (an image capture element incorporated in the camera <b>22</b>) displayed on a monitor <b>23</b>, and presses a button <b>25</b> for execution of image capture when it is confirmed that his or her own iris image on the monitor <b>23</b> is in sharp focus. Upon the pressing of the button, an illuminator <b>24</b> is lit, and an image is taken in synchronization with the lighting timing. The iris image taken is stored in a frame memory <b>32</b> of the cellular phone <b>21</b>. Thereafter, an iris feature extraction program stored in a program memory <b>34</b> is read therefrom into a work memory <b>33</b> for execution of step <b>SB02</b> and the subsequent steps in FIG. <b>6</b>.

0042Although the camera-equipped cellular phone was used in the above description, any other apparatus may be used as long as it is equipped with a camera. For example, a camera-equipped PHS (personal handy phone), a camera-equipped PDA, a camera-equipped interphone, a digital camera having communication function, and the like may be used.

0043In this embodiment, feature extraction is performed by software using an iris feature extraction program. Alternatively, iris feature extraction processing may be implemented by a dedicated circuit or a digital signal processor (DSP) to perform the feature extraction by hardware.

0044Steps <b>SB02</b> and <b>SB03</b> in which the iris region is defined from the iris image taken, step <b>SB04</b> in which the coordinate system is determined, and step <b>SB05</b> in which the iris region is divided into analysis regions are the same as steps <b>SA02</b> to <b>SA05</b> during the registration, respectively. Description of these steps is therefore omitted here.

0045In step <b>SB06</b>, the resolution of the iris image is calculated. Herein, a resolution R is defined as the circumferential length at the iris/pupil boundary or the iris inner boundary calculated in step <b>SB02</b>. The resolution R may be the number of pixels on the iris inner boundary, or may be determined as R = 2π r0 from the radius r0 calculated in step SB02.

0046In step <b>SB07</b>, the frequency band for analysis (analysis frequency band) is determined from the resolution R of the iris image. The resolution R is identical to the sampling frequency Fsp = R (Hz) in the circumferential direction at the iris/pupil boundary. When R points are sampled for one period (360 degrees), the sampling period (sampling interval) is 360/R degrees. Therefore, since sampling is made around 360 degrees at sampling intervals of 360/R degrees, the sampling frequency Fsp = 360 / (360/R) = R (Hz).

0047Thus, according to the sampling theorem, the upper limit Fmp of the frequency effective for analysis of the iris image is determined as <maths id="math0003"><math display="block"><mrow><mi mathvariant="normal">F</mi><mi mathvariant="normal">m</mi><mi mathvariant="normal">p</mi><mo>=</mo><mrow><mrow><mi mathvariant="normal">F</mi><mi mathvariant="normal">s</mi><mi mathvariant="normal">p</mi></mrow><mo>/</mo><mn>2</mn></mrow><mo>=</mo><mrow><mi mathvariant="normal">R</mi><mo>/</mo><mn>2</mn></mrow></mrow></math><img file="EP1237117B1_D0003.tif" /></maths> The upper limit Fmp itself is not an effective analysis frequency. To be precise, therefore, a frequency lower than the upper limit Fmp may be used for the analysis.

0048That is, according to the sampling theorem, even if the analysis is made at a frequency equal to or higher than Fsp/2, a portion of the feature corresponding to such a frequency is useless as a feature. If such a portion is included in the entire feature for comparison, the entire degree of coincidence degrades.

0049The reason why the sampling frequency is determined with respect to the circumference at the iris/pupil boundary is that the pattern (of folds) extending radially from the innermost radius of the iris (near the iris/pupil boundary) works as an effective feature for personal authentication, and that the polar coordinates are applied to the iris separated and the Gabor filters are used for detecting a change in shading in the circumferential direction on the polar coordinate system.

0050When a normal coordinate system is applied to the iris image for the analysis in place of the polar coordinate system, the resolution may be determined from an amount that varies with the size of the iris sampled, such as the radius or diameter of the iris image, the area (number of dots) of the iris region separated, and the area (number of dots) of the pupil, for determination of the analysis frequency band.

0051In some terminals used for authentication, the resolution and the lens system of the image capture device incorporated therein are known, and when the image capture distance is roughly constant (the depth of field is small), the size of the iris taken is roughly the same. In this case, therefore, the analysis frequency band may be determined in advance for each terminal using a method as described above. In this case, a reference table associating the terminal with the analysis frequency band may be prepared, to enable determination of the analysis frequency band from the type of the terminal.

0052There is another case that the image capture distance can be measured with a distance-measuring sensor and camera information such as the lens system (zooming available) and the resolution of the image capture device can be obtained. In this case, the size of the iris taken is roughly predictable. Therefore, a reference table associating the image capture distance and the camera information with the analysis frequency band may be prepared, to enable determination of the analysis frequency band from the image capture distance and the camera information.

0053Assume that during the registration, frequency analysis was performed for four frequency bands having center frequencies F1 to F4 with the frequency F4U as the upper limit, to calculate a feature.

0054In the above case, when the upper limit Fmp of the frequency for the iris image during the authentication is larger than the upper-limit frequency F4U during the registration, frequency analysis is performed for the four frequency bands as was done during the registration. When Fmp &lt; F4U, the maximum n satisfying Fmp ≦ FnU is calculated. If n = 3, frequency analysis for three frequency bands having center frequencies F1 to F3 is performed.

0055In step <b>SB08</b>, a feature is extracted using the Gabor filters represented by expression (1) above. This step is substantially the same as step <b>SA06</b> during the registration, except that the frequency analysis is performed for the frequency bands determined in step <b>SB07.</b> When n = 3 in step <b>SB07,</b> that is, when it has been determined in step <b>SB07</b> that analysis for three frequency bands having center frequencies F1 to F3 should be performed, three Gabor filters (ω1 = 2π F1, ω2 = 2 π F2, ω 3 = 2 π F3 ) corresponding to these frequency bands are used. Extraction of a feature from a Gabor-filtered signal is performed using expressions (2) as in step <b>SA06</b>.

0056In step <b>SB09,</b> as shown in FIG. <b>13</b>, sub-features Fsi (binary codes) corresponding to frequencies Fi, extracted in step <b>SB08</b>, are put together to obtain an identification iris code FT.

0057Once the above processing is completed in the iris authentication terminal <b>15,</b> the generated feature and analysis frequency information are sent to the iris authentication server <b>11</b> via the network <b>13</b>.

0058In this embodiment, the iris authentication terminal <b>15</b> executes the steps until step <b>SB09</b> for feature extraction, and then sends the feature to the iris authentication server <b>11</b> via the network <b>13</b>, so that authentication is made by the iris authentication server <b>11.</b> Alternatively, the iris authentication terminal <b>15</b> may execute only step <b>SB01</b> for iris image capture and send the iris image to the iris authentication server <b>11</b> via the network <b>13,</b> so that the iris authentication server <b>11</b> can execute steps <b>SB02</b> to <b>SB09</b> for feature extraction, as well as steps <b>SB10</b> and <b>SB11</b> for identification.

0059Alternatively, the iris authentication terminal <b>15</b> may execute the series of processing including steps <b>SB01</b> for iris image capture, steps <b>SB02</b> to <b>SB09</b> for feature extraction, and steps <b>SB10</b> and <b>SB11</b> for identification. In this case, the registration iris codes used for comparison are sent to the iris authentication terminal <b>15</b> via the network <b>13</b> for comparison at the terminal <b>15.</b>

0060There are one to N authentication and one to one authentication. In one to N authentication, the person to be authenticated does not inform the system of his or her own ID. Therefore, the feature extracted during the authentication is compared with all reference features in the iris DB <b>12.</b> The person is authenticated when the similarity to (distance from) the most similar reference feature is equal to or more than (less than) a predetermined threshold. In one to one authentication, the person to be authenticated informs the system of his or her own ID. The feature extracted during the authentication is compared with the reference feature corresponding to the ID given by the person. The person is authenticated when the similarity (distance) therebetween is equal to or more than (less than) a predetermined threshold. Thus, in one to N authentication, when the comparison is performed at the iris authentication terminal <b>15</b>, all the features stored in the iris DB <b>12</b> are sent to the terminal <b>15</b> via the network <b>13</b>.

0061In this embodiment, in the iris authentication server <b>11</b>, the iris code for authentication is compared with the reference iris codes stored in the iris DB <b>12</b> for personal authentication. The present invention is applicable to both one to N authentication and one to one authentication described above. Since they are only different in whether the feature is compared with a plurality of features or a single feature, the type of authentication is not specifically mentioned in the following description.

0062In step <b>SB10,</b> the identification iris code generated in step <b>SB09</b> during the authentication is compared with a registration reference iris code stored in step <b>SA07</b> during the registration. In the comparison, as shown in FIG. <b>14</b>, analysis frequency information attached to both the identification and registration iris codes is referred to so that sub-features analyzed at the same frequency are compared with each other. In the illustrated example, where analysis was performed for four frequency bands having center frequencies F1 to F4 during the registration while it was performed for three frequency bands having center frequencies F1 to F3 during the authentication, sub-features FS1 analyzed at the frequency F1, sub-features FS2 analyzed at the frequency F2, and sub-features FS3 analyzed at the frequency F3 are respectively compared with each other.

0063For comparison, the hamming distance is used. In addition, as shown in FIG. <b>15</b>, in calculation of the hamming distance, the features are shifted by a same amount to compensate tilting of the face and rotational movement of the eyeball itself. The minimum hamming distance obtained when the shift amount is varied within a predetermined range (range within which rotation of the iris pattern is allowed) is determined as the final hamming distance.

0064In step <b>SB11,</b> the hamming distance obtained by comparing respective bits of the features is divided by the number of comparison (total number of bits) and normalized, to obtain an authentication distance (authentication score). When the authentication distance is equal to or less than a predetermine threshold, the identity of the person is authenticated. When it exceeds the predetermined threshold, the person is rejected as being a stranger.

0065In this embodiment, the hamming distance (exclusive OR (XOR)) was used because the generated iris code is a binary value. If the feature is a multilevel value, another distance measure (such as Euclid distance and normalized correlation) may be used.

0066Thus, in this embodiment, personal authentication is possible without degrading the authentication precision even when the iris image is taken with an image capture device providing a resolution lower than that obtained during the registration.

Embodiment 2

0067When the image capture device used during the authentication is low in resolution, no sub-feature for a high-resolution portion in multi-resolution analysis (no feature from analysis at a frequency equal to or more than the upper-limit frequency Fmp = Fsp/2) is extracted and thus used for comparison. This reduces the total number of bits (information amount) of features used for comparison. Due to the decrease in information amount, the separability of a comparison score distribution for an identical person from that for different persons may be low, and this may degrade comparison precision.

0068FIG. <b>16</b> conceptually shows a distribution D1 of authentication distances between identical persons and a distribution D2 of authentication distances between different persons, obtained by comparing features in combinations of arbitrary persons. In FIG. <b>16</b>, when a threshold TH for distinguishing the identical person from other persons is set as illustrated, the part of the distribution <b>D2</b> for different persons smaller in authentication distance than the threshold TH (hatched part A1) corresponds to a "false accept" part in which a different person is mistakenly accepted. Therefore, from a distribution as that shown in FIG. <b>16</b>, the false acceptance rate (FAR), the rate at which different persons are mistakenly accepted, can be trial-calculated to some extent.

0069In this embodiment, therefore, FAR is trial-calculated in advance for each combination of sub-features Fsi by referring to the iris DB <b>12.</b> In addition, a reference FAR value, which should be guaranteed, is determined. Thus, in the event that the reference FAR fails to be satisfied when the feature FT is obtained with a low-resolution image capture device without including a sub-feature analyzed at a high frequency (without analysis at a high frequency) during the authentication as shown in Embodiment 1, it is possible to take measures such as inhibiting proceeding to the authentication step.

0070A false rejection rate (FRR) is another indicator of authentication precision. When the image capture device used during the authentication is low in resolution, the feature of a high-resolution portion obtained during multi-resolution analysis is not used. Therefore, since the information amount of the feature used for comparison is small, the distribution D1 for an identical person itself may expand, causing debasement in FRR. In FIG. <b>16</b>, the part of the distribution D1 for an identical person greater in authentication distance than the threshold TH (hatched part A2) corresponds to the "false rejection" part.

0071FIG. <b>17</b> is a flowchart of processing in a personal authentication method of Embodiment 2 of the present invention. The processing shown in FIG. <b>17</b> is executed after the storing of the feature in the iris DB <b>12</b> (see FIG. <b>5</b>).

0072As in Embodiment 1, an iris image is analyzed with 2D Gabor filters having a plurality of frequency passbands, to prepare sub-features FSi corresponding to respective frequencies Fi. In this embodiment, assume that four frequency bands are used for analysis as shown in FIG. 12 to obtain four sub-features FSi. This operation is repeated by the number of persons entered in the iris DB <b>12.</b> Assume herein that the number of persons entered in the iris DB <b>12</b> is N, at least one iris image is taken for each person, and a feature is extracted from each iris image.

0073In step <b>SA08</b>, the reference FAR (= T) and a feature (combination of sub-features) used for trial-calculation of FAR are determined. First, suppose all the sub-features FS1 to FS4 are selected. That is, all the sub-features FS1 to FS4 are put together to generate the feature FT. In step <b>SA09,</b> for the feature FT, calculated are the authentication distance between features extracted from a person P and the authentication distances between a feature extracted from the person P and features extracted from all of the persons other than P. The resultant distributions are shown in FIG. <b>18A</b>.

0074When there are a plurality of registered features for the person P, an authentication distance distribution DA1 can be prepared as shown in FIG. <b>18A</b> by comparing features of the identical person P with one another. An authentication distance distribution DA2 can also be prepared by comparing the person P with the other persons. A threshold ThA is determined from the distributions DA1 and DA2. Herein, the threshold ThA is set so that FAR and FRR are equal to each other.

0075The threshold ThA is set for each combination of the person P and the feature, and the setting may be done in various ways depending on the purpose of the authentication. For example, if it is desired to reject others without fail, the threshold ThA may be set lower trading off a more or less increase in false rejection rate (FRR). If rejection of the identification of the person P will cause much difficulty from the standpoint of user interface, the threshold ThA may be set higher trading off the false acceptance rate (FAR). If only one registered feature is available for the person P, no distribution for the identical person is generated. The threshold Th is therefore determined referring to only FAR.

0076If the registered data amount in the iris DB <b>12</b> is not sufficiently large, the two distributions may not overlap each other, unlike those shown in FIG. <b>18A</b>. In this case, the two distributions may be approximated with an appropriate function to form an overlap portion thereof.

0077In step <b>SA10</b>, the FAR is trial-calculated using the threshold ThA. In the case shown in FIG. <b>18A</b>, the FAR corresponds to a hatched part AA1. Whether or not the part AA1 is smaller than a predetermined reference FAR = T is determined. If AA1 &lt; T, it is judged that analysis at frequencies F1 to F4 is possible for identification of the person P.

0078Likewise, the FAR in the combination of sub-features FS1 to FS4 is trial-calculated for all the persons, and whether or not the trial-calculated FAR is lower than a predetermined threshold T is judged. The reason why the FAR is trial-calculated for each person is as follows. Some persons can be adequately distinguished from others only with a feature analyzed at a low frequency while other persons cannot. Therefore, combinations of sub-features suitable for respective persons can be obtained by trial-calculation for each person.

0079Persons distinguishable from others only with a feature analyzed at a low frequency are those unique in the low-frequency portion of the iris pattern. Naturally, unlike the method of this embodiment, the authentication precision may be trial-calculated for an identical person and for different persons to be used for all the persons in common. In this case, the entire distribution of authentication distances between identical persons and the entire distribution of authentication distances between different persons are compared with each other, to trial-calculate FAR. This eliminates the necessity of holding effective combinations of sub-features for respective persons in the iris DB <b>12,</b> and thus the effect that the memory capacity can be reduced is attained.

0080Subsequently, steps <b>SA08</b> to <b>SA10</b> are repeated for another combination of sub-features. For example, suppose the sub-features FS1 to FS3 were selected as the second combination, and as a result of similar processing, authentication with adequate precision was successfully done for all the persons.

0081Thereafter, suppose the sub-features FS1 and FS2 are selected as the third combination. FIG. <b>18B</b> shows a distribution DB1 of authentication distances between the identical persons P and a distribution DB2 of authentication distances between the person P and all of the persons other than the person P obtained in this case. The FAR corresponds to the area of a hatched part AB1. When AB1 ≧ T, adequate authentication precision is not obtainable in analysis at frequencies F1 and F2 for identification of the person P. Thus, it is found from the calculation using the iris DB <b>12</b> that adequate authentication precision will not be obtained for authentication of the person P unless at least the frequencies F1 to F3 are used for analysis.

0082Therefore, in this embodiment, if the person P attempts to be authenticated using an iris image with a resolution with which the highest frequency used for analysis will be lower than F3, the person P can be warned or notified of inability of authentication by the iris authentication server 11.

0083In this embodiment, the FAR was used as an indicator of authentication precision, which however varies with the setting of the threshold Th. As an alternative indicator of authentication precision independent of the threshold Th, therefore, the sum of FAR and FRR, for example, may be used. In this case, the authentication precision is independent of the threshold Th, and determined only by two distributions as shown in FIG. <b>16</b>.

0084In this embodiment, the distribution for the identical person is used for each person. Actually, however, since the number of registered features of one person is limited (one to several pieces), the distribution as described above may not be obtained in some cases. In these cases, an iris moving picture may be taken for several seconds during the registration of the iris image, and frame images of the moving picture may be extracted to be used as registered images. For example, when a moving picture of an iris is taken for three seconds with a 30 frames/sec progressive scanning camera, a total of 90 iris images are obtained. After defective images such as that taken when the person blinks are removed, the remaining images can be used as images for registration. If a moving picture is taken with an illuminator of varying lightness, images with various pupil sizes can be obtained. In this case, it is possible to obtain a distribution for the identical person using images taken under various conditions.

0085The distribution for the identical person can be updated by adding duly-authenticated features to the database every time the person is authenticated.

0086Steps <b>SA08</b> to <b>SA10</b> are processing performed in advance during the registration, not during the comparison. In the event that data is newly entered in the iris DB <b>12</b> or a registered feature for a certain person is updated, this processing may be executed again. If execution of the processing every time of registration is undesirable, the processing may be executed periodically such as once a week or a month.

0087The combinations of analysis frequencies are not limited to those of frequencies in series such as frequencies F1 to Fk, but may be combinations of discrete frequencies such as frequencies F1, F3 and F5. The latter combinations are effective in such cases that use of frequencies F1, F3 and F5 increases uniqueness (distinction from others) but additional use of frequencies F2 and F4 rather decreases distinction from distributions for different persons.

0088In this embodiment, whether or not a person should be finally authenticated is judged depending on the estimated authentication precision. The authentication precision can also be used to control the power bestowed on the person authenticated after the authentication. For example, in personal authentication for electronic commerce, the ceiling of the amount of money allowed for transaction may be set depending on the estimated authentication precision. In personal authentication for entrance/exit control, the room a person is allowed to enter may be determined depending on the estimated authentication precision for the person. In personal authentication for logging in to a PC or the like, executable commands (browsing only, rewritable, erasable) and the like may be controlled depending on the estimated authentication precision.

0089Whether or not re-authentication is allowed may be determined depending on the estimated authentication precision. For example, if the camera of the iris authentication terminal <b>15</b> is zooming-adjustable, instruction may be made to take an image again with a larger zoom value. If the camera is focusing-adjustable, instruction may be made to take an image again with the iris being closer to the camera.

0090In Embodiment 1, the selection of frequencies during the authentication was made based on the resolution of the iris image taken during the authentication. In place of or in addition to the resolution, the selection of frequencies during the authentication may be made based on the authentication precision for each combination of frequencies. For example, among combinations of frequencies providing a FAR equal to or less than a predetermined value, a combination providing the smallest FAR or a combination composed of the smallest number of frequencies may be used during the authentication.

0091In this embodiment, the threshold of the authentication distance was determined in advance, and the FAR was trial-calculated using the threshold. In reverse, the threshold can be calculated from the minimum guaranteed FAR. Therefore, if the FAR fails to reach the minimum guaranteed value, the threshold may be made tighter for next authentication.

0092The authentication precision can be estimated from the feature obtained during the registration as described above. Therefore, it is possible, for example, to take images of both right and left irises, trial-calculate authentication precision for these images, and register one of the iris images better in authentication precision (that is, more distinguishable from the features for others). Alternatively, if both iris images are low in authentication precision, both of them may be registered to be used for authentication.

0093Likewise, in the case of fingerprint authentication, for example, a plurality of fingers may be registered to finally select a fingerprint with high authentication precision. Otherwise, if a predetermined level of authentication precision is not obtainable with a single finger, a plurality of fingers may be registered to be used together during authentication.

Embodiment 3

0094In Embodiment 1, an iris image was taken with an image capture device providing a comparatively high resolution capable of taking an iris image with an adequate resolution during the registration. Thereafter, multi-resolution analysis was performed to extract a sub-feature for each analysis frequency band. During the authentication, at least one frequency band for analysis is determined from the iris image taken. Sub-features obtained for the same frequency band during the registration and during the authentication were compared with each other, to perform personal authentication.

0095In Embodiment 3 of the present invention, a plurality of iris images with different resolutions are acquired during the registration.

0096Such iris images with different resolutions may be acquired by actually taking a plurality of iris images with different resolutions. This can be done using a plurality of cameras of which image capture elements provide different resolutions, or by taking images a plurality of times with one camera while varying the zoom value of the camera.

0097As another acquiring method, an iris image may be taken at an adequate resolution, and, as shown in FIG. <b>19,</b> the iris image taken may be subjected to a plurality of (three in FIG. <b>19</b>) low-pass filters LPF1 to LPF3 to generate a plurality of iris images with different resolutions. In this case, the resultant iris images are the same in size, but different in frequency components contained in the images (high-frequency components are cut).

0098Thereafter, frequency bands to be used for analysis are determined for the acquired iris images with different resolutions in the following manner.

0099As for the iris images with different resolutions actually taken, the frequency bands are determined based on the number of pixels on the circumference at the boundary between the iris and the pupil, as was done during the authentication in Embodiment 1.

0100Suppose, for example, iris images are taken at three resolutions R1, R2 and R3, and respective upper-limit frequencies Fsp1, Fsp2 and Fsp3 for analysis are determined according to the sampling theorem. Also, n (n = 1 to 4) Gabor filters each having a passband having a center frequency Fn (lower-limit frequency FnL and upper-limit frequency FnU) are prepared as shown in FIG. <b>20</b>. The Gabor filters are selected so that the upper limit FnL of each passband does not exceed a half of the sampling frequency Fs for the iris image. As a result, the image with the resolution R1 is analyzed using only the Gabor filter 1, the image with the resolution R2 is analyzed using the Gabor filters 1 to 3, and the image with the resolution R3 is analyzed using the Gabor filters 1 to 4.

0101As in Embodiment 1, the passbands of the adjacent Gabor filters overlap each other at positions where the Gaussian function is halved.

0102As for the iris images acquired using a plurality of low-pass filters, the frequency bands are determined from the properties of the low-pass filters used.

0103FIGS. <b>21A</b> to <b>21D</b> are views of an iris image taken with a camera and iris images generated using a plurality of low-pass filters, represented in terms of the frequency range using Fourier transform. Note that although 2D Fourier transform is actually used because the images are two-dimensional, one-dimensional representations (dimension in the circumferential direction of the iris on polar coordinates) are shown in FIGS. <b>21A</b> to <b>21D</b> for simplification of description as in Embodiment 1.

0104From FIGS. <b>21A</b> to <b>21D</b>, it is found that the low-pass filter LPFi (i = 1 to 3) allows pass of frequencies equal to and lower than Fci. Herein, the following relationship is established: <maths id="math0004"><math display="block"><mrow><mrow><mrow><mi mathvariant="normal">F</mi><mi mathvariant="normal">s</mi><mi mathvariant="normal">p</mi></mrow><mo>/</mo><mn>2</mn></mrow><mo>></mo><mi mathvariant="normal">F</mi><mi mathvariant="normal">c</mi><mn>1</mn><mo>></mo><mi mathvariant="normal">F</mi><mi mathvariant="normal">c</mi><mn>2</mn><mo>></mo><mi mathvariant="normal">F</mi><mi mathvariant="normal">c</mi><mn>3</mn></mrow></math><img file="EP1237117B1_D0004.tif" /></maths> where Fsp is the sampling frequency of the original image. The frequency band for analysis can be determined uniquely from the properties of the low-pass filter used. This will be described more specifically referring to FIG. <b>22</b>. Suppose the Gabor filters are selected so that the upper-limit frequency FnU of the passband of each Gabor filter does not exceed the blocking frequency Fci of the low-pass filter used for the image in question. Then, the image subjected to the low-pass filter LPF1 is analyzed using the Gabor filters 1 to 3, the image subjected to the low-pass filter LPF2 is analyzed using the Gabor filters 1 and 2, and the image subjected to the low-pass filter LPF3 is analyzed using only the Gabor filter 1. The original image is analyzed using the Gabor filters 1 to 4 as in Embodiment 1.

0105Extraction of features is the same as that described in Embodiment 1 and thus description is omitted here. As shown in FIG. <b>23,</b> extracted features are stored in iris DBs <b>12a</b> to <b>12d</b> separately by combination of analysis frequency bands.

0106During the authentication, as in Embodiments 1 and 2, the frequency bands for analysis are determined from an iris image taken. A combination of frequency bands identical to those for analysis is selected from the plurality of iris DB <b>12a</b> to <b>12d</b> shown in FIG. <b>23</b>, and authentication is performed using the features stored in the selected iris DB in the manner described in Embodiment <b>1.</b> Details of the authentication are omitted here.

0107By the above processing in this embodiment, substantially the same effects as those in Embodiment 1 can be obtained.

0108In this embodiment, the frequency bands for analysis were selected based on the upper-limit frequency of information contained in the image. Alternatively, authentication precision may be estimated in advance as in Embodiment 2, so that a combination of analysis frequencies can be selected based on the estimated authentication precision. In this case, the analysis frequency bands are not limited to those adjacent to one another (for example, F1, F2 and F3 in FIG. <b>22</b>), but may be discrete ones (for example, F1 and F3 in FIG. <b>22</b>) as discussed in Embodiment 2.

0109Thus, according to the present invention, degradation in authentication precision is suppressed even when authentication is done using iris images taken with image capture devices providing different resolutions. In addition, measures suitable for estimated authentication precision can be taken even when the information amount of a feature is low.

0110While the present invention has been described in a preferred embodiment, it will be apparent to those skilled in the art that the disclosed invention may be modified in numerous ways and may assume many embodiments other than that specifically set out and described above.