Information processing device, information processing method, and recording medium

Abstract

This information processing device is provided with an estimation unit for learning and an update unit. The estimation unit for learning uses an estimation model for estimating ellipse parameters for the contour or edge of the iris included in an iris image to estimate the ellipse parameters relating to an iris image for learning. The update unit updates model parameters included in the estimation model by using the estimated ellipse parameters and ground-truth ellipse parameters that are correct ellipse parameters relating to the iris image for learning.

Description

Information processing device, information processing method, and recording medium

[0001] The present disclosure relates to an information processing device, an information processing method, and a recording medium.

[0002] For example, Patent Document 1 describes that for robust feature detection using an eye shape model, pupil detection can be improved by removing erroneous pupil boundary points. In Patent Document 1, erroneous pupil boundary points can be created when the eyelid partially occludes the pupil. The erroneous pupil boundary points reflect the position of the eyelid rather than the true boundary of the pupil.

[0003] The patent document 1 describes that once erroneous pupil boundary points have been identified and removed, an ellipse may be fitted to the pupil using the remaining pupil boundary points, and that algorithms that may be implemented for such ellipse fitting include differential and integral operators, least squares methods, random sample consensus (RANSAC), or ellipse or curve fitting algorithms.

[0004] Special Publication No. 2022-538669

[0005] The present disclosure aims to improve upon the techniques described in the prior art documents mentioned above.

[0006] The information processing device of the present disclosure includes: a training estimation means for estimating ellipse parameters relating to a training iris image using an estimation model for estimating the ellipse parameters relating to the contour or edge of the iris included in the iris image; and an update means for updating model parameters included in the estimation model using estimated ellipse parameters that are the estimated ellipse parameters and correct ellipse parameters that are correct answers to the ellipse parameters relating to the training iris image.

[0007] The information processing method of the present disclosure includes one or more computers: estimating ellipse parameters for a training iris image using an estimation model for estimating ellipse parameters for the contour or edge of an iris included in the iris image; and updating model parameters included in the estimation model using estimated ellipse parameters, which are the estimated ellipse parameters, and correct ellipse parameters, which are correct solutions to the ellipse parameters for the training iris image.

[0008] The recording medium in the present disclosure is a recording medium having recorded thereon a program for causing one or more computers to execute the following: estimating ellipse parameters for a training iris image using an estimation model for estimating ellipse parameters for the contour or edge of the iris included in the iris image; and updating model parameters included in the estimation model using estimated ellipse parameters, which are the estimated ellipse parameters, and correct ellipse parameters, which are correct answers to the ellipse parameters for the training iris image.

[0009] 1 is a diagram illustrating an example configuration of a first information processing system according to the present disclosure. FIG. 2 is a block diagram illustrating an example configuration of a first information processing device according to the present disclosure. FIG. 3 is a flowchart illustrating an example processing operation of the first information processing device according to the present disclosure. FIG. 4 is a block diagram illustrating an example configuration of a first update unit according to the present disclosure. FIG. 5 is a flowchart illustrating an example processing operation of the first update unit according to the present disclosure. FIG. 6 is a diagram illustrating an example physical configuration of a first information processing device according to the present disclosure. FIG. 7 is a diagram illustrating an example of ellipses having the same shape as a figure according to the present disclosure. FIG. 8 is a diagram illustrating an example of characteristic parts of an upper eyelid and a lower eyelid according to the present disclosure. FIG. 9 is a diagram illustrating another example of characteristic parts of an upper eyelid and a lower eyelid according to the present disclosure. FIG. 10 is a diagram illustrating an example in which constraints on a reference ellipse axis, an angle range, a reference direction, and a positive rotation direction are applied to an ellipse rotation angle θ of a first ellipse according to the present disclosure. FIG. 11 is a diagram illustrating an example in which the gaze direction is shifted leftward and rightward from the imaging direction. FIG. 12 is a diagram illustrating an example in which the gaze direction is shifted upward from the imaging direction. FIG. 13 is a diagram illustrating an example in which the gaze direction is shifted downward from the imaging direction. FIG. 14 is a diagram illustrating the relationship between the angle range of the ellipse rotation angle and an error. FIG. 15 is a diagram illustrating an example configuration of a second information processing system according to the present disclosure. FIG. 16 is a block diagram illustrating an example configuration of a second information processing device according to the present disclosure. 1 is a flowchart showing an example of processing operation of a second information processing device according to the present disclosure. FIG. 2 is a block diagram showing an example of a configuration of a second learning estimation unit according to the present disclosure. FIG. 3 is a flowchart showing an example of processing operation of the second learning estimation unit according to the present disclosure. FIG. 4 is a block diagram showing an example of a configuration of a second update unit according to the present disclosure. FIG. 5 is a flowchart showing an example of processing operation of the second update unit according to the present disclosure. FIG. 6 is a block diagram showing an example of a configuration of a first constraint unit according to the present disclosure. FIG. 7 is a flowchart showing an example of processing operation of the first constraint unit according to the present disclosure. FIG. 8 is a block diagram showing an example of a configuration of a third information processing device according to the present disclosure. FIG. 9 is a flowchart showing an example of processing operation of a third information processing device according to the present disclosure. FIG. 10 is a block diagram showing an example of a configuration of a third learning estimation unit according to the present disclosure. FIG. 11 is a flowchart showing an example of processing operation of the third learning estimation unit according to the present disclosure. FIG. 12 is a block diagram showing an example of a configuration of a third update unit according to the present disclosure. FIG. 13 is a flowchart showing an example of processing operation of the third update unit according to the present disclosure. FIG. 14 is a block diagram showing an example of a configuration of a second constraint unit according to the present disclosure.1 is a flowchart illustrating an example of processing operation of a second constraint unit according to the present disclosure. FIG. 2 is a diagram illustrating an example of a normalized image according to the present disclosure. FIG. 3 is a block diagram illustrating an example of a configuration of a fourth information processing device according to the present disclosure. FIG. 4 is a flowchart illustrating an example of processing operation of the fourth information processing device according to the present disclosure. FIG. 5 is a block diagram illustrating an example of a configuration of a fourth learning estimation unit according to the present disclosure. FIG. 6 is a flowchart illustrating an example of processing operation of the fourth learning estimation unit according to the present disclosure. FIG. 7 is a block diagram illustrating an example of a configuration of a fourth update unit according to the present disclosure. FIG. 8 is a flowchart illustrating an example of processing operation of the fourth update unit according to the present disclosure. FIG. 9 is a block diagram illustrating an example of a configuration of a third constraint unit according to the present disclosure. FIG. 10 is a flowchart illustrating an example of processing operation of the third constraint unit according to the present disclosure. FIG. 11 is a diagram illustrating an example of a configuration of a third information processing system according to the present disclosure. FIG. 12 is a block diagram illustrating an example of a configuration of a fifth information processing device according to the present disclosure. FIG. 13 is a flowchart illustrating an example of processing operation of the fifth information processing device according to the present disclosure. FIG. 14 is a block diagram illustrating an example of a configuration of a sixth information processing device according to the present disclosure. FIG. 15 is a flowchart illustrating an example of processing operation of the sixth information processing device according to the present disclosure.

[0010] Hereinafter, embodiments according to the present disclosure will be described with reference to the drawings. In all drawings, similar components are denoted by similar reference numerals, and descriptions thereof will be omitted as appropriate. In the present disclosure, the drawings relate to one or more embodiments.

[0011] [First embodiment] (Overview) Generally, in iris authentication, an iris image including an iris is acquired by photographing an object to be authenticated, and the iris is detected from the iris image. The iris included in the iris image is often elliptical, not necessarily circular.

[0012] Patent Document 1 describes that, as described above, an algorithm may be used to fit an ellipse to the pupil using the remaining pupil boundary points after removing the erroneous pupil boundary points.

[0013] However, there is no description of a technique for estimating the outline or edge of the iris contained in an iris image. Furthermore, the algorithm for fitting an ellipse to the pupil is merely described as a list of its constituent elements. Therefore, even if the technique described in Patent Document 1 is used, it may not be possible to accurately estimate the outline or edge of the iris contained in an iris image.

[0014] One of the objectives of the present disclosure is to accurately estimate the contour or edge of the iris included in an iris image.

[0015] (Configuration Example of Information Processing System S1) As shown in FIG. 1, the information processing system S1 includes a training data storage device 90 that stores training iris images and correct answer data related to the training iris images, and an information processing device 100.

[0016] (Configuration Example of Information Processing Device 100) As shown in FIG. 2, the information processing device 100 includes a learning estimation unit 110 and an update unit 120.

[0017] The training estimation unit 110 estimates ellipse parameters for the training iris image using an estimation model for estimating ellipse parameters for the contour or edge of the iris included in the iris image.

[0018] The update unit 120 updates the model parameters included in the estimation model using estimated ellipse parameters, which are estimated ellipse parameters, and correct ellipse parameters, which are correct solutions to the estimated ellipse parameters related to the training iris image.

[0019] According to this information processing system S1, an estimation model can be constructed that uses an ellipse to estimate the contour or edge of the iris included in the iris image, thereby enabling the contour or edge of the iris included in the iris image to be estimated with high accuracy.

[0020] According to the information processing device 100, it is possible to construct an estimation model that uses an ellipse to estimate the contour or edge of the iris included in the iris image, thereby making it possible to accurately estimate the contour or edge of the iris included in the iris image.

[0021] (Example of Processing Operation of Information Processing Device 100) The information processing device 100 executes information processing as shown in FIG.

[0022] The training estimation unit 110 estimates ellipse parameters for the training iris image using an estimation model for estimating ellipse parameters for the contour or edge of the iris included in the iris image (step S120).

[0023] The update unit 120 updates the model parameters included in the estimation model using estimated ellipse parameters, which are estimated ellipse parameters, and correct ellipse parameters, which are correct answers to the ellipse parameters related to the training iris image (step S120).

[0024] According to this information processing system S1, it is possible to construct an estimation model that uses an ellipse to estimate the shape of the iris contour or edge included in an iris image, thereby enabling the iris contour or edge included in the iris image to be estimated with high accuracy.

[0025] (Details) The training data storage device 90 and the information processing device 100 are connected to each other via a communication network NT. As a result, the training data storage device 90 and the information processing device 100 transmit and receive information to each other via the communication network NT. The communication network NT may be wired, wireless, or a combination of these.

[0026] (Training Data Storage Device 90) For example, training data is stored in advance in the training data storage device 90. The training data includes, for example, training iris images and correct answer data related to the training iris images.

[0027] The training iris images are iris images used for training the estimation model.

[0028] An iris image is an image that includes the iris of a target. In the following, an example in which the target is a human will be described. Note that the target is not limited to a human, but may be an animal such as a dog, cat, pig, cow, bird, or snake.

[0029] 4 is a diagram showing an example of an iris image, which includes the pupil, upper and lower eyelids, and the white of the eye in addition to the iris.

[0030] As described above, the iris image is not limited to the iris image shown in FIG. 4 as long as it includes the iris. That is, the iris image may include, for example, only the iris. Furthermore, for example, the iris image may include only the iris and the pupil. Furthermore, for example, the iris image may include the iris and at least one of the pupil, the upper and lower eyelids, the white of the eye including the conjunctiva and the ocular membrane, the eyelashes, the eyebrows, etc. Furthermore, for example, the iris image may include both eyes. Furthermore, for example, the iris image may further include parts of the face, such as the forehead, between the eyebrows, temples, nose, cheeks, mouth, and hair, or the entire face, or other parts or the entire body.

[0031] The iris image may be, for example, an image obtained by photographing the iris of a subject in a wavelength band in the visible light region or the near-infrared region. The iris image may be, for example, an image obtained by processing an image obtained by photographing the iris of a subject. For example, general image processing that can substantially maintain the shape of the iris included in the iris image may be used for the processing. In detail, for example, the processing may include, but is not limited to, downsampling, cutting out a region including the iris, etc.

[0032] The number of training iris images may be one, but typically multiple. Generally, it is difficult to obtain a large number of training iris images by photographing them all. Therefore, the training iris images may be iris images that have been processed from photographed images, etc.

[0033] The correct answer data will be described in detail later.

[0034] The information processing system S1 may include an imaging device connected to the communication network NT instead of or in addition to the training data storage device 90. This imaging device may be, for example, a camera or the like, and is used to capture an image of the iris of a subject. When an imaging device is included, an iris image captured by the imaging device may be used as the training iris image. Furthermore, the training data storage device 90 may have a function for performing image processing to obtain training iris images, and may store the training iris images obtained by this image processing.

[0035] A storage unit for storing training data may be provided in the information processing device 100. In this case, the information processing system S1 does not need to include a training data storage device 90 separate from the information processing device 100.

[0036] (Regarding the training estimation unit 110) The training estimation unit 110 estimates ellipse parameters related to the contour or edge of the iris included in the training iris image using an estimation model currently being trained. The estimation model is, for example, a machine learning model for estimating ellipse parameters. By training the estimation model, an estimation model that estimates ellipse parameters related to the contour or edge of the iris included in the iris image is constructed. The estimation model includes one or more model parameters, and is configured using, for example, a neural network (hereinafter also referred to as "NN") that uses a convolutional layer, an attention mechanism, etc.

[0037] In more detail, for example, the training estimation unit 110 includes an estimation model currently being trained. The training estimation unit 110 acquires training iris images in response to, for example, a user instruction. For example, the training estimation unit 110 acquires the training iris images from the training data storage device 90. Note that the training estimation unit 110 may further acquire ground truth data.

[0038] For example, the training estimation unit 110 inputs the acquired training iris image to the estimation model. In response to this input, the estimation model estimates ellipse parameters for the input training iris image and outputs the estimated ellipse parameters. The estimated ellipse parameters are ellipse parameters estimated by the estimation model.

[0039] When training the estimation model using multiple training iris images, the training estimation unit 110 may acquire multiple training iris images. In this case, the training estimation unit 110 may, for example, input each of the acquired training iris images to the estimation model. For example, the training estimation unit 110 may input each of the training iris images to the estimation model sequentially. In response to this input, the estimation model may, for example, estimate ellipse parameters for each of the input training iris images and output estimated ellipse parameters for each of the training iris images.

[0040] An example of training an estimation model using multiple training iris images is so-called mini-batch training, which is a machine learning model training method that repeatedly trains an estimation model using a group of images (mini-batch) consisting of, for example, 128 training iris images.

[0041] Note that the batch size used in mini-batch learning (i.e., the number of training iris images constituting a mini-batch) is not limited to the above-mentioned 128, but may be appropriately set to any value equal to or greater than 1. Furthermore, the method for learning the estimation model is not limited to mini-batch learning.

[0042] (Regarding Ellipse Parameters) Ellipse parameters are parameters for specifying an ellipse that represents, for example, the outline or edge of an iris included in an iris image. Hereinafter, an "ellipse that represents the outline or edge of an iris included in an iris image" will also be referred to as an "iris ellipse."

[0043] In general, an iris ellipse is often a double ellipse consisting of a first ellipse, which is an outer iris ellipse, and a second ellipse, which is an inner iris ellipse. The outer iris ellipse (first ellipse) is an ellipse that represents, for example, the boundary between the iris and the white of the eye, the boundary between the cornea and the sclera, etc. The inner iris ellipse (second ellipse) is a second ellipse that represents, for example, the outline or edge of the pupil, the boundary between the iris and the pupil, etc.

[0044] Therefore, the ellipse parameters may include a first ellipse parameter and a second ellipse parameter corresponding to the first ellipse and the second ellipse, respectively. The first ellipse parameter is an ellipse parameter for specifying a first ellipse representing an outer contour or edge of an iris included in the iris image. The second ellipse parameter is an ellipse parameter for specifying a second ellipse representing an inner contour or edge of the iris included in the iris image.

[0045] The iris ellipse may be either the first ellipse or the second ellipse, that is, the ellipse parameters may include either the first ellipse parameter or the second ellipse parameter.

[0046] In more detail, for example, the ellipse parameters may be parameters for specifying the position, size, and shape of an iris ellipse in an iris image. Examples of the configuration of such ellipse parameters are given below. Note that the configuration of the ellipse parameters is not limited to the following examples.

[0047] (Configuration Example 1 of Ellipse Parameters) Ellipse parameters may include, for example, center coordinates (e.g., x and y coordinates of the center), two axis radii, and an ellipse rotation angle. The axis radius means half the length (axis diameter) of the major and minor axes passing through the center of an ellipse. An ellipse with two equal axis radii corresponds to a circle. Therefore, this parameter can also be used to represent a circle.

[0048] The ellipse parameters may be applied to each of the first ellipse parameter and the second ellipse parameter. The ellipse parameters may include, for example, components P1 to P5 constituting the first ellipse parameter and components Q1 to Q5 constituting the second ellipse parameter.

[0049] In this case, for example, P1 and P2 may be the x-coordinate and y-coordinate of the center of the first ellipse, P3 and P4 may be the two axis radii of the first ellipse, and P5 may be the ellipse rotation angle of the first ellipse.

[0050] Furthermore, Q1 and Q2 may be the x-coordinate and y-coordinate of the center of the second ellipse, respectively. Q3 and Q4 may be the two axis radii of the second ellipse. Q5 may be the ellipse rotation angle of the second ellipse.

[0051] In this example, the shaft diameter may be used instead of the shaft radius.

[0052] (Ellipse Rotation Angle θ) The ellipse rotation angle θ is an ellipse parameter that indicates the tilt of the iris ellipse. The ellipse rotation angle θ is, for example, an angle that indicates the rotation angle of an ellipse that represents the outline or edge of the iris included in the iris image.

[0053] The ellipse rotation angle θ is expressed, for example, as the angle of the reference ellipse axis relative to the reference direction.

[0054] (Regarding the Reference Ellipse Axis) The reference ellipse axis is an axis set with respect to the iris ellipse to represent the ellipse rotation angle θ.

[0055] (Regarding Reference Direction) The reference direction is a direction that is determined in advance as a reference for representing a direction.

[0056] The reference direction may be, for example, a predetermined direction in the iris image. In detail, the reference direction may be, for example, the right direction, the upward direction, or the like in the iris image.

[0057] The reference direction may be, for example, the horizontal or vertical direction of the eye. The horizontal and vertical directions of the eye may be directions that intersect with each other, for example, directions that are perpendicular to each other.

[0058] In detail, for example, the lateral and vertical directions of the eye may be the horizontal and vertical directions, respectively. In this case, if the iris image includes a fixed object that fixes the eye position, face position, etc., the directions corresponding to the horizontal and vertical directions in the iris image may be set using a known direction related to the fixed object. For example, the directions corresponding to the horizontal and vertical directions in the iris image may be set using the shooting direction of the image capturing device used to generate the iris image. For example, the shooting direction may be stored in advance in association with the learning iris image. Furthermore, for example, the directions corresponding to the horizontal and vertical directions in the iris image may be stored in advance in association with the iris image. Note that the method for setting the directions corresponding to the horizontal and vertical directions in the iris image is not limited to the method exemplified here.

[0059] For example, if an iris image includes an eye, the horizontal or vertical direction of the eye may be determined from the iris image. A method for determining the horizontal and vertical directions of the eye from the iris image will be described in another embodiment.

[0060] Generally, the shape of the eye has a length in the horizontal direction. Therefore, the horizontal direction of the eye can be said to be a direction whose main component is the horizontal direction. Also, the vertical direction of the eye can be said to be a direction whose main component is the vertical direction. Here, the main component means a component whose magnitude is larger than the magnitude of other components.

[0061] (Regarding the Positive Rotation Direction) The positive rotation direction is a reference rotation direction for expressing the ellipse rotation angle θ. The positive rotation direction may be, for example, either counterclockwise or clockwise. The ellipse rotation angle θ may be expressed, for example, as the magnitude of the angle of the reference ellipse axis in the positive rotation direction relative to the reference direction, or as a positive or negative value corresponding to the magnitude of the positive rotation direction and the reverse rotation direction, respectively.

[0062] (Configuration Example 2 of Ellipse Parameters) Ellipse parameters may include, for example, center coordinates (e.g., x and y coordinates of the center), major axis radius, minor axis radius, and ellipse rotation angle. The major axis radius means half the length of the major axis (major axis diameter). The minor axis radius means half the length of the minor axis (minor axis diameter). An ellipse with equal major axis radius and minor axis radius corresponds to a circle. Therefore, these parameters can also be used to represent a circle.

[0063] The ellipse parameters may be applied to each of the first ellipse parameter and the second ellipse parameter. The ellipse parameters may include, for example, elements P1 to P5 constituting the first ellipse parameter and elements Q1 to Q5 constituting the second ellipse parameter.

[0064] In this case, for example, P1 and P2 may be the x-coordinate and y-coordinate of the center of the first ellipse, respectively. P3 may be the major axis radius of the first ellipse. P4 may be the minor axis radius of the first ellipse. P5 may be the ellipse rotation angle of the first ellipse.

[0065] Furthermore, Q1 and Q2 may be the x-coordinate and y-coordinate of the center of the second ellipse, respectively. Q3 may be the major axis radius of the second ellipse. Q4 may be the minor axis radius of the second ellipse. Q5 may be the ellipse rotation angle of the second ellipse.

[0066] In this example, the major axis diameter and the minor axis diameter may be used instead of the major axis radius and the minor axis radius, respectively.

[0067] (Configuration Example 3 of Ellipse Parameters) The ellipse parameters may include, for example, center coordinates (e.g., x and y coordinates of the center), major axis radius, ellipticity, and ellipse rotation angle. The ellipticity is the ratio of the major axis radius to the minor axis radius, and is generally a value obtained using the calculation formula "minor axis radius / major axis radius." An ellipse with an ellipticity of 1 corresponds to a circle. Therefore, this parameter can also be used to represent a circle.

[0068] The ellipse parameters may be applied to each of the first ellipse parameter and the second ellipse parameter. The ellipse parameters may include, for example, elements P1 to P5 constituting the first ellipse parameter and elements Q1 to Q5 constituting the second ellipse parameter.

[0069] In this case, for example, P1 and P2 may be the x-coordinate and y-coordinate of the center of the first ellipse, respectively. P3 may be the major axis radius of the first ellipse. P4 may be the ellipticity. P5 may be the ellipse rotation angle of the first ellipse.

[0070] Furthermore, Q1 and Q2 may be the x-coordinate and y-coordinate of the center of the second ellipse, respectively. Q3 may be the major axis radius of the second ellipse. Q4 may be the ellipticity of the second ellipse. Q5 may be the ellipse rotation angle of the second ellipse.

[0071] Note that the minor axis diameter and the major axis diameter may be used instead of the minor axis radius and the major axis radius, respectively. As a parameter instead of the ellipticity, for example, "(major axis radius - minor axis radius) / major axis radius", "(major axis radius - minor axis radius) / minor axis radius", the reciprocal of these values, the reciprocal of the ellipticity, etc. may be used.

[0072] Although configuration examples of ellipse parameters have been described above, the configuration of ellipse parameters is not limited to the above examples. For example, the ellipse parameters may further include a parameter indicating a reference direction. For example, the ellipse parameters may further include a parameter indicating a reference ellipse axis. The parameter indicating the reference ellipse axis may be, for example, a parameter indicating whether the reference ellipse axis is a major axis or a minor axis. For example, the ellipse parameters may further include a parameter indicating a positive rotation direction. For example, the ellipse rotation angle θ may be expressed as an angle, or may be expressed using a position (cos θ, sin θ) on a unit circle having the same center as the ellipse.

[0073] (Example of Representation of Ellipse Parameters) The ellipse parameters may be represented by, for example, a vector. When the ellipse parameters are represented by a vector, each element included in the ellipse parameters is set to, for example, a predetermined component of the vector.

[0074] In more detail, for example, the elements included in the ellipse parameters are P1 to P5 and Q1 to Q5. In this case, the ellipse parameters may be represented by a vector in which P1 to P5 are set as the first to fifth components, and Q1 to Q5 are set as the sixth to tenth components, for example.

[0075] When this example is applied to Configuration Example 1, it is not determined whether the major and minor radii of the first ellipse are set to P3 or P4. Also, it is not determined whether the major and minor radii of the second ellipse are set to Q3 or Q4. Therefore, whether the major and minor radii of the first ellipse are set to P3 or P4 may be determined by comparing P3 and P4. Similarly, whether the major and minor radii of the second ellipse are set to Q3 or Q4 may be determined by comparing Q3 and Q4.

[0076] In contrast, when applied to Configuration Example 2, the major axis radius of the first ellipse is set to P3. The minor axis radius of the first ellipse is set to P4. The major axis radius of the second ellipse is set to Q3. The minor axis radius of the second ellipse is set to Q4. As a result, unlike when applied to Configuration Example 1 described above, the major axis radius and minor axis radius of the first ellipse and the second ellipse can be distinguished and identified without comparing P3 and P4 and comparing Q3 and Q4. Therefore, the number of calculations using vectors representing ellipse parameters can be reduced compared to Configuration Example 1.

[0077] Note that the components of the vector to which each component included in the ellipse parameters is set are not limited to the above and may be changed as appropriate. Furthermore, the ellipse parameters are not limited to being represented by vectors and may be represented by any appropriate method.

[0078] (Regarding the correct data) The correct data includes correct ellipse parameters. The correct ellipse parameters include correct ellipse parameters related to the training iris images. When there are multiple training iris images, the correct data may include correct ellipse parameters associated with each of the training iris images. To associate the correct ellipse parameters with the training iris images, for example, information for identifying the training iris images (image ID (identification)) may be used. Note that the method for associating the correct ellipse parameters with the training iris images is not limited to this.

[0079] For example, the correct ellipse parameters may be composed of correct values ​​of components corresponding to the components included in the estimated ellipse parameters. In particular, for example, if the estimated ellipse parameters estimated by the estimation model are composed of P1 to P5 and Q1 to Q5 as described above, the correct ellipse parameters may be composed of correct values ​​of P1 to P5 and Q1 to Q5 for identifying the corresponding iris ellipse included in the training iris image. The correct ellipse parameters may be represented as a vector.

[0080] (Regarding the Update Unit 120) The update unit 120, for example, acquires estimated ellipse parameters and ground truth data included in the training data. The update unit 120, for example, updates one or more model parameters included in the estimation model using the acquired estimated ellipse parameters and ground truth parameters included in the ground truth data. Updating the model parameters enables the estimation model to be trained.

[0081] As shown in FIG. 5 , the update unit 120 includes a loss function calculation unit 121 , a gradient calculation unit 122 , and a parameter update unit 123 .

[0082] The loss function calculation unit 121 calculates the loss using the estimated ellipse parameters and the correct ellipse parameters.

[0083] The gradient calculation unit 122 uses the loss calculated by the loss function calculation unit 121 to calculate the gradient for each of the model parameters.

[0084] The parameter update unit 123 updates each of the model parameters using the gradient calculated by the gradient calculation unit 122 .

[0085] The update unit 120 executes an update process (step S120) as shown in FIG. 6, for example.

[0086] The loss function calculation unit 121 calculates a loss using the estimated ellipse parameters estimated in step S110 and the ground truth ellipse parameters (step S121).

[0087] The gradient calculation unit 122 calculates the gradient for each of the model parameters using the loss calculated in step S121 (step S122).

[0088] The parameter update unit 123 updates each of the model parameters using the gradient calculated in step S122 (step S123).

[0089] Below, detailed examples of the loss function calculation unit 121, the gradient calculation unit 122, and the parameter update unit 123 will be described.

[0090] (Regarding the Loss Function Calculation Unit 121) For example, when the loss function calculation unit 121 acquires estimated ellipse parameters from the learning estimation unit 110, it acquires ground truth data from the training data storage device 90. The loss function calculation unit 121 calculates the loss using, for example, the estimation result of the estimation model and the acquired ground truth data. In detail, for example, the loss function calculation unit 121 calculates the loss using the estimated ellipse parameters and ground truth ellipse parameters included in the acquired ground truth data.

[0091] The loss is, for example, a value corresponding to the difference between the estimation result of the estimation model and the correct answer. In detail, for example, it may be a value corresponding to the distance between the estimation result and the correct answer. The distance may be, for example, the L1 norm (sum of absolute values ​​of differences) or the L2 norm (sum of squares of differences). Examples of values ​​corresponding to the distance include the L1 mean obtained by dividing the L1 norm by the number of components, and the L2 mean obtained by dividing the L2 norm by the number of components. The number of components is, for example, the number of elements included in the estimation result of the estimation model or the correct answer.

[0092] In more detail, for example, the loss is a value corresponding to the difference between the estimated ellipse parameters and the correct ellipse parameters. The value corresponding to the difference may be, for example, a value corresponding to the distance between the estimated ellipse parameters and the correct ellipse parameters.

[0093] The value according to the distance may be either the L1 mean obtained by dividing the L1 norm by the number of components, or the L2 mean obtained by dividing the L2 norm by the number of components, where the number of components is, for example, the number of dimensions when the ellipse parameters are expressed as a vector.

[0094] The distance is not limited to the L1 norm or the L2 norm, and the loss is not limited to a value according to the distance, the L1 average, or the L2 average, and a general loss function may be used, for example.

[0095] When mini-batch learning is performed, the loss function calculation unit 121 may calculate the loss using the estimated ellipse parameters and the correct ellipse parameters for each of the multiple training iris images that make up the mini-batch.

[0096] In this case, the loss function calculation unit 121 may calculate the average value of the losses calculated for each of the multiple training iris images that make up the mini-batch as the loss for that mini-batch. In other words, when the loss is related to a mini-batch, it may be, for example, the sum of the losses such as the L1 norm and L2 norm calculated for each of the multiple training iris images that make up the mini-batch divided by the number of training iris images that make up the mini-batch.

[0097] (Regarding the gradient calculation unit 122) The gradient calculation unit 122 calculates the gradient for each model parameter by backpropagation using, for example, the loss calculated by the loss function calculation unit 121. Note that a method other than backpropagation may be used to calculate the gradient.

[0098] (Parameter Updater 123) The parameter updater 123 updates each of the model parameters using, for example, the gradient calculated by the gradient calculator 122 and a predetermined learning rate.

[0099] For example, the parameter update unit 123 may update each of the model parameters included in the estimation model using a value obtained by multiplying the gradient of each model parameter by a learning rate. The learning rate may be a predetermined value or may be changed from an initial value depending on the learning stage.

[0100] The learning rate or its initial value may be determined using an appropriate general method such as an LR range test, which determines the learning rate based on the change in the accuracy rate or loss when the learning rate is gradually increased by a predetermined amount. When changing the learning rate, an appropriate general method such as CLR (Cyclic Learning Rate), which changes the learning rate for each batch between an upper limit learning rate and a lower limit learning rate, may be used. The parameter update unit 123 may include at least one of a function for determining such a learning rate or its initial value, and a function for changing the learning rate.

[0101] As described above, the learning estimation unit 110 executes the estimation process (step S120), and the update unit 120 executes the update process (step S120), thereby enabling learning of the estimation model.

[0102] When mini-batch learning is used, steps S120 to S120 may be repeated for each mini-batch.

[0103] The condition for terminating the learning of the estimation model (termination condition) may be, for example, repeating steps S110 to S120 a predetermined number of times. The termination condition may be, for example, that the change in the loss value is equal to or less than a predetermined value. The termination condition may be, for example, that steps S110 to S120 are repeated for all training iris images stored in the training data storage device 90. Note that the termination conditions are not limited to those exemplified here. Furthermore, multiple termination conditions may be combined.

[0104] By training the estimation model, a trained estimation model can be constructed, and the trained estimation model can be used to accurately estimate the contour or edge of the iris contained in the iris image.

[0105] The model parameters of the trained estimation model may be stored in the training estimation unit 110 together with the estimation model, or may be stored in a storage unit (not shown) for storing model parameters. This storage unit may be included in the information processing device 100 or in an external device (not shown). In this case, the training estimation unit 110 may estimate the ellipse parameters using the model parameters stored in the storage unit. Furthermore, the update unit 120 may update the model parameters stored in the storage unit.

[0106] (Example of physical configuration of information processing device 100) The information processing device 100 physically includes a bus 1010, a processor 1020, a memory 1030, a storage device 1040, a network interface 1050, an input interface 1060, and an output interface 1070, as shown in FIG.

[0107] The bus 1010 is a data transmission path for transmitting and receiving data among the processor 1020, memory 1030, storage device 1040, network interface 1050, input interface 1060, and output interface 1070. However, the method of connecting the processor 1020 and the like to each other is not limited to bus connection.

[0108] The processor 1020 is implemented by a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), or the like.

[0109] The memory 1030 is a main storage device realized by a RAM (Random Access Memory) or the like.

[0110] The storage device 1040 is an auxiliary storage device realized by a hard disk drive (HDD), a solid state drive (SSD), a memory card, a read only memory (ROM), or the like. The storage device 1040 stores program modules for realizing the functions of the information processing device 100. The processor 1020 reads each of these program modules into the memory 1030 and executes them to realize the function corresponding to the program module.

[0111] The network interface 1050 is an interface for connecting the information processing device 100 to a network. The network is a communication network for transmitting and receiving information to and from other devices (not shown), and may be wired, wireless, or a combination of these.

[0112] The input interface 1060 is an interface for the user to input information, and is composed of, for example, a touch panel, a keyboard, a mouse, and the like.

[0113] The output interface 1070 is an interface for presenting information to the user, and is configured, for example, by a liquid crystal panel, an organic EL (Electro-Luminescence) panel, etc. The output interface 1070 may also include a speaker.

[0114] Although the example in which the information processing device 100 is physically configured as one device (e.g., a computer, etc.) has been described, the information processing device 100 may also be configured as a plurality of devices (e.g., computers, etc.) that transmit and receive information to each other via, for example, a network NT. In this case, the plurality of devices may cooperate to perform information processing. The training data storage device 90 may also be physically configured in the same way as the information processing device 100. Furthermore, the training data storage device 90 and the information processing device 100 may also be physically configured as a single device (e.g., a computer, etc.).

[0115] (Actions and Effects) As described above, according to this embodiment, the information processing device 100 includes the training estimation unit 110 and the updating unit 120. The training estimation unit 110 estimates ellipse parameters related to the training iris image using an estimation model for estimating ellipse parameters related to the contour or edge of the iris included in the iris image. The updating unit 120 updates the model parameters included in the estimation model using estimated ellipse parameters, which are the estimated ellipse parameters, and correct ellipse parameters, which are correct solutions to the ellipse parameters related to the training iris image.

[0116] This makes it possible to construct an estimation model that uses an ellipse to estimate the contour or edge of the iris included in the iris image, thereby enabling the contour or edge of the iris included in the iris image to be estimated with high accuracy.

[0117] According to this embodiment, the ellipse parameters include a first ellipse parameter and a second ellipse parameter. The first ellipse parameter is a parameter for specifying an ellipse that represents an outer contour or edge of an iris included in the iris image. The second ellipse parameter is a parameter for specifying an ellipse that represents an inner contour or edge of an iris included in the iris image.

[0118] Here, the inner contour or edge of the iris corresponds to the contour or edge of the pupil, as described above. Therefore, by including the first ellipse parameter and the second ellipse parameter in the ellipse parameters, it is possible to construct an estimation model that estimates the contour or edge of the iris and pupil contained in the iris image using an ellipse. Therefore, it is possible to accurately estimate the contour or edge of the iris and pupil contained in the iris image.

[0119] According to this embodiment, the ellipse parameters include the center coordinate, the major axis radius, the minor axis radius, and the ellipse rotation angle, or the center coordinate, the major axis radius, the ellipticity, and the ellipse rotation angle.

[0120] By using such ellipse parameters, it is possible to construct an estimation model that uses an ellipse to estimate the contour or edge of the iris included in the iris image, thereby enabling accurate estimation of the contour or edge of the iris and pupil included in the iris image.

[0121] According to this embodiment, it is determined in advance which component of the vector each component included in the ellipse parameters is set to.

[0122] This reduces the number of calculations using vectors representing ellipse parameters, thereby enabling faster processing using ellipse parameters.

[0123] [Embodiment 2] The ellipse parameters may include an ellipse rotation angle θ that indicates the rotation angle of an ellipse that represents the outline or edge of the iris included in the iris image, as described in embodiment 1. The correct ellipse parameters may include a correct ellipse rotation angle that is a correct answer to the ellipse rotation angle θ.

[0124] In the first embodiment, an example has been described in which no constraints are imposed on the angular range of the ellipse rotation angle θ. Also, an example has been described in which constraints are imposed on the reference ellipse axis, reference direction, and positive rotation direction of the ellipse rotation angle θ. However, if the angular range of the ellipse rotation angle θ, the reference ellipse axis, the reference direction, and the positive rotation direction are not appropriately constrained, different ellipse rotation angles θ may be specified for ellipses with the same geometric shape.

[0125] FIG. 8 shows examples of ellipses C1 to C4 that have the same geometric shape. The X and Y coordinates of the centers of the ellipses C1 to C4 are X0 and Y0, respectively, and the major and minor radii are RA and RB, respectively. Therefore, the ellipses C1 to C4 have the same geometric shape, position, and size. However, the ellipse rotation angles θ of the ellipses C1 to C4 are α [degrees], α + 90 [degrees], α + 180 [degrees], and α + 270 [degrees], respectively. Thus, the ellipse rotation angles θ of the ellipses C1 to C4 are expressed as different angles.

[0126] Here, the ellipses C1 and C3 have their major axes as their reference ellipse axes and their ellipse rotation angles θ are expressed, and the ellipses C2 and C4 have their minor axes as their reference ellipse axes and their ellipse rotation angles θ are expressed.

[0127] The ellipse rotation angle θ of the ellipses C1 to C4 has an angle range of 0 degrees to 360 degrees, with the right direction in the figure (e.g., right direction in the iris image) being the reference direction and the counterclockwise direction in the figure (e.g., counterclockwise in the iris image) being the positive direction of rotation. This counterclockwise direction is the rotation direction that can point the right direction upward in the shortest time. If there are no constraints on these, the ellipses can be expressed with many more different angles.

[0128] In this way, if no constraints are set in advance on the angle range, reference direction, reference ellipse axis, and positive rotation direction, and the ellipse rotation angle θ of the iris ellipse for the training iris image is expressed by any method, there will be multiple correct answers for the ellipse rotation angle θ. As illustrated in Figure 8, even if the angle range is set to 0 degrees to 360 degrees, the reference direction is set to the horizontal direction of the iris image, and the positive rotation direction is set to the counterclockwise direction of the iris image, there will be multiple correct answers for the ellipse rotation angle θ. Furthermore, even if the reference ellipse axis is set in advance to be the major axis or the minor axis, there will be multiple correct answers for the ellipse rotation angle θ, as illustrated by ellipses C1 and C3, and ellipses C2 and C4 in Figure 8.

[0129] Generally, when there are multiple correct answers for a single input, the error function used in training a machine learning model converges to the average value of the multiple correct answers, which often results in poor estimation accuracy. Therefore, if an estimation model is trained using a correct ellipse rotation angle that contains a mixture of ellipse rotation angles θ expressed using different methods, the accuracy of the ellipse rotation angle θ estimated by the trained estimation model will be poor.

[0130] Therefore, a constraint may be imposed on the ellipse rotation angle θ included in the ellipse parameters. An example of the constraint on the ellipse rotation angle θ included in the ellipse parameters will be described below. Note that, for the sake of brevity, descriptions that overlap with other embodiments will be omitted as appropriate.

[0131] The constraint on the ellipse rotation angle θ is set with respect to, for example, at least one of the reference ellipse axis, the angle range, the reference direction, and the positive rotation direction.

[0132] The constraint on the ellipse rotation angle θ may be applied to, for example, the correct ellipse rotation angle. That is, correct ellipse parameters including the constrained correct ellipse rotation angle may be used to train the estimation model. For example, the constrained correct ellipse rotation angle may be prepared in advance, and the training data storage device 90 may store correct data including the constrained correct ellipse rotation angle.

[0133] This allows the estimation model to be trained using the constrained correct ellipse rotation angle. As a result of this training, the estimation model estimates an ellipse rotation angle that is constrained in the same way as the correct ellipse rotation angle. Therefore, it is possible to construct an estimation model that estimates ellipse parameters including the constrained ellipse rotation angle.

[0134] (Example of Constraints on Reference Ellipse Axis) The reference ellipse axis may be, for example, a predetermined one of the major and minor axes of the iris ellipse.

[0135] More specifically, for example, the major axis may be predetermined as the reference ellipse axis. In this case, the ellipse rotation angle θ is expressed as the angle of the major axis of the iris ellipse with respect to the reference direction.

[0136] Alternatively, for example, the minor axis may be predetermined as the reference ellipse axis. In this case, the ellipse rotation angle θ is expressed as the angle of the minor axis of the ellipse representing the iris with respect to the reference direction.

[0137] (Example of Constraint on Angular Range) The angular range is the range of the ellipse rotation angle θ, and may be determined in advance.

[0138] The size of the angle range may be, for example, 180 degrees. That is, when the lower limit of the angle range is θmin and the upper limit of the angle range is θmax, θmax and θmin may be predetermined to satisfy θmax - θmin = 180 degrees. θmax and θmin may be either positive or negative values. In this case, the ellipse rotation angle θ is the angle of one of the major and minor axes of an ellipse representing the contour or edge of the iris relative to a reference direction, and is expressed within an angular range of 180 degrees.

[0139] In more detail, for example, the angular range may include the reference direction and have a magnitude of 180 degrees. As described above, the reference direction is a direction predetermined as a reference for the ellipse rotation angle θ, and therefore the angle of the reference direction is 0 degrees. Therefore, θmax and θmin may be predetermined to satisfy θmax - θmin = 180 degrees and θmin < 0 < θmax. In this case, the ellipse rotation angle θ is the angle of a predetermined one of the major and minor axes of an ellipse representing the contour or edge of the iris with respect to the reference direction, and is expressed as an angle range that includes the reference direction and has a magnitude of 180 degrees.

[0140] More specifically, the angle range may be, for example, from −90 degrees to +90 degrees, i.e., θmin and θmax may be −90 degrees and +90 degrees, respectively.

[0141] The ellipse rotation angle θ may be expressed, for example, using a position (cos2φ, sin2φ) on a unit circle having the same center as the ellipse. That is, the estimation model may calculate the ellipse rotation angle using (cos2φ, sin2φ) instead of the ellipse rotation angle θ. In this case, the ellipse rotation angle θ may be output as arctan(sin2φ / cos2φ) / 2. This allows the estimation model to perform calculations using -180 degrees to +180 degrees, while obtaining the ellipse rotation angle θ, which is a continuous value between -90 degrees and +90 degrees.

[0142] The angular range may also be set from the iris image according to predetermined constraints, an example of which will be described in other embodiments.

[0143] (Regarding constraints on the reference direction) As described above, the reference direction may be the horizontal or vertical direction of the eye identified from the iris image. In more detail, for example, the reference direction may be constrained so as to be identified from information about the eye included in the iris image. The information about the eye may be a characteristic position of the eye, a gaze direction, etc. An example of using the gaze direction will be described in another embodiment.

[0144] When the reference direction is the horizontal direction of the eye, a direction intersecting the reference direction at a predetermined angle, such as 90 degrees, may be set as the vertical direction of the eye. When the reference direction is the vertical direction of the eye, a direction intersecting the reference direction at a predetermined angle, such as 90 degrees, may be set as the horizontal direction of the eye.

[0145] (Regarding Feature Positions) Feature positions are positions of characteristic parts of the eyes that are predetermined. Characteristic parts are generally also referred to as feature points, key points, etc. The characteristic parts may include, for example, at least one of the left and right ends of the eyes, the upper and lower ends of the eyes, a part related to the upper eyelid, a part related to the lower eyelid, the center of the eye, etc. An example including the center of the eye will be described in another embodiment.

[0146] The location on the upper or lower eyelid may be a curve representing the entire eyelid, or may be a predetermined location (point) on the eyelid, as exemplified in Figures 9 and 10. The predetermined location (point) on the eyelid may include at least one of the outer corner of the eye, the inner corner of the eye, the center of the eyelid, the top end of the upper eyelid, the bottom end of the lower eyelid, etc.

[0147] 9 is a diagram showing an example of characteristic parts of the upper and lower eyelids. In the example shown in the figure, the characteristic parts of the upper and lower eyelids include the inner corner A of the eye, the center B of the upper eyelid, the outer corner C of the eye, and the center D of the lower eyelid.

[0148] 10 is a diagram showing another example of characteristic points on the upper eyelid and the lower eyelid, which may be one or more predetermined points on the upper eyelid and the lower eyelid, as indicated by dots in the figure.

[0149] (Example 1 of constraints related to the reference direction) The reference direction may be, for example, a direction along a line passing through the left and right ends of the eye. In more detail, for example, the reference direction may be a direction from either the left or right end of the eye to the other. The direction from either the left or right end of the eye to the other is, for example, a direction from the outer corner C of the eye to the inner corner A of the eye. Note that the direction from either the left or right end of the eye to the other is not limited to this, and may be, for example, a direction from the inner corner A of the eye to the outer corner C of the eye.

[0150] In this case, the reference direction may be the horizontal direction of the eye, and the direction intersecting the reference direction at a predetermined angle may be the vertical direction of the eye.

[0151] (Example 2 of constraints related to the reference direction) The reference direction may be, for example, a direction along a line passing through the upper and lower ends of the eye. In more detail, for example, the reference direction may be a direction from one of the upper and lower ends of the eye to the other. The direction from one of the upper and lower ends of the eye to the other is, for example, a direction from the center D of the lower eyelid to the center B of the upper eyelid. Note that the direction from one of the upper and lower ends of the eye to the other is not limited to this, and may be, for example, a direction from the center B of the upper eyelid to the center D of the lower eyelid.

[0152] In this case, the reference direction may be the vertical direction of the eye, and the direction intersecting the reference direction at a predetermined angle may be the horizontal direction of the eye.

[0153] (Example 3 of constraints related to the reference direction) The reference direction may be, for example, a predetermined one of the horizontal and vertical directions of the eye. In detail, for example, the horizontal direction of the eye may be a direction from either the left or right end of the eye to the other. Furthermore, the vertical direction of the eye may be a direction from either the top or bottom end of the eye to the other. The reference direction may be a predetermined one of the horizontal and vertical directions of the eye.

[0154] (Regarding Constraints on Positive Rotation Direction) For example, if an iris image includes an eye, the positive rotation direction may be constrained so as to be specified from information on the eye included in the iris image.

[0155] In more detail, for example, the positive rotation direction may be a direction based on the horizontal and vertical directions of the eye included in the iris image. More specifically, for example, the positive rotation direction may be a rotation direction in which the horizontal direction of the eye faces the vertical direction of the eye within an angle range or at a minimum rotation angle. The positive rotation direction may be a rotation direction in which the vertical direction of the eye faces the horizontal direction of the eye within an angle range or at a minimum rotation angle.

[0156] In detail, for example, the positive rotation direction may be a direction in which the horizontal direction of the eye is directed toward a specific location among the characteristic locations. The specific location may be predetermined, for example, from among the characteristic locations. More specifically, for example, the positive rotation direction may be a direction in which the horizontal direction of the eye is directed toward the specific location within an angle range or at a minimum rotation angle. In this case, the specific location may be, for example, the upper eyelid or may be predetermined with respect to the upper eyelid, such as the center B of the upper eyelid. The positive rotation direction may be a direction in which the vertical direction of the eye is directed toward the specific location within an angle range or at a minimum rotation angle. In this case, the specific location may be, for example, the left end of the eye (for example, the outer corner C of the eye), the right end of the eye (for example, the inner corner A of the eye), or the like.

[0157] The above-mentioned examples of constraints on the reference ellipse axis, angle range, reference direction, and positive rotation direction may be combined. This combination may be determined appropriately, and an example is shown in FIG.

[0158] FIG. 11 is a diagram showing an example in which constraints on the reference ellipse axis, angle range, reference direction, and positive rotation direction are applied to the ellipse rotation angle θ of a first ellipse. In the example shown in the figure, the reference ellipse axis is constrained to the minor axis of the ellipse. In the example shown in the figure, the angle range is constrained to -90 degrees to +90 degrees. In the example shown in the figure, the reference direction is constrained to the horizontal direction of the eye from the outer corner C of the eye to the inner corner A of the eye. In the example shown in the figure, the vertical direction of the eye is constrained to the direction from the center D of the lower eyelid to the center B of the upper eyelid. In the example shown in the figure, the positive rotation direction is constrained to the horizontal direction toward the center B of the upper eyelid.

[0159] In the example of Fig. 11, the reference direction is the horizontal direction of the eye, and the reference ellipse axis is the minor axis. This is also an example of a preferable relationship between the reference direction and the reference ellipse axis when the angle range is -90 to +90 degrees. For example, when the angle range is -90 to +90 degrees and the reference direction is the vertical direction of the eye, the reference ellipse axis is preferably the major axis.

[0160] This is because the smaller the ellipticity, the flatter the iris ellipse becomes, and if an error occurs in the ellipse rotation angle θ, the difference between the area of ​​the iris contained in the iris image and the area of ​​the iris ellipse becomes larger. In other words, it is desirable that the error in the ellipse rotation angle θ be smaller as the ellipticity is smaller. The reason for this will be explained further below.

[0161] The ellipticity of the iris ellipse often decreases when, for example, the gaze direction deviates from the imaging direction. Fig. 12 is a diagram showing an example in which the gaze direction deviates leftward and rightward from the imaging direction. Fig. 13 is a diagram showing an example in which the gaze direction deviates upward from the imaging direction. Fig. 14 is a diagram showing an example in which the gaze direction deviates downward from the imaging direction. As shown in Fig. 12, when the gaze direction deviates leftward and rightward, the iris ellipse becomes a vertically elongated ellipse whose major axis is oriented in a direction whose main component is the vertical direction of the eye. In contrast, as shown in Figs. 13 and 14, when the gaze direction deviates up and down, the iris ellipse becomes a horizontally elongated ellipse whose major axis is oriented in a direction whose main component is the horizontal direction of the eye.

[0162] Furthermore, the eye generally has a horizontal length. Due to this structure of the eye, the range of movement of the iris is larger in the left-right direction than in the up-down direction. Therefore, the ellipticity of the iris ellipse is often smaller when the gaze direction deviates left-right from the imaging direction (see FIG. 12) than when the gaze direction deviates up-down from the imaging direction (see FIGS. 13 and 14).

[0163] Therefore, it is desirable that the magnitude of the error in the ellipse rotation angle θ be smaller when the iris ellipse is vertically long than when the iris ellipse is horizontally long.

[0164] However, if restrictions are placed on the angle range, depending on how the reference direction and reference ellipse axis are defined, the magnitude of the absolute error in the ellipse rotation angle θ for a vertically elongated ellipse may be larger than the magnitude of the absolute error in the ellipse rotation angle θ for a horizontally elongated ellipse.

[0165] For example, suppose that the ellipse rotation angle θ is −90 degrees≦+90 degrees, the reference direction is the horizontal direction of the eye, and the reference ellipse axis is the major axis. In this case, even if the true value of the ellipse rotation angle θ is −88 degrees, there is a possibility that the ellipse rotation angle θ will approach +90 degrees due to an error, resulting in a large absolute error. Here, the absolute error is the absolute value of the difference between the estimated or correct value of the ellipse rotation angle θ and the true value.

[0166] In more detail, for example, as shown in Figure 15, assume that the true value of the ellipse rotation angle θ, with the reference direction being the horizontal direction of the eye and the major axis being the reference ellipse axis, is -88 degrees. If the relative error is -3 degrees, and no constraints are placed on the angle range, the ellipse rotation angle θ can be expressed as -91 degrees. Here, the relative error is the value obtained by subtracting the true value from the estimated or correct value of the ellipse rotation angle θ. If the relative error is -3 degrees, the magnitude of the absolute error is 3 degrees.

[0167] In this detailed example, an ellipse rotation angle θ of -91 degrees means that the angle between the reference direction and the major axis of the reference ellipse is -91 degrees when the angle is in the negative direction. This angle can also be expressed as +89 degrees when the angle is in the positive direction. Therefore, in the above detailed example, if the angle range is constrained to -90 degrees ≦ ellipse rotation angle θ ≦ +90 degrees, the ellipse rotation angle θ becomes +89 degrees. The magnitude of the absolute error in this case is 177 degrees (= |+89 degrees - (-88 degrees)|). Here, |x| means the absolute value of x.

[0168] In other words, when the true value of the ellipse rotation angle θ is near the ends (upper and lower limits) of the angle range, the error may cause it to be expressed as an angle in the opposite direction to the true value, which may result in a large absolute error in the ellipse rotation angle θ.

[0169] Such errors may occur, for example, in the calculation of the estimation model and may be included in the correct ellipse rotation angle. Therefore, when the true value of the ellipse rotation angle θ is near the end of the angle range, the magnitude of the absolute error of the ellipse rotation angle θ in the estimation result may be large as a result of errors occurring in the calculation of the estimation model. Furthermore, when training is performed using a correct ellipse rotation angle that is near the end of the angle range, an estimation model with poor estimation accuracy for the ellipse rotation angle θ may be constructed due to the aforementioned tendency of the error function to converge to the average value of multiple correct answers. Thus, when the true value of the ellipse rotation angle θ is near the end of the angle range, the estimation accuracy of the ellipse rotation angle θ may be poor due to the influence of errors in the calculation of the estimation model and the correct ellipse rotation angle.

[0170] As mentioned above, when the iris ellipse is vertically elongated due to the structure of the eye, the ellipticity often becomes small, so in this case it is particularly desirable that the magnitude of the error be small.

[0171] Therefore, the horizontal direction of the eye may be predetermined as the reference direction, and the minor axis may be predetermined as the reference ellipse axis. Alternatively, the vertical direction of the eye may be predetermined as the reference direction, and the major axis may be predetermined as the reference ellipse axis. That is, the ellipse rotation angle θ may be expressed as the orientation of the minor axis of an ellipse representing the outline or edge of the iris, with the horizontal direction of the eye as the reference direction. Alternatively, the ellipse rotation angle θ may be expressed as the orientation of the major axis of an ellipse representing the outline or edge of the iris, with the vertical direction of the eye as the reference direction.

[0172] As a result, for example, when the ellipse rotation angle θ is -90 degrees≦the ellipse rotation angle θ≦+90 degrees, the ellipse rotation angle θ of a vertically elongated iris ellipse, which often results in a small ellipticity in an actual iris image, is more likely to be near the median of the angle range, 0 degrees, rather than near the end of the angle range. This makes it possible to prevent the magnitude of the absolute error in the ellipse rotation angle θ in the estimation result or correct answer from increasing. This makes it possible to construct an estimation model that uses an ellipse to more accurately estimate the contour or edge of the iris contained in an iris image.

[0173] (Actions and Effects) As described above, according to this embodiment, the ellipse parameters include an ellipse rotation angle θ that indicates the angle of rotation of an ellipse that represents the outline or edge of the iris included in the iris image. The ellipse rotation angle θ is the angle of a predetermined one of the major and minor axes of the ellipse that represents the outline or edge of the iris with respect to a reference direction, and is expressed within an angular range of 180 degrees that includes the reference direction.

[0174] This reduces the possibility that ellipses with the same geometric shape are estimated as ellipses with different ellipse rotation angles θ, and makes it possible to construct an estimation model that accurately estimates the contour or edge of the iris contained in an iris image using an ellipse, thereby making it possible to estimate the contour or edge of the iris and pupil contained in an iris image with even greater accuracy.

[0175] According to this embodiment, the reference direction is the horizontal or vertical direction of the eye contained in the iris image.

[0176] This allows the direction of the ellipse representing the contour or edge of the iris included in the iris image to be expressed based on the reference direction, making it possible to construct an estimation model that uses the ellipse to accurately estimate the contour or edge of the iris included in the iris image, thereby making it possible to estimate the contour or edge of the iris and pupil included in the iris image with even greater accuracy.

[0177] In this embodiment, the horizontal direction is a direction whose main component is the horizontal direction, and the vertical direction is a direction whose main component is the vertical direction. The ellipse rotation angle θ is expressed as an angle within an angle range relative to the reference direction, with the rotation direction in which the horizontal direction is directed vertically upward being the positive direction and the rotation direction in which the horizontal direction is directed vertically downward being the negative direction.

[0178] As a result, the ellipse rotation angle θ in an actual iris image is more likely to be near the median of the angle range, i.e., 0, rather than near the edge of the angle range. This makes it possible to construct an estimation model that uses an ellipse to more accurately estimate the contour or edge of the iris contained in the iris image. This makes it possible to more accurately estimate the contour or edge of the iris and pupil contained in the iris image.

[0179] According to this embodiment, the angle range of the ellipse rotation angle θ is from −90 degrees to +90 degrees.

[0180] This reduces the possibility of estimating different ellipse rotation angles θ for ellipses with the same geometric shape, making it possible to construct an estimation model that accurately estimates the contour or edge of the iris contained in an iris image using an ellipse, thereby making it possible to estimate the contour or edge of the iris and pupil contained in an iris image with even greater accuracy.

[0181] According to this embodiment, the ellipse rotation angle θ is expressed by the orientation of the minor axis of the ellipse representing the outline or edge of the iris, with the horizontal direction of the eye as the reference direction, or by the orientation of the major axis of the ellipse representing the outline or edge of the iris, with the vertical direction of the eye as the reference direction.

[0182] As a result, the ellipse rotation angle θ in an actual iris image is more likely to be near the median of the angle range, i.e., 0, rather than near the edge of the angle range. This makes it possible to construct an estimation model that uses an ellipse to more accurately estimate the contour or edge of the iris contained in the iris image. This makes it possible to more accurately estimate the contour or edge of the iris and pupil contained in the iris image.

[0183] [Embodiment 3] In embodiment 2, an example was described in which the training data storage device 90 stores correct answer data including a constrained correct answer ellipse rotation angle. In this embodiment, an example will be described in which an information processing device generates correct answer data including a constrained correct answer ellipse rotation angle based on source data prepared in advance. Note that, for the sake of brevity, descriptions that overlap with other embodiments will be omitted as appropriate.

[0184] As shown in FIG. 16, for example, the information processing system S2 includes a training data storage device 92 and an information processing device 200 instead of the training data storage device 90 and the information processing device 100, respectively.

[0185] (Regarding the training data storage device 92) For example, training data is stored in advance in the training data storage device 92. The training data includes, for example, training iris images similar to those in the first embodiment and original data.

[0186] The original data is data that is the basis of the correct answer data regarding the training iris image.

[0187] For example, the original data may include unconstrained rotation angles that replace the solution ellipse rotation angles of the ground truth data. The unconstrained rotation angles are angles that represent the ground truth ellipse rotation angles for the training iris images, but are not subject to the constraints of the ground truth ellipse rotation angles used to train the estimation model.

[0188] For example, the reference ellipse axis, reference direction, angle range, and positive rotation direction in the unconstrained rotation angle may be the minor axis, the right direction in the iris image, 0 degrees to 360 degrees, and the counterclockwise direction in the iris image, respectively. Note that the reference ellipse axis, reference direction, angle range, and positive rotation direction in the unconstrained rotation angle are not limited to those exemplified here. The original data may include information indicating at least one of the reference ellipse axis, reference direction, angle range, and positive rotation direction.

[0189] For example, the original data may further include solution ellipse parameters that are similar to the solution data except for the solution ellipse rotation angle.

[0190] For example, the original data may further include a correct position. The correct position is a correct answer for the feature position related to the eye included in the training iris image. In other words, the correct position is a correct answer for the feature position related to the training iris image.

[0191] 17 , the information processing device 200 includes a learning estimation unit 210 and an updating unit 220. The iris image and the learning iris image according to this embodiment include an eye. The eye includes, for example, a pupil, an iris, and upper and lower eyelids.

[0192] The training estimation unit 210 estimates ellipse parameters and feature positions for the training iris image using an estimation model for further estimating ellipse parameters related to the contour or edge of the iris included in the iris image and feature positions of the eye included in the iris image.

[0193] The update unit 220 updates the model parameters included in the estimation model using the estimated ellipse parameters, the correct ellipse parameters, the estimated feature positions that are feature positions estimated for the training iris image, and the correct positions that are the correct answers for the feature positions for the training iris image.

[0194] (Example of Processing Operation of Information Processing Device 200) The information processing device 200 executes information processing as shown in FIG.

[0195] The training estimation unit 210 estimates ellipse parameters and feature positions for the training iris image using an estimation model for estimating ellipse parameters related to the contour or edge of the iris included in the iris image and feature positions of the eye included in the iris image (step S210).

[0196] The update unit 220 further uses the ellipse parameters estimated for the training iris image, the correct ellipse parameters, the estimated feature positions that are the feature positions estimated for the training iris image, and the correct positions that are the correct answers for the feature positions for the training iris image to update the model parameters included in the estimation model (step S220).

[0197] (Regarding the training estimation unit 210) The training estimation unit 210 estimates ellipse parameters and feature positions for a training iris image, for example, using an estimation model currently being trained. The estimation model currently being trained is a machine learning model for estimating ellipse parameters related to the contour or edge of the iris included in the iris image and feature positions of the eye included in the iris image. The estimation model includes one or more model parameters and is configured, for example, using a neural network that uses a convolutional layer, an attention mechanism, etc.

[0198] The estimation model may be composed of, for example, one or more machine learning models. The machine learning model may include, for example, an output layer, a hidden layer, and another output layer. When composed of multiple machine learning models, the estimation model may include, for example, a common model including an input layer, an ellipse estimation model including an output layer for ellipse parameters, and a feature position estimation model including an output layer for feature positions. Each of the ellipse estimation model and the feature position estimation model may perform calculations by acquiring output (e.g., intermediate feature amounts) from the common model. In this case, the ellipse parameters are estimated by the common model and the ellipse estimation model, and the feature positions are estimated by the common model and the feature position estimation model. Note that configuration examples of the estimation models are not limited to these.

[0199] The learning estimation unit 210 includes a first estimation unit 211 and a first output unit 212, as shown in FIG. 19, for example.

[0200] The first estimation unit 211 estimates ellipse parameters and feature positions for the training iris image.

[0201] The first output unit 212 outputs the estimated ellipse parameters and feature positions (i.e., estimated ellipse parameters and estimated feature positions).

[0202] The learning estimation unit 210 executes an estimation process (step S210) as shown in FIG. 20, for example.

[0203] The first estimation unit 211 estimates ellipse parameters and feature positions for the training iris image (step S211).

[0204] The first output unit 212 outputs the estimated ellipse parameters and feature positions (i.e., estimated ellipse parameters and estimated feature positions) (step S212).

[0205] (First Estimation Unit 211) The first estimation unit 211 acquires training iris images in response to, for example, a user instruction. For example, the first estimation unit 211 acquires training iris images from the training data storage device 92. Note that the first estimation unit 211 may further acquire training data, original data, etc.

[0206] For example, the first estimation unit 211 inputs the acquired training iris image to the estimation model. In response to this input, the estimation model estimates ellipse parameters and feature positions related to the input training iris image.

[0207] (Regarding the first output unit 212) The first output unit 212 outputs output data including, for example, estimated ellipse parameters and estimated feature positions. The estimated feature positions are feature positions estimated by the estimation model in response to an input iris image. The estimated ellipse parameters and the estimated feature positions may be represented by vectors.

[0208] (When Multiple Training Iris Images Are Used) When training the estimation model using multiple training iris images, the training estimation unit 210, for example, acquires multiple training iris images. In this case, the training estimation unit 210 may, for example, input each of the acquired training iris images to the estimation model. For example, the training estimation unit 210 may input each of the training iris images sequentially to the estimation model. In response to this input, the estimation model outputs ellipse parameters and feature positions for each of the input training iris images. In this way, the training estimation unit 210 may estimate ellipse parameters for each of the training iris images.

[0209] (Regarding the updating unit 220) The updating unit 220 generates correct answer data including correct answers for the constrained ellipse rotation angles based on, for example, the original data. The updating unit 220 updates the model parameters included in the estimation model using, for example, the estimated ellipse parameters and estimated feature positions included in the output data and the correct answer data.

[0210] In detail, for example, as shown in FIG. 21 , the update unit 220 includes a constraint unit 221, a loss function calculation unit 222, a gradient calculation unit 223, and a parameter update unit 224.

[0211] The constraint unit 221 generates correct data including the constrained correct ellipse rotation angle based on the original data. Note that the training data may include the correct data instead of the original data. In this case, the update unit 220 does not need to include the constraint unit 221.

[0212] The loss function calculation unit 222 calculates a loss using the ellipse parameters and feature positions estimated by the training estimation unit 210 (estimated ellipse parameters and estimated feature positions) and the ground truth data related to the training iris image.

[0213] The gradient calculation unit 223 uses the loss calculated by the loss function calculation unit 222 to calculate the gradient for each of the model parameters included in the estimation model.

[0214] The parameter update unit 224 updates each of the model parameters included in the estimation model using the gradient calculated by the gradient calculation unit 223 .

[0215] The update unit 220 executes an update process (step S220) as shown in FIG. 22, for example.

[0216] The constraint unit 221 generates correct data including the constrained correct ellipse rotation angle based on the original data (step S221).

[0217] The loss function calculation unit 222 calculates a loss using the ellipse parameters and feature positions estimated in step S210 (estimated ellipse parameters and estimated feature positions) and the ground truth data related to the training iris image (step S222).

[0218] The gradient calculation unit 223 calculates the gradient for each of the model parameters included in the estimation model using the loss calculated by the loss function calculation unit 222 (step S223).

[0219] The parameter update unit 224 updates each of the model parameters included in the estimation model using the gradient calculated by the gradient calculation unit 223 (step S224).

[0220] First, a detailed example of the constraint unit 221 will be described.

[0221] (Regarding the Constraint Unit 221) The constraint unit 221 acquires a constrained correct ellipse rotation angle by applying a predetermined constraint to an unconstrained rotation angle included in the original data, and then generates correct data including the acquired correct ellipse rotation angle.

[0222] As shown in FIG. 23, the constraint unit 221 includes a reference direction setting unit 221a, a rotational direction setting unit 221b, a correction unit 221c, and a correct data generation unit 221d.

[0223] The reference direction setting unit 221a sets the reference direction using the correct position included in the original data.

[0224] The positive rotation direction setting unit 221b sets the positive rotation direction using the correct position included in the original data.

[0225] The correction unit 221c corrects the unconstrained rotation angle included in the original data to a constrained elliptical rotation angle using the set reference direction and rotation direction.

[0226] The supervised data generating unit 221d generates supervised data including the constrained ellipse rotation angle.

[0227] The constraint unit 221 executes a constraint process (step S221) as shown in FIG. 24, for example.

[0228] The reference direction setting unit 221a sets the reference direction using the correct position included in the original data (step S221a).

[0229] The positive rotation direction setting unit 221b sets the positive rotation direction using the correct position included in the original data (step S221b).

[0230] The correction unit 221c corrects the unconstrained rotation angle included in the original data to a constrained ellipse rotation angle using the set reference direction and rotation direction (step S221c).

[0231] The supervised data generating unit 221d generates supervised data including the constrained ellipse rotation angle (step S221d).

[0232] (Regarding the reference direction setting unit 221a) The reference direction setting unit 221a sets the horizontal and vertical directions of the eye based on, for example, the correct position and constraints regarding the reference direction, and sets the reference direction to be along a predetermined one of the set horizontal and vertical directions.

[0233] In more detail, for example, the reference direction setting unit 221a may set the direction from the outer corner C of the eye to the inner corner A of the eye as the horizontal direction and the reference direction.The reference direction setting unit 221a may also set the direction from the center D of the lower eyelid to the center B of the upper eyelid, for example.

[0234] (Regarding the positive rotation direction setting unit 221b) The positive rotation direction setting unit 221b sets the positive rotation direction based on, for example, the correct position and constraints regarding the positive rotation direction.

[0235] In more detail, for example, the positive rotation direction setting unit 221b sets the positive rotation direction to the direction in which the horizontal direction faces the vertical direction at the smallest rotation angle.

[0236] (Regarding the Correction Unit 221c) The correction unit 221c corrects the unconstrained rotation angle included in the original data to a constrained ellipse rotation angle using the set reference direction and rotation direction.

[0237] The correction unit 221c applies the constraint of the correct ellipse angle used for training the estimation model using the set reference direction and rotation direction, and corrects the unconstrained rotation angle to a constrained ellipse rotation angle. In detail, for example, the correction unit 221c corrects the unconstrained rotation angle so that it complies with the constraint of the reference ellipse axis, for example, with respect to the reference ellipse axis, reference direction, angle range, and positive rotation direction applied to the unconstrained rotation angle that differ from the constraint of the reference ellipse axis.

[0238] A detailed example will now be described using the constraints illustrated in FIG. 11 . The constraint on the reference ellipse axis is the minor axis. The constraint on the angle range is -90 degrees≦ellipse rotation angle θ≦+90 degrees. The constraint on the reference direction is the horizontal direction of the eye from the outer corner C of the eye to the inner corner A of the eye. The constraint on the positive rotation direction is the horizontal direction toward the center B of the upper eyelid. The vertical direction of the eye is the direction from the center D of the lower eyelid to the center B of the upper eyelid.

[0239] Furthermore, the reference ellipse axis, reference direction, angle range, and positive rotation direction applied to the unconstrained rotation angle are respectively the minor axis, the right direction in the iris image, 0 degrees to 360 degrees, and the counterclockwise direction in the iris image.

[0240] In this example, the constraints on the angle range, reference direction, and positive rotation direction are different from those applied to the unconstrained rotation angle. The correction unit 221c may correct the unconstrained rotation angle, for example, so that the reference direction and positive rotation direction applied to the unconstrained rotation angle become the set reference direction and positive rotation direction. The correction unit 221c may then correct the unconstrained rotation angle, for example, after correcting the reference direction and positive rotation direction, so that the angle range is -90 degrees ≦ ellipse rotation angle θ ≦ +90 degrees. This allows the correction unit 221c to correct the unconstrained rotation angle to a constrained ellipse rotation angle.

[0241] Note that if the ellipse reference axes are different, the unconstrained rotation angles may be modified, for example, first so that the ellipse reference axes conform to the constraints. The method of modifying the unconstrained rotation angles to the constrained ellipse rotation angles is not limited to the example given here.

[0242] (Regarding the supervised data generating unit 221d) The supervised data generating unit 221d generates supervised data including a constrained ellipse rotation angle. This supervised data includes, for example, the constrained ellipse rotation angle in place of the unconstrained rotation angle and the supervised data included in the original data other than the unconstrained rotation angle.

[0243] In detail, for example, the correct data generating unit 221d may generate correct data by replacing the unconstrained rotation angle included in the original data with the constrained ellipse rotation angle. For example, if the original data includes the reference ellipse axis, reference direction, angle range, and positive rotation direction applied to the unconstrained rotation angle, the correct data generating unit 221d may generate correct data excluding these.

[0244] Next, detailed examples of the loss function calculation unit 222, the gradient calculation unit 223, and the parameter update unit 224 will be described.

[0245] (Regarding the loss function calculation unit 222) The loss function calculation unit 222 calculates a loss using, for example, estimated ellipse parameters and estimated feature positions, and correct ellipse parameters and correct positions, for a training iris image. The correct data may include the correct ellipse parameters and the correct feature point positions.

[0246] The loss may be, for example, a value corresponding to the difference between the estimated ellipse parameters and estimated feature positions and the correct ellipse parameters and correct position. The value corresponding to the difference may be, for example, a value corresponding to the distance between the estimated ellipse parameters and estimated feature positions and the correct ellipse parameters and correct position. The distance may be, for example, an L1 norm, an L2 norm, or the like, but is not limited to these. In more detail, the loss may be, for example, an L1 mean, an L2 mean, or the like, but is not limited to these.

[0247] When mini-batch learning is performed, the loss function calculation unit 222 may calculate a loss for each of the multiple training iris images that make up the mini-batch using the estimated ellipse parameters and estimated feature positions, as well as the correct ellipse parameters and correct position. Then, the loss function calculation unit 222 may calculate the average value of the losses calculated for each of the multiple training iris images that make up the mini-batch as the loss for that mini-batch.

[0248] (Regarding the gradient calculation unit 223) The gradient calculation unit 223 calculates the gradient for each model parameter by backpropagation using, for example, the loss calculated by the loss function calculation unit 222. Note that a method other than backpropagation may be used to calculate the gradient.

[0249] (Parameter Updater 224) The parameter updater 224 may update each of the model parameters using, for example, the gradient calculated by the gradient calculator 223 and a predetermined learning rate.

[0250] For example, the parameter update unit 224 may update each of the model parameters included in the estimation model using a value obtained by multiplying the gradient of each model parameter by a learning rate. As described above, the learning rate may be a predetermined value, or may be changed from an initial value depending on the learning stage.

[0251] As described above, the estimation model can be learned by executing the estimation process (step S210) performed by the learning estimation unit 210 and the update process (step S220) performed by the update unit 220.

[0252] When mini-batch learning is used, steps S210 to S220 may be repeated for each mini-batch.

[0253] The condition for terminating the learning of the estimation model (termination condition) may be, for example, repeating steps S210 to S220 a predetermined number of times. The termination condition may be, for example, that the change in the loss value is equal to or less than a predetermined value. The termination condition may be, for example, that steps S210 to S220 are repeated for all training iris images stored in the training data storage device 90. Note that the termination conditions are not limited to those exemplified here. Furthermore, multiple termination conditions may be combined.

[0254] This allows the estimation model to be trained simultaneously (end-to-end) to construct a trained estimation model that estimates feature positions and ellipse parameters. Furthermore, since the estimation model can be trained using a constrained correct ellipse rotation angle, it is possible to construct an estimation model that accurately estimates the ellipse rotation angle θ of the iris ellipse. Therefore, it becomes possible to construct an estimation model that can estimate the contour or edge of the iris included in an iris image with even greater accuracy. Then, using the trained estimation model, it becomes possible to estimate the contour or edge of the iris included in an iris image with even greater accuracy.

[0255] (Operations and Effects) As described above, according to this embodiment, the training estimation unit 210 estimates feature positions related to training iris images using an estimation model. The reference direction is set using the estimated feature positions.

[0256] This allows a reference direction based on the feature position to be set, and the ellipse rotation angle θ can be expressed using an appropriate reference direction in the training iris image. This makes it possible to construct an estimation model that uses an ellipse to more accurately estimate the contour or edge of the iris contained in the iris image. This makes it possible to more accurately estimate the contour or edge of the iris contained in the iris image.

[0257] According to this embodiment, the characteristic positions include positions corresponding to at least one of the upper eyelid and the lower eyelid, which are predetermined as characteristic parts of the eye. The ellipse rotation angle θ is expressed as an angle within an angle range with respect to a reference direction, with the rotation directions in which the horizontal direction faces the upper eyelid and the lower eyelid, respectively, being positive and negative directions.

[0258] This allows the positive and negative directions of the ellipse rotation angle θ to be set from the training iris image, making it possible to represent the ellipse rotation angle θ using an appropriate reference direction in the training iris image. This makes it possible to construct an estimation model that uses an ellipse to more accurately estimate the contour or edge of the iris contained in the iris image. This makes it possible to more accurately estimate the contour or edge of the iris contained in the iris image.

[0259] According to this embodiment, the update unit 220 updates the model parameters included in the estimation model using estimated feature positions, which are estimated feature positions, and correct positions, which are correct answers for the feature positions related to the training iris images. The correct answer data includes the correct answers for the feature positions related to the training iris images.

[0260] This allows the use of a common training iris image to train an estimation model for estimating ellipse parameters and feature positions, thereby enabling the estimation model to be constructed efficiently.

[0261] Fourth Embodiment In this embodiment, an example will be described in which the reference direction is set using the eye center. Note that, for the sake of brevity, descriptions that overlap with other embodiments will be omitted as appropriate.

[0262] In this embodiment, an example will be described in which the iris image and the training iris image include the eye. Also, an example will be described in which the characteristic part of the eye includes the eye center. Therefore, the characteristic position and the correct position according to this embodiment further include the eye center position and the correct eye center position, respectively.

[0263] The eye center is the center of the eye. The eye center may be, for example, the intersection of a line connecting the outer corner and inner corner of the eye with a line connecting the centers of the upper and lower eyelids. The eye center may be, for example, the intersection of a line connecting the outer corner and inner corner of the eye with a line connecting the upper and lower ends of the upper and lower eyelids. The eye center may be, for example, the center of gravity of the eye area surrounded by the upper and lower eyelids. Note that the eye center is not limited to the examples given here and may be determined as appropriate.

[0264] The eye center position is the position of the eye center. The correct eye center position is the correct eye center position of the eye included in the training iris image.

[0265] 25 , the information processing device 300 includes a learning estimation unit 310 and an update unit 320. The information processing device 300 may be included in an information processing system S3 instead of the information processing device 200.

[0266] The information processing device 300 may be provided in the information processing system S1 instead of the information processing device 100. The training data storage device 90 may store correct answer data including a constrained correct ellipse rotation angle, in which case the constraint unit 321 described later may not be provided.

[0267] The training estimation unit 310 estimates ellipse parameters and feature positions for the training iris image using an estimation model for estimating ellipse parameters related to the contour or edge of the iris included in the iris image and feature positions of the eye included in the iris image.

[0268] The update unit 320 updates the model parameters included in the estimation model using the estimated ellipse parameters, the correct ellipse parameters, the estimated feature positions, and the correct positions.

[0269] (Example of Processing Operation of Information Processing Device 300) The information processing device 300 executes information processing as shown in FIG.

[0270] The training estimation unit 310 estimates ellipse parameters and feature positions for the training iris image using an estimation model for estimating ellipse parameters related to the contour or edge of the iris included in the iris image and feature positions of the eye included in the iris image (step S310).

[0271] The update unit 320 updates the model parameters included in the estimation model using the estimated ellipse parameters, the correct ellipse parameters, the estimated feature positions, and the correct positions (step S320).

[0272] (Regarding the training estimation unit 310) The training estimation unit 310 estimates ellipse parameters and feature positions related to a training iris image using an estimation model currently being trained. The feature positions according to this embodiment include the eye center position. The estimation model currently being trained is a machine learning model for estimating ellipse parameters related to the contour or edge of the iris included in the iris image and feature positions of the eye included in the iris image. The estimation model includes one or more model parameters and is configured using, for example, a neural network using a convolutional layer, an attention mechanism, etc.

[0273] As shown in FIG. 27 , the learning estimation unit 310 includes a first estimation unit 311 and a first output unit 312 .

[0274] The first estimation unit 311 estimates ellipse parameters, feature positions, and gaze direction for the training iris image using an estimation model for estimating ellipse parameters for the contour or edge of the iris included in the iris image and feature positions of the eye included in the iris image.

[0275] The first output unit 312 outputs the estimated ellipse parameters and feature positions (i.e., estimated ellipse parameters and estimated feature positions).

[0276] The learning estimation unit 310 executes an estimation process (step S310) as shown in FIG. 28, for example.

[0277] The first estimation unit 311 estimates ellipse parameters and feature positions for the training iris image using an estimation model for estimating ellipse parameters related to the contour or edge of the iris included in the iris image and feature positions of the eye included in the iris image (step S311).

[0278] The first output unit 312 outputs the estimated ellipse parameters, feature positions, and gaze direction (that is, the estimated ellipse parameters, the estimated feature positions, and the estimated gaze direction) (step S312).

[0279] (First Estimation Unit 311) The first estimation unit 311 acquires training iris images in response to, for example, a user instruction. For example, the first estimation unit 311 acquires training iris images from the training data storage device 92. Note that the first estimation unit 311 may further acquire data included in the training data, such as the original data.

[0280] For example, the first estimation unit 311 inputs the acquired training iris image to the estimation model. In response to this input, the estimation model estimates ellipse parameters and feature positions related to the input training iris image.

[0281] The characteristic positions include the center position of the eye. The characteristic positions may further include at least one of the left and right ends of the eye, the upper and lower ends of the eye, a region related to the upper eyelid, and a region related to the lower eyelid.

[0282] (First Output Unit 312) The first output unit 312 outputs output data including, for example, estimated ellipse parameters and estimated feature positions. The estimated feature positions include the estimated eye center position.

[0283] (When Multiple Training Iris Images Are Used) When training the estimation model using multiple training iris images, the training estimation unit 310, for example, acquires multiple training iris images. In this case, the training estimation unit 310 may, for example, input each of the acquired training iris images to the estimation model. For example, the training estimation unit 310 may input each of the training iris images sequentially to the estimation model. In response to this input, the estimation model outputs ellipse parameters and feature positions for each of the input training iris images. In this way, the training estimation unit 310 may estimate ellipse parameters for each of the training iris images.

[0284] (Regarding the updating unit 320) The updating unit 320 generates correct answer data including a correct answer for the constrained ellipse rotation angle based on, for example, the original data. The updating unit 320 updates the model parameters included in the estimation model using, for example, the estimated ellipse parameters and estimated feature positions included in the output data and the correct answer data.

[0285] In detail, for example, as shown in FIG. 29, the update unit 320 includes a constraint unit 321, a loss function calculation unit 222, a gradient calculation unit 223, and a parameter update unit 224 similar to those in the third embodiment.

[0286] The constraint unit 321 generates correct data including the constrained correct ellipse rotation angle based on the original data. Note that the training data may include the correct data instead of the original data. In this case, the update unit 320 does not need to include the constraint unit 321.

[0287] The update unit 320 executes an update process (step S320) as shown in FIG. 30, for example.

[0288] The constraint unit 321 generates correct data including the constrained correct ellipse rotation angle based on the original data (step S321).

[0289] Steps S222 to S224 are executed in the same manner as in the third embodiment.

[0290] (Regarding the Constraint Unit 321) The constraint unit 321 acquires a constrained correct ellipse rotation angle by applying a predetermined constraint to an unconstrained rotation angle included in the original data. The constraint unit 321 according to this embodiment acquires a correct ellipse rotation angle constrained using the eye center position. Then, the constraint unit 321 generates correct data including the acquired correct ellipse rotation angle.

[0291] As shown in FIG. 31, the constraint unit 321 includes a reference direction setting unit 321a, a rotational normal direction setting unit 221b, a correction unit 221c, and a correct data generation unit 221d similar to those in the third embodiment.

[0292] The reference direction setting unit 321a sets the reference direction using the correct position included in the original data.

[0293] The constraint unit 321 executes a constraint process (step S221) as shown in FIG. 32, for example.

[0294] The reference direction setting unit 221a sets the reference direction using the correct position included in the original data (step S221a).

[0295] Steps S221b to S221d similar to those in the third embodiment are executed.

[0296] (Regarding the Reference Direction Setting Unit 321a) The reference direction setting unit 221a sets a reference direction based on, for example, a correct position and a constraint using the eye center. An example in which the eye center is used as a constraint on the reference direction will be described.

[0297] (Example 4 of Constraint on Reference Direction) The reference direction may be, for example, a direction from the center of the eye to a predetermined specific point among the characteristic points. The specific point may be, for example, the characteristic point closest to the center of the eye.

[0298] In detail, for example, the characteristic locations include the left and right ends of the eye (points A and C in FIG. 9), as shown in FIG. 9. The specific locations may be, for example, the left and right ends of the eye.

[0299] In this case, the reference direction setting unit 221a may, for example, identify a feature position closest to the center position of the eye from among the left and right ends of the eye. Then, the reference direction setting unit 221a may, for example, set the direction from the center position of the eye toward the identified feature position as the reference direction. The reference direction set in this case is the direction from the center of the eye included in the iris image toward the feature position closest to the center position of the eye from among the left and right ends of the eye.

[0300] Note that if the constraint on the angle range does not include the reference direction in the angle range (for example, if the reference direction is included and the size of the angle range is 180 degrees), the reference direction setting unit 221a may set the angle range. In detail, for example, if the right end of the eye (point A in FIG. 9) is identified, the reference direction setting unit 221a may set the angle range from −90 degrees to +90 degrees. Also, for example, if the left end of the eye (point C in FIG. 9) is identified, the reference direction setting unit 221a may set the angle range of the ellipse rotation angle θ to +90 degrees.

[0301] (Method 5 for Setting Reference Direction) The reference direction may be, for example, a direction from the center of the eye to a predetermined specific point among characteristic points in the normalized image.

[0302] The normalized image is an image in which the iris included in the iris image is deformed into a predetermined shape. The predetermined shape may be, for example, a shape in which the normalized points are at the same reference distance RL from the center of the eye. During normalization, the iris image may be deformed, for example, so that the orientation of each normalized point relative to the center of the eye remains the same as before normalization.

[0303] The normalized portions may be all of the characteristic portions whose positions are estimated by the estimation model, or may be only a portion of the characteristic portions whose positions are estimated by the fixed model. Fig. 33 is a diagram showing an example of a normalized image. In this figure, when the normalized portions are points A to D (see Fig. 9), an example of an iris image is shown that has been deformed so that points A to D are equidistant from the center of the eye.

[0304] The reference distance RL may be, for example, the distance between the center of the eye in an iris image in which the gaze direction is close to the imaging direction (the deviation between the gaze direction and the imaging direction is within a predetermined range) and the reference position. The reference position may be, for example, the upper end of the eye (point B in FIG. 33 ), but is not limited to this. The reference position may be appropriately determined in advance from among characteristic points whose positions are estimated by the estimation model. The reference distance RL may also be a predetermined value.

[0305] The specified location may be, for example, a characteristic location closest to the center of the eye. In this case, the reference direction setting unit 221a may specify the characteristic location closest to the center of the eye in the normalized image. Then, the reference direction setting unit 221a may set the direction from the center of the eye toward the specified characteristic location in the normalized image as the reference direction.

[0306] As described above, the estimation model can be learned by executing the estimation process (step S320) performed by the learning estimation unit 310 and the update process (step S220) performed by the update unit 220.

[0307] The processing when using a plurality of learning iris images and the termination conditions may be substantially the same as those in the third embodiment, for example.

[0308] As a result, similar to the third embodiment, learning for estimating ellipse parameters and feature positions in the estimation model can be performed simultaneously (end-to-end), allowing a trained estimation model to be constructed. By using the estimated feature positions, a model capable of estimating the contour or edge of the iris contained in an iris image with even greater accuracy can be constructed. Then, using the trained iris estimation model, it becomes possible to estimate the contour or edge of the iris contained in an iris image with even greater accuracy.

[0309] (Actions and Effects) As described above, according to this embodiment, the iris image includes an eye. The estimation model further estimates the feature position of the eye included in the iris image. The reference direction is set based on the center position of the eye included in the iris image and the estimated feature position.

[0310] This allows a reference direction based on the feature position to be set, and the ellipse rotation angle θ can be expressed using an appropriate reference direction in the training iris image. This makes it possible to construct an estimation model that uses an ellipse to more accurately estimate the contour or edge of the iris contained in the iris image. This makes it possible to more accurately estimate the contour or edge of the iris contained in the iris image.

[0311] According to this embodiment, the reference direction is a direction from the eye center position toward the feature position closest to the eye center position.

[0312] This allows the ellipse rotation angle θ to be expressed using an appropriate reference direction in the training iris image by setting a reference direction based on the feature position closest to the eye center position among the multiple feature positions. This makes it possible to construct an estimation model that uses an ellipse to more accurately estimate the contour or edge of the iris contained in the iris image. This makes it possible to more accurately estimate the contour or edge of the iris contained in the iris image.

[0313] According to this embodiment, the feature positions include the left and right edges of the eye, and the reference direction is a direction from the center position of the eye included in the iris image toward the feature position of either the left or right edge of the eye that is closest to the center position of the eye.

[0314] This allows a reference direction corresponding to the lateral direction of the eye to be set based on the feature position, and therefore the ellipse rotation angle θ can be expressed using an appropriate reference direction in the training iris image. This makes it possible to construct an estimation model that uses an ellipse to more accurately estimate the contour or edge of the iris contained in the iris image. This makes it possible to more accurately estimate the contour or edge of the iris contained in the iris image.

[0315] [Embodiment 5] In embodiment 5, an example will be described in which the reference direction is restricted to the gaze direction of an eye included in an iris image. Note that, for the sake of brevity, descriptions that overlap with other embodiments will be omitted as appropriate.

[0316] In this embodiment, an example will be described in which the iris image and the training iris image include eyes, and the correct answer data further includes a correct gaze direction that is the correct answer for the gaze direction of the eye included in the training iris image.

[0317] (Configuration example of information processing device 400) As shown in Fig. 34 , the information processing device 400 includes a learning estimation unit 410 and an update unit 420. The information processing device 400 may be included in an information processing system S3 instead of the information processing device 200.

[0318] The information processing device 400 may be provided in the information processing system S1 instead of the information processing device 100. The training data storage device 90 may store correct answer data including a constrained correct ellipse rotation angle, in which case the constraint unit 421 described later may not be provided.

[0319] The training estimation unit 410 estimates ellipse parameters, feature positions, and gaze direction for the training iris image using an estimation model for estimating ellipse parameters for the contour or edge of the iris included in the iris image and feature positions and gaze direction of the eye included in the iris image.

[0320] The update unit 420 updates the model parameters included in the estimation model using the estimated ellipse parameters, the correct ellipse parameters, the estimated feature position, the correct position, the estimated gaze direction that is the estimated gaze direction, and the correct gaze direction that is the correct gaze direction for the training iris image.

[0321] (Example of Processing Operation of Information Processing Device 400) The information processing device 400 executes information processing as shown in FIG.

[0322] The training estimation unit 410 estimates the ellipse parameters, feature positions, and gaze direction for the training iris image using an estimation model for estimating the ellipse parameters for the contour or edge of the iris included in the iris image and the feature positions and gaze direction of the eye included in the iris image (step S410).

[0323] The update unit 420 updates the model parameters included in the estimation model using the estimated ellipse parameters, the correct ellipse parameters, the estimated feature position, the correct position, the estimated gaze direction, and the correct gaze direction, which is the correct gaze direction for the training iris image (step S420).

[0324] (Regarding the training estimation unit 410) The training estimation unit 410 estimates ellipse parameters, feature positions, and gaze directions for training iris images, for example, using an estimation model currently being trained. The estimation model currently being trained is a machine learning model for estimating ellipse parameters for the contour or edge of the iris included in the iris image, feature positions of the eye included in the iris image, and the gaze direction of the eye included in the iris image. The estimation model includes one or more model parameters and is configured, for example, using a neural network that uses a convolutional layer, an attention mechanism, etc.

[0325] The estimation model may be composed of, for example, one or more machine learning models. The machine learning model may include, for example, an output layer, a hidden layer, and another output layer. When composed of multiple machine learning models, the estimation model may include, for example, a common model including an input layer, an ellipse estimation model including an output layer for ellipse parameters, a feature position estimation model including an output layer for feature positions, and a gaze direction estimation model including an output layer for gaze direction. Each of the ellipse estimation model, the feature position estimation model, and the gaze direction estimation model may perform calculations by acquiring outputs (e.g., intermediate features) from the common model. In this case, the ellipse parameters are estimated by the common model and the ellipse estimation model, the feature positions are estimated by the common model and the feature position estimation model, and the gaze direction is estimated by the common model and the gaze direction estimation model. Note that configuration examples of the estimation models are not limited to these. As described above, the training estimation unit 410 estimates ellipse parameters and feature positions related to the training iris image using the estimation models.

[0326] The learning estimation unit 410 includes a first estimation unit 411 and a first output unit 412, as shown in FIG. 36, for example.

[0327] The first estimation unit 411 further uses an estimation model for estimating ellipse parameters relating to the contour or edge of the iris included in the iris image, and the feature position and gaze direction of the eye included in the iris image, to estimate the ellipse parameters, feature position and gaze direction relating to the training iris image.

[0328] The first output unit 412 outputs the estimated ellipse parameters, feature positions, and gaze direction (i.e., estimated ellipse parameters, estimated feature positions, and estimated gaze direction).

[0329] The learning estimation unit 410 executes an estimation process (step S410) as shown in FIG. 37, for example.

[0330] The first estimation unit 411 further uses a gaze estimation model for estimating ellipse parameters related to the contour or edge of the iris included in the iris image and the feature position and gaze direction of the eye included in the iris image to estimate ellipse parameters, feature position, and gaze direction related to the training iris image (step S411).

[0331] The first output unit 412 outputs the estimated ellipse parameters, feature positions, and gaze direction (that is, the estimated ellipse parameters, the estimated feature positions, and the estimated gaze direction) (step S412).

[0332] (First Estimation Unit 411) The first estimation unit 411 acquires training iris images in response to, for example, a user instruction. For example, the first estimation unit 411 acquires training iris images from the training data storage device 92. Note that the first estimation unit 211 may further acquire training data, original data, etc.

[0333] For example, the first estimation unit 311 inputs the acquired training iris image to the estimation model. In response to this input, the estimation model estimates ellipse parameters, feature positions, and gaze directions related to the input training iris image.

[0334] (Regarding the first output unit 312) The first output unit 312 outputs output data including, for example, estimated ellipse parameters, estimated feature positions, and estimated gaze directions. The estimated gaze directions are gaze directions estimated by the estimation model in response to input of an iris image. The estimated ellipse parameters, estimated feature positions, and estimated gaze directions may be represented by vectors.

[0335] (When Multiple Training Iris Images Are Used) When training the estimation model using multiple training iris images, the training estimation unit 410 may, for example, acquire multiple training iris images. In this case, the training estimation unit 410 may, for example, input each of the acquired training iris images to the estimation model. For example, the training estimation unit 410 may input each of the training iris images sequentially to the estimation model. In response to this input, the estimation model outputs ellipse parameters and feature positions for each of the input training iris images. In this way, the training estimation unit 410 may estimate ellipse parameters for each of the training iris images.

[0336] (Regarding the updating unit 420) The updating unit 420 generates correct answer data including a correct answer for the constrained ellipse rotation angle based on, for example, the original data. The updating unit 420 updates the model parameters included in the estimation model and the gaze estimation model using, for example, the estimated ellipse parameters, estimated feature positions, and estimated gaze directions included in the output data, and the correct answer data.

[0337] As shown in FIG. 38 , the update unit 420 includes a constraint unit 421 , a loss function calculation unit 422 , a gradient calculation unit 423 , and a parameter update unit 424 .

[0338] The constraint unit 421 generates correct data including a constrained correct ellipse rotation angle based on the original data.

[0339] The loss function calculation unit 422 calculates the loss using the ellipse parameters, feature positions, and gaze directions estimated by the training estimation unit 410, and ground truth data related to the training iris image.

[0340] The gradient calculation unit 423 uses the loss calculated by the loss function calculation unit 422 to calculate the gradient for each of the model parameters included in the estimation model.

[0341] The parameter update unit 424 uses the gradient calculated by the gradient calculation unit 423 to update each of the model parameters included in the estimation model.

[0342] The update unit 420 executes an update process (step S420) as shown in FIG. 39, for example.

[0343] The constraint unit 421 generates correct data including the constrained correct ellipse rotation angle based on the original data (step S421).

[0344] The loss function calculation unit 422 calculates the loss using the ellipse parameters and gaze direction estimated by the training estimation unit 410 and the ground truth data related to the training iris image (step S422).

[0345] The gradient calculation unit 423 calculates the gradient for each of the model parameters included in the estimation model using the loss calculated by the loss function calculation unit 422 (step S423).

[0346] The parameter update unit 424 updates each of the model parameters included in the estimation model using the gradient calculated by the gradient calculation unit 423 (step S424).

[0347] First, a detailed example of the constraint unit 421 will be described.

[0348] (Regarding the Constraint Unit 421) The constraint unit 421 acquires a constrained correct ellipse rotation angle by applying a predetermined constraint to an unconstrained rotation angle included in the original data. The constraint unit 421 according to this embodiment acquires a correct ellipse rotation angle constrained using the line of sight. Then, the constraint unit 321 generates correct data including the acquired correct ellipse rotation angle.

[0349] As shown in FIG. 40, the constraint unit 421 includes a reference direction setting unit 421a, a rotational normal direction setting unit 221b, a correction unit 221c, and a correct data generation unit 221d similar to those in the third embodiment.

[0350] The reference direction setting unit 321a sets the reference direction using the correct position included in the original data.

[0351] The constraint unit 421 executes a constraint process (step S421) as shown in FIG. 41, for example.

[0352] The reference direction setting unit 421a sets the reference direction using the correct position included in the original data (step S421a).

[0353] Steps S221b to S221d similar to those in the third embodiment are executed.

[0354] (Reference direction setting unit 421a) The reference direction setting unit 421a sets a reference direction based on, for example, a correct gaze direction and a constraint using the gaze direction. An example in which the gaze direction is used as a constraint on the reference direction will be described.

[0355] (Example 6 of Constraint on Reference Direction) The reference direction may be, for example, a line of sight direction. The reference direction setting unit 421a may set the correct line of sight direction as the reference direction, for example.

[0356] Below, detailed examples of the loss function calculation unit 422, the gradient calculation unit 423, and the parameter update unit 424 will be described.

[0357] (Regarding the loss function calculation unit 422) The loss function calculation unit 422 calculates a loss using, for example, the ellipse parameters, the feature position, and the gaze direction estimated by the training estimation unit 410 for the training iris image, and the correct data. The correct data may include the correct ellipse parameters, the correct position, and the correct gaze direction.

[0358] The loss may be, for example, a value corresponding to a difference between the estimated ellipse parameters, estimated feature position, and estimated gaze direction and the correct ellipse parameters, correct position, and correct gaze direction. The value corresponding to the difference may be, for example, a value corresponding to a distance between the estimated ellipse parameters, estimated feature position, and estimated gaze direction and the correct ellipse parameters, correct position, and correct gaze direction. The distance may be, for example, an L1 norm, an L2 norm, or the like, but is not limited to these. In more detail, the loss may be, for example, an L1 mean, an L2 mean, or the like, but is not limited to these.

[0359] When mini-batch learning is performed, the loss function calculation unit 422 may calculate a loss for each of the multiple training iris images constituting the mini-batch using the estimated ellipse parameters, estimated feature position, and estimated gaze direction, as well as the correct ellipse parameters, correct position, and correct gaze direction. Then, the loss function calculation unit 422 may calculate the average value of the losses calculated for each of the multiple training iris images constituting the mini-batch as the loss for that mini-batch.

[0360] (Regarding the gradient calculation unit 423) The gradient calculation unit 423 calculates the gradient for each model parameter by backpropagation using, for example, the loss calculated by the loss function calculation unit 422. Note that a method other than backpropagation may be used to calculate the gradient.

[0361] (Parameter Updater 424) The parameter updater 424 updates each of the model parameters using the gradient calculated by the gradient calculator 423 and a predetermined learning rate.

[0362] For example, the parameter update unit 424 may update each of the model parameters included in the estimation model using a value obtained by multiplying the gradient of each model parameter by a learning rate. As described above, the learning rate may be a predetermined value, or may be changed from an initial value depending on the learning stage.

[0363] As described above, the estimation model can be learned by executing the estimation process (step S410) performed by the learning estimation unit 410 and the update process (step S420) performed by the update unit 420.

[0364] If multiple training iris images are used, steps S410 to S420 may be repeated for each training iris image. If mini-batch training is used, steps S410 to S420 may be repeated for each mini-batch.

[0365] The condition for terminating the learning of the estimation model (termination condition) may be, for example, repeating steps S410 to S420 a predetermined number of times. The termination condition may be, for example, that the change in the loss value is equal to or less than a predetermined value. The termination condition may be, for example, that steps S410 to S420 are repeated for all training iris images stored in the training data storage device 90. Note that the termination conditions are not limited to those exemplified here. Furthermore, multiple termination conditions may be combined.

[0366] This allows the estimation models to be learned simultaneously (end-to-end) to construct a trained estimation model. The gaze direction often affects the ellipticity of the ellipse parameters, the ellipse rotation angle θ, and the like. Therefore, by using the estimation results of the gaze estimation model, it is possible to construct an estimation model that can estimate the contour or edge of the iris contained in an iris image with even greater accuracy. Then, using the trained estimation model, it becomes possible to estimate the contour or edge of the iris contained in an iris image with even greater accuracy.

[0367] As described above, according to this embodiment, the training estimation unit 410 estimates the gaze direction related to the training iris image by further using a gaze estimation model for estimating the gaze direction of the eye included in the iris image. The reference direction is the estimated gaze direction.

[0368] This allows the gaze direction to be set as the reference direction. Generally, the gaze direction often affects the ellipsoid rotation angle θ of the iris. Therefore, by setting the gaze direction as the reference direction, the ellipsoid rotation angle θ, etc. can be expressed using an appropriate reference direction in the training iris image. This makes it possible to construct an estimation model that uses an ellipse to more accurately estimate the contour or edge of the iris included in the iris image. Therefore, it becomes possible to more accurately estimate the contour or edge of the iris included in the iris image.

[0369] According to this embodiment, the update unit 420 updates the model parameters included in the gaze estimation model using the estimated gaze direction and the correct answer data related to the training iris image. The correct answer data includes the correct answer of the gaze direction related to the training iris image.

[0370] This allows the estimation model to be learned end-to-end simultaneously, and allows the estimation model to be constructed to estimate the ellipse parameters, feature positions, and gaze direction. This makes it possible to construct the estimation model efficiently.

[0371] [Embodiment 6] In embodiment 6, an example of performing iris authentication using ellipse parameters will be described as an example of using a trained estimation model. Note that, for the sake of brevity, descriptions that overlap with other embodiments will be omitted as appropriate.

[0372] (Configuration Example of Information Processing System S5) The information processing system S5 includes, for example, a photographing device 95, the training data storage device 90 described above, and an information processing device 500, as shown in FIG.

[0373] The photographing device 95 photographs the subject to be authenticated and acquires an iris image. The photographing device 95 is, for example, a camera.

[0374] The photographing device 95 photographs an object to be authenticated, for example, to obtain an iris image including the iris of the object to be authenticated. The photographing device 95 may also be used to obtain a learning iris image.

[0375] 43 , the information processing device 500 includes the above-described learning estimation unit 110, update unit 120, estimation unit 530, and authentication unit 540. Note that the information processing device 500 does not necessarily have to include the learning estimation unit 110 and the update unit 120.

[0376] The estimation unit 530 uses the trained estimation model to estimate ellipse parameters for an iris image including the iris of the authentication target.

[0377] The authentication unit 540 authenticates the authentication target using the ellipse parameters estimated for the authentication target.

[0378] (Example of Processing Operation of Information Processing Device 500) The information processing device 500 further executes information processing as shown in FIG.

[0379] The estimation unit 530 uses the trained estimation model to estimate ellipse parameters for the iris image including the iris of the authentication target (step S530).

[0380] The authentication unit 540 authenticates the authentication target using the ellipse parameters estimated for the authentication target (step S540).

[0381] (Regarding the Estimation Unit 530) The estimation unit 530 includes, for example, a trained estimation model. The estimation unit 530 acquires an authentication iris image captured in response to, for example, a user instruction, from the imaging device 95. The authentication iris image is an iris image that includes the iris of the authentication target.

[0382] For example, the estimation unit 530 inputs the acquired authentication iris image into the trained estimation model. In response to this input, the trained estimation model outputs ellipse parameters related to the input authentication iris image. In this way, the estimation unit 530 estimates ellipse parameters related to the authentication iris image.

[0383] The function of the estimation unit 530 corresponds to, for example, the function of the learning estimation units 110, 210, 310, and 410 after learning an estimation model to use an authentication iris image as input to the estimation model instead of training data (learning iris image). The estimation unit 530 may acquire the authentication iris image from a storage unit (not shown) that stores the authentication iris image instead of or in addition to the imaging device 95.

[0384] Furthermore, the model parameters of the trained estimation model may be stored in the training estimation unit 110, 210, 310, or 410 together with the estimation model, or may be stored in a storage unit (not shown) for storing model parameters. This storage unit may be included in the information processing devices 100, 200, 300, or 400, or may be included in an external device (not shown). In this case, the training estimation unit 110, 210, 310, or 410 may estimate ellipse parameters using the model parameters stored in the storage unit. Furthermore, the update unit 120, 220, 320, or 420 may update the model parameters stored in the storage unit. The estimation unit 530 may estimate ellipse parameters, etc., by using the estimation model with reference to the model parameters stored in the storage unit after training.

[0385] The trained estimation model may be an estimation model that further estimates at least one of the feature position and the gaze direction during training. In this case, the output of the trained estimation model does not need to include at least one of the feature position and the gaze direction.

[0386] (Regarding the authentication unit 540) The authentication unit 540, for example, acquires an iris image for authentication and the ellipse parameters estimated by the estimation unit 530. The authentication unit 540, for example, uses the ellipse parameters estimated by the estimation unit 530 to cut out an iris portion from the iris image for authentication. The authentication unit 540, for example, uses the cut-out iris portion to extract iris feature amounts. The iris feature amounts may be, for example, a feature vector. The authentication unit 540, for example, acquires the feature vector by inputting the cut-out iris portion into an extraction model that has been trained to extract the feature vector from the image of the iris portion.

[0387] The authentication unit 540, for example, compares the acquired feature vector with a pre-registered feature vector and performs authentication based on the comparison result. In detail, for example, the authentication unit 540 may determine that authentication is successful when the similarity (e.g., cosine similarity) between the acquired feature vector and the pre-registered feature vector is equal to or greater than a threshold. The authentication unit 540 may determine that authentication is unsuccessful when the similarity is less than the threshold.

[0388] Note that authentication may be performed using any common technique and is not limited to the method described here. Furthermore, the use of the trained estimation model is not limited to iris authentication.

[0389] (Operations and Effects) As described above, according to this embodiment, the information processing device 500 includes the estimation unit 530 that uses a trained estimation model to estimate ellipse parameters related to an iris image that includes the iris of the authentication target.

[0390] This allows iris authentication and the like to be performed using highly accurate ellipse parameters, thereby enabling highly accurate iris authentication and the like.

[0391] Seventh Embodiment In a seventh embodiment, an example in which a processed learning iris image and an authentication iris image are used will be described. Note that, for the sake of brevity, descriptions that overlap with other embodiments will be omitted as appropriate.

[0392] The information processing system S5 may include an information processing device 600 instead of the information processing device 500.

[0393] (Configuration example of information processing device 600) As shown in Fig. 45 , for example, the information processing device 600 includes a learning estimation unit 610, the above-mentioned update unit 120, an estimation unit 630, the above-mentioned authentication unit 540, and an image processing unit 650. Note that the information processing device 600 may not include the learning estimation unit 610 and the update unit 120, but may include the estimation unit 630, the authentication unit 540, and the image processing unit 650. Furthermore, the information processing device 600 may not include the estimation unit 630 and the authentication unit 540, but may include the learning estimation unit 610, the update unit 120, and the image processing unit 650.

[0394] The image processing unit 650 processes the iris image.

[0395] The training estimation unit 610 estimates ellipse parameters for the training iris image using an estimation model for estimating ellipse parameters for the contour or edge of the iris included in the iris image.

[0396] The estimation unit 630 uses the trained estimation model to estimate ellipse parameters for the authentication iris image.

[0397] (Example of Processing Operation of Information Processing Device 600) The information processing device 600 executes information processing such as that shown in FIG.

[0398] The image processing unit 650 processes the iris image (step S650).

[0399] The training estimation unit 610 estimates ellipse parameters for the training iris image using an estimation model for estimating ellipse parameters for the contour or edge of the iris included in the iris image (step S610).

[0400] Step S120 is executed in the same manner as in the first embodiment.

[0401] The information processing device 600 also executes information processing such as that shown in FIG.

[0402] The image processing unit 650 processes the iris image (step S650).

[0403] The estimation unit 630 uses the trained estimation model to estimate ellipse parameters for the authentication iris image (step S640).

[0404] Step S540 is executed in the same manner as in the sixth embodiment.

[0405] (Image Processing Unit 650) The image processing unit 650 acquires an iris image from the training data storage device 90 or the photographing device 95 in response to, for example, a user instruction, etc. The image processing unit 650 processes the acquired iris image.

[0406] The processing may be, for example, downsampling the iris image to an image of a predetermined size. In this case, the image processing unit 650 downsamples the iris image to an image of a predetermined size. Note that the processing is not limited to downsampling, and may be, for example, cutting out an area including the iris. The image processing unit 650 may acquire the iris image from a storage device (not shown) that stores the iris image, instead of or together with the image capturing device 95.

[0407] (Regarding the training estimation unit 610) The training estimation unit 610 estimates ellipse parameters for a training iris image, for example, by using an estimation model currently being trained and the training iris image. The training iris image is, for example, an iris image processed by the image processing unit 650. In more detail, for example, the training iris image is an iris image downsampled by the image processing unit 650.

[0408] (Regarding the estimation unit 630) The estimation unit 630 estimates ellipse parameters for the authentication iris image using a trained estimation model and the authentication iris image. The authentication iris image is an iris image that includes the iris of the authentication target, and is, for example, an iris image that the image processing unit 650 has processed from an image acquired from the imaging device 95. In more detail, for example, the authentication iris image is an iris image that has been downsampled by the image processing unit 650.

[0409] The function of the estimation unit 630 corresponds to, for example, the function of the training estimation units 110, 210, 310, and 410 after training an estimation model, using authentication iris images instead of training data (training iris images) as input to the estimation model. The model parameters of the trained estimation model may be stored in the training estimation units 110, 210, 310, and 410 together with the estimation model, or may be stored in a storage unit (not shown) for storing model parameters. This storage unit may be included in the information processing devices 100, 200, 300, and 400, or in an external device (not shown). In this case, the training estimation units 110, 210, 310, and 410 may estimate ellipse parameters using the model parameters stored in the storage unit. Furthermore, the update units 120, 220, 320, and 420 may update the model parameters stored in the storage unit. The estimation unit 630 may estimate ellipse parameters, etc., by using the estimation model with reference to the model parameters stored in the storage unit after training.

[0410] The trained estimation model may be an estimation model that further estimates at least one of a feature position and a gaze direction during training. In this case, the output of the trained estimation model does not need to include at least one of the estimated feature position and the gaze direction.

[0411] (Operations and Effects) As described above, according to this embodiment, the information processing device 600 further includes an image processing unit 650 that downsamples the iris image to an image of a predetermined size. Each of the learning iris image and the authentication iris image (iris image including the iris to be authenticated) is a downsampled iris image.

[0412] As a result, the training iris image and the authentication iris image are each downsampled, making the data size smaller than the unprocessed image. This reduces the amount of calculation required for the fixed model. This makes it possible to quickly estimate ellipse parameters. As a result, it becomes possible to quickly train the estimation model, perform iris authentication, and so on.

[0413] Although the present disclosure has been described above with reference to the embodiments, the present disclosure is not limited to the above-described embodiments. Various modifications that can be understood by those skilled in the art can be made to the configuration and details of the present disclosure within the scope of the present disclosure. Furthermore, each embodiment can be combined with other embodiments as appropriate.

[0414] In addition, although the flowcharts used in the above description show a sequence of steps (processes), the order of steps executed in each embodiment is not limited to the sequence shown in the flowcharts. In each embodiment, the order of steps shown in the diagrams can be changed as long as it does not cause any problems in terms of the content.

[0415] Some or all of the above embodiments may be described as, but are not limited to, the following supplementary notes. 1. An information processing device comprising: a training estimation means that estimates ellipse parameters for a training iris image using an estimation model for estimating the ellipse parameters for the contour or edge of the iris included in the iris image, and an update means that updates model parameters included in the estimation model using estimated ellipse parameters that are the estimated ellipse parameters and correct ellipse parameters that are correct solutions to the ellipse parameters for the training iris image. 2. The information processing device described in 1., wherein the ellipse parameters include an ellipse rotation angle that indicates a rotation angle of an ellipse that represents the contour or edge of the iris included in the iris image, and the correct ellipse parameters include a correct ellipse rotation angle that is a correct solution to the ellipse rotation angle, and the ellipse rotation angle is an angle of a predetermined one of the major and minor axes of the ellipse that represents the contour or edge of the iris with respect to a reference direction, and a magnitude including the reference direction is expressed within an angular range of 180 degrees. 3. The information processing device according to 2., wherein the iris image and the training iris image include an eye, and the reference direction is a horizontal direction of the eye or a vertical direction intersecting the horizontal direction. 4. The information processing device according to 2. or 3., wherein the iris image and the training iris image include an eye, and the training estimation means estimates the feature position of the training iris image further using an estimation model for estimating a feature position of the eye included in the iris image, and the reference direction is a direction based on the estimated feature point position. 5. The information processing device according to 4., wherein the feature position includes a position corresponding to at least one of an upper eyelid and a lower eyelid that are predetermined as a characteristic part of the eye, and the ellipse rotation angle is expressed as an angle within an angle range relative to the reference direction, with the rotation directions in which the horizontal direction faces the upper eyelid and the lower eyelid, respectively, being positive and negative directions. 6. 3. The information processing device described in Item 3, wherein the horizontal direction is a direction whose main component is the horizontal direction, and the vertical direction is a direction whose main component is the vertical direction, and the ellipse rotation angle is expressed as an angle within an angle range with respect to the reference direction, with a rotation direction in which the horizontal direction is directed vertically upward being a positive direction and a rotation direction in which the horizontal direction is directed vertically downward being a negative direction.7. The information processing device according to any one of 2. to 6, wherein the angle range is from -90 degrees to +90 degrees. 8. The information processing device according to 4. or 5, wherein the reference direction is a direction based on an eye center position that is the center of the eye included in the iris image and at least one of the left and right edges and the top and bottom edges of the eye. 9. The information processing device according to 8., wherein the reference direction is a direction from the eye center position regarding the iris image to a part of the left and right edges and the top and bottom edges of the eye that is closest to the eye center position. 10. The information processing device according to 9., wherein the reference direction is a direction from the eye center position regarding the iris image to a part of the left and right edges of the eye that is closest to the eye center position. 11. The information processing device according to any one of 4., 5, and 8 to 10, wherein the updating means updates model parameters included in the estimation model by further using the estimated position and a correct position that is a correct answer for the feature position regarding the training iris image. 12. The information processing device described in 2., wherein the training estimation means estimates the gaze direction for the training iris image using the estimation model that further estimates the gaze direction of the eye included in the iris image, and the reference direction is the estimated gaze direction. 13. The information processing device described in 12., wherein the updating means updates model parameters included in the gaze estimation model by further using the estimated gaze direction and a correct gaze direction that is a correct answer for the gaze direction for the training iris image. 14. The information processing device described in any one of 1. to 13., wherein the ellipse parameters include a first ellipse parameter for specifying a first ellipse that represents an outer contour or edge of the iris included in the iris image, and a second ellipse parameter for specifying a second ellipse that represents an inner contour or edge of the iris included in the iris image. 15. The information processing device described in any one of 1. to 14., further comprising estimation means that uses the trained estimation model to estimate the ellipse parameters for a target iris image that is an iris image that includes an iris to be authenticated.16. The information processing device according to any one of 1. to 15., wherein the ellipse parameters include center coordinates, a major axis radius, a minor axis radius, and an ellipse rotation angle, or center coordinates, a major axis radius, an ellipticity, and an ellipse rotation angle. 17. The information processing device according to 15., further comprising image processing means for downsampling an iris image to an image of a predetermined size, wherein each of the training iris image and the target iris image is the downsampled iris image. 18. An information processing method, wherein one or more computers estimate ellipse parameters related to the training iris image using an estimation model for estimating ellipse parameters related to the contour or edge of the iris included in the iris image, and update model parameters included in the estimation model using estimated ellipse parameters that are the estimated ellipse parameters and correct ellipse parameters that are correct solutions to the ellipse parameters related to the training iris image. 19. The information processing method according to 18., wherein the ellipse parameters include an ellipse rotation angle indicating a rotation angle of an ellipse representing the outline or edge of the iris included in the iris image, and the correct ellipse parameters include a correct ellipse rotation angle that is a correct answer to the ellipse rotation angle, and the ellipse rotation angle is an angle of a predetermined one of the major and minor axes of the ellipse representing the outline or edge of the iris with respect to a reference direction, and the magnitude including the reference direction is expressed within an angular range of 180 degrees. 20. The information processing method according to 19., wherein the iris image and the training iris image include an eye, and the reference direction is a horizontal direction of the eye or a vertical direction intersecting with the horizontal direction. 21. The information processing method according to 19. or 20, wherein the iris image and the training iris image include an eye, and the training estimation means estimates the feature positions for the training iris image further using an estimation model for estimating feature positions of the eye included in the iris image, and the reference direction is a direction based on the estimated feature point positions.22. The information processing method described in 21., wherein the characteristic position includes a position corresponding to at least one of the upper eyelid and the lower eyelid, which are predetermined as characteristic parts of the eye, and the ellipse rotation angle is expressed as an angle within an angle range relative to the reference direction, with the rotation direction of the horizontal direction toward the upper eyelid and the lower eyelid, respectively, being positive and negative directions. 23. The information processing method described in 20., wherein the horizontal direction is a direction whose main component is the horizontal direction and the vertical direction is a direction whose main component is the vertical direction, and the ellipse rotation angle is expressed as an angle within an angle range relative to the reference direction, with the rotation direction of the horizontal direction facing vertically upward being the positive direction and the rotation direction of the horizontal direction facing vertically downward being the negative direction. 24. The information processing method described in any one of 19. to 23, wherein the angle range is from -90 degrees to +90 degrees. 25. The information processing method according to 21. or 22., wherein the reference direction is a direction based on an eye center position, which is the center of the eye included in the iris image, and at least one of the left and right edges and the top and bottom edges of the eye. 26. The information processing method according to 25., wherein the reference direction is a direction from the eye center position regarding the iris image to a part of the left and right edges and the top and bottom edges of the eye that is closest to the eye center position. 27. The information processing method according to 26., wherein the reference direction is a direction from the eye center position regarding the iris image to a part of the left and right edges of the eye that is closest to the eye center position. 28. The information processing method according to any one of 21., 22, 25. to 27., wherein updating the model parameters further uses the estimated position and a correct position that is a correct answer to the feature position regarding the training iris image to update model parameters included in the estimation model. 29. 19. The information processing method according to claim 19, wherein estimating the ellipse parameters for the training iris image includes estimating the gaze direction for the training iris image using the estimation model that further estimates the gaze direction of an eye included in the iris image, and the reference direction is the estimated gaze direction.30. The information processing method described in 29., wherein updating the model parameters further uses the estimated gaze direction and a correct gaze direction that is a correct answer for the gaze direction related to the training iris image to update the model parameters included in the gaze estimation model. 31. The information processing method described in any one of 18. to 30., wherein the ellipse parameters include a first ellipse parameter for specifying a first ellipse that represents an outer contour or edge of the iris included in the iris image, and a second ellipse parameter for specifying a second ellipse that represents an inner contour or edge of the iris included in the iris image. 32. The information processing method described in any one of 18. to 31., further comprising estimating the ellipse parameters related to a target iris image that is an iris image that includes an iris to be authenticated, using the trained estimation model. 33. The ellipse parameters include center coordinates, major axis radius, minor axis radius, and ellipticity, and ellipse rotation angle, or center coordinates, major axis radius, ellipticity, and ellipse rotation angle, 18. to 32. 34. The information processing method according to any one of 32., further comprising downsampling an iris image to an image of a predetermined size, and wherein each of the learning iris image and the target iris image is the downsampled iris image. 35. A program for causing one or more computers to execute the information processing method according to any one of 18. to 34. 36. A recording medium having recorded thereon a program for causing one or more computers to execute the information processing method according to any one of 18. to 34.

[0416] S1, S2, S5 Information processing system 90, 92 Training data storage device 95 Imaging device 100, 200, 300, 400, 500, 600 Information processing device 110, 210, 310, 410, 610 Learning estimation unit 120, 220, 320, 420 Update unit 121, 222, 422 Loss function calculation unit 122, 223, 423 Gradient calculation unit 123, 224, 424 Parameter update unit 211, 311, 411 First estimation unit 212, 312, 412 First output unit 221, 321, 421 Constraint unit 221a, 321a, 421a Reference direction setting unit 221b Positive rotation direction setting unit 221c Correction unit 221d Correct data generation unit 530, 630 Estimation unit 540 Authentication unit 650 Image processing unit

Claims

1. An information processing device comprising: a training estimation means for estimating ellipse parameters relating to a training iris image using an estimation model for estimating ellipse parameters relating to the contour or edge of an iris included in the iris image; and an update means for updating model parameters included in the estimation model using estimated ellipse parameters that are the estimated ellipse parameters and correct ellipse parameters that are correct solutions to the ellipse parameters relating to the training iris image.

2. The information processing device of claim 1, wherein the ellipse parameters include an ellipse rotation angle indicating the rotation angle of an ellipse representing the contour or edge of the iris contained in the iris image, the correct ellipse parameters include a correct ellipse rotation angle that is the correct answer to the ellipse rotation angle, and the ellipse rotation angle is the angle of a predetermined one of the major and minor axes of the ellipse representing the contour or edge of the iris relative to a reference direction, and the size including the reference direction is expressed within an angle range of 180 degrees.

3. The information processing device according to claim 2, wherein the iris image and the learning iris image include an eye, and the reference direction is a horizontal direction of the eye or a vertical direction intersecting the horizontal direction.

4. An information processing device as described in claim 2 or 3, wherein the iris image and the training iris image include eyes, the training estimation means further uses an estimation model for estimating feature positions of the eyes included in the iris image to estimate the feature positions for the training iris image, and the reference direction is a direction based on the estimated feature point positions.

5. An information processing device as described in claim 4, wherein the characteristic position includes a position corresponding to at least one of the upper eyelid and the lower eyelid, which are predetermined as characteristic parts of the eye, and the ellipse rotation angle is expressed as an angle within an angle range relative to the reference direction, with the rotation directions in which the horizontal direction is directed toward the upper eyelid and the lower eyelid, respectively, being positive and negative directions.

6. The information processing device of claim 3, wherein the horizontal direction is a direction whose main component is the horizontal direction, and the vertical direction is a direction whose main component is the vertical direction, and the ellipse rotation angle is expressed as an angle within an angle range relative to the reference direction, with the rotation direction of the horizontal direction facing vertically upward being the positive direction and the rotation direction of the horizontal direction facing vertically downward being the negative direction.

7. The information processing device according to any one of claims 2 to 6, wherein the angle range is from -90 degrees to +90 degrees.

8. An information processing device according to claim 4 or 5, wherein the reference direction is a direction based on the eye center position, which is the center of the eye included in the iris image, and at least one of the left and right ends and the top and bottom ends of the eye.

9. The information processing device according to claim 8, wherein the reference direction is a direction from the center position of the eye in the iris image to the part of the left, right, upper or lower end of the eye that is closest to the center position of the eye.

10. An information processing device according to claim 9, wherein the reference direction is a direction from the eye center position in the iris image toward the part of the left or right end of the eye that is closest to the eye center position.

11. An information processing device according to any one of claims 4, 5, 8 to 10, wherein the update means further uses the estimated position and a correct position that is the correct answer for the feature position regarding the training iris image to update the model parameters included in the estimation model.

12. The information processing device of claim 2, wherein the learning estimation means estimates the gaze direction for the learning iris image using the estimation model that further estimates the gaze direction of the eye included in the iris image, and the reference direction is the estimated gaze direction.

13. The information processing device according to claim 12, wherein the update means further uses the estimated gaze direction and a correct gaze direction that is the correct answer for the gaze direction related to the training iris image to update model parameters included in the gaze estimation model.

14. An information processing device according to any one of claims 1 to 13, wherein the ellipse parameters include a first ellipse parameter for specifying a first ellipse representing the outer contour or edge of the iris included in the stored iris image, and a second ellipse parameter for specifying a second ellipse representing the inner contour or edge of the iris included in the stored iris image.

15. An information processing device according to any one of claims 1 to 14, further comprising an estimation means for estimating the ellipse parameters for a target iris image, which is an iris image including an iris to be authenticated, using the trained estimation model.

16. The information processing device according to any one of claims 1 to 15, wherein the ellipse parameters include a center coordinate, a major axis radius, a minor axis radius, and an ellipse rotation angle, or a center coordinate, a major axis radius, an ellipticity, and an ellipse rotation angle.

17. An information processing device according to claim 15, further comprising image processing means for downsampling an iris image to an image of a predetermined size, wherein each of the learning iris image and the target iris image is the downsampled iris image.

18. An information processing method in which one or more computers estimate ellipse parameters for a training iris image using an estimation model for estimating ellipse parameters for the contour or edge of the iris included in the iris image, and update model parameters included in the estimation model using estimated ellipse parameters, which are the estimated ellipse parameters, and correct ellipse parameters, which are the correct answers to the ellipse parameters for the training iris image.

19. A recording medium having recorded thereon a program for causing one or more computers to execute the following steps: estimating ellipse parameters for a training iris image using an estimation model for estimating ellipse parameters for the contour or edge of the iris contained in the iris image; and updating model parameters contained in the estimation model using estimated ellipse parameters, which are the estimated ellipse parameters, and correct ellipse parameters, which are the correct solutions to the ellipse parameters for the training iris image.

Images