Authentication device, authentication system, program, and method for generating learning data

Abstract

This technology provides a significant reduction in the effort required to generate training data. [Solution] An exemplary authentication device displays an authentication image with multiple bounding boxes on a display device, requests the user to select a specific bounding box from the multiple bounding boxes that contains an image of a specific class corresponding to an image of a specific class, authenticates the user as a person based on the score value corresponding to the specific class of the image in the selected bounding box selected by the user, and generates training data associating the specific class as the correct label for the image in the selected bounding box selected by the authenticated user.

Description

Technical Field

[0001] The present invention relates to an authentication device, an authentication system, a program, and a learning data generation method for authenticating that a user is a person.

Background Art

[0002] Image recognition technology for recognizing a recognition target object in an image using an artificial intelligence model has been put into practical use. In the supervised learning of an artificial intelligence model (object detection artificial intelligence model) used in such image recognition technology, it is necessary to assign a correct label indicating the position, size, and object name (class) of the recognition target object in the image to the image.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] Currently, mainly manual operations are performed for assigning correct labels, confirming the contents of the assigned correct labels, and correcting the contents of the assigned correct labels, which requires a very large amount of man-hours.

[0005] For example, Patent Document 1 discloses a technique for assisting the correction work of the contents of the assigned correct labels. However, in Patent Document 1, the work related to the correct label is only for generating learning data, so the man-hours only for generating learning data could not be significantly reduced.

[0006] In view of the above circumstances, an object of the present invention is to provide a technique capable of significantly reducing the man-hours only for generating learning data.

Means for Solving the Problems

[0007] An exemplary authentication device of the present invention displays an authentication image with a plurality of bounding boxes on a display device, requests a user to select a specific bounding box from the plurality of bounding boxes that contains an image of a specific class corresponding to an image of a specific class, authenticates the user as a person based on the score value corresponding to the specific class of the image of the selected bounding box selected by the user, and generates training data associating the specific class as the correct label for the image of the selected bounding box selected by the authenticated user. [Effects of the Invention]

[0008] According to an exemplary version of the present invention, the user is required to perform an annotation task on an image (for example, selecting a specific bounding box that contains a specific class image corresponding to an image) as a task necessary for authenticating that the user is a person, thereby significantly reducing the effort required to generate training data. [Brief explanation of the drawing]

[0009] [Figure 1] A diagram showing an example configuration of the authentication system according to the embodiment. [Figure 2] A diagram showing an example server configuration. [Figure 3] A diagram showing the relationship between an image and the XY coordinate plane. [Figure 4] Diagram illustrating the data used to identify the position and shape of the bounding box set for an image. [Figure 5] A diagram showing an example of an image. [Figure 6] Figure 5 shows the annotation information added to the image. [Figure 7] Authentication process flowchart [Figure 8] A diagram showing an example screen of a display device included in a user terminal device. [Figure 9] This figure shows other examples of screens on the display device included in the user terminal device. [Figure 10] Figure showing additional information of an image [Figure 11] Figure showing other additional information of an image [Figure 12] Flowchart of annotation information update process

Mode for Carrying Out the Invention

[0010] Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the drawings.

[0011] [Authentication System] FIG. 1 is a diagram showing a configuration example of an authentication system SYS1 according to an embodiment. As shown in FIG. 1, the authentication system SYS1 includes a user terminal device 1 and a server 2.

[0012] The user terminal device 1 may be, for example, a single terminal shared by a plurality of users, or may be a plurality of terminals respectively owned by a plurality of users

[0013] The user terminal device 1 includes a communication unit capable of communicating with the server 2 via a network NT1 not shown, a display unit capable of displaying an image or the like, and an operation unit operated by the user. As the display unit, for example, a thin display such as a liquid crystal display or an organic EL (Electro Luminescence) display can be used. As the input unit, for example, a keyboard, a pointing device, a touch panel, a joystick, or the like can be used.

[0014] As the user terminal device 1, for example, a personal computer, a smartphone, or the like can be used.

[0015] Server 2 is located in a location separate from the user terminal device 1 (for example, a service center building). Server 2 may be a physical server or a virtual server. Server 2 may consist of one server or multiple servers. Server 2 communicates with the user terminal device 1 via network NT1.

[0016] The authentication system SYS1 can provide specific services to users who have successfully authenticated. The content of these specific services is not particularly limited, but in this embodiment, the specific service is a literature search service.

[0017] [server] Figure 2 shows an example configuration of Server 2. Server 2 comprises a communication unit 21, a storage unit 22, and a controller 23.

[0018] The communication unit 21 transmits and receives arbitrary signals with the user terminal device 1. The controller 23 can also transmit and receive arbitrary information with the user terminal device 1 using the communication unit 21; however, the description of the communication unit 21 may be omitted below.

[0019] The storage unit 22 is configured to have non-volatile memory such as ROM (Read-only memory) or flash memory, and volatile memory such as RAM (Random Access Memory). The storage unit 22 stores a plurality of images 221, a program PG1, and a literature search database DB1.

[0020] Each of the multiple images 221 has annotation information attached to it, which includes the inference results of a trained artificial intelligence model for object detection (in this embodiment, a trained artificial intelligence model for traffic light detection) applied to each of the multiple images 221, as well as label type information.

[0021] Here, the relationship between image 221 and the XY coordinate plane is shown in Figure 3. Image 221 is a two-dimensional still image. Assuming a two-dimensional plane in which image 221 is defined, we define the two axes parallel to this two-dimensional plane as the X axis and the Y axis. The X axis and the Y axis are orthogonal to each other. The X axis is parallel to the horizontal direction of image 221, and the Y axis is parallel to the vertical direction of image 221. Hereafter, the two-dimensional plane parallel to the X axis and the Y axis will be referred to as the XY coordinate plane. Image 221 consists of multiple pixels arranged in a matrix on the XY coordinate plane (arranged in the X axis direction and the Y axis direction, respectively).

[0022] Figure 4 is an explanatory diagram of the data used to identify the position and shape of the bounding box BB set in Image 221.

[0023] The bounding box BB is part of the entire image area of ​​image 221, and the position and size of the bounding box BB are relative to the origin coordinates P. BB , width W BB and height H BB It is determined by the following: The bounding box BB has a rectangular outline. The four sides of the rectangle are parallel to the X-axis or Y-axis, and the length of the side of the rectangle parallel to the X-axis is the width W. BB The length of the side parallel to the Y-axis is the height H. BB Here, the X-axis direction is considered the width direction, and the Y-axis direction is considered the height direction. The coordinates (coordinates on the XY coordinate plane) of the vertices of the rectangle that make up the outer shape of the bounding box BB, located on the negative side of both the X-axis and the Y-axis, are the origin coordinates P. BB The bounding box BB is at the origin coordinates P. BB It is a rectangular region that extends along the positive X-axis and positive Y-axis from the perspective of [the object].

[0024] Figure 5 shows an example of image 221. Figure 6 shows the annotation information added to image 221 shown in Figure 5.

[0025] The image 221 shown in Figure 5 contains images of a traffic light F1, a traffic light F2, a person F3, and a vehicle F4. Furthermore, annotation information added to the image 221 shown in Figure 5 sets the first to seventh bounding boxes BB[1] to BB[7] to the image 221 shown in Figure 5. The traffic light image F1 is located within the third bounding box BB[3], the traffic light image F2 is located within the fourth bounding box BB[4], the person F3 is located within the fifth bounding box BB[5], and the vehicle image F4 is located within the seventh bounding box BB[7].

[0026] The annotation information shown in Figure 6 contains information about the specific class type, label type, origin coordinates, width, height, and score value (confidence level of being of a specific class) for each of the first to seventh bounding boxes BB[1] to BB[7]. The label type indicates whether the specific class type of the detected bounding box BB is correct or not. Initial values ​​of "correct label" and "incorrect label" are entered by comparing the class value with a predetermined threshold, and these are changed by the annotation information update process described later.

[0027] Program PG1 is a program that should be executed by controller 23.

[0028] The literature search database DB1 is a database that associates multiple documents with their respective attributes.

[0029] Controller 23 comprehensively controls the operation of each component in Server 2. Controller 23 is equipped with a processing unit including a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit) as hardware resources. Controller 23 has functional blocks 231 to 233.

[0030] Controller 23 is a program execution device (computer) capable of executing any program. The controller 23's various functions (including the functions of function blocks 231 to 233) are realized by the controller 23 executing program PG1. All operations of the controller 23 described in this embodiment may be operations realized by the controller 23 executing program PG1. Program PG1 may consist of multiple programs.

[0031] The functional blocks of controller 23 are described below. Functional blocks 231, 232, and 233 are the authentication unit, the search unit, and the update unit, respectively.

[0032] The authentication unit 231 authenticates that the user operating the user terminal device 1 is a human being.

[0033] The search unit 232 extracts documents from the document search database DB1 that match the search criteria sent from the user terminal device 1, and sends information about the extracted documents to the user terminal device 1.

[0034] The update unit 233 updates the annotation information attached to image 221.

[0035] [Authentication process] Figure 7 is a flowchart of the authentication process executed by the controller 23. This authentication process is achieved when the controller 23 executes the program PG1 described above. The authentication process shown in Figure 7 starts when a login request is sent from the user terminal device 1.

[0036] First, in step S10, the controller 23 (authentication unit 231) displays an image 221 with multiple bounding boxes BB on the display device included in the user terminal device 1, and requests the user to select a specific bounding box from among the multiple bounding boxes BB that contains a specific class image corresponding to an image of a specific class (traffic light). After the processing in step S10, the screen of the display device included in the user terminal device 1 will look like screen 11 shown in Figure 8, for example. After step S10, the process proceeds to step S20.

[0037] The screen 11 shown in Figure 8 includes a message M1 requesting the user to select a specific bounding box containing a specific class image corresponding to an image of a specific class (traffic light), and an image 221. The screen 11 shown in Figure 8 also includes an image update button B1 and a login execution button B2.

[0038] Message M1 is a message that says, "Please select N bounding boxes that surround the traffic lights." Here, N is less than the number of bounding boxes whose score value is equal to or greater than a threshold (e.g., 0.5) in the annotation information attached to image 221 included in screen 11. However, since the setting of N is not always appropriate, it is desirable to provide an image update button B1 for switching the images displayed on screen 11, as in this embodiment.

[0039] Furthermore, it is desirable to make it easier to distinguish bounding boxes selected by the user from those not selected by the user by displaying an icon nearby to indicate that they have been selected, or by graying them out.

[0040] In step S20, the controller 23 (authentication unit 231) determines whether it is necessary to switch the image 221 included in the screen 11 shown in Figure 8 to another image 221. If the image update button B1 included in the screen 11 shown in Figure 8 is pressed by the user, the controller 23 (authentication unit 231) determines that it is necessary to switch the image 221 included in the screen 11 shown in Figure 8 to another image 221, and in step S30, it switches the image 221 included in the screen 11 shown in Figure 8 to another image 221. After step S30, the process returns to step S10.

[0041] If the image update button B1 has not been pressed by the user, the process proceeds to step S40.

[0042] In step S40, the controller 23 (authentication unit 231) determines whether the selection of traffic signals has been completed. If the login execution button B2 included in the screen 11 shown in Figure 8 is pressed by the user, the controller 23 (authentication unit 231) determines that the selection of traffic signals has been completed and proceeds to step S50.

[0043] If the login execution button B2 has not been pressed by the user, the process returns to step S10.

[0044] In step S50, the controller 23 (authentication unit 231) displays an image 221 with multiple bounding boxes BB on the display device included in the user terminal device 1, and requests the user to select a bounding box other than a specific bounding box from among the multiple bounding boxes BB. After the processing in step S50, the screen of the display device included in the user terminal device 1 will look like screen 11 shown in Figure 9, for example. After step S50, the process proceeds to step S60.

[0045] The screen 11 shown in Figure 9 includes a message M2 requesting the user to select a bounding box other than the specified bounding box, and an image 221. The screen 11 shown in Figure 9 also includes an image update button B1 and a login execution button B2.

[0046] Message M2 is a message that says, "Please select M bounding boxes that enclose something other than a traffic light." Here, M is less than the number of bounding boxes whose score value is less than a threshold (e.g., 0.5) in the annotation information attached to image 221 included in screen 11. However, since the setting of M is not always appropriate, it is desirable to provide an image update button B1 for switching the images displayed on screen 11, as in this embodiment.

[0047] Furthermore, as with the processing in step S10, it is desirable to make it easier to distinguish bounding boxes selected by user interaction from bounding boxes not selected by user interaction by displaying an icon indicating selection nearby or by graying them out.

[0048] In step S60, the controller 23 (authentication unit 231) determines whether it is necessary to switch the image 221 included in the screen 11 shown in Figure 9 to another image 221. If the image update button B1 included in the screen 11 shown in Figure 9 is pressed by the user, the controller 23 (authentication unit 231) determines that it is necessary to switch the image 221 included in the screen 11 shown in Figure 9 to another image 221, and in step S70, it switches the image 221 included in the screen 11 shown in Figure 9 to another image 221. After step S70, it returns to step S50. Note that although the sudden change in the selection target may cause some discomfort to the user, it is also possible to return to step S10 instead of step S50 after step S70. By returning to step S10 instead of step S50, it is possible to secure training data that can be used for learning efficiently with a small number of users.

[0049] If the image update button B1 has not been pressed by the user, the process proceeds to step S80.

[0050] In step S80, the controller 23 (authentication unit 231) determines whether the selection of anything other than traffic lights has been completed. If the login execution button B2 included in the screen 11 shown in Figure 9 is pressed by the user, the controller 23 (authentication unit 231) determines that the selection of anything other than traffic lights has been completed and proceeds to step S90.

[0051] If the login execution button B2 has not been pressed by the user, the process returns to step S50.

[0052] In step S90, the controller 23 (authentication unit 231) determines whether the total number of images 221 in which both traffic light selection and non-traffic light selection have been completed has reached L. L can be set to any number. That is, L may be 1 or 2 or more. However, by setting L to 2 or more, it is possible to secure training data that can be efficiently used for learning with a small number of users.

[0053] If the number of images 221 in which both the selection of traffic lights and the selection of non-traffic lights have been completed has not reached L, the process proceeds to step S100. In step S100, the controller 23 (authentication unit 231) selects an image 221 to be included in the screen 11 shown in Figure 8, which is different from the image 221 that was most recently included in the screen 11 shown in Figure 9. After step S100, the process returns to step S10.

[0054] If the total number of images 221 in which both traffic light selections and non-traffic light selections have been completed reaches L, proceed to step S110.

[0055] In step S110, the controller 23 (authentication unit 231) calculates the accuracy rate of the user's selection based on the score value. The accuracy rate of the user's selection is defined as "the ratio of correctly selected frames to the total number of frames to be selected". Whether a selection is correct or not is determined according to the score value. For example, if a bounding box with a score value of 0.5 or higher is selected as "a frame surrounded by a traffic light", then "a frame surrounded by a traffic light" is considered to have been correctly selected. Also, if a bounding box with a score value less than 0.5 is selected as "a frame surrounded by something other than a traffic light", then "a frame surrounded by something other than a traffic light" is considered to have been correctly selected.

[0056] Here, we take the example of a case where the above N items are replaced by 2 items, and the above M items are replaced by 3 items.

[0057] For example, if person A correctly selects two frames containing traffic lights and three frames containing things other than traffic lights, person A's accuracy rate is (2+2) / (2+3) = 100%.

[0058] For example, if person B correctly selects one frame containing a traffic light and two frames containing something other than a traffic light, then person B's accuracy rate is (1+2) / (2+3) = 60%.

[0059] For example, if person C correctly selects 0 frames containing traffic lights and 1 frame containing something other than traffic lights, person B's accuracy rate will be (0+1) / (2+3) = 20%.

[0060] After step S110, proceed to step S120.

[0061] In step S120, the controller 23 (authentication unit 231) determines whether the authentication of the user as a human is successful or not based on the accuracy rate calculated in step S110. For example, if the accuracy rate is 20% or less, authentication fails, meaning the user is not a human. For example, if the accuracy rate is between 20% and 60%, re-authentication is required. For example, if the accuracy rate is above 60%, authentication is successful, meaning the user is a human.

[0062] The accuracy rate is calculated based on a score value corresponding to a specific class (traffic light), so the controller 23 (authentication unit 231) authenticates that the user is a human based on the score value corresponding to the specific class (traffic light).

[0063] If it is determined that re-authentication is required, for example, the controller 23 (authentication unit 231) may display a message indicating that re-authentication is required on the display device included in the user terminal device 1, then reset the contents of the selection process (display image and information of the selected bounding box BB), and then return to step S10.

[0064] If authentication is determined to be successful, the process proceeds to step S130, where the controller 23 (authentication unit 231) authorizes the search unit 232 to begin processing. This initiates the literature search service provided by the server 2. Once the literature search service is in operation, the controller 23 (search unit 232) displays, for example, a search criteria input screen on the display device included in the user terminal device 1. After step S130, the authentication process ends.

[0065] If authentication is determined to have failed, for example, the controller 23 (authentication unit 231) will prohibit access from the same IP address as the user terminal device 1 that made the login request, and then terminate the authentication process.

[0066] Each time authentication is successful and the authentication process shown in Figure 7 is completed, the controller 23 (authentication unit 231) updates the information shown in Figure 10 regarding the number of people who have successfully authenticated (including those who have successfully authenticated through re-authentication) for each image 221 in which both the selection of traffic lights and the selection of non-traffic lights have been completed in the authentication process shown in Figure 7. The information on people who have successfully authenticated, shown in Figure 10, is added to image 221 as additional information.

[0067] Each time authentication is successful and the authentication process shown in Figure 7 is completed, the controller 23 (authentication unit 231) also updates the information on the number of people who selected each bounding box shown in Figure 11 for each image 221 where both the selection of traffic lights and the selection of non-traffic lights have been completed in the authentication process shown in Figure 7. Note that the number of people who selected each bounding box in the information shown in Figure 11 is limited to people who have successfully authenticated (including those who have successfully authenticated through re-authentication). The information on the number of people who selected each bounding box shown in Figure 11 is added to image 221 as additional information.

[0068] [Annotation information update process] Figure 12 is a flowchart of the annotation information update process executed by the controller 23. This annotation information update process is achieved when the controller 23 executes the program PG1 described above. The annotation information update process shown in Figure 12 starts when authentication is successful and the authentication process shown in Figure 7 is completed.

[0069] First, in step S210, the controller 23 (update unit 233) determines whether there are any images 221 in which the number of successfully authenticated individuals who have undergone the selection process (number of people V0 shown in Figure 10) exceeds a predetermined value. If there are no images 221 in which the number of successfully authenticated individuals who have undergone the selection process exceeds a predetermined value, the annotation information update process is terminated without updating the annotation information.

[0070] If there is an image 221 in which the number of successfully authenticated individuals who have undergone the selection process exceeds a predetermined value, the process proceeds to step S220.

[0071] In step S220, the controller 23 (update unit 233) updates the annotation information of image 221 when the number of successful authentications who have undergone the selection process (number of people V0 shown in Figure 10) exceeds a predetermined value.

[0072] Specifically, the controller 23 (update unit 233) determines, regardless of the score value of each bounding box, that bounding boxes for which the number of people selected as a specific bounding box is less than the first judgment value will be labeled as incorrect, and that bounding boxes for which the number of people selected as a specific bounding box is greater than the second judgment value will be labeled as correct. The second judgment value is a value greater than or equal to the first judgment value.

[0073] Furthermore, the controller 23 (update unit 233) sets the labels of bounding boxes selected by fewer than the third judgment value as correct labels, regardless of the score value of each bounding box, and the labels of bounding boxes selected by more than the fourth judgment value as incorrect labels. The fourth judgment value is a value greater than or equal to the third judgment value. The third judgment value may be the same as the first judgment value, or it may be different from the first judgment value. The fourth judgment value may be the same as the second judgment value, or it may be different from the second judgment value.

[0074] By allowing users to select classes other than a specific class, in addition to selecting a specific class, we can expect to improve the accuracy of updating annotation information on training images.

[0075] After step S220, proceed to step S230.

[0076] In step S230, the controller 23 (update unit 233) updates (generates) the training data by adding the image 221 with the updated annotation information (label type) to the training data.

[0077] After step S230, the annotation information update process is terminated.

[0078] Furthermore, the artificial intelligence model for object detection is trained using the updated training data, thereby updating image 221.

[0079] When training an artificial intelligence model, it is acceptable to use only "correct labels" as training data. By training only with data that has been detected with the correct label, the model becomes more likely to detect the correct label.

[0080] The processing in steps S220 and S230 generates training data in which a specific class (traffic light) is associated as the correct label with the bounding box image selected by the authenticated user. In other words, the user's selection during the authentication process updates (corrects) the annotation information of image 221, generating highly accurate training data. This significantly reduces the effort required to generate training data.

[0081] Furthermore, statistical processing is performed in step S220, and training data is generated using the results of statistical processing of bounding boxes selected by individuals who successfully authenticated (including those who successfully authenticated through re-authentication). As a result, the individual characteristics of individuals who successfully authenticated (including those who successfully authenticated through re-authentication) are less likely to be reflected in the training data, and an improvement in the accuracy of the training data can be expected.

[0082] <Notes, etc.> The various technical features disclosed in the embodiments for carrying out the invention as specified herein can be modified in various ways without departing from the spirit of the technical creation. Furthermore, the multiple embodiments and modifications disclosed in the embodiments for carrying out the invention as specified herein may be combined to the extent possible.

[0083] For example, in the flowchart shown in Figure 7, steps S50 to S80 may be omitted. If steps S50 to S80 are omitted, in step S90 the controller 23 (authentication unit 231) only needs to determine whether the number of images 221 for which the traffic light selection has been completed has reached L. This reduces the burden on the user's selection work.

[0084] Furthermore, in the flowchart shown in Figure 7, for example, either steps S10-S40 or steps S50-S80 may be omitted for each user. In other words, for example, server 2 may alternately perform the process shown in Figure 7 with steps S10-S40 omitted and the process shown in Figure 7 with steps S50-S80 omitted. This reduces the burden on each user's selection process while suppressing a decrease in the accuracy of annotation information updates.

[0085] In the embodiment described above, login was permitted only after successful authentication of the user's identity. However, it is also possible to first perform personal authentication using an ID (identification) and password, and then authenticate the user's identity only after successful personal authentication.

[0086] Human identity may be authenticated using methods other than those disclosed in the embodiments described above. Even when human identity is authenticated using methods other than those disclosed in the embodiments described above, it is preferable to implement a training data generation method that, similar to the embodiments described above, requires the user to perform an annotation task on an image as a task necessary for authenticating the user as a person, authenticates the user as a person based on the results of the task performed by the user, and generates training data based on the results of the task performed by the user. With this training data generation method, since the user is required to perform an annotation task on an image as a task necessary for authenticating the user as a person, the amount of work required to generate training data can be significantly reduced.

[0087] In addition to the methods disclosed in the embodiments described above, an annotation process for authenticating human identity could involve, for example, assigning a correct label indicating a specific class to an image. In this case, the user could be asked to select whether or not the image contains an image of an object corresponding to the specific class.

[0088] In this case, for example, the screen of the display device included in user terminal device 1 could contain an image and the message, "If there is a traffic light in the image, please check the checkbox." Here, the method for authenticating that the user is human can be to authenticate that the user is human by analyzing the behavior of the user's actions to check the checkbox.

[0089] Furthermore, the correct label may also indicate the position and size of the object in the image corresponding to the specific class, in addition to the specific class itself.

[0090] In this case, for example, the screen of the display device included in user terminal device 1 could contain an image and the message, "If there is a traffic light in the image, please surround the traffic light with a frame." Here, the method for authenticating that the user is a person can be to authenticate that the user is a person by analyzing the behavior of the user's operation to surround the traffic light with a frame. [Explanation of Symbols]

[0091] 1. User terminal device 2. Server SYS1... Authentication System

Claims

1. An authentication image with multiple bounding boxes is displayed on the display device. The user is asked to select a specific bounding box from among the multiple bounding boxes that contains an image of a specific class that corresponds to an image of a specific class. Based on the score value corresponding to the specific class of the image of the selection bounding box selected by the user, the system authenticates that the user is a person. Authentication device that generates training data associating the specific class as the correct label with the image of the selection bounding box selected by the user whose identity has been authenticated as a human.

2. The authentication device according to claim 1, which switches the authentication images and requests the user to select a specific bounding box from the plurality of bounding boxes for each of the authentication images.

3. The authentication device according to claim 1, which calculates the accuracy rate of the user's selection based on the score value and authenticates that the user is a person based on the accuracy rate.

4. The authentication device according to claim 1, which also requires the user to select a bounding box other than the specified bounding box from among the plurality of bounding boxes.

5. The authentication device according to any one of claims 1 to 4, which generates the learning data using the results of statistical processing of selection bounding boxes selected by each of the multiple users whose identity has been authenticated as a person.

6. An authentication system including a user terminal device and an authentication device, The user terminal device is Including a display unit and an operating unit, The authentication device is An authentication image with multiple bounding boxes is displayed on the display unit of the user terminal device. The user is requested via the user terminal device to select a specific bounding box from among the plurality of bounding boxes that contains an image of a specific class that corresponds to an image of a specific class. The selection information of the selection bounding box selected by the user's operation of the control unit is obtained. Based on the score value corresponding to the specific class of the image of the selection bounding box based on the acquired selection information, the user is authenticated as a person. The system generates training data in which the image of the selection bounding box selected by the user, whose identity has been verified as a human, is associated with the specific class as the correct label. Authentication system.

7. Displaying an authentication image with multiple bounding boxes on a display device, The user is asked to select a specific bounding box from among the multiple bounding boxes that contains an image of a specific class that corresponds to an image of a specific class, Authenticating that the user is a person based on the score value corresponding to the specific class of the image of the selection bounding box selected by the user, To generate training data in which the specific class is associated as the correct label with the image of the selection bounding box selected by the user whose identity has been verified as a human. A program that causes a computer to execute something.

8. A method for generating training data that a computer can perform, The user is required to perform the task of annotating an image as a necessary step to verify that the user is a person. Based on the results of the work performed by the aforementioned user, the system authenticates that the user is a person. Based on the results of the work performed by the aforementioned user, training data is generated. Method for generating training data.

9. The method for generating training data according to claim 8, wherein the annotation work is the work of assigning correct labels indicating a specific class to the image.

10. The method for generating training data according to claim 9, wherein the correct label indicates, in addition to the specific class, the position and size of the object corresponding to the specific class in the image.

Images