WO2023073954A1 - 虹彩認証装置、虹彩認証システム、虹彩認証方法、及び、記録媒体 - Google Patents
虹彩認証装置、虹彩認証システム、虹彩認証方法、及び、記録媒体 Download PDFInfo
- Publication number
- WO2023073954A1 WO2023073954A1 PCT/JP2021/040123 JP2021040123W WO2023073954A1 WO 2023073954 A1 WO2023073954 A1 WO 2023073954A1 JP 2021040123 W JP2021040123 W JP 2021040123W WO 2023073954 A1 WO2023073954 A1 WO 2023073954A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- iris
- image
- resolution
- feature amount
- magnification
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims description 82
- 238000006243 chemical reaction Methods 0.000 claims abstract description 255
- 238000000605 extraction Methods 0.000 claims abstract description 87
- 230000006870 function Effects 0.000 claims description 62
- 238000001914 filtration Methods 0.000 claims description 31
- 239000000284 extract Substances 0.000 claims description 25
- 230000008569 process Effects 0.000 claims description 18
- 238000013139 quantization Methods 0.000 claims description 10
- 230000002194 synthesizing effect Effects 0.000 claims description 10
- 238000004590 computer program Methods 0.000 claims description 9
- 230000001131 transforming effect Effects 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 abstract description 45
- 238000012545 processing Methods 0.000 description 100
- 238000010586 diagram Methods 0.000 description 16
- 238000001514 detection method Methods 0.000 description 13
- 238000010801 machine learning Methods 0.000 description 13
- 230000007246 mechanism Effects 0.000 description 12
- 238000012549 training Methods 0.000 description 10
- 230000009466 transformation Effects 0.000 description 10
- 238000013528 artificial neural network Methods 0.000 description 9
- 238000004891 communication Methods 0.000 description 9
- 230000000694 effects Effects 0.000 description 9
- 239000013598 vector Substances 0.000 description 9
- 230000004913 activation Effects 0.000 description 7
- 230000009467 reduction Effects 0.000 description 7
- 238000013075 data extraction Methods 0.000 description 6
- 230000015572 biosynthetic process Effects 0.000 description 5
- 238000003786 synthesis reaction Methods 0.000 description 5
- 210000001747 pupil Anatomy 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 239000000470 constituent Substances 0.000 description 3
- 238000010276 construction Methods 0.000 description 3
- 238000013527 convolutional neural network Methods 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 238000003909 pattern recognition Methods 0.000 description 2
- 238000007639 printing Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 210000004087 cornea Anatomy 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 230000006866 deterioration Effects 0.000 description 1
- 230000002542 deteriorative effect Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011478 gradient descent method Methods 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 208000016339 iris pattern Diseases 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000012887 quadratic function Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
- G06F21/30—Authentication, i.e. establishing the identity or authorisation of security principals
- G06F21/31—User authentication
- G06F21/32—User authentication using biometric data, e.g. fingerprints, iris scans or voiceprints
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
Definitions
- This disclosure relates to technical fields of iris authentication devices, iris authentication systems, iris authentication methods, and recording media.
- Non-Patent Document 1 describes a technique for super-resolving an image so that it has more information for matching by machine learning using a loss function for matching.
- Non-Patent Document 2 describes a technique of performing super-resolution corresponding to various enlargement ratios with one network by estimating a filter according to the enlargement ratio of upsampling.
- the object of this disclosure is to provide an iris authentication device, an iris authentication system, an iris authentication method, and a recording medium aimed at improving the techniques described in prior art documents.
- the iris authentication device includes iris image acquisition means for acquiring an iris image including the iris of a living body, and a magnification for the iris image is calculated from the size of the iris region included in the iris image and a desired size.
- generating means for generating a resolution-converted image obtained by converting the resolution of the iris image according to the magnification; and post-conversion feature extracting means for extracting a post-conversion feature that is a feature of the resolution-converted image.
- One aspect of the iris authentication system includes iris image acquisition means for acquiring an iris image including the iris of a living body, and a magnification for the iris image calculated from the size of the iris region included in the iris image and a desired size.
- generating means for generating a resolution-converted image obtained by converting the resolution of the iris image according to the magnification; and post-conversion feature extracting means for extracting a post-conversion feature that is a feature of the resolution-converted image.
- an iris image including the iris of a living body is acquired, a magnification for the iris image is calculated from the size of the iris region included in the iris image and a desired size, and the magnification is Accordingly, a resolution-converted image obtained by converting the resolution of the iris image is generated, and a post-conversion feature quantity, which is a feature quantity of the resolution-converted image, is extracted.
- an iris image including the iris of a living body is obtained in a computer, a magnification for the iris image is calculated from the size of the iris region included in the iris image and a desired size, and the An iris authentication method is executed for generating a resolution-converted image obtained by converting the resolution of the iris image in accordance with a magnification, and extracting a post-conversion feature quantity, which is a feature quantity of the resolution-converted image.
- FIG. 1 is a block diagram showing the configuration of an iris authentication device according to the first embodiment.
- FIG. 2 is a block diagram showing the configuration of an iris authentication device according to the second embodiment.
- FIG. 3 is a flow chart showing the flow of iris authentication operation performed by the iris authentication device in the second embodiment.
- FIG. 4 is a block diagram showing the configuration of an iris authentication device according to the third embodiment.
- FIG. 5 is a flow chart showing the flow of learning operations performed by the iris authentication device according to the third embodiment.
- FIG. 6 is a block diagram showing the configuration of an iris authentication device according to the fifth embodiment.
- FIG. 7 is a flowchart showing the flow of super-resolution processing performed by the iris authentication device according to the fifth embodiment.
- FIG. 8 is a block diagram showing the configuration of an iris authentication device according to the sixth embodiment.
- FIG. 9 is a flowchart showing the flow of super-resolution processing performed by the iris authentication device according to the sixth embodiment.
- FIG. 10 is a block diagram showing the configuration of an iris authentication device according to the seventh embodiment.
- FIG. 11 is a flowchart showing the flow of super-resolution processing performed by the iris authentication device according to the seventh embodiment.
- FIG. 12 is a block diagram showing the configuration of an iris authentication device according to the eighth embodiment.
- FIG. 13 is a flow chart showing the flow of iris authentication operation performed by the iris authentication device in the eighth embodiment.
- FIG. 14 is a block diagram showing the configuration of an iris authentication system according to the ninth embodiment.
- a first embodiment of an iris authentication device, an iris authentication method, and a recording medium will be described.
- the first embodiment of the iris authentication device, the iris authentication method, and the recording medium will be described below using the iris authentication device 1 to which the first embodiment of the iris authentication device, the iris authentication method, and the recording medium is applied. explain. [1-1: Configuration of iris authentication device 1]
- FIG. 1 is a block diagram showing the configuration of the iris authentication device 1 according to the first embodiment.
- the iris authentication device 1 includes an iris image acquisition unit 11 , a calculation unit 12 , a generation unit 13 , and a post-conversion feature amount extraction unit 14 .
- the iris image acquisition unit 11 acquires the iris image LI including the iris of the living body.
- the iris refers to the ring-shaped area around the pupil inside the pupil of the eye.
- the iris has a pattern unique to each individual. Also, since the iris is covered with the cornea and is less likely to be damaged, it is a site suitable for biometric authentication.
- the calculator 12 calculates the magnification for the iris image LI from the size of the iris region included in the iris image LI and the desired size.
- the size of the iris region included in the iris image LI may be represented, for example, by the number of pixels in the iris region in the iris image, the diameter of the iris region in the iris image, the area of the iris region in the iris image, or the like. good.
- a case where the size of the iris area included in the iris image LI is represented by the number of pixels of the iris area in the iris image will be described below as an example.
- iris authentication authentication is performed using an iris pattern.
- the desired number of pixels may be the number of pixels suitable for the iris authentication. Since the iris is substantially circular, the number of pixels may be represented by the radius of the corresponding area. A substantially circular iris region is also called an iris circle. Also, the number of pixels may correspond to the resolution. For example, the desired number of pixels may be 100 pixels or more, 100 pixels, 125 pixels, and so on.
- iris authentication it is preferable to use an iris image HI with relatively high resolution.
- iris detection detects edges of the pupil region and the iris region, it can be performed even with a low-resolution iris image LI.
- Information obtained by this iris detection such as the position of the pupil circle, the iris circle, and the number of pixels, can be used in super-resolution processing for increasing the resolution of the low-resolution iris image LI.
- the super-resolution process is a process of increasing the resolution of a low-resolution image to generate a high-resolution image, and refers to a process capable of generating a relatively high-quality high-resolution image.
- the calculation unit 12 may calculate the ratio of the radius of the detected iris circle and the radius of the area of the desired number of pixels to calculate the magnification. That is, the calculator 12 can calculate the magnification based on information obtained by iris detection.
- the magnification is not limited to 1x or more, and may be less than 1x.
- the calculator 12 may calculate the magnification as 2.
- the calculator 12 may calculate the magnification as 0.5 times.
- the generating unit 13 generates a resolution-converted image RI by converting the resolution of the iris image LI according to the magnification. For example, if the magnification calculated by the calculator 12 from the radius of the iris circle is 2, the generator 13 may generate the resolution-converted image RI by doubling the resolution of the iris image LI.
- the post-conversion feature amount extraction unit 14 extracts the post-conversion feature amount OC, which is the feature amount of the resolution-converted image RI.
- the post-conversion feature quantity extraction unit 14 may be constructed so as to be able to extract a feature quantity from an image having a desired number of pixels.
- the calculation unit 12 may calculate the magnification so that the feature can be extracted appropriately, and the generation unit 13 may generate the resolution-converted image RI according to the magnification.
- the feature amount here is a value representing the feature of the iris necessary for performing iris authentication.
- the post-transformation feature amount extraction unit 14 may be constructed by, for example, a convolutional neural network. [1-2: Technical Effects of Iris Authentication Device 1]
- the iris authentication device 1 in the first embodiment can convert the iris image LI into an image with a desired number of pixels regardless of the number of pixels of the iris image LI.
- the iris authentication device 1 according to the first embodiment can obtain a high-resolution iris image HI by performing super-resolution processing to increase the resolution of the low-resolution iris image LI.
- the resolution of the iris image LI on which the iris authentication device 1 in the first embodiment performs super-resolution processing may be any resolution, and is not limited to a specific resolution.
- the iris authentication device 1 in the first embodiment can perform iris authentication using iris images LI with various resolutions.
- the calculation unit 12 calculates a scale factor so that the post-conversion feature quantity extraction unit 14 can appropriately extract the feature quantity, and the generation unit 13 calculates the scale factor according to the scale factor. Generate a resolution-converted image RI. That is, in the iris authentication device 1 according to the first embodiment, it is not necessary to change the mechanism for iris authentication. Therefore, the iris authentication device 1 according to the first embodiment can be applied to a mechanism constructed to enable iris authentication using an iris image HI with a desired number of pixels. [2: Second embodiment]
- a second embodiment of an iris authentication device, an iris authentication method, and a recording medium will be described.
- the second embodiment of the iris authentication device, the iris authentication method, and the recording medium will be described below using the iris authentication device 2 to which the second embodiment of the iris authentication device, the iris authentication method, and the recording medium is applied. explain. [2-1: Configuration of iris authentication device 2]
- FIG. 2 is a block diagram showing the configuration of the iris authentication device 2 according to the second embodiment.
- the same reference numerals are assigned to the components that have already been described, and detailed description thereof will be omitted.
- the iris authentication device 2 includes an arithmetic device 21 and a storage device 22. Furthermore, the iris authentication device 2 may include a communication device 23 , an input device 24 and an output device 25 . However, the iris authentication device 2 does not have to include at least one of the communication device 23 , the input device 24 and the output device 25 . Arithmetic device 21 , storage device 22 , communication device 23 , input device 24 and output device 25 may be connected via data bus 26 .
- the computing device 21 includes, for example, at least one of a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), and an FPGA (Field Programmable Gate Array). Arithmetic device 21 reads a computer program. For example, arithmetic device 21 may read a computer program stored in storage device 22 . For example, the computing device 21 reads a computer program stored in a computer-readable non-temporary recording medium to a recording medium reading device (not shown) provided in the iris authentication device 2 (for example, an input device 24 to be described later). can be read using The computing device 21 may acquire (that is, download) a computer program from a device (not shown) arranged outside the iris authentication device 2 via the communication device 23 (or other communication device). may be read).
- a CPU Central Processing Unit
- GPU Graphics Processing Unit
- FPGA Field Programmable Gate Array
- Arithmetic device 21 executes the read computer program. As a result, logical functional blocks for executing the operations to be performed by the iris authentication device 2 are realized in the arithmetic device 21 .
- the arithmetic device 21 can function as a controller for implementing logical functional blocks for executing the operations (in other words, processing) that the iris authentication device 2 should perform.
- FIG. 2 shows an example of logical functional blocks implemented within the computing device 21 to perform the iris authentication operation.
- the computing device 21 includes an iris image acquisition unit 211 as a specific example of the ⁇ iris image acquisition means'', a calculation unit 212 as a specific example of the ⁇ calculation means'', and a ⁇ generation A generation unit 213 that is a specific example of the "means”, a post-conversion feature amount extraction unit 214 that is a specific example of the "post-conversion feature amount extraction unit", and a specific example of the "determination means” and the "authentication means”
- An authenticator 215 is implemented.
- the calculator 212 may include an iris circle detector 2121 and an enlargement factor calculator 2122 .
- the computing device 21 does not have to include the authentication unit 215 .
- the storage device 22 can store desired data.
- the storage device 22 may temporarily store computer programs executed by the arithmetic device 21 .
- the storage device 22 may temporarily store data temporarily used by the arithmetic device 21 while the arithmetic device 21 is executing a computer program.
- the storage device 22 may store data that the iris authentication device 2 saves over a long period of time.
- the storage device 22 may include at least one of RAM (Random Access Memory), ROM (Read Only Memory), hard disk device, magneto-optical disk device, SSD (Solid State Drive), and disk array device. good. That is, the storage device 22 may include non-transitory recording media.
- the storage device 22 may store the super-resolution model SM, the feature quantity generation model GM, and the matching feature quantity CC. However, the storage device 22 may not store at least one of the super-resolution model SM, the feature amount generation model GM, and the matching feature amount CC. The details of the super-resolution model SM, the feature quantity generation model GM, and the matching feature quantity CC will be described later.
- the communication device 23 can communicate with devices external to the iris authentication device 2 via a communication network (not shown).
- the input device 24 is a device that accepts input of information to the iris authentication device 2 from outside the iris authentication device 2 .
- the input device 24 may include an operation device (for example, at least one of a keyboard, a mouse and a touch panel) that can be operated by the operator of the iris authentication device 2 .
- the input device 24 may include a reading device capable of reading information recorded as data on a recording medium that can be externally attached to the iris authentication device 2 .
- the output device 25 is a device that outputs information to the outside of the iris authentication device 2 .
- the output device 25 may output information as an image.
- the output device 25 may include a display device (so-called display) capable of displaying an image showing information to be output.
- the output device 25 may output information as voice.
- the output device 25 may include an audio device capable of outputting audio (so-called speaker).
- the output device 25 may output information on paper. That is, the output device 25 may include a printing device (so-called printer) capable of printing desired information on paper. [2-2: Iris authentication operation performed by iris authentication device 2]
- FIG. 3 is a flow chart showing the flow of iris authentication operation performed by the iris authentication device 2 in the second embodiment.
- the number of pixels in the iris area is smaller than the desired number of pixels and the magnification is the enlargement rate.
- the iris image acquisition unit 211 acquires an iris image including the iris of the living body (step S21).
- the iris circle detection unit 2121 detects an iris circle from the iris image (step S22).
- the iris circle detection unit 2121 may calculate a vector representing the center position and radius of the iris circle from the input iris image.
- the iris circle detection unit 2121 may be composed of, for example, a regression neural network.
- a regression neural network includes multiple convolutional layers and multiple activation layers, extracts the feature values of the input image, and converts the extracted feature values into vectors representing the center position and radius of the corresponding region using a linear layer. can do.
- the iris image LI input to the iris circle detection unit 2121 and the vector output from the iris circle detection unit 2121 may be normalized.
- a neural network with any structure can be used as long as it meets the requirements.
- Examples of the structure of the neural network include those similar to structures such as VGG, ResNet (Residual neural network), etc., which are models trained with large-scale image datasets. good too.
- a normalization layer such as batch normalization may be used as an intermediate layer of the neural network.
- ReLU Rectified Linear Unit
- the iris circle detection unit 2121 may be an image processing mechanism that is not composed of a neural network.
- the magnification calculator 2122 calculates the magnification for the iris image LI from the radius of the iris circle included in the iris image LI detected by the iris circle detector 2121 and the desired radius (step S23).
- the magnification may be the ratio of the radius of the iris circle included in the iris image LI to the radius of the iris circle of desired size.
- the enlargement ratio is not a simple ratio of the radius of the iris circle included in the iris image LI and the radius of the iris circle of a desired size, but may be, for example, a value converted to a logarithm or power of the ratio. good.
- the enlargement ratio calculation unit 2122 similarly to the calculation unit 12 of the first embodiment, uses a parameter may be calculated as
- the iris circle detection unit 2121 may calculate the diameter of the iris circle from the input iris image. In this case, the magnification calculator 2122 calculates the magnification for the iris image LI from the diameter of the iris circle included in the iris image LI detected by the iris circle detector 2121 and the desired diameter. Also, the iris circle detection unit 2121 may calculate the area of the iris circle from the input iris image. In this case, the magnification calculator 2122 calculates the magnification for the iris image LI from the area of the iris circle included in the iris image LI detected by the iris circle detector 2121 and the desired area.
- the generation unit 213 generates a resolution-converted image RI, which is a super-resolution image obtained by increasing the resolution of the iris image LI, according to the enlargement ratio (step S24).
- the generation unit 213 may use the enlargement ratio calculated by the enlargement ratio calculation unit 2122 as it is, or may use the expansion ratio calculated by the expansion ratio calculation unit 2122 after normalizing it.
- the generation unit 213 may generate the resolution-converted image RI, which is a super-resolution image, using the super-resolution model SM.
- the super-resolution model SM is a model constructed by machine learning to output a resolution-converted image RI for an input iris image LI.
- a specific example of the construction method of the super-resolution model SM will be described in detail in third and fourth embodiments. Further, specific examples of the constructed super-resolution model SM will be described in detail in fifth to seventh embodiments.
- the post-conversion feature amount extraction unit 214 extracts the post-conversion feature amount OC, which is the feature amount of the resolution-converted image RI (step S25).
- the post-conversion feature amount extraction unit 214 may extract the post-conversion feature amount OC from the resolution-converted image RI using the feature amount generation model GM.
- the feature quantity generation model GM extracts the feature quantity of the corresponding iris image HI. It is a model that can be generated.
- the feature quantity generation model GM may be constructed by machine learning so as to be able to output a feature quantity suitable for iris authentication when the iris image HI is input. Specifically, the feature quantity generation model GM is set so that the loss function set based on the error between a plurality of feature quantities generated from the iris image HI of the same individual is small (preferably minimized).
- the feature quantity generation model GM may be constructed by adjusting the included learning parameters.
- the feature amount generation model GM may be constructed as a convolutional neural network that generates feature amounts by, for example, convolution processing.
- the feature quantity generation model GM may be any model capable of generating feature quantities with high accuracy, and may be another trained neural network.
- Matching data may be input to the constructed feature quantity generation model GM to generate a matching feature quantity CC, which is the feature quantity of the matching data.
- the generated matching feature amount CC may be registered in the storage device 22 .
- the authentication unit 215 performs authentication using a score indicating the degree of similarity between the post-conversion feature amount OC and the feature amount prepared in advance (step S26).
- authentication refers to at least one of identifying a person and determining that the person is the person.
- the authenticating unit 215 may determine that the person is the real person when the matching score indicating the degree of similarity between the post-conversion feature amount OC and the matching feature amount CC prepared in advance is equal to or greater than a threshold.
- the authentication unit 215 may calculate the matching score using, for example, the cosine similarity between the transformed feature quantity OC and the matching feature quantity CC.
- the authentication unit 215 determines whether each feature amount is similar. You may Alternatively, the authentication unit 215 may calculate the matching score using, for example, the L2 distance function or the L1 distance function between the converted feature amount OC and the matching feature amount CC. L2 distance function, L1 distance function, or the like, feature amounts of data related to the same individual tend to be close to each other, and authentication unit 215 determines whether each feature amount is similar. may
- the output device 25 outputs the authentication result of the authentication unit 215, the enlargement ratio calculated by the enlargement ratio calculation unit 2122, and the resolution-converted image RI generated by the generation unit 213 to the outside of the iris authentication device 2 (step S27). .
- the output from the output device 25 may be confirmed by a person to be authenticated, a manager, a security guard, or the like. Further, the output device 25 may output an alert when the magnification is greater than or equal to a predetermined size. If the generation unit 213 enlarges the image with a magnification greater than a predetermined size, the authentication accuracy may decrease. , can pay attention to the applicable certification. [2-3: Technical Effects of Iris Authentication Device 2]
- Iris authentication often requires a relatively high-resolution iris image HI with an iris radius of 100 pixels or more.
- a low-resolution image LI with less than 100 pixels is used for iris authentication, if a certain degree of accuracy can be achieved, for example, one relatively low-resolution camera can be used for authentication simultaneously with other biometric authentication. be able to.
- the iris authentication device 2 in the second embodiment can convert a low-resolution iris image LI into a high-resolution resolution-converted image RI, which is a super-resolution image, regardless of the resolution of the iris image LI. Therefore, iris authentication can be performed with high accuracy. Therefore, if the iris authentication device 2 in the second embodiment is applied, an authentication that performs both biometric authentication and iris authentication, for example, using an image captured using a single relatively inexpensive camera. can be realized.
- a third embodiment of an iris authentication device, an iris authentication method, and a recording medium will be described.
- the third embodiment of the iris authentication device, the iris authentication method, and the recording medium will be described below using an iris authentication device 3 to which the third embodiment of the iris authentication device, the iris authentication method, and the recording medium is applied. explain. [3-1: Configuration of iris authentication device 3]
- FIG. 4 is a block diagram showing the configuration of the iris authentication device 3 in the third embodiment.
- the same reference numerals are assigned to the components that have already been described, and detailed description thereof will be omitted.
- the iris authentication device 3 includes an arithmetic device 21 and a storage device 22.
- the computing device 21 includes a learning image acquisition unit 316 as a specific example of the "learning image acquisition means” and an input image generation unit 317 as a specific example of the "input image acquisition means”. and an iris information estimation unit 300 including a learning unit 318, which is a specific example of the “learning means”, and an iris image acquisition unit 211, a calculation unit 212, a generation unit 213, and a post-transform feature extraction unit 214.
- the input image generator 317 includes a batch data extractor 3171 and a resolution converter 3172 .
- the learning unit 318 includes a loss function calculator 3181 , a gradient calculator 3182 and a parameter updater 3183 .
- the storage device 22 may store the learning images TI. However, the storage device 22 does not have to store the learning images TI. If the storage device 22 does not store the learning image TI, the communication device 23 may acquire the learning image TI from a device external to the iris authentication device 2, or the input device 24 may acquire the learning image TI from a device external to the iris authentication device 2. An input of the learning image TI from may be accepted.
- the training image TI may be an iris image containing an iris area with a desired number of pixels.
- the learning image acquiring unit 316, the input image generating unit 317, the learning unit 318, and the iris information estimating unit 300 perform machine learning using the learning images TI to obtain the supersolution used by the generating unit 213. Construct an image model SM. Details of the operations of the learning image acquiring unit 316, the input image generating unit 317, the learning unit 318, and the iris information estimating unit 300 will be described with reference to FIG. [3-2: Learning operation performed by iris authentication device 3]
- FIG. 5 is a flow chart showing the flow of the learning operation performed by the iris authentication device 3 in the third embodiment.
- the learning image acquiring unit 316 acquires a data set of learning images TI including an iris region with a desired number of pixels, which are stored in, for example, the storage device 22 (step S31).
- the learning image TI may be an image having the same resolution as the iris image HI having a resolution suitable for authentication by the iris authentication device 3 .
- the batch data extraction unit 3171 randomly extracts batch size batch data of the learning images TI from the data set of the learning images TI acquired by the learning image acquisition unit 316 (step S32). For example, when the batch size is 32, the batch data extraction unit 3171 extracts 32 training images TI. Batch sizes of 32, 64, 128, etc. may be used. The batch size value is not particularly limited, and any value can be used.
- the resolution conversion unit 3172 generates an input image II by converting the resolution of the learning image TI according to the reciprocal of an arbitrary enlargement ratio (step S33). That is, the resolution conversion unit 3172 generates the low-resolution input image II from the high-resolution learning image TI.
- the input image generation unit 317 can prepare an image obtained by reducing the resolution of the training image TI as the input image II.
- the resolution conversion unit 3172 may reduce the learning image TI to generate a low-resolution input image II.
- the resolution conversion unit 3172 may reduce the learning image TI by thinning out the pixels of the learning image TI. That is, by reducing the training image TI by the resolution conversion unit 3172, it is possible to generate the input image II in which the resolution of the training image TI is lowered.
- the resolution conversion unit 3172 reduces the resolution of each of the training images TI extracted by the batch data extraction unit 3171, for example, using the reciprocal of an arbitrary enlargement ratio selected according to the uniform random number distribution, and generates the input image II. You may In this case, the resolution conversion unit 3172 can generate batch data that uniformly includes input images II of various resolutions.
- the resolution conversion unit 3172 generates an input image II by reducing the resolution of all the batch data of the batch-sized learning images TI extracted by the batch data extraction unit 3171 at the same timing using the same reciprocal of the enlargement ratio. You can At this time, the resolution converter 3172 can generate batch data including the input images II with the same resolution. In this case, the resolution conversion unit 3172 lowers the resolution of the batch data of the batch-sized training images TI extracted by the batch data extraction unit 3171 at different timings using different reciprocals of the magnification ratios, and generates the input image II. may If the input image generation unit 317 can prepare the input images II so that the input images II with various resolutions are uniformly included in the entire data set of the learning images TI acquired by the learning image acquisition unit 316, good.
- steps S34, S35, and S36 may be the same as the operations of steps S21, S24, and S25 described using FIG.
- the data used for the operation in the second embodiment is the iris image LI for authentication, but the data used for the operation in the third embodiment is the input image II prepared for learning.
- the iris image acquisition unit 211 acquires one input image II from the batch data of the batch size input images II (step S34).
- the generating unit 213 generates a resolution-converted input image RII by converting the resolution of the input image II using the enlargement factor used when reducing the learning image TI in the resolution converting unit 3172 (step S35).
- the resolution-converted input image RII has the same resolution as the learning image TI.
- the post-conversion feature amount extraction unit 214 extracts the input feature amount OIC, which is the feature amount of the resolution-converted input image RII (step S36).
- the post-conversion feature amount extraction unit 214 extracts the learning feature amount TC, which is the feature amount of the learning image TI, or stores the learning feature amount TC, which is the feature amount of the learning image TI, in advance together with the learning image TI. It may be stored in device 22 .
- the iris image acquisition unit 211 determines whether or not all the input images II of the batch data of the batch size input images II have been processed (step S39). If all the batch data of the input image II of the batch size have not been processed (step S39: No), the process proceeds to step S34.
- the iris information estimation unit 300 performs the operations of steps S34 to S38 for all input images II of batch data of input images II of batch size. When calculations are performed by a GPU or multithreading, estimation calculations for input images of each batch size may be performed in parallel. Part of the processing may be parallel, and serial and parallel processing may be mixed.
- the learning unit 318 causes the generation unit 213 to learn the generation method of the resolution-converted images RI. . Specifically, the learning unit 318 causes the super-resolution model SM used by the generation unit 213 to learn the method of generating the resolution-converted image RI, and constructs the super-resolution model SM. More specifically, the learning unit 318 adjusts learning parameters included in the super-resolution model SM.
- the learning unit 318 uses a first loss function that increases the loss as the learning feature amount TC and the input feature amount OIC are dissimilar, and a second loss function that increases the loss as the training image TI and the resolution-converted input image RII are dissimilar. Based on at least one of the loss functions of , the generation unit 213 is made to learn the method of generating the resolution-converted image RI. The learning unit 318 may optimize the iris information estimation unit 300 based on the loss function.
- the loss function calculator 3181 uses a first loss function in which the loss increases as the learning feature amount TC and the input feature amount OIC dissimilar, and a loss increases as the learning image TI and the resolution-converted input image RII dissimilar. Calculation using at least one of the increasing second loss functions is performed (step S40).
- the loss function calculation unit 3181 inputs the learned feature amount TC, which is the correct individual label, and the input feature amount OIC of the resolution-converted input image RII extracted by the post-conversion feature amount extraction unit 214, and calculates the degree of dissimilarity between them.
- the indicated first loss value may be output.
- the loss function calculation unit 3181 converts the one-hot vector generated from the learned feature amount TC, which is the correct personal label, and the feature vector as the input feature amount OIC extracted by the post-transform feature amount extraction unit 214 to cross entropy loss. A function may be compared to obtain a first loss.
- the loss function calculator 3181 inputs the training image TI, which is a high-resolution image, and the resolution-converted input image RII generated by the generator 213, and outputs a second loss value indicating the degree of dissimilarity between them. You may The loss function calculator 3181 may compare the learning image TI and the resolution-converted input image RII generated by the generator 213 using the L1 distance loss function to obtain a second loss.
- the loss function calculator 3181 is not limited to the cross-entropy loss function and the L1 distance loss function, and may use other loss functions such as the KL divergence function and the L2 distance function.
- the loss function calculator 3181 may weight the calculated loss according to the expansion rate calculated by the calculator 212 .
- super-resolution processing with a large magnification ratio is often more difficult than super-resolution processing with a small magnification ratio.
- authentication processing using a super-resolution image obtained by super-resolution processing when the magnification is large is equivalent to authentication using a super-resolution image obtained by super-resolution processing when the magnification is small.
- Authentication accuracy is often inferior to processing. Therefore, the loss function calculator 3181 may use a loss function that gives a large weight to the loss resulting from the super-resolution processing when the magnification is large.
- the learning unit 318 arbitrarily selects the second enlargement ratio in which the loss weight corresponding to the input image II generated by using the first enlargement ratio as an arbitrary enlargement ratio is smaller than the first enlargement ratio.
- the generating unit 213 may be caused to perform learning based on a loss function that is larger than the weight of the loss corresponding to the generated input image II by using it as the enlargement ratio of .
- the learning unit 318 instructs the generation unit 213 based on a loss function in which the weight of the loss corresponding to the resolution-converted input image RII generated by using the enlargement ratio increases as the enlargement ratio increases. Let them learn. By doing so, the learning contribution of the super-resolution processing when the magnification is large becomes large. Then, the learning unit 318 can construct a super-resolution model SM whose authentication performance does not easily depend on the enlargement ratio.
- the loss function calculator 3181 may weight the first loss and the second loss separately according to the expansion rate. Alternatively, the loss function calculator 3181 may weight the first loss and the second loss respectively according to the expansion ratio, take the sum of the weights, and output a single loss. .
- the loss function calculation unit 3181 calculates the loss of each input image II as , may be weighted according to different resolutions to calculate the loss of batch data for the batch size.
- the loss function calculator 3181 may calculate the average value of each weighted loss and output it as the batch data loss of the batch size.
- the resolution conversion unit 3172 weights the loss of each input image II according to the enlargement factor used in step S33 to generate each input image II, and adjusts the batch size. Batch data loss may be calculated.
- the resolution conversion unit 3172 determines that the first magnification used to generate the first input image II is greater than the second magnification used to generate the second input image II.
- the magnification used in step S33 to generate each input image II is such that the weight for the loss of the first input image II is greater than the weight for the loss of the second input image II.
- the loss of batch data for batch size may be calculated with weighting accordingly.
- the loss function calculation unit 3181 calculates the loss of each input image II as , may be weighted identically to compute the batch size loss of batch data.
- the loss function calculator 3181 may calculate a loss average value, which is the average loss value of each input image II. In this case, since batch data of batch sizes generated by the resolution conversion unit 3172 at different timings have different resolutions, the loss function calculation unit 3181 may weight the loss average value according to the resolution.
- the gradient calculator 3182 uses the value of the loss output by the loss function calculator 3181 and uses the error backpropagation method to calculate the gradient of the learning parameters included in the super-resolution model SM (step S41).
- the parameter updating unit 3183 updates the values of the learning parameters included in the super-resolution model SM using the calculated gradients of the learning parameters (step S42). Updating the value of the learning parameter in step S42 corresponds to learning of the super-resolution model SM.
- the parameter updating unit 3183 may optimize the value of the learning parameter so that the value of the loss function is minimized. Examples of the optimization method used by the parameter updating unit 3183 include stochastic gradient descent, Adam, and the like, but are not limited to these.
- the parameter updating unit 3183 may update the learning parameters using hyperparameters such as weight decay and momentum even when the stochastic gradient descent method is used.
- the input image generator 317 determines whether batch data has been extracted from the predetermined learning image TI (step S43). If the batch data has not been extracted from the predetermined learning image TI (step S43: No), the process proceeds to step S32. For example, when the learning image acquiring unit 316 acquires 320 data sets of learning images TI and the batch size is 32, the iris information estimating unit 300 performs steps S32 to S42 ten times. Just do it. If batch data has already been extracted from the predetermined training image TI (step S43: Yes), the learning unit 318 saves the optimized super-resolution model SM including the optimally updated learning parameters in the storage device 22. (Step S44). [3-3: Technical Effects of Iris Authentication Device 3]
- the iris authentication device 3 teaches the generating unit 213 how to generate the resolution-converted image RI based on a loss function that increases the loss as the learning image TI and the resolution-converted input image RII are dissimilar. Therefore, the accuracy of super-resolution processing can be increased. Further, the iris authentication device 3 in the third embodiment causes the generation unit 213 to learn a generation method of the resolution-converted image RI based on a loss function in which the loss increases as the learned feature amount TC and the input feature amount OIC are dissimilar. is performed, a resolution-converted image RI suitable for iris authentication can be generated.
- the resolution-converted image RI can be generated from which the feature amount suitable for iris authentication can be extracted.
- a resolution model SM can be constructed. Since the post-conversion feature amount OC output by the iris information estimation unit 300 is a feature amount used for authenticating the person, the resolution-converted image RI requires extraction of the post-conversion feature amount OC suitable for authenticating the person. It is desirable that the image is capable of In other words, the resolution-converted image RI generated by the super-resolution model SM is an image that has undergone super-resolution processing with high accuracy, and is also an image suitable for matching.
- the degree of difficulty of super-resolution processing differs depending on the magnification ratio, so if the loss function is calculated uniformly regardless of the magnification ratio, the accuracy of super-resolution processing may not be maintained. That is, there is a possibility that the precision of the super-resolution processing will deteriorate when the magnification is large.
- the iris authentication device 3 in the third embodiment uses a loss function that is weighted according to the enlargement ratio. can be done.
- the resolution conversion unit 3172 determines that the first magnification used to generate the first input image II is greater than the second magnification used to generate the second input image II.
- the magnification used in step S33 to generate each input image II is such that the weight for the loss of the first input image II is greater than the weight for the loss of the second input image II. Compute the batch data loss for the batch size, weighting accordingly.
- a super-modified model SM capable of generating a resolution-converted image RI capable of maintaining the accuracy of iris authentication even when a relatively low-resolution iris image LI is input.
- the generation unit 213 using the super-resolution model SM constructed by the iris authentication device 3 of the third embodiment generates a high-resolution resolution-converted image suitable for highly accurate matching. RI generation can be achieved.
- the iris authentication device 3 in the third embodiment devises a learning method for super-resolution processing. Accuracy can be maintained.
- the iris authentication device 3 in the third embodiment can construct a super-resolution model SM whose authentication performance does not easily depend on the enlargement ratio.
- the iris authentication device 3 according to the third embodiment can perform super-resolution processing of an iris image corresponding to various magnifications while maintaining authentication accuracy.
- a fourth embodiment of an iris authentication device, an iris authentication method, and a recording medium will be described.
- the fourth embodiment of the iris authentication device, the iris authentication method, and the recording medium will be described below using an iris authentication device 3 to which the fourth embodiment of the iris authentication device, the iris authentication method, and the recording medium is applied. explain.
- the iris authentication device 3 in the fourth embodiment may have the same configuration as the iris authentication device 3 in the third embodiment described above. Compared to the iris authentication device 3 in the third embodiment, the iris authentication device 3 in the fourth embodiment has an input image II generation process by a resolution conversion unit 3172 and a loss function calculation process by a loss function calculation unit 3181. is different. That is, the iris authentication device 3 in the fourth embodiment differs from the iris authentication device 3 in the third embodiment in the operations of step S33 and step S40 shown in FIG. Other features of the iris authentication device 3 in the fourth embodiment may be the same as other features of the iris authentication device 3 in the third embodiment. [4-1: Learning operation performed by iris authentication device 3]
- the resolution conversion unit 3172 uses a first enlargement ratio and a second enlargement ratio smaller than the first enlargement ratio as arbitrary enlargement ratios.
- the resolution conversion unit 3172 converts a plurality of input images II so that the frequency of generating the input image II according to the reciprocal of the first enlargement ratio is higher than the frequency of generating the input image II according to the reciprocal of the second enlargement ratio. to generate an input image II (step S33). That is, the resolution conversion unit 3172 generates a plurality of input images II such that the larger the magnification used, the more input images II are generated.
- the plurality of input images II generated by the resolution conversion unit 3172 include more input images II generated using the reciprocal of a large enlargement factor.
- the selection of the enlargement rate used by the resolution conversion unit 3172 becomes more frequent as the value of the enlargement rate increases.
- the resolution conversion unit 3172 may be configured such that the smaller the value of the reciprocal of the enlargement ratio, the easier it is to select the value to be used for the low-resolution processing of the learning image TI.
- the resolution conversion unit 3172 may select the enlargement factor to be used with a probability distribution that facilitates generation of the input image II with a lower resolution.
- the resolution converter 3172 may select the enlargement factor to be used according to a weighted probability distribution. In order to learn many super-resolution processing with a large enlargement ratio, the resolution conversion unit 3172 can select the enlargement ratio to be used with a probability distribution that facilitates generation of a low-resolution image with a large enlargement ratio. .
- the resolution conversion unit 3172 may be constructed so that the lower the resolution of the input image II, the easier it is to generate. As a result, a large enlargement ratio is used more in subsequent super-resolution processing by the generation unit 213 .
- the probability distribution used by the resolution conversion unit 3172 to select the enlargement ratio may be created using a linear function, a quadratic function, or the like. There are no other restrictions on the probability distribution to be used as long as it facilitates the selection of low-resolution images with a large magnification ratio.
- the operation of the resolution converter 3172 in the fourth embodiment plays the same role as the weighting calculation by the loss function calculator 3181 in the third embodiment. Therefore, in the fourth embodiment, the loss function calculator 3181 does not need to weight the loss in calculating the loss. Therefore, in the fourth embodiment, the loss function calculator 3181 does not need to perform weighting according to the enlargement ratio (step S40). [4-2: Technical Effects of Iris Authentication Device 4]
- the iris authentication device 3 of the fourth embodiment performs machine learning weighted according to the enlargement ratio in order to generate a high-resolution resolution-converted image RI suitable for high-precision matching regardless of the resolution of the iris image LI. to build a super-resolution model SM.
- the generation unit 213 using the super-resolution model SM constructed by the iris authentication device 3 of the fourth embodiment generates a high-resolution resolution-converted image RI suitable for highly accurate matching regardless of the resolution of the iris image LI. can be realized.
- the iris authentication device 3 of the fourth embodiment also instructs the generation unit 213 to learn the generation method of the resolution-converted image RI based on the loss function in which the loss increases as the learning image TI and the resolution-converted input image RII are dissimilar. is performed, the accuracy of the super-resolution processing can be increased. Also, in the iris authentication device 3 of the fourth embodiment, the generation unit 213 learns the generation method of the resolution-converted image RI based on the loss function in which the loss increases as the learned feature amount TC and the input feature amount OIC are dissimilar. is performed, a resolution-converted image RI suitable for iris authentication can be generated.
- the iris authentication device 3 of the fourth embodiment can also construct a super-resolution model SM whose authentication performance is less dependent on the enlargement ratio. Also, the iris authentication device 3 of the fourth embodiment can perform super-resolution processing of an iris image corresponding to various magnifications while maintaining authentication accuracy.
- Both the iris authentication device 3 in the third embodiment and the iris authentication device 3 in the fourth embodiment can realize high-precision super-resolution processing regardless of the enlargement ratio.
- the iris authentication device 3 in the third embodiment has the effect of enabling highly accurate iris authentication
- the iris authentication device 3 in the fourth embodiment has a super-resolution model SM construction process. is simple.
- the iris authentication device 3 in the fourth embodiment weights the distribution of the resolution of the input image II and directly manipulates the input image II to be input, the iris authentication device 3 in the third embodiment Compared to , the contribution of weighting to the construction process is large, and it is possible to further prevent the accuracy from deteriorating due to the enlargement rate. [5: Fifth embodiment]
- a fifth embodiment of an iris authentication device, an iris authentication method, and a recording medium will be described.
- the fifth embodiment of the iris authentication device, the iris authentication method, and the recording medium will be described below using the iris authentication device 5 to which the fifth embodiment of the iris authentication device, the iris authentication method, and the recording medium is applied. explain. [5-1: Configuration of iris authentication device 5]
- FIG. 6 is a block diagram showing the configuration of the iris authentication device 5 in the fifth embodiment.
- the iris authentication device 5 includes an iris image acquisition unit 211 , a calculation unit 212 , a generation unit 513 , and a post-conversion feature amount extraction unit 214 .
- the generation unit 513 uses the super-resolution model SM to perform super-resolution processing for generating a resolution-converted image RI obtained by converting the resolution of the iris image LI according to the enlargement ratio.
- Generation unit 513 includes feature amount extraction unit 5131 , filter generation unit 5132 , and conversion unit 5133 . Details of the operations of the feature amount extraction unit 5131, the filter generation unit 5132, and the conversion unit 5133 will be described with reference to FIG. [5-2: Super-resolution processing performed by generation unit 513]
- FIG. 7 is a flowchart showing the flow of super-resolution processing performed by the iris authentication device 5 according to the fifth embodiment.
- the iris image acquisition unit 211 acquires an iris image LI including the iris of the living body (step S51).
- the calculator 212 calculates the enlargement ratio for the iris image LI (step S52).
- the feature quantity extraction unit 5131 extracts the pre-conversion feature quantity PC, which is the feature quantity of the iris image LI (step S53).
- the feature amount extraction unit 5131 may extract the pre-conversion feature amount PC from the low-resolution iris image LI using the low-resolution feature amount extraction model included in the super-resolution model SM.
- the low-resolution feature quantity extraction model may be a model capable of outputting a feature quantity suitable for filtering, which will be described later, when a low-resolution iris image LI is input.
- the low-resolution feature quantity extraction model may be constructed, for example, by machine learning so that when an iris image LI is input, a feature quantity suitable for filtering, which will be described later, can be output.
- the feature quantity extraction unit 5131 may input the iris image LI to the low-resolution feature quantity extraction model and output the pre-conversion feature quantity PC.
- the filter generation unit 5132 generates one or more conversion filters for converting the pre-conversion feature amount PC according to the enlargement ratio calculated by the calculation unit 212 (step S54).
- the filter generation unit 5132 may generate one or more conversion filters according to the magnification using the conversion filter generation model included in the super-resolution model SM.
- the conversion filter generation model may be a model capable of generating a conversion filter suitable for filtering, which will be described later, when an enlargement factor is input.
- the conversion filter generation model may be configured, for example, by machine learning so that when an enlargement factor is input, a conversion filter suitable for filtering to be described later can be output.
- the filter generation unit 5132 may input the enlargement factor calculated by the calculation unit 212 to the conversion filter generation model and output one or more conversion filters.
- the filter generation unit 5132 may generate a conversion filter for convolution processing.
- the filter generation unit 5132 may generate a conversion filter with a size of 3 ⁇ 3, for example.
- the size of the conversion filter is not limited to 3 ⁇ 3, and may be 5 ⁇ 5.
- the size of the conversion filter can be arbitrarily determined according to requirements such as processing speed and processing accuracy.
- the filter generation unit 5132 may determine the size of the conversion filter.
- the filter generation unit 5132 may generate (Cin ⁇ Cout) conversion filters, for example.
- Cin may be, for example, a number corresponding to the number of channels of the pre-transform feature PC. Cin may be, for example, 3 if the iris image LI is a color image, and may be 1 if the iris image LI is a gray image.
- Cout may be 3 when the resolution-converted image RI output by the filtering process is a color image, and may be 1 when the resolution-converted image RI output by the filtering process is a gray image. you can
- the conversion filter generated by the filter generation unit 5132 may be used to increase the resolution of the pre-conversion feature amount PC extracted from the low-resolution iris image LI.
- the pre-conversion feature amount PC extracted from the low-resolution iris image LI may have a size of (Cin ⁇ h ⁇ w), for example.
- the feature quantity extraction unit 5131 may generate Cin pre-conversion feature quantities PC having a size of (h ⁇ w).
- the feature quantity that has been high-resolution using the conversion filter may have a size of (Cout ⁇ H ⁇ W), for example. More specifically, Cout high-resolution feature quantities having a size of (H ⁇ W) may be generated.
- the conversion filter generation model may be input with an enlargement factor composed of the one-dimensional vector and output a conversion filter having a size of (Cin ⁇ Cout ⁇ 3 ⁇ 3). More specifically, the conversion filter generation model may be input with an enlargement factor configured by the one-dimensional vector, and output Cin ⁇ Cout conversion filters having a size of (3 ⁇ 3). Alternatively, the conversion filter generation model may be input with an enlargement factor configured by the one-dimensional vector, and output Cin ⁇ Cout conversion filters having a size of (h ⁇ w).
- the filter generation unit 5132 may generate a conversion filter other than the filter for convolution processing.
- the filter generation unit 5132 may generate a conversion filter having the same size as the feature amount extracted by the feature amount extraction unit 5131 .
- the size of the feature amount may be, for example, (Cin ⁇ h ⁇ w).
- the conversion unit 5133 converts the pre-conversion feature amount PC by filtering using one or more conversion filters to generate the resolution-converted image RI (step S55).
- the conversion unit 5133 may use the conversion filter generated by the filter generation unit 5132 to perform filter processing on the pre-conversion feature amount PC. Further, the conversion unit 5133 converts the low-resolution iris image LI using the conversion filter generated by the filter generation unit 5132 to generate a resolution-converted image RI, which is a super-resolution image with high resolution. good too.
- the conversion unit 5133 may adjust the size of the pre-conversion feature amount PC according to the enlargement ratio before performing the filtering process. For example, when the enlargement ratio is 2, the conversion unit 5133 may double the size of the pre-conversion feature PC by inserting zeros between pixels of the pre-conversion feature PC. . Further, for example, when the enlargement ratio is 1.5 times, the conversion unit 5133 inserts zeros between every two pixels of the pre-conversion feature amount PC to increase the size of the pre-conversion feature amount PC. It may be magnified 1.5 times. The conversion unit 5133 may insert a value other than zero between pixels to increase the size of the pre-conversion feature amount PC.
- the conversion unit 5133 may insert a value obtained by copying the value of an adjacent pixel between pixels to increase the size of the pre-conversion feature amount PC.
- the conversion unit 5133 may adjust the size of the pre-conversion feature amount PC using other methods.
- the conversion unit 5133 may expand the size of the pre-conversion feature amount PC by interpolation using the nearest neighbor method, linear interpolation method, bilinear method, bicubic method, or the like.
- the conversion unit 5133 may perform convolution processing with a stride of 1 using a conversion filter on the complemented feature amount.
- the stride refers to the application interval of convolution
- the convolution processing with stride 1 refers to performing convolution processing by moving the transform filter at intervals of one pixel.
- the conversion unit 5133 may perform convolution processing using the filtering model included in the super-resolution model SM.
- the filtering model may be a model that can output a resolution-converted image RI using a conversion filter when the pre-conversion feature amount PC is input.
- the filter processing model may be configured to be able to output the resolution-converted image RI using a conversion filter when the pre-conversion feature amount PC is input by machine learning, for example.
- the conversion unit 5133 may input the pre-conversion feature amount PC to the filtering model and output the resolution-converted image RI.
- the number of convolution layers realized by the filtering model is not limited to one, and may be multiple layers. In this case, an activation layer such as a ReLU function may be inserted after each convolutional layer.
- the conversion unit 5133 may perform filter processing other than convolution processing.
- the transformation unit 5133 may generate a filter feature amount having the same size as the pre-transformation feature amount PC, and output the element product of the pre-transformation feature amount PC and the filter feature amount.
- the number of layers realized by the filtering model is not limited to one, and may be multiple layers. Also, a plurality of layers in which these layers and an activation layer are combined may be used.
- the post-conversion feature amount extraction unit 214 extracts the post-conversion feature amount OC, which is the feature amount of the resolution-converted image RI (step S56). [5-3: Technical Effects of Iris Authentication Device 5]
- the iris authentication device 5 of the fifth embodiment estimates and generates a conversion filter for each magnification of super-resolution processing. Therefore, a single super-resolution model SM can perform super-resolution processing corresponding to various enlargement ratios.
- the iris authentication device 5 of the fifth embodiment is particularly effective when the resolution of the resolution-converted image RI is fixed. That is, the super-resolution model SM used in the iris authentication device 5 of the fifth embodiment can output a resolution-converted image RI with a desired resolution regardless of the resolution of the iris image LI.
- the existing iris authentication mechanism can perform iris authentication even when an iris image LI of any resolution is input.
- iris authentication device an iris authentication method, and a recording medium according to a sixth embodiment
- the sixth embodiment of the iris authentication device, the iris authentication method, and the recording medium will be described below using an iris authentication device 6 to which the sixth embodiment of the iris authentication device, the iris authentication method, and the recording medium is applied. explain. [6-1: Configuration of iris authentication device 6]
- FIG. 8 is a block diagram showing the configuration of the iris authentication device 6 in the sixth embodiment.
- the iris authentication device 6 includes an iris image acquisition unit 211 , a calculation unit 212 , a generation unit 613 , and a post-conversion feature amount extraction unit 214 .
- the generation unit 613 uses the super-resolution model SM to perform super-resolution processing to generate a resolution-converted image RI obtained by converting the resolution of the iris image LI according to the enlargement ratio.
- the generation unit 613 includes a feature amount extraction unit 6131 , an enlargement ratio feature amount extraction unit 6132 , a synthesis unit 6133 , and a conversion unit 6134 . Note that the generation unit 613 does not have to include the conversion unit 6134 . [6-2: Super-resolution processing performed by generation unit 613]
- FIG. 9 is a flow chart showing the flow of super-resolution processing performed by the iris authentication device 6 in the sixth embodiment.
- the iris image acquisition unit 211 acquires an iris image including the iris of the living body (step S61).
- the calculation unit 212 calculates the enlargement ratio for the iris image (step S62).
- the feature quantity extraction unit 6131 extracts the pre-conversion feature quantity PC, which is the feature quantity of the iris image LI (step S63).
- the feature amount extraction unit 6131 may extract the pre-conversion feature amount PC from the low-resolution iris image LI using the low-resolution feature amount extraction model included in the super-resolution model SM.
- the low-resolution feature quantity extraction model is a model capable of generating a feature quantity suitable for at least one of feature quantity synthesis processing and filtering processing, which will be described later, when a low-resolution iris image LI is input. good.
- the low-resolution feature quantity extraction model is constructed by, for example, machine learning so that when an iris image LI is input, a feature quantity suitable for at least one of feature quantity synthesis processing and filter processing, which will be described later, can be output.
- the feature amount extraction unit 6131 may input the iris image LI to the low-resolution feature amount extraction model and output the pre-conversion feature amount PC.
- the enlargement rate feature amount extraction unit 6132 extracts the enlargement rate feature amount RC, which is the feature amount of the enlargement rate (step S64).
- the enlargement rate feature amount extraction unit 6132 may generate an enlargement rate feature amount map that is the feature amount of the enlargement rate.
- the enlargement factor feature quantity extraction unit 6132 may extract the enlargement factor feature quantity RC using the enlargement factor feature quantity extraction model included in the super-resolution model SM.
- the enlargement rate feature amount extraction model may be constructed so as to be capable of outputting an enlargement rate feature amount RC suitable for at least one of feature amount synthesis processing and filter processing, which will be described later, when an enlargement rate is input. .
- the enlargement rate feature amount extraction unit 6132 may input the enlargement rate to the enlargement rate feature amount extraction model and output the enlargement rate feature amount RC.
- the enlargement ratio feature quantity extraction unit 6132 may extract the enlargement ratio feature quantity RC having the same size as the pre-conversion feature quantity PC.
- the synthesizing unit 6133 synthesizes the pre-transformation feature amount PC and the enlargement ratio feature amount RC, and transforms the pre-transformation feature amount PC (step S65).
- the synthesizing unit 6133 may synthesize the pre-conversion feature quantity PC and the enlargement ratio feature quantity RC to generate a synthesized feature quantity.
- the synthesizing unit 6133 may transform the pre-conversion feature amount PC into a feature amount that does not depend on the enlargement ratio.
- the synthesizing unit 6133 may perform any of combination, element sum, and element multiplication.
- the synthesizing unit 6133 may synthesize the feature quantity map of the iris image LI and the magnification feature quantity map.
- the enlargement ratio feature quantity map generated by the enlargement ratio feature quantity extraction unit 6132 can have a size of (Cf ⁇ h ⁇ w).
- Cf may be, for example, the same number as the number of channels of the pre-transform feature PC.
- the synthesizing unit 6133 may combine the feature map of the iris image LI and the enlargement ratio feature map by channels to obtain the synthesized feature map as the synthesized feature.
- the conversion unit 6134 generates a resolution-converted image RI (step S66).
- the conversion unit 6134 may use the filtering model included in the super-resolution model SM to generate the resolution-converted image RI.
- the filtering model may be a model that can output a resolution-converted image RI using a conversion filter when a converted pre-conversion feature amount PC (combined feature amount) is input.
- the filter processing model is configured to be able to output a resolution-converted image RI using a conversion filter when a pre-conversion feature value PC (combined feature value) converted by machine learning, for example, is input. good too.
- the conversion filter may be a filter that does not depend on the enlargement ratio, and can be used regardless of the number of pixels in the iris image LI.
- the conversion unit 6134 may input the converted pre-conversion feature amount PC (combined feature amount) to the filter processing model and output the resolution-converted image RI.
- the conversion unit 6134 may output the resolution-converted image RI by performing convolution processing on the synthesized feature amount.
- the transformation unit 6134 may perform convolution processing using one convolution layer.
- the convolution layers may be multiple layers, and the conversion unit 6134 may perform convolution processing using multiple layers in which the convolution layers and the activation layers are combined.
- the generation unit 613 does not have to include the independent conversion unit 6134 .
- the synthesizing unit 6133 synthesizes the pre-transformation feature amount PC and the enlargement ratio feature amount RC, transforms the pre-transformation feature amount PC, and performs convolution processing on the post-transformation pre-transformation feature amount PC to obtain a resolution-converted image.
- RI may be generated.
- the synthesizing unit 6133 may generate the resolution-converted image RI using the filtering model described above.
- the post-conversion feature amount extraction unit 214 extracts the post-conversion feature amount OC, which is the feature amount of the resolution-converted image RI (step S67). [6-3: Technical effect of iris authentication device 6]
- the iris authentication device 6 of the sixth embodiment combines the pre-conversion feature quantity PC and the enlargement ratio feature quantity RC to perform super-resolution processing corresponding to various enlargement ratios using a single super-resolution model SM. It can be performed.
- the enlargement factor feature amount extraction unit 6132 can extract the enlargement factor feature amount RC according to the enlargement factor.
- a transform image RI can be generated.
- the iris authentication device 6 of the sixth embodiment is particularly effective when the resolution of the resolution-converted image RI is fixed.
- the super-resolution model SM used in the iris authentication device 6 of the sixth embodiment is a resolution-converted image with a desired resolution regardless of the resolution of the iris image LI, as in the iris authentication device 5 of the fifth embodiment.
- RI can be output. Therefore, by applying to the existing iris authentication mechanism a super-resolution model SM that has been learned and constructed so as to be able to output a resolution-converted image RI corresponding to each existing iris authentication mechanism, the existing iris authentication mechanism can accurately perform iris authentication even when an iris image LI of any resolution is input.
- iris authentication device an iris authentication method, and a recording medium according to a seventh embodiment
- the seventh embodiment of the iris authentication device, the iris authentication method, and the recording medium will be described below using the iris authentication device 7 to which the seventh embodiment of the iris authentication device, the iris authentication method, and the recording medium is applied. explain. [7-1: Configuration of iris authentication device 7]
- FIG. 10 is a block diagram showing the configuration of the iris authentication device 7 in the seventh embodiment.
- the iris authentication device 7 includes an iris image acquisition unit 211 , a calculation unit 212 , a generation unit 713 , and a post-conversion feature amount extraction unit 214 .
- the generating unit 713 uses the super-resolution model SM to perform super-resolution processing for generating a resolution-converted image RI by converting the resolution of the iris image LI according to the magnification.
- the generation unit 713 includes a feature quantity extraction unit 7131 , a quantization unit 7132 , a filter generation unit 7133 , a conversion unit 7134 and a reduction unit 7135 .
- the iris authentication device 7 according to the seventh embodiment differs from the iris authentication device 5 according to the fifth embodiment in that a quantization unit 7132 is provided before the filter generation unit 7133 and a reduction unit 7135 is provided after the conversion unit 7134. .
- a quantization unit 7132 is provided before the filter generation unit 7133 and a reduction unit 7135 is provided after the conversion unit 7134.
- FIG. 11 is a flow chart showing the flow of super-resolution processing performed by the iris authentication device 7 in the seventh embodiment.
- the iris image acquisition unit 211 acquires an iris image LI including the iris of the living body (step S71).
- the calculator 212 calculates the enlargement ratio for the iris image LI (step S72).
- the feature quantity extraction unit 7131 extracts the pre-conversion feature quantity PC, which is the feature quantity of the iris image LI (step S73).
- the feature quantity extraction unit 7131 may extract the pre-conversion feature quantity PC from the low-resolution iris image LI using the low-resolution feature quantity extraction model included in the super-resolution model SM.
- the low-resolution feature quantity extraction model may be a model capable of outputting a feature quantity suitable for filtering, which will be described later, when a low-resolution iris image LI is input.
- the low-resolution feature quantity extraction model may be constructed, for example, by machine learning so that when an iris image LI is input, a feature quantity suitable for filtering, which will be described later, can be output.
- the feature quantity extraction unit 7131 may input the iris image LI to the low-resolution feature quantity extraction model and output the pre-conversion feature quantity PC.
- the quantization unit 7132 quantizes the enlargement rate to a predetermined enlargement rate (step S74).
- the quantization section 7132 may quantize the input enlargement factor into powers of 2 such as 2, 4, and 8, for example. In this case, the quantization unit 7132 can output 2 times, for example, when an enlargement factor of 1.5 times is input. Specifically, the quantization section 7132 may search for n satisfying 2 n ⁇ 1 ⁇ R ⁇ 2 n with respect to the enlargement factor R, and output 2 n as the quantized enlargement factor.
- the predetermined enlargement ratio does not have to be a power of 2, and may be an arbitrary power such as a power of 1.5 or a power of 2.5. Also, the predetermined enlargement ratio does not have to be a value represented by a power, and other discrete values such as multiples of 2 may be employed.
- the filter generating unit 7133 generates one or more transform filters for transforming the pre-transform feature amount PC according to the quantized enlargement ratio (step S75).
- the filter generation unit 7133 according to the seventh embodiment is different from the filter generation unit 5132 according to the fifth embodiment, which can receive a continuous value of the enlargement factor, in that a discrete value of the enlargement factor is input.
- the filter generation unit 7133 may generate one or more conversion filters according to the enlargement ratio using a conversion filter generation model included in the super-resolution model SM.
- the conversion filter generation model may be a model capable of generating a conversion filter suitable for filtering, which will be described later, when a quantized enlargement factor is input.
- the conversion filter generation model may be configured to be able to output a conversion filter suitable for filtering, which will be described later, when a quantized enlargement factor is input by, for example, machine learning.
- the filter generation unit 7133 may input the enlargement factor quantized by the quantization unit 7132 to the conversion filter generation model and output one or more conversion filters.
- the filter generation unit 7133 does not generate conversion filters according to various enlargement factors, but generates conversion filters according to quantized enlargement factors. That is, the conversion filter generation model is constructed by learning the generation of conversion filters specialized for limited enlargement ratios. Thus, the conversion filter generation model in the seventh embodiment, since it is built by learning specialized for a limited enlargement rate, use the conversion filter generated by the filter generation unit 7133 using the conversion filter generation model Thus, super-resolution processing with higher precision can be realized.
- the conversion unit 7134 converts the pre-conversion feature amount PC by filtering using one or more conversion filters to generate a first resolution converted image (step S76).
- the conversion unit 7134 preferably adjusts the size of the pre-conversion feature amount PC according to the enlargement ratio before performing the filtering process.
- the conversion unit 7134 may perform convolution processing with stride 1 on the complemented feature amount using a conversion filter.
- the conversion unit 7134 may generate the first resolution converted image using the filtering model included in the super-resolution model SM.
- the filtering model may be a model that can output a first resolution converted image using a conversion filter when the pre-conversion feature amount PC is input.
- the filter processing model may be configured to be able to output the first resolution converted image using a conversion filter when the pre-conversion feature amount PC is input by machine learning, for example.
- the conversion unit 7134 may input the pre-conversion feature amount PC to the filtering model and output the first resolution converted image.
- the number of convolution layers realized by the filtering model is not limited to one, and may be multiple layers. In this case, an activation layer such as a ReLU function may be inserted after each convolutional layer.
- the reduction unit 7135 reduces the first resolution-converted image to generate a second resolution-converted image in which the number of pixels in the iris region is the same as the desired number of pixels (step S77). For example, if the enlargement ratio is 1.5 times and the quantized enlargement ratio is 2 times, the reduction unit 7135 converts the iris image LI to the first resolution image obtained by performing the double super-resolution process. The converted image may be down-sampled to a second resolution converted image having 1.5 times the number of pixels of the iris image LI. The reduction unit 7135 may perform downsampling by general thinning processing or the like.
- the post-conversion feature amount extraction unit 214 extracts the post-conversion feature amount OC, which is the feature amount of the resolution-converted image RI (step S78). [7-3: Technical Effects of Iris Authentication Device 7]
- the iris authentication device 7 of the seventh embodiment estimates and generates a conversion filter corresponding to the quantized magnification of super-resolution processing. Therefore, a single super-resolution model SM can perform super-resolution processing corresponding to various enlargement ratios.
- the iris authentication device 7 of the seventh embodiment performs up-sampling to 2, 4, 8 times, etc., using a conversion filter according to the magnification rate, and further down-samples from that size, so that the continuous magnification rate can be obtained with high accuracy. can be realized.
- the iris authentication device 7 of the seventh embodiment is particularly effective when the resolution of the resolution-converted image RI is fixed.
- the super-resolution model SM used in the iris authentication device 7 of the seventh embodiment can achieve the desired resolution regardless of the resolution of the iris image LI, as in the iris authentication devices 5 and 6 of the fifth and sixth embodiments. can output a resolution-converted image RI of Therefore, by applying to the existing iris authentication mechanism a super-resolution model SM that has been learned and constructed so as to be able to output a resolution-converted image RI corresponding to each existing iris authentication mechanism, the existing iris authentication mechanism can accurately perform iris authentication even when an iris image LI of any resolution is input. [8: Eighth Embodiment]
- an eighth embodiment of an iris authentication device, an iris authentication method, and a recording medium will be described.
- the eighth embodiment of the iris authentication device, the iris authentication method, and the recording medium will be described below using an iris authentication device 8 to which the eighth embodiment of the iris authentication device, the iris authentication method, and the recording medium is applied. explain. [8-1: Configuration of iris authentication device 8]
- FIG. 12 is a block diagram showing the configuration of an iris authentication device 8 according to the eighth embodiment.
- the iris authentication device 8 includes an iris image acquisition unit 211 , a calculation unit 212 , a generation unit 213 , a post-conversion feature amount extraction unit 214 , an authentication unit 215 and an adjustment unit 819 .
- FIG. 13 is a flow chart showing the flow of super-resolution processing performed by the iris authentication device 8 in the eighth embodiment.
- the iris authentication device 8 performs operations from step S21 to step S27 in the same manner as in the second embodiment.
- the adjustment unit 819 adjusts the threshold used for authentication by the authentication unit 215 according to the enlargement ratio (step S81). That is, in the eighth embodiment, the degree of difficulty with which the authentication unit 215 authenticates the person is adjusted according to the enlargement ratio.
- the authenticating unit 215 detects that the matching score indicating the degree of similarity between the post-conversion feature quantity extracted by the post-conversion feature quantity extraction unit 214 and the matching feature quantity prepared in advance is equal to or greater than the threshold adjusted by the adjustment unit 819. , the person is authenticated (step S26). [8-3: Technical Effects of Iris Authentication Device 8]
- the iris authentication device 8 in the eighth embodiment can adjust the threshold value used for authentication by the authentication unit 215 according to the enlargement ratio, it is preferable to change the certainty of authentication due to the enlargement ratio. Also, by adjusting the threshold used for authentication, it is possible to adjust the degree of difficulty in authenticating the person. [9: Ninth Embodiment]
- a ninth embodiment of an iris authentication device, an iris authentication system, an iris authentication method, and a recording medium will be described.
- an iris authentication device, an iris authentication system, an iris authentication method, and an authentication system 100 to which the ninth embodiment of the iris authentication device, the iris authentication method, and the recording medium are applied will be described.
- a ninth embodiment of the medium will be described. [9-1: Configuration of Iris Authentication System 100]
- FIG. 14 is a block diagram showing the configuration of the iris authentication system 100 according to the ninth embodiment.
- iris authentication system 100 includes first device 101 and second device 102 .
- the first device 101 includes an iris image acquisition section 11 and a calculation section 12 .
- the second device 102 includes a generator 13 and a post-conversion feature extractor 14 .
- the iris image acquisition unit 11 as a specific example of the ⁇ iris image acquisition means''
- the calculation unit 12 as a specific example of the ⁇ calculation means''
- the generation unit 13 as a specific example of the ⁇ generation means''
- the post-conversion feature quantity extraction unit 14 which is a specific example of the post-conversion feature quantity extraction means, may be provided in a different device.
- the first device 101 may include only the iris image acquisition unit 11
- the second device 102 may include the calculation unit 12 , the generation unit 13 , and the post-transform feature extraction unit 14 .
- the iris image acquisition unit 11, the calculation unit 12, the generation unit 13, and the post-conversion feature amount extraction unit 14 may be provided in different combinations in the first device 101 and the second device 102.
- the first device 101 and the second device 102 are capable of communication and can transmit and receive their respective processing results.
- the first device 101 includes an iris image acquisition unit 11 and a calculation unit 12
- the second device 102 includes a generation unit 13 and a post-transform feature extraction unit 14.
- the first device 10 can transmit the calculation result of the calculation unit 12 to the second device 102
- the second device 102 receives the calculation result
- the generation unit 13 generates an iris image according to the calculation result.
- a resolution-converted image can be generated by converting the resolution of the image.
- the iris authentication system may include three or more devices, and the iris image acquiring unit 11, the calculating unit 12, the generating unit 13, and the post-conversion feature amount extracting unit 14 may be combined in arbitrary combinations. device.
- magnification is 1 or more has been described, but the magnification is not limited to 1 or more, and may be less than 1.
- the iris authentication device in the above embodiment determines the magnification based on the number of pixels in the iris region included in the iris image
- the magnification may be determined regardless of the number of pixels in the iris region.
- the magnification used for resolution conversion by the iris authentication device may be determined according to the distance between the imaging device and the living body when the iris image is captured.
- the magnification used by the iris authentication device may be a magnification that enables resolution conversion to be performed appropriately. [10: Appendix]
- an iris image acquiring means for acquiring an iris image including the iris of a living body; calculating means for calculating a magnification of the iris image from the size of the iris region included in the iris image and a desired size; generating means for generating a resolution-converted image obtained by converting the resolution of the iris image according to the magnification; and post-conversion feature quantity extraction means for extracting a post-conversion feature quantity that is a feature quantity of the resolution-converted image.
- the iris authentication device according to any one of the items.
- the learning means sets the weight of the loss corresponding to the input image generated by using the first magnification as the arbitrary magnification to a second magnification smaller than the first magnification. 5.
- the iris authentication device according to appendix 3 or 4, wherein the generating means performs learning based on a loss function that is larger than the weight of the loss corresponding to the input image generated by using as .
- the input image generation means uses, as the arbitrary magnification, a second magnification lower than the first magnification as the frequency of generating the input image by using the first magnification as the arbitrary magnification. 6.
- the iris authentication device according to any one of appendices 3 to 5, wherein the plurality of input images are generated so as to be higher in frequency than the input images are generated.
- the generating means is a pre-conversion feature amount extraction means for extracting a pre-conversion feature amount that is a feature amount of the iris image; filter generation means for generating one or more conversion filters for converting the pre-conversion feature value according to the magnification; and transforming means for generating the resolution-converted image by transforming the pre-transformed feature amount by one or more filtering processes using the one or more transform filters.
- the generating means is a pre-conversion feature amount extraction means for extracting a pre-conversion feature amount that is a feature amount of the iris image; a magnification feature amount extracting means for extracting a magnification feature amount that is the feature amount of the magnification; and converting means for generating the resolution-converted image by synthesizing the pre-conversion feature amount and the magnification feature amount and converting the pre-conversion feature amount.
- the generating means is a pre-conversion feature amount extraction means for extracting a pre-conversion feature amount that is a feature amount of the iris image; quantization means for quantizing the magnification to a predetermined magnification; a filter generating means for generating one or more transform filters for transforming the pre-transformed feature quantity according to the quantized magnification; conversion means for generating a first resolution-converted image by converting the pre-conversion feature amount by one or more filtering processes using the one or more conversion filters; and reducing means for reducing the first resolution-converted image to generate a second resolution-converted image in which the size of the iris region is the same as the desired size.
- iris authentication device is a pre-conversion feature amount extraction means for extracting a pre-conversion feature amount that is a feature amount of the iris image; quantization means for quantizing the magnification to a predetermined magnification; a filter generating means for generating one or more transform filters for transforming the pre-transformed feature quantity according to
- Appendix 10 a determination means for determining that the person is a person when a matching score indicating the degree of similarity between the post-conversion feature extracted by the post-conversion feature extraction means and a pre-prepared matching feature is equal to or greater than a threshold; 10.
- the iris authentication device according to any one of appendices 1 to 9, further comprising adjusting means for adjusting the threshold according to the magnification.
- the iris authentication device according to 1.
- an iris image acquiring means for acquiring an iris image including the iris of a living body; calculating means for calculating a magnification of the iris image from the size of the iris region included in the iris image and a desired size; generating means for generating a resolution-converted image obtained by converting the resolution of the iris image according to the magnification; an iris authentication system comprising post-conversion feature amount extraction means for extracting a post-conversion feature amount that is a feature amount of the resolution-converted image.
- [Appendix 13] Acquiring an iris image containing the iris of a living body, calculating a magnification for the iris image from the size of the iris region included in the iris image and a desired size; generating a resolution-converted image obtained by converting the resolution of the iris image according to the magnification; An iris authentication method for extracting a post-conversion feature amount that is a feature amount of the resolution-converted image.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Security & Cryptography (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computer Hardware Design (AREA)
- Software Systems (AREA)
- General Engineering & Computer Science (AREA)
- Collating Specific Patterns (AREA)
Abstract
Description
[1:第1実施形態]
[1-1:虹彩認証装置1の構成]
虹彩認証では、虹彩の模様を用いた認証を行う。このため、虹彩認証では、その虹彩認証に適切な画素数の虹彩領域を含む画像を用いる必要がある。所望の画素数とは、その虹彩認証に適切な画素数であってよい。虹彩は略円形であるので、画素数は、該当領域の半径で表してもよい。略円形の虹彩領域を虹彩円とも呼ぶ。また、画素数は、解像度と対応していてもよい。例えば、所望の画素数は、100画素以上であってよく、100画素、125画素等であってよい。
[1-2:虹彩認証装置1の技術的効果]
[2:第2実施形態]
[2-1:虹彩認証装置2の構成]
[2-2:虹彩認証装置2が行う虹彩認証動作]
虹彩円検出部2121は、虹彩画像から虹彩円を検出する(ステップS22)。虹彩円検出部2121は、入力された虹彩画像から、虹彩円の中心位置と半径を表すベクトルを算出してもよい。虹彩円検出部2121は、例えば、回帰ニューラルネットワークで構成されていてもよい。回帰ニューラルネットワークは、複数の畳み込み層と複数の活性化層とを含み、入力画像の特徴量を抽出し、抽出した特徴量を、線形層により、該当領域の中心位置と半径を表すベクトルに変換することができる。虹彩円検出部2121に入力される虹彩画像LIと、虹彩円検出部2121から出力されるベクトルとは正規化されていてもよい。虹彩円検出部2121をニューラルネットワークとして構築する場合、要件にあっていれば、どのような構造のニューラルネットワークも用いることができる。ニューラルネットワークの構造としては、例えば、大規模画像データセットで学習されたモデルであるVGG、ResNet(Residual neural network)等の構造と同様のものを挙げることができるが、これら以外の構造を用いてもよい。ニューラルネットワークの中間層として、バッチノーマライゼーション等の正規化層を用いてもよい。活性化層としては、ReLU(Rectified Linear Unit)を使う場合が多いが、他の活性化関数を用いてもよい。また、虹彩円検出部2121は、ニューラルネットワークで構成されていない画像処理の機構であってもよい。
[2-3:虹彩認証装置2の技術的効果]
[3:第3実施形態]
[3-1:虹彩認証装置3の構成]
[3-2:虹彩認証装置3が行う学習動作]
[3-3:虹彩認証装置3の技術的効果]
また、第3実施形態における虹彩認証装置3は、学習特徴量TCと入力特徴量OICとが類似しないほど損失が大きくなる損失関数に基づいて、生成部213に解像度変換画像RIの生成方法の学習を行わせるので、虹彩認証に適した解像度変換画像RIを生成できる。すなわち、超解像処理の学習に、超解像処理を施された画像から抽出された特徴量を用いるので、虹彩認証に適した特徴量を抽出することのできる解像度変換画像RIを生成できる超解像モデルSMを構築することができる。虹彩情報推定部300が出力する変換後特徴量OCは、本人と認証するために用いる特徴量であるので、解像度変換画像RIは、本人と認証するのに適切な変換後特徴量OCの抽出ができる画像であることが望ましい。言い換えると、超解像モデルSMが生成する解像度変換画像RIは、精度よく超解像処理が施された画像であり、照合に適した画像でもある。
[4:第4実施形態]
[4-1:虹彩認証装置3が行う学習動作]
[4-2:虹彩認証装置4の技術的効果]
[5:第5実施形態]
[5-1:虹彩認証装置5の構成]
[5-2:生成部513が行う超解像処理]
[5-3:虹彩認証装置5の技術的効果]
[6:第6実施形態]
[6-1:虹彩認証装置6の構成]
[6-2:生成部613が行う超解像処理]
[6-3:虹彩認証装置6の技術的効果]
[7:第7実施形態]
[7-1:虹彩認証装置7の構成]
[7-2:生成部713が行う超解像処理]
[7-3:虹彩認証装置7の技術的効果]
[8:第8実施形態]
[8-1:虹彩認証装置8の構成]
[8-2:虹彩認証装置8が行う虹彩認証動作]
認証部215は、変換後特徴量抽出部214が抽出した変換後特徴量と、予め用意された照合特徴量との類似度を示す照合スコアが、調整部819が調整した閾値以上である場合に、本人と認証する(ステップS26)。
[8-3:虹彩認証装置8の技術的効果]
[9:第9実施形態]
[9-1:虹彩認証システム100の構成]
第1装置101と第2装置102とは、通信可能であり、各々の処理結果を送受信することができる。図14に示すように、第1装置101が虹彩画像取得部11と、算出部12とを備え、第2装置102が生成部13と、変換後特徴量抽出部14とを備えている場合を例に挙げて説明する。この場合、第1装置10は、算出部12の算出結果を第2装置102に送信することができ、第2装置102は当該算出結果を受信し、生成部13は、当該算出結果応じて虹彩画像の解像度を変換した解像度変換画像を生成することができる。さらに、虹彩認証システムは、3以上の装置を含んでいてもよく、虹彩画像取得部11と、算出部12と、生成部13と、変換後特徴量抽出部14とは、任意の組み合わせで各々の装置に備わっていてもよい。
[10:付記]
[付記1]
生体の虹彩を含む虹彩画像を取得する虹彩画像取得手段と、
前記虹彩画像に含まれる虹彩領域の大きさと、所望の大きさとから、前記虹彩画像に対する倍率を算出する算出手段と、
前記倍率に応じて前記虹彩画像の解像度を変換した解像度変換画像を生成する生成手段と、
前記解像度変換画像の特徴量である変換後特徴量を抽出する変換後特徴量抽出手段と
を備える虹彩認証装置。
[付記2]
前記虹彩領域の大きさは、前記所望の大きさより小さく、
前記倍率は、拡大率であり、
前記生成手段は、前記拡大率に応じて、前記虹彩画像を高解像化した超解像画像を生成する
付記1に記載の虹彩認証装置。
[付記3]
前記所望の大きさの虹彩領域を含む学習画像を取得する学習画像取得手段と、
任意の倍率の逆数に応じて前記学習画像の解像度を変換した入力画像を生成する入力画像生成手段とを更に備え、
前記生成手段は、前記任意の倍率に応じて前記入力画像の解像度を変換した、前記学習画像と同じ解像度の解像度変換入力画像を生成し、
前記学習画像と前記解像度変換入力画像とが類似しないほど損失が大きくなる損失関数に基づいて、前記生成手段に前記解像度変換画像の生成方法の学習を行わせる学習手段を更に備える
付記1または2に記載の虹彩認証装置。
[付記4]
前記所望の大きさの虹彩領域を含む学習画像を取得する学習画像取得手段と、
任意の倍率の逆数に応じて前記学習画像の解像度を変換した入力画像を生成する入力画像生成手段とを更に備え、
前記生成手段は、前記任意の倍率に応じて前記入力画像の解像度を変換した、前記学習画像と同じ解像度の解像度変換入力画像を生成し、
前記変換後特徴量抽出手段は、前記学習画像の特徴量である学習特徴量と、前記解像度変換入力画像の特徴量である入力特徴量とを抽出し、
前記学習特徴量と前記入力特徴量とが類似しないほど損失が大きくなる損失関数に基づいて、前記生成手段に前記解像度変換画像の生成方法の学習を行わせる学習手段を更に備える
付記1から3の何れか1項に記載の虹彩認証装置。
[付記5]
前記学習手段は、第1の倍率を前記任意の倍率として用いることで生成される前記入力画像に対応する前記損失の重みが、前記第1の倍率よりも小さい第2の倍率を前記任意の倍率として用いることで生成される前記入力画像に対応する前記損失の重みよりも大きくなる損失関数に基づいて、前記生成手段に学習を行わせる
付記3または4に記載の虹彩認証装置。
[付記6]
前記入力画像生成手段は、第1の倍率を前記任意の倍率として用いることで前記入力画像を生成する頻度が、前記第1の倍率よりも小さい第2の倍率を前記任意の倍率として用いることで前記入力画像を生成する頻度よりも高くなるように、複数の前記入力画像を生成する
付記3から5の何れか1項に記載の虹彩認証装置。
[付記7]
前記生成手段は、
前記虹彩画像の特徴量である変換前特徴量を抽出する変換前特徴量抽出手段と、
前記倍率に応じて、前記変換前特徴量を変換するための1つ以上の変換フィルタを生成するフィルタ生成手段と、
前記1つ以上の変換フィルタを用いた1つ以上のフィルタ処理により前記変換前特徴量を変換することで、前記解像度変換画像を生成する変換手段と
を含む
付記1から6の何れか1項に記載の虹彩認証装置。
[付記8]
前記生成手段は、
前記虹彩画像の特徴量である変換前特徴量を抽出する変換前特徴量抽出手段と、
前記倍率の特徴量である倍率特徴量を抽出する倍率特徴量抽出手段と、
前記変換前特徴量と前記倍率特徴量とを合成して、前記変換前特徴量を変換することで、前記解像度変換画像を生成する変換手段と
を含む
付記1から7の何れか1項に記載の虹彩認証装置。
[付記9]
前記生成手段は、
前記虹彩画像の特徴量である変換前特徴量を抽出する変換前特徴量抽出手段と、
前記倍率を所定の倍率に量子化する量子化手段と、
前記量子化された倍率に応じて、前記変換前特徴量を変換するための1つ以上の変換フィルタを生成するフィルタ生成手段と、
前記1つ以上の変換フィルタを用いた1つ以上のフィルタ処理により前記変換前特徴量を変換することで、第1解像度変換画像を生成する変換手段と、
当該第1解像度変換画像を縮小して、前記虹彩領域の大きさが前記所望の大きさと同じである第2解像度変換画像を生成する縮小手段と
を含む
付記1から8の何れか1項に記載の虹彩認証装置。
[付記10]
前記変換後特徴量抽出手段が抽出した変換後特徴量と、予め用意された照合特徴量との類似度を示す照合スコアが閾値以上である場合に、本人と判定する判定手段と、
前記倍率に応じて前記閾値を調整する調整手段と
を更に備える
付記1から9の何れか1項に記載の虹彩認証装置。
[付記11]
前記変換後特徴量と、予め用意された特徴量との類似度を示すスコアを用いて認証を行う認証手段と、
前記認証手段の認証結果、前記倍率、及び前記解像度変換画像を当該虹彩認証装置の外部へ出力し、前記倍率が所定以上の場合はアラートを出力する出力手段と
を更に備える
付記1から10の何れか1項に記載の虹彩認証装置。
[付記12]
生体の虹彩を含む虹彩画像を取得する虹彩画像取得手段と、
前記虹彩画像に含まれる虹彩領域の大きさと、所望の大きさとから、前記虹彩画像に対する倍率を算出する算出手段と、
前記倍率に応じて前記虹彩画像の解像度を変換した解像度変換画像を生成する生成手段と、
前記解像度変換画像の特徴量である変換後特徴量を抽出する変換後特徴量抽出手段と
を備える虹彩認証システム。
[付記13]
生体の虹彩を含む虹彩画像を取得し、
前記虹彩画像に含まれる虹彩領域の大きさと、所望の大きさとから、前記虹彩画像に対する倍率を算出し、
前記倍率に応じて前記虹彩画像の解像度を変換した解像度変換画像を生成し、
前記解像度変換画像の特徴量である変換後特徴量を抽出する
虹彩認証方法。
[付記14]
コンピュータに、
生体の虹彩を含む虹彩画像を取得し、
前記虹彩画像に含まれる虹彩領域の大きさと、所望の大きさとから、前記虹彩画像に対する倍率を算出し、
前記倍率に応じて前記虹彩画像の解像度を変換した解像度変換画像を生成し、
前記解像度変換画像の特徴量である変換後特徴量を抽出する
虹彩認証方法を実行させるためのコンピュータプログラムが記録された記録媒体。
11,211 虹彩画像取得部
12,212 算出部
2121 虹彩円検出部
2122 拡大率算出部
13,213,513,613,713 生成部
14,214 変換後特徴量抽出部
215 認証部
21 演算装置
22 記憶装置
300 虹彩情報推定部
316 学習画像取得部
317 入力画像生成部
3171 バッチデータ抽出部
3172 解像度変換部
318 学習部
3181 損失関数計算部
3182 勾配計算部
3183 パラメータ更新部
5131,6131,7131 特徴量抽出部
5132,7133 フィルタ生成部
5133,6134,7134 変換部
6132 拡大率特徴量抽出部
6133 合成部
7132 量子化部
7135 縮小部
819 調整部
100 虹彩認証システム
SM 超解像モデル
GM 特徴量生成モデル
LI,H1 虹彩画像
RI 解像度変換画像
PC 変換前特徴量
OC 変換後特徴量
TI 学習画像
TC 学習特徴量
II 入力画像
RII 解像度変換入力画像
OIC 入力特徴量
RC 拡大率特徴量
CC 照合特徴量
Claims (14)
- 生体の虹彩を含む虹彩画像を取得する虹彩画像取得手段と、
前記虹彩画像に含まれる虹彩領域の大きさと、所望の大きさとから、前記虹彩画像に対する倍率を算出する算出手段と、
前記倍率に応じて前記虹彩画像の解像度を変換した解像度変換画像を生成する生成手段と、
前記解像度変換画像の特徴量である変換後特徴量を抽出する変換後特徴量抽出手段と
を備える虹彩認証装置。 - 前記虹彩領域の大きさは、前記所望の大きさより小さく、
前記倍率は、拡大率であり、
前記生成手段は、前記拡大率に応じて、前記虹彩画像を高解像化した超解像画像を前記解像度変換画像として生成する
請求項1に記載の虹彩認証装置。 - 前記所望の大きさの虹彩領域を含む学習画像を取得する学習画像取得手段と、
任意の倍率の逆数に応じて前記学習画像の解像度を変換した入力画像を生成する入力画像生成手段とを更に備え、
前記生成手段は、前記任意の倍率に応じて前記入力画像の解像度を変換した、前記学習画像と同じ解像度の解像度変換入力画像を生成し、
前記学習画像と前記解像度変換入力画像とが類似しないほど損失が大きくなる損失関数に基づいて、前記生成手段に前記解像度変換画像の生成方法の学習を行わせる学習手段を更に備える
請求項1または2に記載の虹彩認証装置。 - 前記所望の大きさの虹彩領域を含む学習画像を取得する学習画像取得手段と、
任意の倍率の逆数に応じて前記学習画像の解像度を変換した入力画像を生成する入力画像生成手段とを更に備え、
前記生成手段は、前記任意の倍率に応じて前記入力画像の解像度を変換した、前記学習画像と同じ解像度の解像度変換入力画像を生成し、
前記変換後特徴量抽出手段は、前記学習画像の特徴量である学習特徴量と、前記解像度変換入力画像の特徴量である入力特徴量とを抽出し、
前記学習特徴量と前記入力特徴量とが類似しないほど損失が大きくなる損失関数に基づいて、前記生成手段に前記解像度変換画像の生成方法の学習を行わせる学習手段を更に備える
請求項1から3の何れか1項に記載の虹彩認証装置。 - 前記学習手段は、第1の倍率を前記任意の倍率として用いることで生成される前記入力画像に対応する前記損失の重みが、前記第1の倍率よりも小さい第2の倍率を前記任意の倍率として用いることで生成される前記入力画像に対応する前記損失の重みよりも大きくなる損失関数に基づいて、前記生成手段に学習を行わせる
請求項3または4に記載の虹彩認証装置。 - 前記入力画像生成手段は、第1の倍率を前記任意の倍率として用いることで前記入力画像を生成する頻度が、前記第1の倍率よりも小さい第2の倍率を前記任意の倍率として用いることで前記入力画像を生成する頻度よりも高くなるように、複数の前記入力画像を生成する
請求項3から5の何れか1項に記載の虹彩認証装置。 - 前記生成手段は、
前記虹彩画像の特徴量である変換前特徴量を抽出する変換前特徴量抽出手段と、
前記倍率に応じて、前記変換前特徴量を変換するための1つ以上の変換フィルタを生成するフィルタ生成手段と、
前記1つ以上の変換フィルタを用いた1つ以上のフィルタ処理により前記変換前特徴量を変換することで、前記解像度変換画像を生成する変換手段と
を含む
請求項1から6の何れか1項に記載の虹彩認証装置。 - 前記生成手段は、
前記虹彩画像の特徴量である変換前特徴量を抽出する変換前特徴量抽出手段と、
前記倍率の特徴量である倍率特徴量を抽出する倍率特徴量抽出手段と、
前記変換前特徴量と前記倍率特徴量とを合成して、前記変換前特徴量を変換することで、前記解像度変換画像を生成する変換手段と
を含む
請求項1から7の何れか1項に記載の虹彩認証装置。 - 前記生成手段は、
前記虹彩画像の特徴量である変換前特徴量を抽出する変換前特徴量抽出手段と、
前記倍率を所定の倍率に量子化する量子化手段と、
前記量子化された倍率に応じて、前記変換前特徴量を変換するための1つ以上の変換フィルタを生成するフィルタ生成手段と、
前記1つ以上の変換フィルタを用いた1つ以上のフィルタ処理により前記変換前特徴量を変換することで、第1解像度変換画像を生成する変換手段と、
当該第1解像度変換画像を縮小して、前記虹彩領域の大きさが前記所望の大きさと同じである第2解像度変換画像を生成する縮小手段と
を含む
請求項1から8の何れか1項に記載の虹彩認証装置。 - 前記変換後特徴量抽出手段が抽出した変換後特徴量と、予め用意された照合特徴量との類似度を示す照合スコアが閾値以上である場合に、本人と判定する判定手段と、
前記倍率に応じて前記閾値を調整する調整手段と
を更に備える
請求項1から9の何れか1項に記載の虹彩認証装置。 - 前記変換後特徴量と、予め用意された特徴量との類似度を示すスコアを用いて認証を行う認証手段と、
前記認証手段の認証結果、前記倍率、及び前記解像度変換画像を当該虹彩認証装置の外部へ出力し、前記倍率が所定以上の場合はアラートを出力する出力手段と
を更に備える
請求項1から10の何れか1項に記載の虹彩認証装置。 - 生体の虹彩を含む虹彩画像を取得する虹彩画像取得手段と、
前記虹彩画像に含まれる虹彩領域の大きさと、所望の大きさとから、前記虹彩画像に対する倍率を算出する算出手段と、
前記倍率に応じて前記虹彩画像の解像度を変換した解像度変換画像を生成する生成手段と、
前記解像度変換画像の特徴量である変換後特徴量を抽出する変換後特徴量抽出手段と
を備える虹彩認証システム。 - 生体の虹彩を含む虹彩画像を取得し、
前記虹彩画像に含まれる虹彩領域の大きさと、所望の大きさとから、前記虹彩画像に対する倍率を算出し、
前記倍率に応じて前記虹彩画像の解像度を変換した解像度変換画像を生成し、
前記解像度変換画像の特徴量である変換後特徴量を抽出する
虹彩認証方法。 - コンピュータに、
生体の虹彩を含む虹彩画像を取得し、
前記虹彩画像に含まれる虹彩領域の大きさと、所望の大きさとから、前記虹彩画像に対する倍率を算出し、
前記倍率に応じて前記虹彩画像の解像度を変換した解像度変換画像を生成し、
前記解像度変換画像の特徴量である変換後特徴量を抽出する
虹彩認証方法を実行させるためのコンピュータプログラムが記録された記録媒体。
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/JP2021/040123 WO2023073954A1 (ja) | 2021-10-29 | 2021-10-29 | 虹彩認証装置、虹彩認証システム、虹彩認証方法、及び、記録媒体 |
JP2023556058A JPWO2023073954A1 (ja) | 2021-10-29 | 2021-10-29 |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/JP2021/040123 WO2023073954A1 (ja) | 2021-10-29 | 2021-10-29 | 虹彩認証装置、虹彩認証システム、虹彩認証方法、及び、記録媒体 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2023073954A1 true WO2023073954A1 (ja) | 2023-05-04 |
Family
ID=86157643
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2021/040123 WO2023073954A1 (ja) | 2021-10-29 | 2021-10-29 | 虹彩認証装置、虹彩認証システム、虹彩認証方法、及び、記録媒体 |
Country Status (2)
Country | Link |
---|---|
JP (1) | JPWO2023073954A1 (ja) |
WO (1) | WO2023073954A1 (ja) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2002259981A (ja) * | 2001-02-28 | 2002-09-13 | Matsushita Electric Ind Co Ltd | 個人認証方法および装置 |
JP2008090483A (ja) * | 2006-09-29 | 2008-04-17 | Oki Electric Ind Co Ltd | 個人認証システム及び個人認証方法 |
JP2009282925A (ja) * | 2008-05-26 | 2009-12-03 | Sharp Corp | 虹彩認証支援装置及び虹彩認証支援方法 |
JP2020071627A (ja) * | 2018-10-31 | 2020-05-07 | ソニーセミコンダクタソリューションズ株式会社 | 画像処理装置および画像処理方法 |
-
2021
- 2021-10-29 WO PCT/JP2021/040123 patent/WO2023073954A1/ja active Application Filing
- 2021-10-29 JP JP2023556058A patent/JPWO2023073954A1/ja active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2002259981A (ja) * | 2001-02-28 | 2002-09-13 | Matsushita Electric Ind Co Ltd | 個人認証方法および装置 |
JP2008090483A (ja) * | 2006-09-29 | 2008-04-17 | Oki Electric Ind Co Ltd | 個人認証システム及び個人認証方法 |
JP2009282925A (ja) * | 2008-05-26 | 2009-12-03 | Sharp Corp | 虹彩認証支援装置及び虹彩認証支援方法 |
JP2020071627A (ja) * | 2018-10-31 | 2020-05-07 | ソニーセミコンダクタソリューションズ株式会社 | 画像処理装置および画像処理方法 |
Also Published As
Publication number | Publication date |
---|---|
JPWO2023073954A1 (ja) | 2023-05-04 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US11069047B2 (en) | Image processing method, image processing apparatus, computing device and computer-readable storage medium | |
JP6825269B2 (ja) | 画像認識に用いられるニューラルネットワークの訓練方法及び訓練装置 | |
JP2023545565A (ja) | 画像検出方法、モデルトレーニング方法、画像検出装置、トレーニング装置、機器及びプログラム | |
CN110322400B (zh) | 图像处理方法及装置、图像处理系统及其训练方法 | |
CN109361934B (zh) | 图像处理方法、装置、设备及存储介质 | |
WO2018086354A1 (zh) | 图像升频系统及其训练方法、以及图像升频方法 | |
Liu et al. | Dynast: Dynamic sparse transformer for exemplar-guided image generation | |
US8605118B2 (en) | Matrix generation apparatus, method, and program, and information processing apparatus | |
CN112837224A (zh) | 一种基于卷积神经网络的超分辨率图像重建方法 | |
CN113361251A (zh) | 一种基于多阶段生成对抗网络的文本生成图像方法及系统 | |
KR20210045828A (ko) | 영상 처리 장치 및 방법 | |
Chira et al. | Image super-resolution with deep variational autoencoders | |
WO2021213188A1 (zh) | 图像处理模型的训练方法、图像处理方法及对应的装置 | |
Wei et al. | Improving resolution of medical images with deep dense convolutional neural network | |
CN112508782A (zh) | 网络模型的训练方法、人脸图像超分辨率重建方法及设备 | |
CN112132741A (zh) | 一种人脸照片图像和素描图像的转换方法及系统 | |
US12100071B2 (en) | System and method for generating images of the same style based on layout | |
KR20210116922A (ko) | 초해상도 모델의 메타 러닝을 통한 빠른 적응 방법 및 장치 | |
CN114612289A (zh) | 风格化图像生成方法、装置及图像处理设备 | |
WO2023073954A1 (ja) | 虹彩認証装置、虹彩認証システム、虹彩認証方法、及び、記録媒体 | |
Ismail et al. | Sparse data-based image super-resolution with ANFIS interpolation | |
Chen et al. | Linear-ResNet GAN-based anime style transfer of face images | |
CN114298979B (zh) | 一种局灶性病变征象描述指导的肝核磁影像序列生成方法 | |
Hasan et al. | Single image super-resolution using back-propagation neural networks | |
Li et al. | Image Segmentation with Fuzzy Clustering Based on Generalized Entropy. |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 21962494 Country of ref document: EP Kind code of ref document: A1 |
|
ENP | Entry into the national phase |
Ref document number: 2023556058 Country of ref document: JP Kind code of ref document: A |
|
WWE | Wipo information: entry into national phase |
Ref document number: 18705213 Country of ref document: US |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 21962494 Country of ref document: EP Kind code of ref document: A1 |