KR20240006814A - Method and system for generating training data for deep learning for identification recognition - Google Patents

Abstract

The method of generating training data for deep learning for ID recognition according to an embodiment of the present invention is a method in which a training application executed by at least one processor of a terminal generates training data for deep learning for ID recognition, where a physical ID card is photographed. Obtaining an ID sample image; generating a background template image by removing text from the obtained ID sample image; generating a hologram template image including at least a partial hologram pattern area on the obtained ID sample image; A step of generating a text template image in which various types of ID information randomly generated according to a predetermined standard are arranged according to a predetermined ID card format; and generating a composite image combining the generated background template image, text template image, and hologram template image into one.

Description

Method and system for generating training data for ID recognition deep learning {METHOD AND SYSTEM FOR GENERATING TRAINING DATA FOR DEEP LEARNING FOR IDENTIFICATION RECOGNITION}

The present invention relates to a method and system for generating training data for ID recognition deep learning. More specifically, it relates to a method and system for generating a training data set to improve the performance of an ID recognition deep learning model.

With the development of information and communications technology (ICT), various services that require user identity authentication or information, such as those at banks or securities firms, are being provided online.

For example, in order to use internet banking, a user provides an image of his or her identification card (for example, a resident registration card or driver's license, etc.) and performs identity authentication based on image processing.

In order to provide the above identity recognition service, an automated process is implemented in the prior art to obtain an image of a predetermined ID card, extract a predetermined character from the obtained image, and perform reading through a computer. do.

There are various ways to implement such ID recognition, and in the past, optical character recognition (OCR) technology was mainly used, but with the recent development of artificial intelligence technology, deep learning-based ID recognition technology using deep neural networks has been developed. It is mainly comprised of

In detail, in deep learning-based ID recognition, a certain neural network model is trained to be specialized for ID recognition, and deep learning-based ID recognition is performed using the learned neural network model to implement faster and more accurate data processing compared to general image processing. You can.

In order to perform such deep learning-based ID recognition, a large amount of training data is required to learn the corresponding neural network model.

However, in the conventional method, there is a problem in that it is difficult to collect a large amount of training data (eg, physical ID image, etc.) required for neural network learning due to problems such as personal information issues.

In addition, even if some images of physical ID cards are collected, there is a high possibility that they contain hologram patterns or various noises (e.g., light reflection and/or motion blur, etc.) to prevent forgery/falsification of ID cards, so these obstacles It is difficult to solve the problem because there is a limitation that a large amount of data must be collected to cover the problems.

Because of these problems, a type of sample ID image that can avoid personal information issues has been provided in the past, but this is significantly different from the actual ID card image, so its effectiveness is insufficient.

As such, in the past, it was actually very difficult to collect a large amount of physical ID images, so deep learning learning aimed at improving the performance of the neural network model for ID recognition was not performed smoothly.

Therefore, in order to solve these problems, it is very necessary to develop and introduce new technologies that can easily build a large training data set that can effectively support deep learning learning of neural network models for ID recognition.

KR 10-2142347 B1

The present invention was created to solve the problems described above, and its purpose is to provide a method and system for generating a training data set to improve the performance of an ID card recognition deep learning model.

In detail, the present invention seeks to provide a training data generation method and system for deep learning ID card recognition that generates a virtual ID card image by digitizing each step of physically producing an ID card.

In addition, the present invention provides a training data generation method and system for deep learning ID card recognition that generates a virtual ID card image reflecting the characteristics of the image of the physical ID card (e.g., holographic patterns or various noises that vary depending on the shooting angle). We would like to provide.

In addition, the present invention seeks to provide a method and system for generating training data for ID card recognition deep learning that generates a virtual ID card image that embodies various types of noise that can appear in a physical ID card from various angles.

In addition, the present invention seeks to provide a method and system for generating training data for deep learning for ID card recognition that supports learning a deep learning model for recognizing a given ID card based on training data based on the virtual ID card image generated as above. do.

However, the technical problems to be achieved by the present invention and embodiments of the present invention are not limited to the technical problems described above, and other technical problems may exist.

The method of generating training data for deep learning for ID recognition according to an embodiment of the present invention is a method in which a training application executed by at least one processor of a terminal generates training data for deep learning for ID recognition, where a physical ID card is photographed. Obtaining an ID sample image; generating a background template image by removing text from the obtained ID sample image; generating a hologram template image including at least a partial hologram pattern area on the obtained ID sample image; A step of generating a text template image in which various types of ID information randomly generated according to a predetermined standard are arranged according to a predetermined ID card format; and generating a composite image combining the generated background template image, text template image, and hologram template image into one.

At this time, the step of generating the background template image includes generating a mask image by masking the text area in the ID sample image, and applying the generated mask image and the ID card sample image to a first deep learning model. It includes a step of inputting the background template image and obtaining the background template image from a first deep learning model that has received the image.

Additionally, generating the mask image includes expanding the masked text area based on a predetermined dilation function.

In addition, the step of generating the background template image further includes performing inpainting processing of applying a background color according to the ID sample image to the text area removed from the background template image.

In addition, the step of generating the hologram template image includes generating a hologram RGB table based on the hologram pattern area, and the hologram RGB table includes pixels displaying the hologram pattern area. It is a set of hologram estimated range data that specifies RGB values within a certain range based on RGB values.

In addition, the step of generating the hologram template image includes detecting invalid pixels that do not meet the hologram estimate range in the ID card sample image based on the hologram RGB table, and masking the detected invalid pixel area. It further includes generating a hologram template image.

Additionally, generating the composite image includes generating a background-text combination image by rendering the text template image on the background template image.

In addition, generating the composite image further includes generating the composite image based on an alpha-blending operation based on the background-text combination image and the hologram template image.

In addition, the step of generating the composite image includes generating a plurality of hologram crop images by randomly dividing the hologram template image based on a random crop operation based on the hologram template image, and the generated plurality of hologram crop images. It further includes generating the composite image based on an alpha blending operation based on the hologram crop image and the background-text combination image.

In addition, the step of generating the composite image based on an alpha blending operation based on the plurality of hologram crop images and the background-text combination image includes placing each of the plurality of hologram crop images at a random position on the background-text combination image. Each includes a synthesis step.

In addition, the method of generating training data for deep learning for ID recognition according to an embodiment of the present invention further includes the step of augmenting the amount of data by implementing a predetermined noise on the generated synthetic image.

In addition, the method of generating training data for ID recognition deep learning according to an embodiment of the present invention further includes the step of training a predetermined ID recognition deep learning model based on the generated synthetic image.

Meanwhile, a training data generation system for ID recognition deep learning according to an embodiment of the present invention includes at least one memory storing a training application; And at least one processor that reads the training application stored in the memory to generate training data for ID recognition deep learning, wherein instructions of the training application include: acquiring an ID sample image obtained by photographing a physical ID card; , generating a background template image by removing text in the obtained ID sample image, generating a hologram template image including at least a portion of the hologram pattern area on the obtained ID sample image, and randomly randomizing according to a predetermined standard. A step of generating a text template image in which various types of ID information generated according to a predetermined ID card format are arranged, and a step of generating a composite image combining the generated background template image, text template image, and hologram template image into one. Contains commands.

At this time, the command of the training application includes generating a background-text combination image by rendering the text template image on the background template image, and alpha based on the background-text combination image and the hologram template image. It further includes instructions for generating the composite image based on an alpha-blending operation.

On the other hand, the data generation server according to an embodiment of the present invention includes at least one memory module; and at least one processor module; at least one application stored in the memory module and executed by the processor module, wherein the at least one application acquires an ID sample image obtained by photographing a physical ID card, Creating a background template image by removing the text in the obtained ID sample image, generating a hologram template image including at least a portion of the hologram pattern area on the obtained ID sample image, and various types of ID information randomly generated according to a predetermined standard. A text template image is arranged according to a predetermined ID card format, and a composite image is created by combining the created background template image, text template image, and hologram template image into one.

At this time, the at least one application generates a background-text combination image by rendering the text template image on the background template image, and performs alpha blending (alpha blending) based on the background-text combination image and the hologram template image. The composite image is generated based on Alpha-blending) operation.

The method and system for generating training data for ID recognition deep learning according to an embodiment of the present invention generates a training data set to improve the performance of the ID recognition deep learning model, thereby providing ID recognition deep learning that is difficult to collect due to personal information issues, etc. It has the effect of easily building a training data set for learning learning.

In addition, the training data generation method and system for ID recognition deep learning according to an embodiment of the present invention digitizes each step of physically producing an ID card to generate a virtual ID image, thereby meeting the requirements required for ID recognition deep learning. Even if training data is not individually collected from physical ID cards, it can be obtained and utilized in large quantities in an easy, quick and efficient manner, which has the effect of significantly reducing the time and cost required to build the training data.

In addition, the method and system for generating training data for deep learning for ID recognition according to an embodiment of the present invention is a virtual ID that reflects the characteristics of the image captured by the physical ID card (e.g., holographic patterns or various noises that vary depending on the shooting angle). By creating an ID card image, a virtual ID card image with a more realistic shape can be created and the performance and quality of deep learning learning using this can be improved.

In addition, the training data generation method and system for deep learning for ID recognition according to an embodiment of the present invention generates a virtual ID image that embodies various types of noise that can appear in a physical ID card from various angles, The deep learning model learned based on this can prevent the deterioration of ID recognition performance due to various noises that may occur on the image that is the target of ID recognition, and thus provides a deep learning model for ID recognition that provides more robust recognition performance. It has the effect of providing.

In addition, the method and system for generating training data for deep learning for ID recognition according to an embodiment of the present invention involves learning a deep learning model for recognizing a predetermined ID based on training data based on the virtual ID image generated as above. By supporting this, it is possible to further improve the performance by enabling learning of the deep learning model based on a larger amount of training data, and various ID recognition services based on the deep learning model (e.g., online-based banking and It has the effect of improving the quality, safety, and reliability of (or identity authentication services for securities company work, etc.).

However, the effects that can be obtained from the present invention are not limited to the effects mentioned above, and other effects not mentioned can be clearly understood from the description below.

Figure 1 is a conceptual diagram of a training data generation system for ID card recognition deep learning according to an embodiment of the present invention.
Figure 2 is an internal block diagram of a terminal according to an embodiment of the present invention.
Figure 3 is a flowchart illustrating a method of generating training data for ID card recognition deep learning according to an embodiment of the present invention.
Figure 4 is an example of a diagram for explaining a method of generating a background template image according to an embodiment of the present invention.
Figure 5 is an example of a hologram template image according to an embodiment of the present invention.
Figure 6 is an example of a diagram for explaining a method of generating a hologram RGB table according to an embodiment of the present invention.
Figure 7 is an example of a text template image according to an embodiment of the present invention.
Figure 8 is an example of a drawing for explaining a method of generating a composite image according to an embodiment of the present invention.
Figure 9 is an example of a diagram for explaining a method of generating a composite image based on a plurality of hologram crop images according to an embodiment of the present invention.
10 to 12 are examples of drawings for explaining noise for data enhancement according to an embodiment of the present invention.

Since the present invention can be modified in various ways and can have various embodiments, specific embodiments will be illustrated in the drawings and described in detail in the detailed description. The effects and features of the present invention and methods for achieving them will become clear by referring to the embodiments described in detail below along with the drawings. However, the present invention is not limited to the embodiments disclosed below and may be implemented in various forms. In the following embodiments, terms such as first and second are used not in a limiting sense but for the purpose of distinguishing one component from another component. Additionally, singular expressions include plural expressions unless the context clearly dictates otherwise. Additionally, terms such as include or have mean that the features or components described in the specification exist, and do not preclude the possibility of adding one or more other features or components. Additionally, in the drawings, the sizes of components may be exaggerated or reduced for convenience of explanation. For example, the size and thickness of each component shown in the drawings are shown arbitrarily for convenience of explanation, so the present invention is not necessarily limited to what is shown.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. When describing with reference to the drawings, identical or corresponding components will be assigned the same reference numerals and redundant description thereof will be omitted. .

Figure 1 is a conceptual diagram of a training data generation system for ID card recognition deep learning according to an embodiment of the present invention.

Referring to FIG. 1, the training data generation system 1000 for ID recognition deep learning according to an embodiment of the present invention (hereinafter referred to as training data generation system) is a training data set for improving the performance of a predetermined ID recognition deep learning model. A training data generation service that generates can be provided.

In an embodiment, the training data generation system 1000 that implements the above training data generation service may include a terminal 100, a data generation server 200, and a network 300.

At this time, the terminal 100 and the data generation server 200 may be connected through the network 300.

Here, the network 300 according to the embodiment refers to a connection structure that allows information exchange between each node, such as the terminal 100 and/or the data generation server 200. This network 300 Examples include the 3rd Generation Partnership Project (3GPP) network, Long Term Evolution (LTE) network, World Interoperability for Microwave Access (WIMAX) network, Internet, Local Area Network (LAN), and Wireless Local Area Network (Wireless LAN). , WAN (Wide Area Network), PAN (Personal Area Network), Bluetooth network, satellite broadcasting network, analog broadcasting network, DMB (Digital Multimedia Broadcasting) network, etc., but are not limited thereto.

Hereinafter, the terminal 100 and the data generation server 200 that implement the training data generation system 1000 will be described in detail with reference to the attached drawings.

- Terminal (100: Terminal)

The terminal 100 according to an embodiment of the present invention may be a computing device installed with a training application (hereinafter referred to as application) that provides a training data generation service.

In detail, from a hardware perspective, the terminal 100 may include a mobile type computing device 100-1 and/or a desktop type computing device 100-2 on which an application is installed.

Here, the mobile type computing device 100-1 may be a mobile device such as a smart phone or tablet PC with an application installed.

For example, the mobile type computing device 100-1 includes smart phones, mobile phones, digital broadcasting devices, personal digital assistants (PDAs), portable multimedia players (PMPs), and tablet PCs. may be included.

In addition, the desktop type computing device 100-2 provides a training data generation service based on wired/wireless communication, such as a fixed desktop PC with an application installed, a laptop computer, or a personal computer such as an ultrabook. It may include devices where programs to run are installed.

Additionally, depending on the embodiment, the terminal 100 may further include a server computing device that provides a training data generation service environment.

Figure 2 is an internal block diagram of the terminal 100 according to an embodiment of the present invention.

Meanwhile, referring to FIG. 2, from a functional perspective, the terminal 100 includes a memory 110, a processor assembly 120, a communication processor 130, an interface unit 140, an input system 150, and a sensor system 160. ) and a display system 170. These components may be configured to be included within the housing of the terminal 100.

In detail, the application 111 is stored in the memory 110 , and the application 111 may store one or more of various applications, data, and commands for providing a training data generation service environment.

That is, the memory 110 may store commands and data that can be used to create a training data generation service environment.

Additionally, the memory 110 may include a program area and a data area.

Here, the program area according to the embodiment may be linked between the operating system (OS) and functional elements that boot the terminal 100, and the data area is data generated according to the use of the terminal 100. can be saved.

Additionally, the memory 110 may include at least one non-transitory computer-readable storage medium and a temporary computer-readable storage medium.

For example, the memory 110 may be a variety of storage devices such as ROM, EPROM, flash drive, hard drive, etc., and web storage that performs the storage function of the memory 110 on the Internet. may include.

The processor assembly 120 may include at least one processor capable of executing instructions of the application 111 stored in the memory 110 to perform various tasks for creating a training data generation service environment.

In an embodiment, the processor assembly 120 may control the overall operation of components through the application 111 of the memory 110 to provide a training data generation service.

This processor assembly 120 may be a system-on-chip (SOC) suitable for the terminal 100 that includes a central processing unit (CPU) and/or a graphics processing unit (GPU), and the operating system stored in the memory 110. (OS) and/or application programs can be executed, and each component mounted on the terminal 100 can be controlled.

Additionally, the processor assembly 120 may internally communicate with each component through a system bus and may include one or more bus structures, including a local bus.

In addition, the processor assembly 120 includes application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), and controllers. ), micro-controllers, microprocessors, and other electrical units for performing functions.

The communication processor 130 may include one or more devices for communicating with external devices. This communication processor 130 can communicate through a wireless network.

In detail, the communication processor 130 can communicate with the terminal 100 that stores content sources for implementing a training data generation service environment, and can communicate with various user input components such as a controller that receives user input.

In an embodiment, the communication processor 130 may transmit and receive various data related to the training data generation service to another terminal 100 and/or an external server.

This communication processor 130 uses technical standards or communication methods for mobile communication (e.g., Long Term Evolution (LTE), Long Term Evolution-Advanced (LTE-A), 5G New Radio (NR), WIFI). Alternatively, data can be transmitted and received wirelessly with at least one of a base station, an external terminal 100, and an arbitrary server on a mobile communication network established through a communication device capable of performing short-distance communication.

The sensor system 160 may include various sensors such as an image sensor 161, a position sensor (IMU) 163, an audio sensor 165, a distance sensor, a proximity sensor, and a contact sensor.

Here, the image sensor 161 may capture images and/or videos of the physical space around the terminal 100.

In an embodiment, the image sensor 161 may capture and acquire various images and/or videos related to the training data generation service.

In an embodiment, the image sensor 161 may acquire captured images for ID card sample images, etc.

In addition, the image sensor 161 is disposed on the front or/and rear of the terminal 100 and can acquire an image by photographing the direction in which the terminal 100 is placed. You can photograph space.

This image sensor 161 may include an image sensor device and an image processing module. In detail, the image sensor 161 can process still images or moving images obtained by an image sensor device (eg, CMOS or CCD).

Additionally, the image sensor 161 may use an image processing module to process still images or moving images obtained through the image sensor device to extract necessary information and transmit the extracted information to the processor.

This image sensor 161 may be a camera assembly including at least one camera. The camera assembly may include a general camera that photographs a visible light band, and may further include a special camera such as an infrared camera or a stereo camera.

In addition, the image sensor 161 as described above may be included and operated in the terminal 100, depending on the embodiment, or may be included in an external device (e.g., an external server, etc.) and may be included in the above-described communication processor 130 and/or Alternatively, it may operate through linkage based on the interface unit 140.

The position sensor (IMU, 163) can detect at least one of the movement and acceleration of the terminal 100. For example, it may consist of a combination of various position sensors such as an accelerometer, gyroscope, or magnetometer.

Additionally, the location sensor (IMU) 163 may recognize spatial information about the physical space around the terminal 100 in conjunction with the location communication processor 130, such as the GPS of the communication processor 130.

The audio sensor 165 can recognize sounds around the terminal 100.

In detail, the audio sensor 165 may include a microphone capable of detecting voice input from a user using the terminal 100.

In an embodiment, the audio sensor 165 may receive voice data required for a training data generation service from the user.

The interface unit 140 can connect the terminal 100 to communicate with one or more other devices. In detail, the interface unit 140 may include a wired and/or wireless communication device compatible with one or more different communication protocols.

Through this interface unit 140, the terminal 100 can be connected to various input/output devices.

For example, the interface unit 140 may be connected to an audio output device such as a headset port or a speaker to output audio.

As an example, the audio output device is connected through the interface unit 140, but embodiments in which the audio output device is installed inside the terminal 100 may also be included.

Additionally, for example, the interface unit 140 may be connected to an input device such as a keyboard and/or mouse to obtain user input.

This interface unit 140 connects devices equipped with a wired/wireless headset port, an external charger port, a wired/wireless data port, a memory card port, and an identification module. Port, audio I/O (Input/Output) port, video I/O (Input/Output) port, earphone port, power amplifier, RF circuit, transceiver and other communication circuits It may be configured to include at least one of:

The input system 150 may detect user input (e.g., gesture, voice command, button operation, or other type of input) related to the training data generation service.

In detail, the input system 150 may include a predetermined button, a touch sensor, and/or an image sensor 161 that receives a user motion input.

Additionally, the input system 150 is connected to an external controller through the interface unit 140 and can receive user input.

The display system 170 may output various information related to the training data generation service as a graphic image.

In an embodiment, the display system 170 displays various user interfaces for training data generation services, ID sample images, background template images, hologram template images, text template images, background-text combination images, and/or composite images, etc. can do.

These displays include liquid crystal display (LCD), thin film transistor-liquid crystal display (TFT LCD), organic light-emitting diode (OLED), and flexible display. , a 3D display, or an e-ink display.

The above components may be disposed within the housing of the terminal 100, and the user interface may include a touch sensor 173 on the display 171 configured to receive user touch input.

In detail, the display system 170 may include a display 171 that outputs an image and a touch sensor 173 that detects a user's touch input.

For example, the display 171 may be implemented as a touch screen by forming a mutual layer structure or being integrated with the touch sensor 173. This touch screen can function as a user input unit that provides an input interface between the terminal 100 and the user, and at the same time, can provide an output interface between the terminal 100 and the user.

Meanwhile, the terminal 100 according to an embodiment of the present invention performs deep learning (e.g., ID recognition deep learning, etc.) related to a training data generation service by linking with a predetermined deep-learning neural network. can be performed.

Here, the deep learning neural network according to the embodiment is R-CNN (Regions with CNN features), Fast R-CNN, Faster R-CNN, Mask R-CNN, U-Net network and/or convolutional neural network ( CNN, Convolution Neural Network), etc. may be included, but are not limited thereto.

Additionally, depending on the embodiment, the terminal 100 may further perform at least part of the functional operations performed by the data generation server 200, which will be described later.

- Data generation server (200: Data generation server)

Meanwhile, the data generation server 200 according to an embodiment of the present invention can perform a series of processes to provide a training data generation service.

In detail, in the embodiment, the data generation server 200 exchanges the data necessary to drive the training data generation process for deep learning learning of a predetermined ID card in an external device such as the terminal 100 with the external device. , the training data generation service can be provided.

More specifically, in the embodiment, the data generation server 200 generates the application 111 from an external device (in the embodiment, a mobile type computing device 100-1 and/or a desktop type computing device 100-2, etc.). An environment in which this can operate can be provided.

To this end, the data generation server 200 may include an application program, data, and/or commands for operating the application 111, and may transmit and receive various data based thereon with the external device.

Additionally, in the embodiment, the data generation server 200 may obtain a predetermined ID sample image.

Here, the ID sample image according to the embodiment may mean an image obtained by photographing a physical ID card with a camera.

Additionally, in the embodiment, the data generation server 200 may generate a background template image based on the obtained ID sample image.

Here, the background template image according to the embodiment may mean an image in which the text in the ID sample image has been removed.

Additionally, in the embodiment, the data generation server 200 may generate a hologram template image based on the obtained ID sample image.

Here, the hologram template image according to the embodiment may mean an image including a predetermined hologram pattern area displayed on the ID card to prevent forgery/tampering with the ID card.

Additionally, in the embodiment, the data generation server 200 may generate a text template image.

Here, the text template image according to the embodiment contains various ID information (e.g., license number, name, resident registration number, address, ID issuance date and/or issuing agency, etc.) randomly generated according to a predetermined standard in a predetermined ID format. It may mean an image in an arranged form.

Additionally, in the embodiment, the data generation server 200 may generate a composite image based on the template images created as above.

That is, in the embodiment, the data generation server 200 may generate a composite image based on the generated background template image, text template image, and hologram template image.

Here, the composite image according to the embodiment may mean a virtual ID card image created by combining the background template image, text template image, and hologram template image into one according to a predetermined method.

Additionally, in the embodiment, the data generation server 200 may perform data augmentation based on the generated composite image.

Additionally, in the embodiment, the data generation server 200 may perform learning of a deep learning model for recognizing a predetermined ID card based on the generated synthetic image.

Additionally, in the embodiment, the data generation server 200 performs deep learning (e.g., ID recognition deep learning, etc.) required for the training data generation service in conjunction with a predetermined deep-learning neural network. can do.

In detail, in the embodiment, the data generation server 200 reads a predetermined deep neural network driving program built to perform the deep learning from the memory module 230 and stores it in the read predetermined deep neural network system. You can perform deep learning as described below.

Here, the deep learning neural network according to the embodiment is R-CNN (Regions with CNN features), Fast R-CNN, Faster R-CNN, Mask R-CNN, U-Net network and/or convolutional neural network ( CNN, Convolution Neural Network), etc. may be included, but are not limited thereto.

At this time, depending on the embodiment, the deep learning neural network as described above is directly included in the data generation server 200, or is implemented as a separate device and/or server from the data generation server 200 to provide the training data generation service. Deep learning can be performed.

In the following description, the deep learning neural network is described as being implemented by being included in the data generation server 200, but is not limited thereto.

Additionally, in the embodiment, the data generation server 200 may store and manage various application programs, commands, and/or data for implementing a training data generation service.

In an embodiment, the data generation server 200 may generate at least one ID sample image, background template image, hologram template image, text template image, background-text combination image, composite image, hologram RGB table, hologram crop image, and deep learning. Models, data processing algorithms, and/or various user interfaces can be stored and managed.

However, the functional operations that the data generation server 200 can perform in the embodiment of the present invention are not limited to those described above, and other functional operations may be further performed.

Meanwhile, referring further to FIG. 1, in the embodiment, the data generation server 200 as described above includes at least one processor module (210) for data processing and at least one processor module 210 for data exchange with an external device. It is a predetermined computing device including the above communication module (220) and at least one memory module (230) that stores various applications, data, and/or instructions for providing training data generation services. It can be implemented.

Here, the memory module 230 may store one or more of an operating system (OS), various application programs, data, and commands for providing a training data generation service.

Additionally, the memory module 230 may include a program area and a data area.

Here, the program area according to the embodiment may be linked between the operating system (OS) that boots the server and functional elements, and the data area may store data generated according to the use of the server.

In an embodiment, the memory module 230 may be a variety of storage devices such as ROM, RAM, EPROM, flash drive, hard drive, etc., and may be a web device that performs the storage function of the memory module 230 on the Internet. It could also be storage (web storage).

Additionally, the memory module 230 may be a recording medium that is removable from the server.

Meanwhile, the processor module 210 can control the overall operation of each unit described above in order to implement a training data generation service.

This processor module 210 may be a system-on-chip (SOC) suitable for a server that includes a central processing unit (CPU) and/or graphics processing unit (GPU), and the operating system (OS) stored in the memory module 230 ) and/or application programs, etc., and control each component mounted on the server.

Additionally, the processor module 210 may internally communicate with each component through a system bus and may include one or more bus structures, including a local bus.

In addition, the processor module 210 includes application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), and controllers. ), micro-controllers, microprocessors, and other electrical units for performing functions.

In the above description, it has been explained that the data generation server 200 according to the embodiment of the present invention performs the functional operations as described above. However, depending on the embodiment, at least part of the functional operation performed by the data generation server 200 is externally performed. Various embodiments may be possible, such as performing the device (e.g., terminal 100, etc.), and at least part of the functional operation performed by the external device may be further performed by the data generation server 200. .

- How to generate training data for ID recognition deep learning

Hereinafter, the attached FIGS. 3 to 12 show a method in which the application 111 executed by at least one processor of the terminal 100 according to an embodiment of the present invention generates training data for deep learning for ID card recognition. Please refer to and explain in detail.

In an embodiment of the present invention, at least one processor of the terminal 100 may execute at least one application 111 stored in at least one memory 110 or operate in a background state.

Hereinafter, the method of providing the above-described training data generation service by operating the at least one processor to execute instructions of the application 111 will be briefly described as being performed by the application 111.

Figure 3 is a flowchart illustrating a method of generating training data for ID card recognition deep learning according to an embodiment of the present invention.

Referring to FIG. 3, in the embodiment, the application 111 executed by at least one processor of the terminal 100 or operating in the background may obtain an ID card sample image. (S101)

Here, the ID card sample image according to the embodiment may mean an image obtained by photographing a physical ID card with a camera.

In an embodiment, this ID sample image may include a plurality of hologram sample images obtained by photographing a physical ID card from various angles.

In detail, in the embodiment, the application 111 operates in conjunction with the image sensor 161 and/or a certain external device (e.g., another terminal 100, data generation server 200, and/or an Internet server, etc.) to You can obtain a sample image of the same ID card.

Additionally, in the embodiment, the application 111 may generate a background template image based on the obtained ID sample image . (S103)

Here, the background template image according to the embodiment may mean an image with the text in the ID card sample image removed.

Figure 4 is an example of a diagram for explaining a method of generating a background template image according to an embodiment of the present invention.

In detail, referring to FIG. 4 , in the embodiment, the application 111 may generate a mask image 20 based on the ID card sample image 10.

At this time, in the embodiment, the mask image 20 may be an image obtained by masking the text area within the ID card sample image 10.

In more detail, in the embodiment, the application 111 sets the RGB value for each pixel in the ID sample image 10 to a predetermined threshold indicating black, considering that the text is displayed based on the black RGB value. , for example, 150 or less, etc.) can be determined.

Additionally, in the embodiment, the application 111 may perform masking processing on pixels that meet the threshold, that is, the black text area.

Accordingly, the application 111 can obtain the mask image 20 in which masking of the text area within the ID card sample image 10 is implemented.

At this time, in the embodiment, the application 111 may expand the masking area by applying a predetermined dilation function.

For example, the application 111 can expand and use the masking area described above based on the expansion function (cv2.dilate) of the morphology calculation (Perspective Calculate).

Therefore, the application 111 can prevent problems such as unnatural boundaries occurring when only the text area is accurately masked within the ID card sample image 10 or deteriorating the quality of the infating process to be performed later. You can.

Additionally, in the embodiment, the application 111 may generate the background template image 30 based on the ID sample image 10 and the mask image 20 in conjunction with a predetermined deep learning neural network.

In detail, in the embodiment, the application 111 uses a predetermined ID sample image 10 and a predetermined mask image 20 as input data, and outputs a predetermined background template image 30 based on the input data. It can be linked with a deep learning neural network (hereinafter referred to as background template output model).

At this time, in an embodiment of the present invention, a deep learning neural network algorithm uses the predetermined ID sample image 10 and the predetermined mask image 20 as input data and the predetermined background template image 30 as output data. It is not intended to be limiting, and any algorithm (eg, U-Net style network, etc.) that can perform the above-described functional operations among the disclosed deep learning algorithms may be included and operated in an embodiment of the present invention.

Additionally, in the embodiment, the application 111 may input the above-described ID sample image 10 and mask image 20 into the linked background template output model.

And the application 111 generates the ID card sample image 10 and the mask image 20 through synthesis based on the background template output model that receives the ID card sample image 10 and the mask image 20. A background template image 30 can be obtained.

That is, the application 111 inputs the ID card sample image 10 and the mask image 20 into a predetermined background template output model, and outputs the ID card sample image 10 and the mask image 20. A background template image 30 synthesized can be obtained.

Accordingly, the application 111 may obtain a background template image 30 in which the masking area (i.e., text area) according to the mask image 20 is erased from the ID card sample image 10.

At this time, in the embodiment, the application 111, in conjunction with a predetermined deep learning neural network, naturally overlays the background color according to the ID card sample image 10 on the masking area of the background template image. ) processing can be performed.

For reference, the inpainting may refer to a technology that provides a complete image by reconstructing damaged and/or missing parts in a given image.

In an embodiment of the present invention, the deep learning neural network that performs the inpainting (hereinafter referred to as the inpainting performance model) can apply any disclosed algorithm as long as it is a deep learning algorithm capable of performing the inpainting function operation. In the embodiment of the invention, this algorithm itself is not restricted or limited.

Accordingly, the application 111 can generate the background template image 30 in which the text in the ID sample image 10 is erased more naturally.

Additionally, in the embodiment, the application 111 may generate a hologram template image based on the obtained ID sample image 10. (S105)

Figure 5 is an example of a hologram template image according to an embodiment of the present invention.

Here, referring to FIG. 5, the hologram template image 40 according to the embodiment is an image including a predetermined hologram pattern area (hereinafter referred to as a hologram area) displayed on the ID card to prevent forgery/tampering with the ID card. It can mean.

In an embodiment, the hologram template image 40 includes a predetermined hologram area extracted based on at least one hologram sample image (hereinafter referred to as target hologram image) among the plurality of hologram sample images of the ID card sample image 10. It may refer to the image it contains.

In detail, in the embodiment, the application 111 may extract at least one hologram area (hereinafter referred to as target hologram area) from the target hologram image.

In more detail, in the embodiment, the application 111 may obtain a user input for selecting a predetermined target hologram area on the target hologram image.

And in the embodiment, the application 111 may extract a target hologram area selected according to the obtained user input from the target hologram image.

Additionally, in the embodiment, the application 111 may generate a hologram RGB table based on the extracted target hologram area.

Figure 6 is an example of a diagram for explaining a method of generating a hologram RGB table according to an embodiment of the present invention.

Here, referring to FIG. 6, the hologram RGB table according to the embodiment specifies RGB values (in the embodiment, hologram estimation range) within a predetermined range based on the RGB values of pixels displaying the target hologram area. It can refer to table-type data composed of a set of pre-existing data.

In an embodiment, the application 111 may estimate a pixel having an RGB value that satisfies the hologram estimation range in the above hologram RGB table as a hologram area.

In detail, in the embodiment, the application 111 may obtain RGB values for each pixel included in the target hologram area.

Additionally, in the embodiment, the application 111 operates within a preset range (e.g., etc.) can be determined.

For example, if the RGB value of the first pixel included in the target hologram area is '50', the RGB value of the second pixel is '78', and the RGB value of the third pixel is '120', , based on the RGB value of the first pixel '40~60' range that satisfies, based on the RGB value of the second pixel '68~88' range that satisfies, based on the RGB value of the third pixel The range of '110 to 130' that satisfies can be determined as the hologram estimation range.

Additionally, in the embodiment, the application 111 may generate the hologram RGB table based on the hologram estimation range determined as above.

That is, in the embodiment, the application 111 determines at least one hologram estimation range that satisfies a preset range based on the RGB value for each pixel in the target hologram area, and uses the determined at least one hologram estimation range as data. The branch can generate the hologram RGB table.

Additionally, in the embodiment, the application 111 may perform masking processing on the target hologram image based on the hologram RGB table generated as above.

In detail, in the embodiment, the application 111 may detect at least one pixel (hereinafter referred to as invalid pixel) that does not meet the RGB value according to the hologram estimation range in the hologram RGB table among the pixels implementing the target hologram image. there is.

In other words, the application 111 may detect at least one remaining invalid pixel other than the pixels in the target hologram image that are determined to meet the hologram estimation range and represent the hologram area.

Additionally, in the embodiment, the application 111 may perform masking processing on the at least one detected invalid pixel area.

Accordingly, the application 111 may generate the hologram template image 40 by masking the remaining areas other than the predetermined hologram area on the target hologram image.

In this way, the application 111 generates a hologram RGB table that presents a judgment standard for a predetermined hologram area based on a plurality of hologram sample images taken from various angles of a specific physical ID card, and based on this, the hologram sample image By acquiring a hologram template image 40 (i.e., an image representing the hologram area within the hologram sample image) with the remaining areas other than the hologram area masked, the characteristics of the hologram are expressed in different aspects depending on the shooting angle. The hologram RGB table can be generated by considering , and thus, hologram template images 40 of various shapes can be easily obtained.

Additionally, in the embodiment, the application 111 may create a text template image. (S107)

Figure 7 is an example of a text template image according to an embodiment of the present invention.

Here, referring to FIG. 7, the text template image 50 according to the embodiment contains various ID information randomly generated according to a predetermined standard (e.g., license number, name, resident registration number, address, ID issuance date, and/or It may refer to an image in which the issuing agency, etc.) is arranged according to a prescribed ID card format.

That is, in the embodiment, the text template image 50 may be an image that displays ID information randomly generated according to a predetermined ID card format.

In detail, in the embodiment, the application 111 may randomly generate data values for each parameter based on a preset standard for each parameter in the ID information.

Exemplarily, if the first parameter of the ID information is 'resident registration number' and the preset standard for the first parameter is '6-digit number - 7-digit number format', the application 111 operates according to the preset standard. The first parameter value (eg, '123456-7891011', etc.) may be generated in a random manner.

As another example, if the second parameter of the ID information is 'name' and the preset standard for the second parameter is '3-character format of Korean alphabet combination', the application 111 randomly generates data according to the preset standard. The second parameter value (eg, 'Hong Gil-dong', etc.) can be generated in this way.

Additionally, in the embodiment, the application 111 may generate the text template image 50 by displaying the randomly generated data values for each parameter according to a predetermined ID card format as described above.

In an embodiment, the application 111 may provide a text template image (50) that displays each randomly generated parameter value based on a preset image size, a predetermined identification card form that provides a printing font and/or placement location for each parameter value. ) can be created.

For example, the application 111 sets the value of a predetermined first parameter (e.g., resident registration number, etc.) to a first image size (e.g., an image size with a predetermined aspect ratio, etc.), a first font (e.g., a Gulim font, etc.), ) and the first parameter value (e.g., '123456', which is a randomly generated resident registration number) based on the first ID card form designated to be displayed based on the first arrangement location (e.g., a predetermined (x, y) coordinate value point, etc.) A text template image (50) that displays '-7891011', etc.) can be created.

In this way, the application 111 is a text template in which various randomly generated ID information (e.g., license number, name, resident registration number, address, ID issuance date and/or issuing agency, etc.) is displayed according to a predetermined ID card format. By implementing the image 50, a virtual ID image (in the embodiment, a composite image) generated later using this can be easily implemented in different forms according to various randomly generated ID information as described above. Even if the ID information required for the virtual ID image is not individually collected from the actual ID, it can be obtained and utilized quickly and efficiently in an efficient manner.

Additionally, in an embodiment, application 111 may generate a composite image based on template images created as described above. (S109)

That is, in the embodiment, the application 111 may generate a composite image based on the background template image 30, text template image 50, and hologram template image 40 created as described above.

Figure 8 is an example of a drawing for explaining a method of generating a composite image according to an embodiment of the present invention.

Here, referring to FIG. 8, the composite image 70 according to the embodiment is obtained by combining the background template image 30, the text template image 50, and the hologram template image 40 into one according to a predetermined method. It may refer to a virtual ID image that is created.

In detail, in the embodiment, the application 111 may generate a background-text combination image based on the background template image 30 and the text template image 50.

Here, the background-text combination image 60 according to the embodiment may mean an image that is combined by rendering a predetermined text template image 50 on a predetermined background template image 30. .

That is, in the embodiment, the application 111 renders the text template image 50 on the generated background template image 30 to create the background template image 30 and the text template image 50. A background-text combination image 60 that combines can be obtained.

Additionally, in the embodiment, the application 111 may generate the composite image 70 based on the background-text combination image 60 and the hologram template image 40 obtained as above.

In detail, in the embodiment, the application 111 generates the composite image 70 by performing an alpha-blending operation based on the background-text combination image 60 and the hologram template image 40. You can.

For reference, the alpha blending refers to a computer color expression used to create a transparent effect when compositing by overlaying a second image (foreground image) on a predetermined first image (background image). A new value called Alpha (A) is further assigned to the value RGB, and the RGB value for each pixel in the background image and the RGB value for each pixel in the foreground image above reflect the weight based on the assigned alpha value. It refers to an operation method that mixes and displays.

Specifically, in the embodiment, the application 111 assigns an alpha value to each of the background-text combination image 60 and the hologram template image 40 and performs an alpha blending operation based on this to create the background-text The composite image 70 can be generated by combining the holographic template image 40 on the combination image 60 with a predetermined transparency.

At this time, depending on the embodiment, the application 111 determines the alpha value for each of the background-text combination image 60 and the hologram template image 40 in a predetermined manner (e.g., user input and/or a preset algorithm, etc.). ) can be randomly assigned.

For example, the application 111 assigns the alpha value of the background-text combination image 60 to '0.5' and sets the alpha value of the hologram template image 40 to '0.5' based on a predetermined method. By applying the background-text combination image 60 and the hologram template image 40 at a weight ratio of 0.5, a combined composite image 70 can be generated.

Through this, the application 111 implements a predetermined hologram pattern with a predetermined transparency on the virtual ID card image (in the embodiment, the composite image 70) generated as above, thereby creating a composite image with a more realistic shape. (70) can be generated.

FIG. 9 is an example of a diagram illustrating a method of generating a composite image 70 based on a plurality of hologram crop images according to an embodiment of the present invention.

At this time, referring to FIG. 9, in the embodiment, the application 111 performs a plurality of hologram crops based on a random crop operation on the hologram template image 40 when generating the composite image 70. Generate images (45-1, 45-2, 45-3, ...: 45) and perform an alpha blending operation based on the generated plurality of hologram crop images (45) and the background-text combination image (60). Thus, the composite image 70 may be generated.

For reference, the random crop operation may refer to an operation method that provides a plurality of images by randomly dividing a given image.

In detail, in the embodiment, the application 111 performs a random crop operation based on the hologram template image 40 to generate the plurality of hologram crop images 45, which are a plurality of images obtained by randomly dividing the hologram template image 40. can be created.

Additionally, in the embodiment, the application 111 may perform an alpha blending operation based on the plurality of generated holographic crop images 45 and the background-text combination image 60 described above.

Specifically, in the embodiment, the application 111 may provide an alpha value to each of the plurality of holographic crop images 45 and the background-text combination image 60.

And the application 111 performs an alpha blending operation based on the alpha values assigned to each image so that each of the plurality of hologram crop images 45 has a predetermined transparency on the background-text combination image 60. The composite image 70 can be created in a synthesized form.

At this time, in the embodiment, the application 111 may generate the composite image 70 by combining each of the plurality of holographic crop images 45 at random positions on the background-text combination image 60. .

That is, the application 111 randomly divides the hologram template image 40 into a plurality of images and composites them at random positions on the background-text combination image 60 with a predetermined transparency to create a virtual ID image (implemented). In an example, by generating the composite image 70, the composite image 70 in a form in which the noise effect caused by the hologram is more maximized can be implemented.

Additionally, in the embodiment, the application 111 may perform data augmentation based on the generated composite image 70. (S111)

In detail, in the embodiment, the application 111 implements a predetermined noise on the composite image 70 generated as above to additionally generate different composite images 70 based on the composite image 70. Data augmentation can be performed to increase the number.

At this time, in the embodiment, the noise is light reflection, motion blur, perspective, random angle rotation, salt and pepper noise, and/or shadow ( Shadow) effects, etc. may be included.

In an embodiment, the application 111 may perform image processing using a predetermined kernel to implement noise as described above on the composite image 70, and through this, the composite image 70 Based data augmentation can be performed.

10 to 12 are examples of drawings for explaining noise for data enhancement according to an embodiment of the present invention.

For example, referring to (a) of FIG. 10, the application 111 applies a light reflection effect to implement a predetermined light reflection effect at a predetermined position on the composite image 70 based on the image processing. By doing so, data augmentation can be performed.

For example, the application 111 may augment data by implementing a light reflection effect with size and intensity according to user settings at random locations on the composite image 70.

As another example, referring to (b) of FIG. 10, the application 111 performs the image processing based on a predetermined motion blur algorithm (e.g., Gaussian blur algorithm, etc.) so that the composite image 70 moves in a specific direction. Data augmentation can be performed by applying a motion blur effect that creates a shaking effect.

To this end, the application 111 determines a predetermined kernel (e.g., see kernels (a) to (d) in FIG. 11) according to the motion effect to be implemented, and performs a motion blur operation based on the determined kernel (e.g., Arithmetic mean calculation based on the kernel and each pixel value corresponding to the kernel area, etc.) may be performed.

As another example, referring to FIG. 12, the application 111 implements a perspective effect that converts the composite image 70 to have a predetermined direction and/or angle based on the image processing. By applying it, data augmentation can be performed.

In this way, the application 111 implements data augmentation by applying a predetermined noise to the synthetic image 70 generated as described above, thereby performing deep learning learning for later ID recognition even with a small amount of data. It is possible to build a larger amount of training data to be used, and thus the time and cost required to build the training data can be effectively reduced.

In addition, through this, the application 111 implements various noises that may occur when photographing a physical ID card with a real camera in the code to easily reflect them in the virtual ID card image (i.e., the composite image 70 in the embodiment). Therefore, the performance and quality of deep learning learning for ID recognition that is later performed based on the composite image 70 can be improved.

Additionally, in the embodiment, the application 111 may perform deep learning model learning for ID recognition. (S113)

In detail, in the embodiment, the application 111 may generate a training data set based on the plurality of composite images 70 generated as above.

Additionally, in the embodiment, the application 111 may perform learning on a deep learning model for ID recognition (hereinafter referred to as ID recognition deep learning model) using the generated training data set.

At this time, as an example, the ID recognition deep learning model may implement certain functional operations for various ID recognition services, such as certain ID authentication and/or information extraction.

In this way, the application 111 generates a large amount of virtual identification images (for example, holographic patterns (holographic noise) and/or various noises that vary depending on the shooting angle) that can be expressed in images actually taken of real identification cards. In the example, by creating a synthetic image (70) and supporting the learning of an ID recognition deep learning model using it, it is possible to easily build an ID training data set that is difficult to collect due to personal information issues, and from the physical ID card. Deep learning is performed using a large training data set constructed more realistically by reflecting the characteristics that can be expressed, thereby improving performance and quality, and at the same time improving the quality, safety, and reliability of various ID recognition services using the deep learning model. It can also be improved further.

As above, the method and system for generating training data for ID recognition deep learning according to an embodiment of the present invention generates a training data set to improve the performance of the ID recognition deep learning model, so that ID cards that are difficult to collect due to personal information issues, etc. It has the effect of easily building a training data set for recognition deep learning learning.

In addition, the training data generation method and system for ID recognition deep learning according to an embodiment of the present invention digitizes each step of physically producing an ID card to generate a virtual ID image, thereby meeting the requirements required for ID recognition deep learning. Even if training data is not individually collected from physical ID cards, it can be obtained and utilized in large quantities in an easy, quick and efficient manner, which has the effect of significantly reducing the time and cost required to build the training data.

In addition, the method and system for generating training data for deep learning for ID recognition according to an embodiment of the present invention is a virtual ID that reflects the characteristics of the image captured by the physical ID card (e.g., holographic patterns or various noises that vary depending on the shooting angle). By creating an ID card image, a virtual ID card image with a more realistic shape can be created and the performance and quality of deep learning learning using this can be improved.

In addition, the training data generation method and system for deep learning for ID recognition according to an embodiment of the present invention generates a virtual ID image that embodies various types of noise that can appear in a physical ID card from various angles, The deep learning model learned based on this can prevent the deterioration of ID recognition performance due to various noises that may occur on the image that is the target of ID recognition, and therefore provides an ID recognition deep learning model that provides more robust recognition performance. There is an effect that can be done.

In addition, the method and system for generating training data for deep learning for ID recognition according to an embodiment of the present invention involves learning a deep learning model for recognizing a predetermined ID based on training data based on the virtual ID image generated as above. By supporting this, it is possible to further improve the performance by enabling learning of the deep learning model based on a larger amount of training data, and various ID recognition services based on the deep learning model (e.g., online-based banking and It has the effect of improving the quality, safety, and reliability of (or identity authentication services for securities company work, etc.).

Meanwhile, the embodiments according to the present invention described above can be implemented in the form of program instructions that can be executed through various computer components and recorded on a computer-readable recording medium. The computer-readable recording medium may include program instructions, data files, data structures, etc., singly or in combination. Program instructions recorded on the computer-readable recording medium may be specially designed and configured for the present invention, or may be known and usable by those skilled in the computer software field. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and magneto-optical media such as floptical disks. medium), and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, etc. Examples of program instructions include not only machine language code such as that created by a compiler, but also high-level language code that can be executed by a computer using an interpreter or the like. A hardware device can be converted into one or more software modules to perform processing according to the invention and vice versa.

The specific implementations described in the present invention are examples and do not limit the scope of the present invention in any way. For the sake of brevity of the specification, descriptions of conventional electronic components, control systems, software, and other functional aspects of the systems may be omitted. In addition, the connections or connection members of lines between components shown in the drawings exemplify functional connections and/or physical or circuit connections, and in actual devices, various functional connections or physical connections may be replaced or added. Can be represented as connections, or circuit connections. Additionally, if there is no specific mention such as “essential,” “important,” etc., it may not be a necessary component for the application of the present invention.

In addition, although the detailed description of the present invention has been described with reference to preferred embodiments of the present invention, those skilled in the art or those skilled in the art will understand the spirit of the present invention as described in the patent claims to be described later. It will be understood that the present invention can be modified and changed in various ways without departing from the technical scope. Therefore, the technical scope of the present invention should not be limited to what is described in the detailed description of the specification, but should be defined by the scope of the patent claims.

Claims (16)

A method in which a training application executed by at least one processor of a terminal generates training data for ID recognition deep learning, comprising:
Obtaining an ID sample image obtained by photographing a physical ID card;
generating a background template image by removing text from the obtained ID sample image;
generating a hologram template image including at least a partial hologram pattern area on the obtained ID sample image;
A step of generating a text template image in which various types of ID information randomly generated according to a predetermined standard are arranged according to a predetermined ID card format; and
Comprising the step of generating a composite image combining the generated background template image, text template image, and hologram template image into one.
How to generate training data for ID recognition deep learning. According to claim 1,
The step of generating the background template image is,
Creating a mask image by masking the text area in the ID card sample image;
Inputting the generated mask image and the ID sample image into a first deep learning model;
Comprising the step of obtaining the background template image from a first deep learning model that receives the image.
How to generate training data for ID recognition deep learning. According to clause 2,
The step of generating the mask image is,
Comprising the step of expanding the masked text area based on a predetermined dilation function.
How to generate training data for ID recognition deep learning. According to clause 2,
The step of generating the background template image is,
Further comprising performing inpainting processing to apply a background color according to the ID sample image on the text area removed from the background template image.
How to generate training data for ID recognition deep learning. According to claim 1,
The step of generating the hologram template image is,
Comprising the step of generating a hologram RGB table based on the hologram pattern area,
The hologram RGB table is a set of hologram estimation range data specifying RGB values within a predetermined range based on the RGB values of pixels displaying the hologram pattern area.
How to generate training data for ID recognition deep learning. According to clause 5,
The step of generating the hologram template image is,
detecting invalid pixels that do not meet the hologram estimate range in the ID card sample image based on the hologram RGB table;
Further comprising generating the hologram template image by masking the detected invalid pixel area.
How to generate training data for ID recognition deep learning. According to claim 1,
The step of generating the composite image is,
Comprising the step of generating a background-text combination image by rendering the text template image on the background template image.
How to generate training data for ID recognition deep learning. According to clause 7,
The step of generating the composite image is,
Further comprising generating the composite image based on an alpha-blending operation based on the background-text combination image and the hologram template image.
How to generate training data for ID recognition deep learning. According to clause 8,
The step of generating the composite image is,
Generating a plurality of hologram crop images by randomly dividing the hologram template image based on a random crop operation based on the hologram template image;
Further comprising generating the composite image based on an alpha blending operation based on the plurality of holographic crop images and the background-text combination image.
How to generate training data for ID recognition deep learning. According to clause 9,
Generating the composite image based on an alpha blending operation based on the plurality of holographic crop images and the background-text combination image includes:
Comprising the step of synthesizing each of the plurality of holographic crop images at random positions on the background-text combination image.
How to generate training data for ID recognition deep learning. According to claim 1,
Further comprising the step of augmenting the data amount by implementing a predetermined noise on the generated composite image.
How to generate training data for ID recognition deep learning. According to claim 1,
Further comprising the step of training a predetermined ID recognition deep learning model based on the generated synthetic image.
How to generate training data for ID recognition deep learning. at least one memory storing a training application; and
At least one processor that reads the training application stored in the memory and generates training data for ID recognition deep learning,
The command of the training application is,
A step of obtaining an ID sample image obtained by photographing a physical ID card,
Creating a background template image by removing text from the obtained ID sample image;
generating a hologram template image including at least a partial hologram pattern area on the obtained ID sample image;
A step of generating a text template image in which various types of ID information randomly generated according to a predetermined standard are arranged according to a predetermined ID card format;
A command for generating a composite image combining the generated background template image, text template image, and hologram template image into one.
Training data generation system for ID recognition deep learning. According to claim 13,
The command of the training application is,
generating a background-text combination image by rendering the text template image on the background template image;
Further comprising an instruction for generating the composite image based on an alpha-blending operation based on the background-text combination image and the hologram template image.
Training data generation system for ID recognition deep learning. at least one memory module; and
At least one processor module;
At least one application stored in the memory module and executed by the processor module,
The at least one application is,
Obtain a sample image of an ID card taken from a physical ID card,
Create a background template image by removing the text in the obtained ID sample image,
Generate a hologram template image including at least some hologram pattern areas on the obtained ID sample image,
Generates a text template image in which various ID information randomly generated according to predetermined standards is arranged according to a predetermined ID format,
Creating a composite image that combines the created background template image, text template image, and hologram template image into one.
Data generation server. According to claim 15,
The at least one application is,
Generating a background-text combination image by rendering the text template image on the background template image,
Generating the composite image based on an alpha-blending operation based on the background-text combination image and the hologram template image.
Data generation server.

Images