WO2022131517A1

WO2022131517A1 - Method for separating fingerprint from fingerprint overlay image using deep-learning algorithm, and device therefor

Info

Publication number: WO2022131517A1
Application number: PCT/KR2021/014664
Authority: WO
Inventors: 이병호; 채민석; 이주현; 유동헌; 조재범
Original assignee: 서울대학교산학협력단
Priority date: 2020-12-18
Filing date: 2021-10-20
Publication date: 2022-06-23
Also published as: KR20220088163A; KR102570081B1

Abstract

Disclosed are a method for separating a fingerprint from a fingerprint overlay image using a deep-learning algorithm, and a device therefor. The fingerprint separation device: obtains a fingerprint separation image by separating a detection target fingerprint from a fingerprint overlay image, in which at least two fingerprints are overlaid, through a first deep-learning network trained to separate the overlaid fingerprints; generates a replacement image in which a region of interest, in which the plurality of fingerprints are overlaid in the fingerprint overlay image, is replaced with a region, corresponding to the region of interest, in the fingerprint separation image; and separates the background and the detection target fingerprint in the replacement image through a second deep-learning network trained to separate the background and the fingerprint, and outputs the detection target fingerprint.

Description

Method and device for separating fingerprints from superimposed fingerprint images using deep learning algorithm

An embodiment of the present invention relates to a method and an apparatus for separating a fingerprint from a fingerprint superimposed image in which a plurality of fingerprints are superimposed, and more particularly, a method and an apparatus for separating a fingerprint using several checkpoints of a deep learning algorithm is about

Overlapping oil fingerprints are often found at crime scenes, but in order to use them as court evidence, the image corresponding to each oil fingerprint must be separated from the overlapped image. There are various conventional techniques for separating overlapping fingerprints. A technology that separates the ridges in different directions of each oil fingerprint mixed in the overlapped part by guessing through the direction of the ridges of each oil fingerprint based on the photo taken in the visible light band. After irradiating the overlapping oil fingerprints with a laser of high power There are techniques for detecting the fluorescence intensity of fingerprints by various wavelength bands and separating them based on this, and techniques for separating based on spectra obtained after analyzing samples of overlapping oil fingerprints through mass spectrometry. However, in the case of the overlapping oil fingerprint image separation technology using the orientation field of the ridge, there is a disadvantage that the separation result is not good if there is an ambiguous feature point in the direction of the ridge in the overlapped portion or the direction overlaps. Separation technology based on fluorescence intensity can separate only samples taken with a time difference, and since the fluorescence intensity peaks in a specific wavelength band, it is based on guesswork through separate observation after measurement, so its use is limited. The technique of separating overlapping fingerprints through mass spectrometry has the disadvantage of damaging the sample in the process of ionizing the sample.

The technical problem to be achieved by the embodiment of the present invention is to provide a method and an apparatus for separating overlapping fingerprints using a deep learning algorithm.

An example of a fingerprint separation method according to an embodiment of the present invention for achieving the above technical problem is in a fingerprint superimposed image in which at least two or more fingerprints are superimposed through a first deep learning network learned to separate the superimposed fingerprints Separating the detection target fingerprint to obtain a fingerprint separation image; generating an alternative image in which a region of interest in which a plurality of fingerprints are overlapped in the fingerprint superimposed image is replaced with a region of the fingerprint separated image corresponding to the region of interest; and separating and outputting the background and the detection target fingerprint from the replacement image through the second deep learning network trained to separate the background and the fingerprint.

In order to achieve the above technical problem, an example of a fingerprint separation apparatus according to an embodiment of the present invention is a fingerprint superimposed image in which at least two or more fingerprints are superimposed through a first deep learning network trained to separate the superimposed fingerprints. a fingerprint separation unit for obtaining a fingerprint separation image by separating the detection target fingerprint; a region of interest setting unit generating a replacement image in which the region of interest, where the fingerprint is overlapped in the fingerprint superimposed image, is replaced with the region of the separated fingerprint image corresponding to the region of interest; and a fingerprint output unit that separates and outputs the background and the detection target fingerprint from the alternative image through the second deep learning network trained to separate the background and the fingerprint.

According to an embodiment of the present invention, fingerprint separation performance can be improved by using several checkpoints of a deep learning algorithm. In addition, oil fingerprints obtained at the crime scene can be quickly restored at the scene, which has the advantage of expediting the initial investigation. Although it is not adopted as evidence, it can be helpful for elementary investigations even if the results restored by the present invention are difficult to adopt as evidence, such as a lie detector that is helpful in the investigation, and the restored results are finally verified by an authorized person later It may be accepted as evidence through Through the present invention, an improved target fingerprint image can be obtained compared to fingerprint separation using an existing deep learning algorithm, and unnecessary or incorrect fingerprint information can be reduced to reduce confusion in the initial investigation.

1 is a view showing an example of a fingerprint separation device according to an embodiment of the present invention;

2 is a flowchart illustrating an example of a fingerprint separation method according to an embodiment of the present invention;

3 and 4 are diagrams schematically showing a fingerprint separation method according to an embodiment of the present invention;

5 is a diagram showing an example of a deep learning network according to an embodiment of the present invention;

6 and 7 are diagrams illustrating an example of a method of generating learning data for a deep learning network according to an embodiment of the present invention;

8 is a diagram illustrating an example of a pre-processing process of learning data for a deep learning network according to an embodiment of the present invention;

9 is a view showing an example of a method of generating a checkpoint for each fingerprint type according to an embodiment of the present invention;

10 is a diagram illustrating an example of a fingerprint separation result according to an embodiment of the present invention, and

11 is a diagram illustrating the configuration of an example of a fingerprint separation device according to an embodiment of the present invention.

Hereinafter, a fingerprint separation method and an apparatus according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

1 is a view showing an example of a fingerprint separation device according to an embodiment of the present invention.

Referring to FIG. 1 , the fingerprint separation apparatus 100 separates and outputs a detection target fingerprint using a deep learning algorithm when receiving a fingerprint superimposed image in which at least two or more fingerprints are superimposed. Here, the fingerprint superimposed image refers to an image in which at least two or more oil fingerprints are partially or entirely overlapped with each other on the background image. For example, the fingerprint superimposed image may be an image obtained through a predetermined pre-processing process after photographing an oil fingerprint existing on an object at a crime scene using a camera or the like.

In order to separate the desired detection target fingerprint from the fingerprint superimposed image, it is necessary to separate the overlapping fingerprints and the background. However, when all overlapping fingerprints and background images are removed with one deep learning algorithm, the complexity increases depending on the background of a complex shape or the difficulty of separation, so there is a problem in that the separation result of the detection target fingerprint is not accurate.

Accordingly, this embodiment proposes a fingerprint separation method using a first deep learning network specialized in separating overlapping fingerprints and a second deep learning network specialized in separating background and fingerprints. In other words, the deep learning network has a difference in the performance of fingerprint separation depending on the deep learning network structure, deep learning learning epoch, learning data, etc. The first checkpoint and the second checkpoint of the deep learning network with excellent performance in separating the background and the fingerprint are respectively called and the first deep learning network and the second deep learning network are set up, and then the fingerprint is separated through these deep learning networks suggest how to

2 is a flowchart illustrating an example of a fingerprint separation method according to an embodiment of the present invention.

Referring to FIG. 2 , the fingerprint separation apparatus 100 first separates the detection target fingerprint from the superimposed fingerprint image using the first deep learning network learned to separate the overlapped fingerprints ( S200 ). For example, the fingerprint separation apparatus 100 may set the first deep learning network by loading the first checkpoint of the deep learning network learned to separate the overlapped fingerprints. Here, the checkpoint is data defining the deep learning network, such as various parameters or variable values of the deep learning network at a specific point in the learning process. By loading a checkpoint, it is possible to set up a deep learning network trained up to a specific point in time.

The first deep learning network may be set according to a fingerprint type. For example, if there are a plurality of first checkpoints corresponding to a plurality of fingerprint types as shown in FIG. 9 , the fingerprint separation apparatus 100 selects the first checkpoint according to the type of the detection target fingerprint to be separated from the fingerprint superimposed image. By loading the first deep learning network, it is possible to generate a fingerprint separation image obtained by first separating the detection target fingerprint. Checkpoints according to a plurality of fingerprint types are reviewed again in FIG. 9 .

The fingerprint separation apparatus 100 designates a region where fingerprints are overlapped in the fingerprint superimposed image as a region of interest, and replaces the region of interest with a corresponding region of the separated fingerprint image (S210). For example, the fingerprint separation apparatus 100 designates a region where two fingerprints overlap in the fingerprint superimposed image 430 as the region of interest 432 as shown in FIG. It is possible to generate an alternative image 440 replaced with the corresponding region of the fingerprint separation image 420 including

The fingerprint separation device 100 uses the second deep learning network learned to separate the background and the fingerprint to secondarily separate and output the detection target fingerprint from the alternative image (S220). For example, the fingerprint separation apparatus 100 may set the second deep learning network by loading the second checkpoint of the deep learning network learned to separate the background and the fingerprint. When the first checkpoint salpin is loaded prior to the deep learning network, the first deep learning network is set, and when the second checkpoint is loaded, the second deep learning network is set. Like the first deep learning network, the second deep learning network may also be set according to the fingerprint type. The second checkpoint according to the plurality of fingerprint types will be reviewed again in FIG. 9 .

3 and 4 are diagrams schematically illustrating a fingerprint separation method according to an embodiment of the present invention.

3 and 4 together, the fingerprint separation apparatus 100 inputs the

superimposed fingerprint images

300 and 410 in which at least two or more fingerprints are superimposed to the first deep learning network 310 . The first deep learning network 310 may be set by loading the first checkpoint 315 that has been learned and stored in advance to separate the overlapping fingerprints. The fingerprint separation apparatus 100 acquires the

fingerprint separation images

320 and 420 obtained by first separating the detection target fingerprint 412 through the first deep learning network 310 .

The fingerprint separation apparatus 100 designates a region in which a plurality of fingerprints are overlapped in the fingerprint superimposed

images

300 and 410 as a Region of Intrest (ROI) 432, and sets the region of interest 432 to the first deep learning network. It is replaced with the corresponding region 422 of the

fingerprint separation images

320 and 420 obtained through 310 . The fingerprint separation apparatus 100 may provide a screen interface through which the region of interest 432 may be input from the user.

The fingerprint separation apparatus 100 inputs the

replacement images

330 and 440 in which the region of interest 432 is replaced with the corresponding region 422 of the stake separation image 320 to the second deep learning network 340 . The second deep learning network 340 may be set by loading the second checkpoint 345 that has been previously learned and stored in order to separate the background and the fingerprint.

The fingerprint separation apparatus 100 separates and outputs the detection target fingerprint 412 from the

alternative images

330 and 440 through the second deep learning network 340 . In other words, the second deep learning network 340 identifies all fingerprints other than the detection target fingerprint 412 in the

alternative images

330 and 440 as a background and separates the detection target fingerprint 412 from the

final images

350 and 450. print out

The fingerprint separation apparatus 100 may acquire the fingerprint superimposed image 410 through a pre-processing process. Referring to FIG. 4 , the fingerprint separation apparatus 100 scans the image 400 so that the detection target fingerprint 412 is in a preset direction (eg, longitudinal direction, etc.) The overlapping fingerprint area may be rotated and cropped so that the number of pixels per unit length becomes a predetermined number to match the resolution to generate the fingerprint superimposed image 410 . For example, the fingerprint separation apparatus 100 may adjust the resolution of the fingerprint superimposed image 410 to 500 ppi (pixles per inches). This means that if a fingerprint is an actual 1-inch size, the image will be 500 pixels wide. The unscaled image 400 can be used by adjusting the size to fit 500 ppi. At this time, since it is necessary to find out the actual size of the fingerprint, a distance of 1 cm can be specified through a ruler scale on the original image (4000) before the preposition process. In another embodiment, a vertical line is drawn on the detection target fingerprint 412 to detect the detection target fingerprint ( 412), the resolution can be adjusted by calculating the length of the corresponding fingerprint. Drawing a line on the original image 400 can be done manually by the user. In this case, the fingerprint separation device is positioned on a vertical line. Accordingly, the fingerprint superimposed image 410 may be rotated and the detection target fingerprint 412 may be identified.

5 is a diagram illustrating an example of a deep learning network according to an embodiment of the present invention.

Referring to FIG. 5 , it has been proven by many recent studies that the deep learning algorithm shows excellent performance in image processing tasks such as pattern recognition, restoration, and separation from images. Fingerprint separation technology using a deep learning algorithm builds a learning dataset by synthesizing an artificial fingerprint that reflects the characteristics of an actual field oil fingerprint and a background image, and then trains the deep learning network to extract the original artificial fingerprint. In this process, the deep learning network identifies fingerprints aligned in a certain direction in the fingerprint superimposed image as separation targets, and learns in the direction of removing other fingerprints from the background. Therefore, fingerprint separation using deep learning involves a process of separating fingerprints and a process of separating fingerprints from the background. At this time, the network parameters learned from the deep learning process to the intermediate process are stored as checkpoints, and each checkpoint separates the fingerprints according to the deep learning network structure, the deep learning learning epoch, and the variables of the preprocessing process for synthesizing artificial fingerprints. There is a difference in performance. Therefore, the more the appropriate checkpoint is called and executed according to the fingerprint separation situation, the better the separation performance.

In this embodiment, the first checkpoint with good separation performance of the overlapping fingerprint and the second checkpoint with good separation performance of the background and fingerprint are identified and stored in advance, and then each checkpoint is loaded to separate the fingerprints.

6 and 7 are diagrams illustrating an example of a method of generating learning data for a deep learning network according to an embodiment of the present invention.

6 and 7 , the fingerprint separation apparatus 100 generates learning data 630 in which the background image 600 and at least two

artificial fingerprints

610 and 620 are superimposed. The background image 600 may be a surface image of various objects (eg, various types of paper such as receipts) from which fingerprints can be detected. The

artificial fingerprints

610 and 620 are fingerprints previously generated by various conventional methods.

At least two

artificial fingerprints

610 and 620 may overlap in different directions. For example, the fingerprint separation apparatus 100 arranges the first artificial fingerprint 710 in a predefined direction (eg, vertical direction) and places the second artificial fingerprint 720 in a predetermined vertical direction as shown in FIG. 7 . After the angle is rotated, the training data 730 may be generated by overlapping with the background image 700 . Here, fingerprints arranged in a predefined direction (ie, vertical direction), ie, the first artificial fingerprint 710 , become a target for the deep learning network to separate.

After constructing the training dataset in this way, the deep learning network can be trained to separate and output fingerprints (ie, the first artificial fingerprint 710) arranged in a predefined direction (eg, vertical). have. In this embodiment, it is possible to quickly process even a general laptop with low usage, and the resulting image is simple and clear. A large number of fingerprint images can be processed quickly because only the preprocessing process of rotating the detection target fingerprint in a predefined direction in the fingerprint superimposed image input to the deep learning network is performed.

The deep learning network used in this embodiment is the first deep learning network trained to separate the overlapping fingerprints as shown in FIGS. 2 and 3 and the second deep learning network trained to separate the background and the fingerprint.

Both the first deep learning network and the second deep learning network may be generated by learning using the

learning data

630 and 730 in which at least two

artificial fingerprints

610, 620, 710, 720 are superimposed on the

background images

600 and 700. However, the first deep learning network is trained to specialize in the separation of superimposed fingerprints, and the second deep learning network is trained to specialize in the separation of backgrounds and fingerprints. Depending on the learning epoch of the deep learning network or the type of learning data, the deep learning network may show excellent performance in separating overlapping fingerprints or in separating background and fingerprints. Therefore, the fingerprint separation device 100 stores the first checkpoint of the deep learning network at the point in time showing excellent performance in fingerprint separation in the learning process of the deep learning network, and the deep learning network at the point in time showing excellent performance in separating the background and fingerprint One may store the second checkpoint.

For example, the fingerprint separation apparatus 100 uses a plurality of learning

data

630 and 730 created by superimposing the

background images

600 and 700 and at least two

artificial fingerprints

610, 620, 710, 720 to learn the deep learning network until a certain point in time. A first checkpoint defining a network may be stored. The fingerprint separation device 100 may then load the first checkpoint to set the first deep learning network.

The fingerprint separation device 100 also includes learning data created by superimposing

background images

600 and 700 and one

artificial fingerprint

610 and 710 or a first image and background created by overlapping

background images

600 and 700 and one

artificial fingerprint

610 and 710. A deep learning network using learning data in which a second image made by superimposing the

images

600 and 700 and at least two

artificial fingerprints

610, 620, 710, 720 (for example, the first image: the second image = 5:5) is mixed) It is possible to store the second checkpoint defining the deep learning network trained up to a certain point in time. The fingerprint separation device 100 may then load a second checkpoint to set up a second deep learning network.

8 is a diagram illustrating an example of a preprocessing process of learning data for a deep learning network according to an embodiment of the present invention.

Referring to FIG. 8 , the fingerprint separation apparatus 100 may perform a pre-processing process so that the artificial fingerprint of the learning data set is similar to the oil fingerprint of the actual field. For example, the fingerprint separation device 100 adds a curve to all or a part of the ridge of the artificial fingerprint 800 ( 810 ), modulates the thickness of all or a part of the ridge ( 820 ), or sharpness of all or a part of the ridge. A preprocessing process of adjusting (eg, blurring) 830 may be performed. The fingerprint separation apparatus 100 may automatically perform a pre-processing process or may receive the curvature, thickness, sharpness, etc. of the fingerprint from the user through a user interface and reflect it on the artificial fingerprint 800 . The fingerprint separating apparatus 100 may rotate or symmetrical the preprocessed artificial fingerprint to overlap other artificial fingerprints to generate the learning data 840 .

9 is a diagram illustrating an example of a method of generating a checkpoint for each fingerprint type according to an embodiment of the present invention.

Referring to FIG. 9 , the fingerprint separation apparatus 100 generates a learning dataset 910 for a plurality of fingerprint types 900 and then uses the learning dataset 910 for each fingerprint type to learn a deep learning network. can For example, various types of fingerprints may exist, such as a fingerprint type in which the fingerprint ridges are arranged in a circle, and a fingerprint type in which the fingerprint ridges are arranged in a horizontal direction. It is assumed that the learning dataset 910 for each fingerprint type is predefined.

The fingerprint separation device 100 learns the deep learning network using the first learning dataset of the first fingerprint type, and then stores the checkpoints of the learned deep learning network. In addition, after learning the deep learning network using the second learning dataset of the second fingerprint type, checkpoints of the learned deep learning network are stored. In this way, N checkpoints 920 for N fingerprint types 900 may be generated and stored.

The fingerprint separation apparatus 100 may load any one of the N checkpoints 920 according to the type of the detection target fingerprint of the fingerprint superimposed image. For example, if the detection target fingerprints of the fingerprint superimposed image are arranged horizontally, the fingerprint separation apparatus 100 may load the learned checkpoints using the horizontally arranged fingerprint type learning dataset. The fingerprint separation device 100 displays information on the fingerprint type for each checkpoint (eg, at least one representative image for each fingerprint type, etc.) through the screen interface so that the user can check the fingerprint type similar to the detection target fingerprint. Points can be easily selected.

The learning dataset 910 for each fingerprint type may be defined to generate a plurality of first checkpoints specialized for separation of overlapping stakes, or may be defined to generate a plurality of second checkpoints specialized for separation of background and fingerprints.

10 is a diagram illustrating an example of a fingerprint separation result according to an embodiment of the present invention.

Referring to FIG. 10 , for the fingerprint superimposed image 1000 , a process of separating different overlapping stakes and a process of separating a background and a fingerprint should be performed in a complex manner. If these two processes are not performed properly, there is a problem in that the ridges of the fingerprint are blurred or the ridges that were not in the fingerprint superimposed image appear in the image 1010 of the detection target fingerprint separated. The inaccurate fingerprint image 1010 may provide difficulty in the initial investigation by providing fingerprint information different from the truth (eg, feature points, core, or delta). As in the present embodiment, when the detection target fingerprint is separated using the first checkpoint specialized for separating the overlapping fingerprints and the second checkpoint specialized for separating the background and the background, the fingerprint 1020 can be accurately separated. An inaccurate ridge portion 1015 exists in the fingerprint separation image 1010 using a conventional deep learning algorithm, but an accurate ridge 1020 is expressed in the stake separation image 1020 using this embodiment.

11 is a diagram showing the configuration of an example of a fingerprint separation device according to an embodiment of the present invention.

Referring to FIG. 11 , the fingerprint separating apparatus 100 includes a learning unit 1100 , a preprocessing unit 1110 , a fingerprint separating unit 1120 , an ROI setting unit 1130 , and a fingerprint output unit 1140 .

The learning unit 1100 trains the first deep learning network to specialize in separating the overlapping fingerprints, and trains the second deep learning network to specialize in separating the background and the fingerprint, and then the first deep learning network and the second deep learning network that have been trained Store the first checkpoint and the second checkpoint of , respectively. In another embodiment, checkpoints for each fingerprint type may be generated and stored, and an example of this is shown in FIG. 9 .

The pre-processing unit 1110 rotates the image so that the detection target fingerprint is in a preset direction (eg, longitudinal direction, etc.) in the image obtained by photographing the overlapping fingerprint, and crops the area of the overlapped fingerprint to have a predefined resolution. ) to create a fingerprint superimposed image. An example of the preprocessor is shown in FIG. 4 .

The fingerprint separation unit 1120 generates a fingerprint separation image obtained by first separating a detection target fingerprint from a fingerprint superimposed image through a first deep learning network set by loading a first checkpoint.

The region of interest setting unit 1130 generates an alternative image in which the region of interest in which a plurality of fingerprints are overlapped in the fingerprint superimposed image is replaced with the corresponding region of the fingerprint separation image. For example, when the region of interest 432 is set by the user as shown in FIG. 4 , the region of interest 432 is replaced with the region 422 of the fingerprint separation image 420 for the firstly separated detection target fingerprint. create an image

The fingerprint output unit 1140 loads the second checkpoint and outputs the secondly separated detection target fingerprint from the alternative image through the set second deep learning network. The fingerprint output unit 1140 may normalize the value of each pixel of the detection target fingerprint output through the second deep learning network to a value of 0 to 1. For example, if each pixel has a value of 0 to 255, it can be normalized to a value of 0 to 1 by dividing by 255. The fingerprint output unit 1140 may output a detection target fingerprint in which pixels less than or equal to a predefined value are removed.

The present invention can also be implemented as computer-readable codes on a computer-readable recording medium. The computer-readable recording medium includes all types of recording devices in which data readable by a computer system is stored. Examples of computer-readable recording media include ROM, RAM, CD-ROM, magnetic tape, floppy disk, and optical data storage device. In addition, the computer-readable recording medium is distributed in a network-connected computer system so that the computer-readable code can be stored and executed in a distributed manner.

So far, the present invention has been looked at with respect to preferred embodiments thereof. Those of ordinary skill in the art to which the present invention pertains will understand that the present invention can be implemented in a modified form without departing from the essential characteristics of the present invention. Therefore, the disclosed embodiments are to be considered in an illustrative rather than a restrictive sense. The scope of the present invention is indicated in the claims rather than the foregoing description, and all differences within the scope equivalent thereto should be construed as being included in the present invention.

Claims

obtaining a fingerprint separation image by separating a detection target fingerprint from a fingerprint superimposed image in which at least two or more fingerprints are superimposed through a first deep learning network trained to separate the superimposed fingerprints;

generating an alternative image in which a region of interest in which a plurality of fingerprints are overlapped in the fingerprint superimposed image is replaced with a region of the fingerprint separated image corresponding to the region of interest; and

and outputting the background and the detection target fingerprint separately from the replacement image through a second deep learning network trained to separate the background and the fingerprint.
According to claim 1, wherein the step of obtaining the fingerprint separation image,

Setting the first deep learning network by loading a first checkpoint of the deep learning network learned to separate the overlapping fingerprints; A fingerprint separation method comprising: a.
The method of claim 2, wherein the setting comprises:

and loading a first checkpoint corresponding to a fingerprint type of the detection target fingerprint from among a plurality of first checkpoints corresponding to a plurality of fingerprint types.
According to claim 1, wherein the outputting step,

Setting the second deep learning network by loading a second checkpoint of the deep learning network learned to separate the background and the fingerprint;
The method of claim 4, wherein the setting comprises:

and loading a second checkpoint corresponding to the fingerprint type of the detection target fingerprint from among a plurality of second checkpoints corresponding to the plurality of fingerprint types.
The method of claim 1,

The method of claim 1, further comprising: rotating the superimposed fingerprint image so that the detection target fingerprint is in a vertical direction; before acquiring the fingerprint separated image.
The method of claim 1,

Adjusting the resolution of the fingerprint superimposed image so that the number of pixels per unit length in the longitudinal direction of the detection target fingerprint becomes a predefined value; .
The method of claim 1,

The step of learning the first deep learning network; further comprising before the step of obtaining the fingerprint separation image,

The learning step is

generating a learning dataset in which a plurality of fingerprints are superimposed on various backgrounds using a plurality of background images and a plurality of artificial fingerprints;

training the first deep learning network to separate fingerprints using the training dataset; and

Storing the first checkpoint of the learned first deep learning network; Fingerprint separation method comprising a.
The method of claim 1,

Learning the second deep learning network; further comprising before the step of obtaining the fingerprint separation image,

The learning step is

generating a learning dataset in which a plurality of fingerprints are superimposed on various backgrounds using a plurality of background images and a plurality of artificial fingerprints;

training the second deep learning network to separate a background and a fingerprint using the training dataset; and

Storing a second checkpoint of the learned second deep learning network; Fingerprint separation method comprising: a.
a fingerprint separation unit configured to obtain a separated fingerprint image by separating a detection target fingerprint from a fingerprint superimposed image in which at least two or more fingerprints are superimposed through a first deep learning network trained to separate the superimposed fingerprints;

a region of interest setting unit generating a replacement image in which the region of interest, where the fingerprint is overlapped in the fingerprint superimposed image, is replaced with the region of the separated fingerprint image corresponding to the region of interest; and

and a fingerprint output unit that separates and outputs the background and the detected fingerprint from the alternative image through a second deep learning network trained to separate the background and the fingerprint.
11. The method of claim 10,

Using a plurality of background images and a plurality of artificial fingerprints to generate a learning dataset in which a plurality of fingerprints are superimposed on various backgrounds, and to learn the first deep learning network or the second deep learning network using the learning dataset Fingerprint separation device comprising a; learning unit.
11. The method of claim 10,

A preprocessor that adjusts the resolution of the fingerprint superimposed image to a predetermined value so that the number of pixels per unit length in the longitudinal direction of the detection target fingerprint, and rotates the fingerprint superimposed image so that the detection target fingerprint is in a vertical direction; Fingerprint separation device, characterized in that it further comprises.
A computer-readable recording medium in which a program for performing the method according to claim 1 is recorded.