WO2012099435A2 - Method for discriminating banknotes using a bayesian approach - Google Patents

Method for discriminating banknotes using a bayesian approach Download PDF

Info

Publication number
WO2012099435A2
WO2012099435A2 PCT/KR2012/000548 KR2012000548W WO2012099435A2 WO 2012099435 A2 WO2012099435 A2 WO 2012099435A2 KR 2012000548 W KR2012000548 W KR 2012000548W WO 2012099435 A2 WO2012099435 A2 WO 2012099435A2
Authority
WO
WIPO (PCT)
Prior art keywords
banknote
feature vector
representative
unit cell
bill
Prior art date
Application number
PCT/KR2012/000548
Other languages
French (fr)
Korean (ko)
Other versions
WO2012099435A3 (en
WO2012099435A9 (en
Inventor
최의선
Original Assignee
노틸러스효성 주식회사
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority claimed from KR1020110006280A external-priority patent/KR101232684B1/en
Priority claimed from KR1020110006275A external-priority patent/KR101232683B1/en
Application filed by 노틸러스효성 주식회사 filed Critical 노틸러스효성 주식회사
Publication of WO2012099435A2 publication Critical patent/WO2012099435A2/en
Publication of WO2012099435A9 publication Critical patent/WO2012099435A9/en
Publication of WO2012099435A3 publication Critical patent/WO2012099435A3/en

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/84—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using probabilistic graphical models from image or video features, e.g. Markov models or Bayesian networks
    • G—PHYSICS
    • G07—CHECKING-DEVICES
    • G07D—HANDLING OF COINS OR VALUABLE PAPERS, e.g. TESTING, SORTING BY DENOMINATIONS, COUNTING, DISPENSING, CHANGING OR DEPOSITING
    • G07D7/00—Testing specially adapted to determine the identity or genuineness of valuable papers or for segregating those which are unacceptable, e.g. banknotes that are alien to a currency
    • G07D7/20—Testing patterns thereon
    • G07D7/202—Testing patterns thereon using pattern matching
    • G07D7/206—Matching template patterns
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/50—Extraction of image or video features by performing operations within image blocks; by using histograms, e.g. histogram of oriented gradients [HoG]; by summing image-intensity values; Projection analysis

Definitions

  • the present invention relates to a banknote discrimination method using a Bayesian approach, and more particularly, after dividing a banknote image generated by scanning a banknote into a predetermined number of unit cells, a sensor measured in each unit cell. After calculating the representative value representing each unit cell by using the data, extracting the feature vector of the banknotes using the calculated representative value of each unit cell, and reducing the dimension of the extracted feature vector through the linear feature extraction method.
  • GML Geographicsian Maximum Likelihood
  • Bayesian Bayesian (Bayes) is able to recognize papers relatively accurately and determine the authenticity of paper money. ian) relates to a banknote discrimination method using an approach.
  • banknote recognition automation equipment As banknote recognition automation equipment is popularized in all of you, there is a demand for a high-speed banknote discrimination apparatus capable of processing various bills at high speed.
  • most banknote discrimination devices are designed to be unsuitable for high speed due to low hardware performance, and the time constraints to recognize paper stocks in a short time and to determine the authenticity of bills require a lot of time for complex data processing and algorithms. It cannot be allocated.
  • the paper recognition algorithm is developed and applied in various forms according to the type or configuration of the bill data acquisition sensor.
  • the acquisition sensor has a one-dimensional array structure
  • the acquired one-dimensional multichannel data is processed using a method such as neural network learning.
  • This method has a problem of gain in terms of recognition speed due to a small amount of acquired data, but a decrease in recognition rate when the number of books is increased or when various bills in various countries are recognized.
  • a differential module device using a two-dimensional contact image sensor is widely developed. Since the two-dimensional contact image sensor is configured to take images of both sides of banknotes or both sides, the amount of data acquired is enormous. Therefore, it is necessary to set a template for a region having high discrimination power between the papers from the acquired banknote images.
  • this template setting process requires a lot of time and computational complexity due to the complex process such as quantization, binary code conversion of vector tablets, and coordinates, and high data resolution is required for accurate comparison. Was not possible.
  • counterfeiting is discriminated using optical or magnetic characteristics of the banknotes, and among them, authenticity determination using infrared data is performed by using an IR sensor of a one-dimensional array structure. After measuring the IR signal strength at a specific position, the counterfeit is determined by determining the presence of a signal compared to a reference value.
  • this one-dimensional array structure method is limited in the amount of data compared to the importance of IR authenticity determination, there is a disadvantage that is sensitive to changes in the skew (shift) and shift (change) environment that occurs during the transfer of bills.
  • an object of the present invention is to divide a bill image generated by scanning a bill into a predetermined number of unit cells, and then calculate a representative value representing each unit cell using sensor data measured in each unit cell. After extracting the feature vectors of banknotes using the calculated representative value of each unit cell as a factor, reducing the dimension of the extracted feature vectors through the linear feature extraction method, GML (Gaussian Maximum) Likelihood)
  • GML Garnier Maximum
  • the present invention provides a method for discriminating banknotes using a Bayesian approach that can recognize paper types relatively accurately and determine the authenticity of banknotes.
  • the present invention generating a bill image using the sensor data obtained by scanning the entire bill, dividing the generated bill image into a predetermined number of unit cells, the Calculating the representative value representing each unit cell by using the obtained sensor data for each divided unit cell, extracting the feature vector of the bill using the calculated representative value for each unit cell, the linear feature extraction method Extracting a representative feature vector by reducing the dimension of the feature vector of the extracted banknote by using a method and applying a Gaussian Maximum Likelihood (GML) classification method to the extracted representative feature vector to perform book type recognition on the banknote It provides a banknote discrimination method using a Bayesian approach that is configured to include.
  • GML Gaussian Maximum Likelihood
  • the banknote discrimination method using the Bayesian approach according to the present invention does not require a separate template setting for a region having high discrimination power between banknote images, as it uses the whole banknote image. Compared to this, high-speed data processing is possible, and it can be applied even when the input image has ultra low resolution.
  • the approximate area (or entire area) of the IR authenticity image is first templated to measure not only the intensity of the IR signal but also the similarity of the dominant shape of the pattern. Since the average value calculation is used after applying the block / mesh structure, it is possible to perform a fast recognition process and also apply to a low resolution CIS image.
  • FIG. 1 is a flow chart illustrating a method of recognizing the winding type of a banknote by applying a banknote discrimination method using a Bayesian approach according to a first embodiment of the present invention.
  • FIG. 2 illustrates an example of dividing a scanned banknote image into a predetermined number of unit cells in recognizing a paper type of a banknote by applying a banknote discrimination method using a Bayesian approach according to the first embodiment of the present invention.
  • FIG. 3 illustrates an example in which a representative value is calculated using an average value of sensor data for each unit cell in recognizing the paper currency of a banknote by applying the banknote discrimination method using the Bayesian approach according to the first embodiment of the present invention.
  • FIG. 4 is a flowchart illustrating a method of determining the authenticity of a banknote by applying a banknote discrimination method using a Bayesian approach according to a second embodiment of the present invention.
  • 5 is an example of dividing into a predetermined number of unit cells for each template region obtained in determining the authenticity of a banknote by applying the banknote discrimination method using the Bayesian approach according to the second embodiment of the present invention.
  • FIG. 6 illustrates an example in which a feature vector is calculated using representative values of sensor data for each unit cell in determining authenticity of a banknote by applying a banknote discrimination method using a Bayesian approach according to a second embodiment of the present invention. that.
  • FIG. 1 is a flowchart illustrating a method of recognizing the paper type of a bill by applying a bill discrimination method using a Bayesian approach according to the first embodiment of the present invention.
  • a bill image is generated by scanning the entire bill to be recognized using a contact image sensor (CIS) (S110), The generated paper image is divided into a predetermined number of unit cells (S120).
  • CIS contact image sensor
  • the bill image generation method in order to prevent only a part of the bill image is scanned by the shaking or vibration that may occur in the transfer operation, it is generally formed larger than the bill size. Therefore, the scanned image includes a bill image and a margin image of the surroundings. Accordingly, it is determined whether the banknotes are aligned, and if there is a tilt, the tilt is corrected by the angle, and only the banknote image excluding peripheral margins is extracted.
  • the number of unit cells for dividing the banknote image may vary depending on the size and resolution of the generated banknote image, typically 10 to 40 unit cells in the horizontal direction and 4 to 20 in the vertical direction It is good to divide each into unit cells.
  • Figure 2 shows an example of dividing the scanned banknote image into 14 ⁇ 7 unit cells according to the first embodiment of the present invention.
  • a representative value representing each unit cell is calculated using sensor data measured in each of the divided unit cells (S130), and a feature vector of a banknote having a calculated representative value for each unit cell is extracted. (S140).
  • sensor data measured in each of the divided unit cells (S130)
  • a feature vector of a banknote having a calculated representative value for each unit cell is extracted.
  • S140 As shown in FIG. 3, there are a plurality of pixels respectively corresponding to the CIS array in one unit cell, and each pixel has sensor data measured at each corresponding CIS pixel.
  • the sensor data is used to calculate one scalar value (representative value) representing each unit cell.
  • the representative value representing each unit cell is an average of sensor data constituting each unit cell, Various factors such as variance or maximum may be applied, and it is preferable to use an average value of sensor data in a unit cell that can most effectively reflect the characteristics of sensor data in each unit cell.
  • each unit cell is composed of 42 unit pixels in total, and assuming that sensor data is obtained for each pixel, 42 units present in the unit cells [7, 2] are present. 9, which is an average value of sensor data of one pixel, is calculated as a representative value A 7,2 of the unit cells [7, 2], and the calculated representative value is a factor of the feature vector representing the unit cell.
  • a representative value representing each unit cell is calculated in this manner, and a bill image feature vector is extracted using the representative value calculated for each unit cell, as shown in FIG.
  • the bill image divided into four unit cells extracts a feature vector X having 14 ⁇ 7 factors.
  • the feature vector (X) of the banknote is extracted by the above-described method, it is possible to discriminate the class of paper by comparing the extracted feature vector factors.
  • a step of reducing the dimension of the feature vector is performed (S150). The reason for reducing the dimension of the extracted feature vector is to reduce the object of computation by removing unnecessary parts of the feature vector, and to extract only the representative feature vectors that are important for the recognition of the denomination. Accordingly, the dimension of the extracted feature vector is reduced by applying the linear feature extraction method, and only the predetermined feature vector representing the feature of the bill image is selected.
  • the linear feature extraction method is a method of analyzing statistical characteristics of sensor data, and the principal component analysis (PCA) and linear discriminant analysis (LDA) are representative examples.
  • Principal Component Analysis (PCA) is a non-historical statistical technique that can effectively find image features and is the most optimal technique for performing dimensional reduction, but it is not ideal for classification purposes, such as judging species. Therefore, in the present embodiment, it is effective to reduce the dimension of the feature vector using the linear discrimination method (LDA), and the steps of reducing the dimension of the extracted feature vector using the linear discrimination method (LDA) will be described below. do.
  • Linear Discrimination is a technique of calculating the optimal linear discrimination matrix ⁇ T which maps the extracted feature vectors into the class space and enables class identification based on Euclidean distance.
  • the ratio between the class variance ( ⁇ 2 B : Between class variance) and the class variance ( ⁇ 2 W : Within class variance) for each volume of tens of thousands of note image feature vectors ( ⁇ 2 B / ⁇ 2 W ) Finding the linear discrimination matrix ⁇ T which can maximize, and apply the linear discrimination matrix to the extracted feature vector (X), the dimension of the extracted feature vector (X) as shown in Equation 1 below
  • Equation 1 By reducing the value, a representative feature vector Y that better represents the feature of the banknote image can be obtained.
  • the representative feature vector Y calculated as described above has a relatively lower dimension than the feature vector X originally extracted from the bill image, and at the same time more effectively represents the feature of the bill image.
  • the present invention recognizes the paper-species for the corresponding bill using the Gaussian Maximum Likelihood (GML) classification method on the obtained representative feature vector (S160). That is, using the average vector and variance matrix previously calculated for each class by using the representative feature vector calculated through the above process, it is calculated which probability class is the most likely to be included, and the probability value corresponds to the highest class. Sort the bills. For example, when classifying domestic species, the average vector and variance matrix of each of 1,000 won, 5,000 won, 10,000 won, and 50,000 won bills are used to calculate the probability that the representative feature vector will be included in the four types. The bill is classified by the paper type whose calculated value corresponds to the largest value.
  • GML Gaussian Maximum Likelihood
  • a method of calculating the probability that the representative feature vector is included in the corresponding paper type may be performed by Equation 2 below.
  • Equation 2 ML is a value representing the probability that the representative feature vector is included in the denomination class. According to Equation 2, the higher the probability that the representative feature vector is included in the denomination class, It is more likely that the species is. Therefore, the ML value to be included in the class is calculated by using the representative feature vector and the mean vector and variance matrix of each kind previously stored in the database, and the calculated values are compared with each other to compare the calculated values. As a judgment of the paper species of the bill, the paper sheet recognition is performed.
  • the present invention can perform the scoop recognition by using the Bayesian approach based on the representative value for each unit cell for the bill image, compared to the conventional scoop recognition method using a conventional neural network circuit In addition, even low-resolution bill images can be recognized relatively accurately.
  • FIG. 4 is a flowchart illustrating a method of determining the authenticity of banknotes by applying a banknote discrimination method using a Bayesian approach according to a second embodiment of the present invention.
  • the IR authenticity security element In order to determine the authenticity of a banknote by applying the banknote discrimination method using the Bayesian approach according to the present invention, first, by using the infrared sensor to obtain the sensor data obtained by scanning the whole banknote to determine the authenticity, the IR authenticity security element
  • the template regions in which the infrared pattern exists are specified based on the prior position information of the apparatus (S210), and the specified template regions are divided into a predetermined number of unit cells (S220).
  • the number of unit cells for dividing the template region may vary according to the scan area of the specified template region, and typically 2 to 20 unit cells in the horizontal direction and 2 to 10 units in the vertical direction for each template region. Splitting into cells is fine.
  • FIG. 5 shows an example of dividing a predetermined number of unit cells for each template region acquired according to the second embodiment of the present invention.
  • a representative value representing each unit cell is calculated using the sensor data measured in each divided unit cell (S230), and a feature vector of a banknote having a calculated representative value for each unit cell is extracted. (S240).
  • S230 the sensor data measured in each divided unit cell
  • S240 a feature vector of a banknote having a calculated representative value for each unit cell is extracted.
  • S240 a plurality of pixels respectively corresponding to the IR sensor array exist in one unit cell, and each pixel has sensor data measured by each corresponding IR sensor.
  • the sensor data is used to calculate one scalar value (representative value) representing each unit cell.
  • the representative value representing each unit cell is an average of sensor data constituting each unit cell, Various factors such as variance or maximum may be applied, and it is preferable to use an average value of sensor data in a unit cell that can most effectively reflect the characteristics of sensor data in each unit cell.
  • each unit cell is composed of 42 unit pixels in total, and assuming that sensor data is obtained for each pixel, each unit cell exists in the unit cell A [1,4]. 9, which is an average value of the sensor data of 42 pixels, is calculated as the representative value A 1,4 of the unit cell A [1,4], and the calculated representative value is a factor of the feature vector representing the unit cell.
  • a representative value representing each unit cell is calculated in this manner, and a feature vector having a representative value calculated for each unit cell is extracted. Accordingly, the template region divided as shown in FIG. A feature vector (X) having as many factors as the number of unit cells is extracted.
  • the authenticity of the bill is determined by comparing the extracted feature vector factors.
  • the feature vector factor is too large, it is difficult to perform a robust and fast authenticity determination.
  • a step of reducing the dimension of the feature vector is performed (S250). The reason for reducing the dimension of the extracted feature vectors is to reduce the computational object by removing unnecessary portions of the feature vectors, and to extract only representative feature vectors that are important for authenticity determination. Accordingly, the dimension of the extracted feature vector is reduced by applying the linear feature extraction method, and only a predetermined feature vector representing the characteristic of the infrared pattern is selected.
  • the linear feature extraction method is a method of analyzing statistical characteristics of sensor data, and the principal component analysis (PCA) and linear discriminant analysis (LDA) are representative examples.
  • PCA principal component analysis
  • LDA linear discriminant analysis
  • PCA principal component analysis
  • it is effective to reduce the dimension of the feature vector by using principal component analysis (PCA), which is a non-statistical statistical technique that can effectively find image features.
  • PCA principal component analysis
  • feature vectors extracted by using principal component analysis (PCA) The steps for reducing the dimension of the circuit will be briefly described.
  • Principal component analysis extracts a few major factor values that can represent the feature vector from several factors constituting the extracted feature vector (X), which is composed of the major factor values
  • PCA Principal component analysis
  • the representative feature vector (Y) calculated as described above has a relatively lower dimension than the feature vector (X) extracted initially, and at the same time more effectively represents a feature for authenticity determination of banknotes.
  • One of the characteristics of the principal component analysis described above is that for a data group distributed over the same time, a vector of directions having a large degree of variance can be obtained.
  • the eigenvalues and corresponding eigenvectors can be obtained.
  • the vector with the higher eigenvalue becomes an important element of the data group, and the smaller eigenvalue vector This means that the vector is less important than the first vector. Therefore, when principal component analysis is performed using tens of thousands of pneumatic features, eigenvalues for pneumococcal and corresponding eigenvectors can be obtained. do.
  • the method of calculating the probability that the representative feature vector is included in the corresponding pneumatic class region may be performed through a process similar to that of [Equation 2] described in the case of the first embodiment.
  • ML shown in [Equation 2] is a value representing the probability that the representative feature vector is included in the pneumococcal class of the subject species, and the higher the probability that the representative feature vector is included in the pneumococcal class increases the probability of the pneumoconiosis. . Therefore, the ML value to be included in the authenticity class is calculated using the representative feature vector and the authenticity class average vector and the variance matrix previously stored in the database. When the calculated value is less than or equal to the preset reference value, the banknote is determined as counterfeit.
  • the present invention determines the authenticity of banknotes using a Bayesian approach based on a representative value for each unit cell for the region in which the infrared pattern exists, thereby determining authenticity of the banknote authenticity using a conventional neural network. Not only can it be performed quickly, but the authenticity of banknotes can be accurately determined even with a low resolution banknote image.
  • the banknote discrimination method using the Bayesian approach according to the present invention is capable of processing data at a higher speed than the conventional discrimination method, and can be applied even when the input image has an ultra low resolution. Can be applied.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Multimedia (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Health & Medical Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • Databases & Information Systems (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Inspection Of Paper Currency And Valuable Securities (AREA)

Abstract

The present invention relates to a method for discriminating banknotes using a Bayesian approach. More particularly, the present invention relates to a method for discriminating banknotes using a Bayesian approach, which comprises dividing a banknote image generated by scanning a banknote into a predetermined number of unit cells, calculating a representative value of each unit cell using sensor data measured in each unit cell, extracting a banknote feature vector which uses the calculated representative value of each unit as a factor, reducing the dimension of the extracted feature vector through a linear feature extraction, and performing banknote discrimination on the feature vector, the dimension of which is reduced, using a Gaussian maximum likelihood (GML) classification. According to the present invention, as compared to conventional methods for discriminating banknotes using a neutral network, banknote class recognition and counterfeit banknote discrimination may be quickly performed, banknote class recognition may be performed relatively accurately, and a counterfeit banknote may be discriminated even from a banknote image having a low resolution.

Description

베이시안 접근법을 이용한 지폐 감별 방법Banknote Discrimination Using a Bayesian Approach
본 발명은 베이시안(Bayesian) 접근법을 이용한 지폐 감별 방법에 관한 것으로서, 더욱 상세하게는 지폐를 스캔하여 생성된 지폐이미지를 소정 개수의 단위셀로 분할한 후, 각각의 단위셀 내에서 측정된 센서데이터를 이용하여 각 단위셀을 대표하는 대표값을 산출하고, 산출된 각 단위셀별 대표값을 인자로 하는 지폐의 특징벡터를 추출하여, 선형특징추출법을 통해 추출된 특징벡터의 차원을 감소시킨 후, 차원이 감소된 특징벡터를 대상으로 GML(Gaussian Maximum Likelihood) 분류법을 이용하여 지폐 감별을 수행함에 따라, 종래의 신경망 회로를 이용한 지폐 감별 방법에 비해 권종인식 및 지폐 진위 판별을 신속하게 수행할 수 있을 뿐만 아니라 낮은 해상도의 지폐이미지로도 상대적으로 정확하게 권종을 인식하고, 지폐의 진위를 판별할 수 있는 베이시안(Bayesian) 접근법을 이용한 지폐 감별 방법에 관한 것이다. The present invention relates to a banknote discrimination method using a Bayesian approach, and more particularly, after dividing a banknote image generated by scanning a banknote into a predetermined number of unit cells, a sensor measured in each unit cell. After calculating the representative value representing each unit cell by using the data, extracting the feature vector of the banknotes using the calculated representative value of each unit cell, and reducing the dimension of the extracted feature vector through the linear feature extraction method. As the banknote discrimination is performed by using the GML (Gaussian Maximum Likelihood) classification method for the feature vectors with reduced dimensions, it is possible to perform the recognition of the paper species and the authenticity of bills faster than the banknote discrimination method using the conventional neural network. In addition to low resolution banknote images, Bayesian (Bayes) is able to recognize papers relatively accurately and determine the authenticity of paper money. ian) relates to a banknote discrimination method using an approach.
지폐 인식 자동화기기가 여러분야에 보편화되면서 다양한 지폐를 고속으로 처리할 수 있는 고속 지폐 감별 장치가 요구되고 있다. 하지만 대부분의 지폐 감별 장치는 낮은 하드웨어의 성능으로 고속화에 적합하지 않는 구조로 설계되어 있으며, 짧은 시간 내에 권종을 인식하고, 지폐의 진위를 판별해야 하는 시간적 제약은 복잡한 데이터 처리 및 알고리즘에 많은 시간을 할당할 수 없게 하고 있다.As banknote recognition automation equipment is popularized in all of you, there is a demand for a high-speed banknote discrimination apparatus capable of processing various bills at high speed. However, most banknote discrimination devices are designed to be unsuitable for high speed due to low hardware performance, and the time constraints to recognize paper stocks in a short time and to determine the authenticity of bills require a lot of time for complex data processing and algorithms. It cannot be allocated.
일반적으로 권종인식 알고리즘은 지폐 데이터 취득 센서의 종류나 구성에 따라 다양한 형태로 개발되어 응용된다. 예를 들어, 취득 센서가 1차원 어레이 구조를 갖는 지폐 감별 장치에서는 취득된 1차원 다채널 데이터를 이용, 신경망 학습 등과 같은 방법을 이용하여 처리한다. 이러한 방법은 취득 데이터 양이 적어 인식속도면에서는 이득이나, 권종수가 증가하거나 여러 나라의 다양한 지폐를 인식하는 경우 인식률이 하락하는 문제점이 있다.Generally, the paper recognition algorithm is developed and applied in various forms according to the type or configuration of the bill data acquisition sensor. For example, in the banknote discrimination apparatus in which the acquisition sensor has a one-dimensional array structure, the acquired one-dimensional multichannel data is processed using a method such as neural network learning. This method has a problem of gain in terms of recognition speed due to a small amount of acquired data, but a decrease in recognition rate when the number of books is increased or when various bills in various countries are recognized.
이러한 1차원 어레이 구조의 한계점을 극복하기 위하여 2차원 컨택 이미지 센서를 사용한 감별 모듈 장치가 널리 개발되는 추세이다. 2차원 컨택 이미지 센서의 구성은 지폐 단면 혹은 양면의 이미지를 모두 취하는 형태로 구성됨에 따라 취득되는 데이터 양이 방대하여, 획득된 지폐 영상으로부터 권종간 변별력이 높은 영역에 대한 탬플릿 설정이 필수적이다. 그러나, 이러한 템플릿 설정 과정은 양자화 및 이진부호의 벡터 태블릿 변환작업, 좌표화 등 복잡한 처리과정이 필요함에 따라 많은 시간과 연산량이 요구되었으며, 정확한 비교를 위해서는 상대적으로 높은 해상도가 요구됨에 따라 고속 데이터 처리가 불가능하였다.In order to overcome the limitations of the one-dimensional array structure, a differential module device using a two-dimensional contact image sensor is widely developed. Since the two-dimensional contact image sensor is configured to take images of both sides of banknotes or both sides, the amount of data acquired is enormous. Therefore, it is necessary to set a template for a region having high discrimination power between the papers from the acquired banknote images. However, this template setting process requires a lot of time and computational complexity due to the complex process such as quantization, binary code conversion of vector tablets, and coordinates, and high data resolution is required for accurate comparison. Was not possible.
한편, 지폐의 진위를 판단함에 있어서는, 지폐의 광학적 특성이나 자기적 특성을 이용하여 위폐를 판별하고 있으며, 그 중 적외선 데이터를 이용한 진위 판별은 1차원 어레이 구조의 IR 센서를 이용하여 특정 권종 지폐내의 특정 위치에 있는 IR 신호 세기를 측정 후 기준값 대비 신호 유무를 판단함으로써 위폐를 판별하고 있다. 그러나, 이러한 1차원 어레이 구조 방식은 IR 진위 판별의 중요성에 비하여 데이터의 양이 제한적이여서, 지폐 반송 중 발생하는 스큐(skew) 및 시프트(shift) 환경변화에 민감한 단점이 있다. On the other hand, in determining the authenticity of banknotes, counterfeiting is discriminated using optical or magnetic characteristics of the banknotes, and among them, authenticity determination using infrared data is performed by using an IR sensor of a one-dimensional array structure. After measuring the IR signal strength at a specific position, the counterfeit is determined by determining the presence of a signal compared to a reference value. However, this one-dimensional array structure method is limited in the amount of data compared to the importance of IR authenticity determination, there is a disadvantage that is sensitive to changes in the skew (shift) and shift (change) environment that occurs during the transfer of bills.
이러한 문제점을 해결하기 위하여, 컨택 이미지 센서 형태의 2D IR 영상 데이터를 취득하여 진위를 판별하는 방법이 도입되었다. 2D IR 영상 데이터를 이용한 진위 판별 방법은 1차원 어레이 센서 구조에 비하여 상대적으로 정밀한 진위가 가능하지만, 초당 진위 판별 매수 등 처리속도 제약으로 저해상도를 가짐에 따라 정확한 진위 판별에 어려움이 있었다. 또한, 권종별 IR 진위 판별 요소에 대한 사전정보를 기반으로 한 템플릿 설정 과정이 포함되는데, 저해상도의 이미지로 인하여 부정확한 위치 및 영역크기가 설정됨에 따라 처리속도 및 진위판별 성능을 하락시키는 원인이 되었다. In order to solve this problem, a method of determining authenticity by acquiring 2D IR image data in the form of a contact image sensor has been introduced. The authenticity determination method using 2D IR image data is more accurate than the one-dimensional array sensor structure. However, since the resolution is low due to processing speed constraints such as the number of authenticity determinations per second, it is difficult to accurately determine the authenticity. Also included is a template setting process based on prior information on IR authenticity determination factors for different types of papers. Inaccurate location and area size are set due to low resolution images, which causes processing speed and authenticity discrimination performance. .
또한, 이러한 지폐의 권종인식이나 진위판별에 있어서는 감별 알고리즘 구현이 중요한 변수가 되는데, 종래의 감별 알고리즘으로 사용되는 신경회로망(Neural Network), SVM(support vector machine) 등 기계학습 기법들은 레이어간 비선형 판별 함수의 커널(Kernel) 제약으로 인식 속도면에서 개선이 요구되고 있으며, 새로운 권종이 추가될 경우 다시 학습과정을 거쳐야 하기 때문에 신권 추가가 용이하지 않다는 단점이 있다.In addition, the implementation of the discrimination algorithm becomes an important variable in the paper currency recognition and authenticity discrimination of bills. Machine learning techniques such as neural networks and support vector machines (SVMs), which are used as conventional discrimination algorithms, are nonlinear discrimination between layers. Due to the kernel limitation of the function, improvement in recognition speed is required, and when a new kind is added, it is not easy to add a priesthood because it needs to go through the learning process again.
본 발명은 상기한 종래 기술에 따른 문제점을 해결하기 위한 것이다. 즉, 본 발명의 목적은, 지폐를 스캔하여 생성된 지폐이미지를 소정 개수의 단위셀로 분할한 후, 각각의 단위셀 내에서 측정된 센서데이터를 이용하여 각 단위셀을 대표하는 대표값을 산출하고, 산출된 각 단위셀별 대표값을 인자로 하는 지폐의 특징벡터를 추출하여, 선형특징추출법을 통해 추출된 특징벡터의 차원을 감소시킨 후, 차원이 감소된 특징벡터를 대상으로 GML(Gaussian Maximum Likelihood) 분류법을 이용하여 권종 인식 및 지폐 진위 판별을 수행함에 따라, 종래의 신경망 회로를 이용한 지폐 감별 방법에 비해 권종인식 및 지폐 진위 판별을 신속하게 수행할 수 있을 뿐만 아니라 낮은 해상도의 지폐이미지로도 상대적으로 정확하게 권종을 인식하고, 지폐의 진위를 판별할 수 있는 베이시안(Bayesian) 접근법을 이용한 지폐 감별 방법을 제공함에 있다.The present invention is to solve the above problems according to the prior art. That is, an object of the present invention is to divide a bill image generated by scanning a bill into a predetermined number of unit cells, and then calculate a representative value representing each unit cell using sensor data measured in each unit cell. After extracting the feature vectors of banknotes using the calculated representative value of each unit cell as a factor, reducing the dimension of the extracted feature vectors through the linear feature extraction method, GML (Gaussian Maximum) Likelihood) By using the classification method, it is possible to perform papermaking recognition and banknote authenticity faster than the conventional banknote discrimination method using the neural network circuit, and also to recognize the paper money with low resolution. The present invention provides a method for discriminating banknotes using a Bayesian approach that can recognize paper types relatively accurately and determine the authenticity of banknotes.
상기의 목적을 달성하기 위한 기술적 사상으로서 본 발명은, 지폐 전체를 스캔하여 얻어진 센서데이터를 이용하여 지폐이미지를 생성하는 단계, 생성된 지폐이미지를 미리 정해진 일정 갯수의 단위셀로 분할하는 단계, 상기 분할된 각 단위셀별로, 얻어진 센서데이터를 이용하여 각 단위셀을 대표하는 대표값을 산출하는 단계, 상기 산출된 각 단위셀별 대표값을 인자로 하는 지폐의 특징벡터를 추출하는 단계, 선형특징추출법을 이용하여 상기 추출된 지폐의 특징벡터의 차원을 감소시켜 대표 특징벡터를 추출하는 단계, 상기 추출된 대표 특징벡터에 GML(Gaussian Maximum Likelihood) 분류법을 적용하여 해당 지폐에 대한 권종 인식을 수행하는 단계를 포함하여 구성되는 베이시안 접근법을 이용한 지폐 감별 방법을 제공한다.As a technical idea for achieving the above object, the present invention, generating a bill image using the sensor data obtained by scanning the entire bill, dividing the generated bill image into a predetermined number of unit cells, the Calculating the representative value representing each unit cell by using the obtained sensor data for each divided unit cell, extracting the feature vector of the bill using the calculated representative value for each unit cell, the linear feature extraction method Extracting a representative feature vector by reducing the dimension of the feature vector of the extracted banknote by using a method and applying a Gaussian Maximum Likelihood (GML) classification method to the extracted representative feature vector to perform book type recognition on the banknote It provides a banknote discrimination method using a Bayesian approach that is configured to include.
또한, 적외선 센서를 통해 지폐 전체를 스캔하여 센서데이터를 획득하고, 적외선 패턴이 존재하는 템플릿의 영역을 특정하는 단계, 특정된 템플릿 영역을 미리 정해진 일정 갯수의 단위셀로 분할하는 단계, 상기 분할된 단위셀별로, 얻어진 센서데이터를 이용하여 각 단위셀을 대표하는 대표값을 산출하는 단계, 상기 산출된 각 단위셀별 대표값을 인자로 하는 지폐의 특징벡터를 추출하는 단계, 선형특징추출법을 이용하여 상기 추출된 지폐의 특징벡터의 차원을 감소시켜 대표 특징벡터를 추출하는 단계, 상기 추출된 대표 특징벡터에 GML(Gaussian Maximum Likelihood)분류법을 적용하여 해당 지폐에 대한 진위 판별을 수행하는 단계를 포함하여 구성되는 베이시안 접근법을 이용한 지폐 감별 방법을 제공한다.In addition, scanning the entire bill through the infrared sensor to obtain the sensor data, specifying a region of the template in which the infrared pattern exists, dividing the specified template region into a predetermined number of unit cells, the divided For each unit cell, using the obtained sensor data to calculate a representative value representing each unit cell, extracting a feature vector of a banknote having the calculated representative value for each unit cell as a factor, using a linear feature extraction method Extracting a representative feature vector by reducing the dimension of the extracted feature vector of the banknote, and applying authenticity likelihood (GML) classification to the extracted representative feature vector to perform authenticity determination on the banknote; Provided is a banknote discrimination method using a constructed Bayesian approach.
본 발명에 따른 베이시안 접근법을 이용한 지폐 감별 방법은, 지폐 영상 전체를 사용하여 권종 인식을 수행함에 따라, 지폐 영상으로부터 권종간 변별력이 높은 영역에 대한 별도의 탬플릿 설정이 필요치 않아 종래의 인식방법에 비해 고속의 데이터 처리가 가능하며, 입력 영상이 초 저해상도를 가질 경우에도 적용이 가능하다.The banknote discrimination method using the Bayesian approach according to the present invention does not require a separate template setting for a region having high discrimination power between banknote images, as it uses the whole banknote image. Compared to this, high-speed data processing is possible, and it can be applied even when the input image has ultra low resolution.
또한, 실제 위조권 정보가 없는 경우일지라도, 1차적으로 IR 진위 영상에 대한 대략적인 영역(혹은 전체영역)을 템플릿화하여 IR 신호의 세기뿐만 아니라 패턴 유효 형상(dominant shape)의 유사도 측정이 가능하며, 블록/메쉬 구조 적용 후 평균값 연산을 이용하므로, 고속의 인식처리가 가능하고, 저해상도 CIS 영상일 경우에도 적용가능하다.In addition, even if there is no actual counterfeit right information, the approximate area (or entire area) of the IR authenticity image is first templated to measure not only the intensity of the IR signal but also the similarity of the dominant shape of the pattern. Since the average value calculation is used after applying the block / mesh structure, it is possible to perform a fast recognition process and also apply to a low resolution CIS image.
또한, 통계적 정규 분포 추정 기법에 기반하므로, 샘플 분포 추정의 신뢰도를 측정하여 인식률에 대한 정량적 예측이 가능하며, 특징 벡터 변환은 선형 연산으로만 구성됨에 따라, 기존 신경망에서의 비선형 판별함수 사용시 보다 고속으로 권종 인식 및 지폐 진위 판별이 가능하다.In addition, since it is based on a statistical normal distribution estimation technique, it is possible to quantitatively predict the recognition rate by measuring the reliability of the sample distribution estimation, and the feature vector transformation is composed only of linear operations, so that it is faster than using a nonlinear discriminant function in an existing neural network. As a result, it is possible to recognize the paper type and determine the authenticity of paper money.
더불어, 신권 추가 시 다시 학습과정을 거쳐야 하는 기계학습 방법과 달리 신권의 특징값을 기기에 추가하기만 하면 되므로 신권 추가 시의 권종 인식 및 진위 판별이 용이하고, 인식률 제어나 감별에러 복구 등을 위하여 알고리즘을 직접 수정할 필요가 없이 비교 레퍼런스 데이터만 갱신하면 되는 장점이 있어 유지 보수가 용이하다.In addition, unlike the machine learning method, which requires a re-learning process when adding the priesthood, it is only necessary to add the characteristic values of the priesthood to the device, so that it is easy to recognize the denomination and authenticity when adding the priesthood, and to control the recognition rate or recover the discrimination error. Maintenance is easy because the comparison reference data only needs to be updated without the need to modify the algorithm directly.
도 1은 본 발명의 제 1실시예에 따른 베이시안 접근법을 이용한 지폐 감별 방법을 적용하여 지폐의 권종을 인식하는 방법을 나타낸 순서도.1 is a flow chart illustrating a method of recognizing the winding type of a banknote by applying a banknote discrimination method using a Bayesian approach according to a first embodiment of the present invention.
도 2는 본 발명의 제 1실시예에 따른 베이시안 접근법을 이용한 지폐 감별 방법을 적용하여 지폐의 권종을 인식함에 있어서 스캔된 지폐이미지를 소정개수의 단위셀로 분할한 예를 도시한 것.FIG. 2 illustrates an example of dividing a scanned banknote image into a predetermined number of unit cells in recognizing a paper type of a banknote by applying a banknote discrimination method using a Bayesian approach according to the first embodiment of the present invention.
도 3은 본 발명의 제 1실시예에 따른 베이시안 접근법을 이용한 지폐 감별 방법을 적용하여 지폐의 권종을 인식함에 있어서 각 단위셀별 센서데이터들의 평균값을 이용하여 대표값을 산출한 예를 도시한 것.FIG. 3 illustrates an example in which a representative value is calculated using an average value of sensor data for each unit cell in recognizing the paper currency of a banknote by applying the banknote discrimination method using the Bayesian approach according to the first embodiment of the present invention. .
도 4는 본 발명의 제 2실시예에 따른 베이시안 접근법을 이용한 지폐 감별 방법을 적용하여 지폐의 진위를 판별하는 방법을 나타낸 순서도.4 is a flowchart illustrating a method of determining the authenticity of a banknote by applying a banknote discrimination method using a Bayesian approach according to a second embodiment of the present invention.
도 5는 본 발명의 제 2실시예에 따른 베이시안 접근법을 이용한 지폐 감별 방법을 적용하여 지폐의 진위를 판별함에 있어서 획득한 템플릿 영역별로 각각 미리 정해진 소정 갯수의 단위셀로 분할한 예.5 is an example of dividing into a predetermined number of unit cells for each template region obtained in determining the authenticity of a banknote by applying the banknote discrimination method using the Bayesian approach according to the second embodiment of the present invention.
도 6은 본 발명의 제 2실시예에 따른 베이시안 접근법을 이용한 지폐 감별 방법을 적용하여 지폐의 진위를 판별함에 있어서 각 단위셀별 센서데이터들의 대표값을 이용하여 특징벡터를 산출한 예를 도시한 것.FIG. 6 illustrates an example in which a feature vector is calculated using representative values of sensor data for each unit cell in determining authenticity of a banknote by applying a banknote discrimination method using a Bayesian approach according to a second embodiment of the present invention. that.
이하, 본 발명의 바람직한 실시예를 첨부 도면에 의거하여 상세하게 설명하기로 한다.Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings.
도 1은 본 발명의 제 1실시예에 따른 베이시안 접근법을 이용한 지폐 감별 방법을 적용하여 지폐의 권종을 인식하는 방법을 나타낸 순서도이다.1 is a flowchart illustrating a method of recognizing the paper type of a bill by applying a bill discrimination method using a Bayesian approach according to the first embodiment of the present invention.
본 발명에 따른 베이시안 접근법을 이용한 지폐 감별 방법을 적용하여 지폐의 권종을 인식하기 위해서는, 먼저 컨택 이미지 센서(CIS)를 이용하여 인식을 원하는 지폐 전체를 스캔하여 지폐이미지를 생성하고(S110), 상기 생성된 지페이미지를 소정개수의 단위셀로 분할한다(S120).In order to recognize the bill type of the bill by applying the bill discrimination method using the Bayesian approach according to the present invention, a bill image is generated by scanning the entire bill to be recognized using a contact image sensor (CIS) (S110), The generated paper image is divided into a predetermined number of unit cells (S120).
이때, 지폐이미지 생성 방법은, 이송동작에서 발생할 수 있는 흔들림이나 진동에 의해 지폐이미지의 일부만이 스캔되는 것을 방지하기 위하여, 지폐크기보다 크게 형성되는 것이 일반적이다. 따라서, 스캔된 이미지에는 지폐이미지와 주변의 여백이미지가 함께 포함되어 있다. 이에 따라, 지폐의 정렬 여부를 판단하여 만약 기울기가 있을 경우, 그 각도만큼 기울기를 보정하고, 주변 여백을 제외한 지폐이미지만을 추출한다.At this time, the bill image generation method, in order to prevent only a part of the bill image is scanned by the shaking or vibration that may occur in the transfer operation, it is generally formed larger than the bill size. Therefore, the scanned image includes a bill image and a margin image of the surroundings. Accordingly, it is determined whether the banknotes are aligned, and if there is a tilt, the tilt is corrected by the angle, and only the banknote image excluding peripheral margins is extracted.
또한, 지폐이미지를 분할하는 단위셀의 개수는 생성된 지폐이미지의 크기와 해상도에 따라 달라질 수 있으며, 통상 생성된 지폐이미지를 가로 방향으로 10 내지 40개의 단위셀과, 세로 방향으로 4 내지 20개의 단위셀로 각각 분할하는 것이 무난하다. 도 2는 본 발명의 제 1실시예에 따라 스캔된 지폐이미지를 14×7개의 단위셀로 분할한 예를 보여주고 있다.In addition, the number of unit cells for dividing the banknote image may vary depending on the size and resolution of the generated banknote image, typically 10 to 40 unit cells in the horizontal direction and 4 to 20 in the vertical direction It is good to divide each into unit cells. Figure 2 shows an example of dividing the scanned banknote image into 14 × 7 unit cells according to the first embodiment of the present invention.
이어서, 분할된 각각의 단위셀 내에서 측정된 센서데이터를 이용하여 각 단위셀을 대표하는 대표값을 산출하며(S130), 산출된 각 단위셀별 대표값을 인자로 하는 지폐의 특징벡터를 추출한다(S140). 도 3에 도시된 바와 같이, 하나의 단위셀 내에는 CIS 어레이에 각각 대응하는 다수개의 픽셀이 존재하고, 각각의 픽셀은 각각의 대응하는 CIS 픽셀에서 측정된 센서데이터를 가진다. 본 발명에서는 이러한 센서데이터를 이용하여 각 단위셀을 대표하는 하나의 스칼라값(대표값)을 산출하게 되는데, 이때 각 단위셀을 대표하는 대표값으로는 각 단위셀을 구성하는 센서데이터들의 평균, 분산 또는 최대값 등 다양한 인자를 적용할 수 있으며, 바람직하게는 각 단위셀 내 센서데이터의 특징을 가장 효과적으로 반영할 수 있는 단위셀내 센서데이터의 평균값을 사용하는 것이 좋다.Subsequently, a representative value representing each unit cell is calculated using sensor data measured in each of the divided unit cells (S130), and a feature vector of a banknote having a calculated representative value for each unit cell is extracted. (S140). As shown in FIG. 3, there are a plurality of pixels respectively corresponding to the CIS array in one unit cell, and each pixel has sensor data measured at each corresponding CIS pixel. In the present invention, the sensor data is used to calculate one scalar value (representative value) representing each unit cell. In this case, the representative value representing each unit cell is an average of sensor data constituting each unit cell, Various factors such as variance or maximum may be applied, and it is preferable to use an average value of sensor data in a unit cell that can most effectively reflect the characteristics of sensor data in each unit cell.
예를 들어, 도 3에 도시된 바와 같이 각각의 단위셀은 총 42개의 단위 픽셀로 이루어지고, 이들 각 픽셀별로 센서데이터가 얻어진다고 가정할 때, 상기 단위셀 [7, 2]내에 존재하는 42개 픽셀의 센서데이터의 평균값인 9가 단위셀 [7, 2]의 대표값(A7,2)으로 산출되고, 산출된 대표값은 해당 단위셀을 대표하는 특징벡터의 인자가 된다. For example, as shown in FIG. 3, each unit cell is composed of 42 unit pixels in total, and assuming that sensor data is obtained for each pixel, 42 units present in the unit cells [7, 2] are present. 9, which is an average value of sensor data of one pixel, is calculated as a representative value A 7,2 of the unit cells [7, 2], and the calculated representative value is a factor of the feature vector representing the unit cell.
즉, 이와 같은 방법으로 각 단위셀을 대표하는 대표값을 산출하여, 각각의 단위셀별 산출된 대표값을 인자로 하는 지폐이미지 특징벡터를 추출하게 되며, 도 3에 도시된 바와 같이, 14×7개의 단위셀로 분할된 지폐이미지는 14×7개의 인자를 갖는 특징벡터(X)를 추출하게 된다.That is, a representative value representing each unit cell is calculated in this manner, and a bill image feature vector is extracted using the representative value calculated for each unit cell, as shown in FIG. The bill image divided into four unit cells extracts a feature vector X having 14 × 7 factors.
상술한 방법으로 지폐의 특징벡터(X)가 추출되면 추출된 특징벡터 인자들을 비교하여 권종을 감별하게 되는데, 특징벡터의 인자가 너무 많을 경우 견실하고 빠른 권종 인식을 수행하기 어려우므로, 본 발명에서는 특징벡터의 차원을 감소시키는 단계를 수행한다(S150). 추출된 특징벡터의 차원을 감소시키는 이유는 특징벡터의 불필요한 부분을 제거하여 연산대상을 줄이고, 권종 인식에 있어 중요한 대표 특징벡터만을 추출하기 위함이다. 이에 따라, 선형특징추출법을 적용하여 추출된 특징벡터의 차원을 감소시키고, 지폐이미지의 특징을 잘 나타내는 소정의 대표 특징벡터(dominant feature vector)만을 선별한다.When the feature vector (X) of the banknote is extracted by the above-described method, it is possible to discriminate the class of paper by comparing the extracted feature vector factors. However, when the number of feature vectors is too large, it is difficult to perform robust and fast class recognition. A step of reducing the dimension of the feature vector is performed (S150). The reason for reducing the dimension of the extracted feature vector is to reduce the object of computation by removing unnecessary parts of the feature vector, and to extract only the representative feature vectors that are important for the recognition of the denomination. Accordingly, the dimension of the extracted feature vector is reduced by applying the linear feature extraction method, and only the predetermined feature vector representing the feature of the bill image is selected.
선형특징추출법은 센서데이터의 통계적인 특성을 분석하는 방법으로서 주성분 분석법(Principal component Analysis, PCA)과 선형 판별법(Linear Discriminant Analysis, LDA) 등이 그 대표적인 예이다. 주성분 분석(PCA)은 이미지 특징을 효과적으로 찾을 수 있는 비교사 통계학적 기법이며, 차원축소를 수행할 때 가장 최적인 기법이기는 하나, 권종 판별처럼 분류 목적을 위해서는 이상적이지 않다. 따라서, 본 실시예에서는 선형 판별법(LDA)을 이용하여 특징벡터의 차원을 감소시키는 것이 효과적이며, 이하에서는 선형 판별법(LDA)을 이용하여 추출된 특징벡터의 차원을 감소시키는 단계에 대해 설명하기로 한다.The linear feature extraction method is a method of analyzing statistical characteristics of sensor data, and the principal component analysis (PCA) and linear discriminant analysis (LDA) are representative examples. Principal Component Analysis (PCA) is a non-historical statistical technique that can effectively find image features and is the most optimal technique for performing dimensional reduction, but it is not ideal for classification purposes, such as judging species. Therefore, in the present embodiment, it is effective to reduce the dimension of the feature vector using the linear discrimination method (LDA), and the steps of reducing the dimension of the extracted feature vector using the linear discrimination method (LDA) will be described below. do.
선형 판별법(LDA)은 상기 추출된 특징백터를 클래스 공간으로 맵핑하여 유클리디안 거리 등에 기반하여 클래스 확인을 할 수 있도록 하는 최적의 선형판별 매트릭스(ΦT)를 산출하는 기법이다. 즉, 수만개의 지폐이미지 특징벡터에 대하여 각각의 권종별 클래스간 분산(σ2 B:Between class variance)과 클래스내 분산(σ2 W:Within class variance)의 비율(σ2 B/σ2 W)을 최대화 할 수 있는 선형판별 매트릭스(ΦT)를 찾고, 그 선형판별 매트릭스를 상기 추출된 특징벡터(X)에 적용하여, 아래의 [수학식 1]과 같이 추출된 특징벡터(X)의 차원을 감소시켜, 지폐이미지의 특징을 보다 잘 나타내는 대표 특징벡터(Y)를 얻을 수 있다.Linear Discrimination (LDA) is a technique of calculating the optimal linear discrimination matrix Φ T which maps the extracted feature vectors into the class space and enables class identification based on Euclidean distance. In other words, the ratio between the class variance (σ 2 B : Between class variance) and the class variance (σ 2 W : Within class variance) for each volume of tens of thousands of note image feature vectors (σ 2 B / σ 2 W ) Finding the linear discrimination matrix Φ T which can maximize, and apply the linear discrimination matrix to the extracted feature vector (X), the dimension of the extracted feature vector (X) as shown in Equation 1 below By reducing the value, a representative feature vector Y that better represents the feature of the banknote image can be obtained.
수학식 1
Figure PCTKR2012000548-appb-M000001
 Equation 1
Figure PCTKR2012000548-appb-M000001
이와 같이 산출된 대표 특징벡터(Y)는 지폐이미지로부터 최초에 추출된 특징벡터(X)보다 상대적으로 낮은 차원을 갖게 되고, 동시에 지폐이미지의 특징을 보다 효과적으로 대표하게 된다. The representative feature vector Y calculated as described above has a relatively lower dimension than the feature vector X originally extracted from the bill image, and at the same time more effectively represents the feature of the bill image.
이어서, 획득한 대표 특징벡터에 GML(Gaussian Maximum Likelihood) 분류법을 이용하여 해당 지폐에 대한 권종인식을 수행한다(S160). 즉, 상술한 과정을 통해 산출된 대표 특징벡터를 권종 클래스별로 기 산출되어 있는 평균벡터 및 분산행렬을 이용하여 어느 권종 클래스에 포함될 확률이 가장 높은가를 산출한 뒤, 그 확률값이 가장 높은 권종으로 해당 지폐를 분류한다. 예를 들어, 국내의 권종을 분류할 경우, 천원권, 오천원권, 만원권 및 오만원권 각각의 평균벡터 및 분산행렬을 이용하여, 대표 특징벡터가 상기 4가지의 권종에 포함될 확률을 각각 산출하며, 그 산출된 결과값이 가장 큰 값에 해당하는 권종으로 해당 지폐를 분류한다.Subsequently, the present invention recognizes the paper-species for the corresponding bill using the Gaussian Maximum Likelihood (GML) classification method on the obtained representative feature vector (S160). That is, using the average vector and variance matrix previously calculated for each class by using the representative feature vector calculated through the above process, it is calculated which probability class is the most likely to be included, and the probability value corresponds to the highest class. Sort the bills. For example, when classifying domestic species, the average vector and variance matrix of each of 1,000 won, 5,000 won, 10,000 won, and 50,000 won bills are used to calculate the probability that the representative feature vector will be included in the four types. The bill is classified by the paper type whose calculated value corresponds to the largest value.
상기 대표 특징벡터가 해당 권종에 포함될 확률을 산출하는 방법은 하기의 [수학식 2]에 의해서 수행될 수 있다. A method of calculating the probability that the representative feature vector is included in the corresponding paper type may be performed by Equation 2 below.
수학식 2
Figure PCTKR2012000548-appb-M000002
 Equation 2
Figure PCTKR2012000548-appb-M000002
여기서, here,
ML : 유사도(Maximum Likelyhood),ML: Maximum Likelyhood,
i : 권종의 종류,i: type of winding species,
Y : 대표 특징벡터,Y: representative feature vector,
Figure PCTKR2012000548-appb-I000001
: 권종별 분산행렬,
Figure PCTKR2012000548-appb-I000001
: Distributed matrix by volume,
Figure PCTKR2012000548-appb-I000002
: 권종별 평균벡터
Figure PCTKR2012000548-appb-I000002
: Average vector by volume type
이다.to be.
상기의 [수학식 2]에서 ML은 상기 대표 특징벡터가 해당 권종 클래스에 포함될 확률을 대표하는 값으로서, 상기의 [수학식 2]에 따르면 대표 특징벡터가 해당 권종 클래스에 포함될 확률이 높을 수록, 해당 권종일 확률이 높아진다. 따라서, 대표 특징벡터와 데이터베이스에 기 저장된 권종별 평균벡터 및 분산행렬을 이용하여 권종 클래스에 포함될 ML값을 각각 산출하고, 그 산출값들을 비교하여, 비교한 결과값 중 가장 높은 ML값을 가지는 권종으로 해당 지폐의 권종을 판단함으로써 권종인식을 수행한다.In Equation 2, ML is a value representing the probability that the representative feature vector is included in the denomination class. According to Equation 2, the higher the probability that the representative feature vector is included in the denomination class, It is more likely that the species is. Therefore, the ML value to be included in the class is calculated by using the representative feature vector and the mean vector and variance matrix of each kind previously stored in the database, and the calculated values are compared with each other to compare the calculated values. As a judgment of the paper species of the bill, the paper sheet recognition is performed.
상술한 바와 같이, 본원 발명은 지폐이미지에 대한 단위셀별 대표값을 바탕으로 베이시안 접근법을 이용하여 권종 인식을 수행함으로써, 종래의 신경망 회로를 이용한 권종 인식 방법에 비해 권종인식을 신속하게 수행할 수 있을 뿐만 아니라 낮은 해상도의 지폐이미지로도 상대적으로 정확하게 권종을 인식할 수 있다. As described above, the present invention can perform the scoop recognition by using the Bayesian approach based on the representative value for each unit cell for the bill image, compared to the conventional scoop recognition method using a conventional neural network circuit In addition, even low-resolution bill images can be recognized relatively accurately.
도 4는 본 발명의 제 2실시예에 따른 베이시안 접근법을 이용한 지폐 감별 방법을 적용하여 지폐의 진위를 판별하는 방법을 나타낸 순서도이다.4 is a flowchart illustrating a method of determining the authenticity of banknotes by applying a banknote discrimination method using a Bayesian approach according to a second embodiment of the present invention.
본 발명에 따른 베이시안 접근법을 이용한 지폐 감별 방법을 적용하여 지폐의 진위를 판별하기 위해서는, 먼저 적외선센서를 이용하여 진위 판별을 원하는 지폐 전체를 스캔하여 얻어진 센서데이터를 획득하고, IR 진위 보안요소에 대한 사전 위치정보를 기반으로 하여 적외선 패턴이 존재하는 템플릿 영역들을 특정하며(S210), 상기 특정된 템플릿 영역들을 미리 정해진 일정 갯수의 단위셀로 분할한다(S220).In order to determine the authenticity of a banknote by applying the banknote discrimination method using the Bayesian approach according to the present invention, first, by using the infrared sensor to obtain the sensor data obtained by scanning the whole banknote to determine the authenticity, the IR authenticity security element The template regions in which the infrared pattern exists are specified based on the prior position information of the apparatus (S210), and the specified template regions are divided into a predetermined number of unit cells (S220).
이때, 템플릿 영역을 분할하는 단위셀의 개수는 특정된 템플릿 영역의 스캔 면적에 따라 달라질 수 있으며, 통상 각각의 템플릿 영역별로 가로 방향으로 2 내지 20개의 단위셀과, 세로 방향으로 2 내지 10개의 단위셀로 분할하는 것이 무난하다. 도 5는 본 발명의 제 2실시예에 따라 획득한 템플릿 영역별로 각각 미리 정해진 소정 갯수의 단위셀로 분할한 예를 보여주고 있다.In this case, the number of unit cells for dividing the template region may vary according to the scan area of the specified template region, and typically 2 to 20 unit cells in the horizontal direction and 2 to 10 units in the vertical direction for each template region. Splitting into cells is fine. FIG. 5 shows an example of dividing a predetermined number of unit cells for each template region acquired according to the second embodiment of the present invention.
이어서, 분할된 각각의 단위셀 내에서 측정된 센서데이터를 이용하여 각 단위셀을 대표하는 대표값을 산출하며(S230), 산출된 각 단위셀별 대표값을 인자로 하는 지폐의 특징벡터를 추출한다(S240). 도 6에 도시된 바와 같이, 하나의 단위셀 내에는 IR센서 어레이에 각각 대응하는 다수개의 픽셀이 존재하고, 각각의 픽셀은 각각의 대응하는 IR센서에서 측정된 센서데이터를 가진다. 본 발명에서는 이러한 센서데이터를 이용하여 각 단위셀을 대표하는 하나의 스칼라값(대표값)을 산출하게 되는데, 이때 각 단위셀을 대표하는 대표값으로는 각 단위셀을 구성하는 센서데이터들의 평균, 분산 또는 최대값 등 다양한 인자를 적용할 수 있으며, 바람직하게는 각 단위셀 내 센서데이터의 특징을 가장 효과적으로 반영할 수 있는 단위셀내 센서데이터의 평균값을 사용하는 것이 좋다.Subsequently, a representative value representing each unit cell is calculated using the sensor data measured in each divided unit cell (S230), and a feature vector of a banknote having a calculated representative value for each unit cell is extracted. (S240). As shown in FIG. 6, a plurality of pixels respectively corresponding to the IR sensor array exist in one unit cell, and each pixel has sensor data measured by each corresponding IR sensor. In the present invention, the sensor data is used to calculate one scalar value (representative value) representing each unit cell. In this case, the representative value representing each unit cell is an average of sensor data constituting each unit cell, Various factors such as variance or maximum may be applied, and it is preferable to use an average value of sensor data in a unit cell that can most effectively reflect the characteristics of sensor data in each unit cell.
예를 들어, 도 6에 도시된 바와 같이 각각의 단위셀은 총 42개의 단위 픽셀로 이루어지고, 이들 각 픽셀별로 센서데이터가 얻어진다고 가정할 때, 상기 단위셀 A[1,4]내에 존재하는 42개 픽셀의 센서데이터의 평균값인 9가 단위셀 A[1,4]의 대표값(A1,4)으로 산출되고, 산출된 대표값은 해당 단위셀을 대표하는 특징벡터의 인자가 된다. For example, as shown in FIG. 6, each unit cell is composed of 42 unit pixels in total, and assuming that sensor data is obtained for each pixel, each unit cell exists in the unit cell A [1,4]. 9, which is an average value of the sensor data of 42 pixels, is calculated as the representative value A 1,4 of the unit cell A [1,4], and the calculated representative value is a factor of the feature vector representing the unit cell.
즉, 이와 같은 방법으로 각 단위셀을 대표하는 대표값을 산출하여, 각각의 단위셀별 산출된 대표값을 인자로 하는 특징벡터를 추출하게 되며, 이에 따라 도 6에 도시된 바와 같이 분할된 템플릿 영역별 단위셀 개수만큼의 인자를 갖는 특징벡터(X)가 추출된다.That is, a representative value representing each unit cell is calculated in this manner, and a feature vector having a representative value calculated for each unit cell is extracted. Accordingly, the template region divided as shown in FIG. A feature vector (X) having as many factors as the number of unit cells is extracted.
상술한 방법으로 특징벡터(X)가 추출되면 추출된 특징벡터 인자들을 비교하여 지폐의 진위를 판별하게 되는데, 특징벡터의 인자가 너무 많을 경우 견실하고 빠른 진위 판별을 수행하기 어려우므로, 본 발명에서는 특징벡터의 차원을 감소시키는 단계를 수행한다(S250). 추출된 특징벡터의 차원을 감소시키는 이유는 특징벡터의 불필요한 부분을 제거하여 연산대상을 줄이고, 진위 판별에 있어 중요한 대표 특징벡터만을 추출하기 위함이다. 이에 따라, 선형특징추출법을 적용하여 추출된 특징벡터의 차원을 감소시키고, 적외선 패턴의 특징을 잘 나타내는 소정의 대표 특징벡터(dominant feature vector)만을 선별한다.When the feature vector (X) is extracted by the above-described method, the authenticity of the bill is determined by comparing the extracted feature vector factors. However, when the feature vector factor is too large, it is difficult to perform a robust and fast authenticity determination. A step of reducing the dimension of the feature vector is performed (S250). The reason for reducing the dimension of the extracted feature vectors is to reduce the computational object by removing unnecessary portions of the feature vectors, and to extract only representative feature vectors that are important for authenticity determination. Accordingly, the dimension of the extracted feature vector is reduced by applying the linear feature extraction method, and only a predetermined feature vector representing the characteristic of the infrared pattern is selected.
선형특징추출법은 센서데이터의 통계적인 특성을 분석하는 방법으로서 주성분 분석법(Principal component Analysis, PCA)과 선형 판별법(Linear Discriminant Analysis, LDA) 등이 그 대표적인 예이다. 본 실시예에서는 이미지 특징을 효과적으로 찾을 수 있는 비교사 통계학적 기법인 주성분 분석법(PCA)을 이용하여 특징벡터의 차원을 감소시키는 것이 효과적이며, 이하에서는 주성분 분석법(PCA)을 이용하여 추출된 특징벡터의 차원을 감소시키는 단계에 대해 간략하게 설명하기로 한다.The linear feature extraction method is a method of analyzing statistical characteristics of sensor data, and the principal component analysis (PCA) and linear discriminant analysis (LDA) are representative examples. In this embodiment, it is effective to reduce the dimension of the feature vector by using principal component analysis (PCA), which is a non-statistical statistical technique that can effectively find image features. Hereinafter, feature vectors extracted by using principal component analysis (PCA) The steps for reducing the dimension of the circuit will be briefly described.
주성분 분석법(PCA)은 추출된 특징벡터(X)를 구성하는 여러 인자들로부터 해당 특징벡터를 대표할 수 있는 소수의 주요 인자값들을 추출하여, 이를 주요 인자값들로 구성되어 최초의 특징벡터보다 감소된 차원을 갖는 대표 특징벡터를 추출해 내는데 사용되는 방법으로서, 본 발명의 경우, 수만개의 지폐로부터 얻어지는 지폐 특징벡터 데이터들을 이용하여 얻어진 공분산 매트릭스를 활용하여 대표 특징벡터의 추출을 위한 고유 매트릭스(ΦT)를 산출한다.Principal component analysis (PCA) extracts a few major factor values that can represent the feature vector from several factors constituting the extracted feature vector (X), which is composed of the major factor values As a method used to extract a representative feature vector having a reduced dimension, in the present invention, a unique matrix for extracting a representative feature vector using a covariance matrix obtained using banknote feature vector data obtained from tens of thousands of bills Calculate T ).
상술한 공분산 매트릭스를 이용한 고유 매트릭스(ΦT)의 산출과정은 이미 넓게 쓰이고 있는 통계적 추출방법의 하나로서 여기에서는 그에 대한 상세한 설명은 생략하기로 한다.The calculation process of the eigen matrix Φ T using the covariance matrix described above is one of widely used statistical extraction methods, and a detailed description thereof will be omitted.
즉, 추출된 특징벡터(X)에 상술한 과정을 통해 산출된 고유 매트릭스(ΦT)를 적용하여, 앞의 제 1실시예의 경우에서 설명한 [수학식 1]과 같이 특징벡터(X)의 차원을 감소시켜, 적외선 패턴의 특징을 보다 잘 나타내는 대표 특징벡터(Y)를 얻을 수 있다.That is, by applying the eigen matrix Φ T calculated through the above process to the extracted feature vector (X), the dimension of the feature vector (X) as shown in [Equation 1] described in the case of the first embodiment By reducing this, a representative feature vector Y that better represents the characteristics of the infrared pattern can be obtained.
이와 같이 산출된 대표 특징벡터(Y)는 최초에 추출된 특징벡터(X)보다 상대적으로 낮은 차원을 갖게 되고, 동시에 지폐의 진위 판별을 위한 특징을 보다 효과적으로 대표하게 된다. The representative feature vector (Y) calculated as described above has a relatively lower dimension than the feature vector (X) extracted initially, and at the same time more effectively represents a feature for authenticity determination of banknotes.
이어서, 획득한 대표 특징벡터에 GML(Gaussian Maximum Likelihood) 분류법을 이용하여 해당 지폐에 대한 진위 판별을 수행한다(S260). 즉, 상술한 과정을 통해 산출된 대표 특징벡터가 기 산출되어 있는 진폐 클래스 영역에 포함될 확률을 산출한 뒤, 그 확률값이 기 설정된 기준 보다 낮을 경우 해당 지폐를 위폐로 판별한다.Subsequently, authenticity of the corresponding bill is performed using the Gaussian Maximum Likelihood (GML) classification method on the obtained representative feature vector (S260). That is, after calculating the probability that the representative feature vector calculated through the above-described process is included in the pre-calculated pneumococcal class region, and if the probability value is lower than the preset reference, the banknote is determined as counterfeit.
상술한 주성분 분석법의 특징 중의 하나는 동일한 시간상에 분포하는 데이터 집단의 경우, 분산의 정도가 큰 방향들의 벡터를 구할 수 있다는 것이다. 즉, 공분산 행렬의 행렬 계산으로 고유값과 이에 해당하는 각각의 고유벡터들을 구할 수 있게 되는데, 이중 고유값이 높은 것에 해당하는 벡터가 그 데이터 집단의 중요한 요소가 되며, 보다 작은 고유값의 벡터는 첫번째 벡터보다는 중요도가 적은 벡터라는 의미를 가지게 된다. 따라서, 수만개의 진폐 특징벡터들을 이용하여 주성분 분석을 수행할 경우, 진폐에 대한 고유값들과 이에 해당하는 고유벡터들을 구할 수 있게 되며, 이를 이차원 그래프로 표시할 경우 진폐 클래스 영역을 산출할 수 있게 된다.One of the characteristics of the principal component analysis described above is that for a data group distributed over the same time, a vector of directions having a large degree of variance can be obtained. In other words, by calculating the matrix of the covariance matrix, the eigenvalues and corresponding eigenvectors can be obtained.The vector with the higher eigenvalue becomes an important element of the data group, and the smaller eigenvalue vector This means that the vector is less important than the first vector. Therefore, when principal component analysis is performed using tens of thousands of pneumatic features, eigenvalues for pneumococcal and corresponding eigenvectors can be obtained. do.
이때, 상기 대표 특징벡터가 해당 진폐 클래스 영역에 포함될 확률을 산출하는 방법은 앞의 제 1실시예의 경우에서 설명한 [수학식 2]와 유사한 과정을 통해 수행될 수 있다.In this case, the method of calculating the probability that the representative feature vector is included in the corresponding pneumatic class region may be performed through a process similar to that of [Equation 2] described in the case of the first embodiment.
이 경우, 상기 [수학식 2]에 나타나는 ML은 상기 대표 특징벡터가 해당 권종의 진폐 클래스에 포함될 확률을 대표하는 값으로서, 대표 특징벡터가 해당 진폐 클래스에 포함될 확률이 높을수록 진폐일 확률이 높아진다. 따라서, 대표 특징벡터와 데이터베이스에 기 저장된 진위클래스 평균벡터 및 분산행렬을 이용하여 진위 클래스에 포함될 ML값을 산출하고, 그 산출값이 기 설정된 기준값 이하일 경우, 해당 지폐를 위폐로 판단한다.In this case, ML shown in [Equation 2] is a value representing the probability that the representative feature vector is included in the pneumococcal class of the subject species, and the higher the probability that the representative feature vector is included in the pneumococcal class increases the probability of the pneumoconiosis. . Therefore, the ML value to be included in the authenticity class is calculated using the representative feature vector and the authenticity class average vector and the variance matrix previously stored in the database. When the calculated value is less than or equal to the preset reference value, the banknote is determined as counterfeit.
상술한 바와 같이, 본원 발명은 적외선 패턴이 존재하는 영역에 대한 단위셀별 대표값을 바탕으로 베이시안 접근법을 이용하여 지폐의 진위를 판별함으로써, 종래의 신경망 회로를 이용한 지폐 진위 판별 방법에 비해 진위 판단을 신속하게 수행할 수 있을 뿐만 아니라 낮은 해상도의 지폐이미지로도 상대적으로 정확하게 지폐의 진위를 판별할 수 있다.As described above, the present invention determines the authenticity of banknotes using a Bayesian approach based on a representative value for each unit cell for the region in which the infrared pattern exists, thereby determining authenticity of the banknote authenticity using a conventional neural network. Not only can it be performed quickly, but the authenticity of banknotes can be accurately determined even with a low resolution banknote image.
이상에서 설명한 본 발명은 전술한 실시예 및 첨부된 도면에 의해 한정된 것은 아니며, 본 발명의 기술적 사상을 벗어나지 않는 범위 내에서 여러 가지 치환, 변경 및 변환이 가능하다는 것은 본 발명이 속하는 기술분야에서 통상의 지식을 가진 자에게 있어 명백하다 할 것이다.The present invention described above is not limited to the above-described embodiments and the accompanying drawings, and it is common in the art that various substitutions, changes, and conversions are possible without departing from the technical spirit of the present invention. It will be clear to those who have knowledge of God.
본 발명에 따른 베이시안 접근법을 이용한 지폐 감별 방법은, 종래의 감별방법에 비해 고속의 데이터 처리가 가능하며, 입력 영상이 초 저해상도를 가질 경우에도 적용이 가능하여, 다양한 종류의 지폐처리장치에 폭넓게 적용될 수 있다.The banknote discrimination method using the Bayesian approach according to the present invention is capable of processing data at a higher speed than the conventional discrimination method, and can be applied even when the input image has an ultra low resolution. Can be applied.
더불어, 신권 추가 시 다시 학습과정을 거쳐야 하는 종래의 기계학습 방법과 달리 신권의 특징값을 기기에 추가하기만 하면 되므로 신권 추가 시의 권종 인식 및 진위 판별이 용이하고, 인식률 제어나 감별에러 복구 등을 위하여 알고리즘을 직접 수정할 필요가 없이 비교 레퍼런스 데이터만 갱신하면 되는 장점이 있어 다양한 권종을 사용하는 모든 금융자동화기기에 범용적으로 적용될 수 있다.In addition, unlike the conventional machine learning method that requires a re-learning process when adding a priesthood, it is easy to recognize priesthood type and authenticity when adding a priesthood, and to recognize recognition rate and to recover discrimination errors. For this reason, it is only necessary to update the comparison reference data without the need to directly modify the algorithm.

Claims (10)

  1. 지폐를 감별하는 방법에 있어서,In the method of discriminating a bill,
    지폐 전체를 스캔하여 얻어진 센서데이터를 이용하여 지폐이미지를 생성하는 단계;Generating a bill image using the sensor data obtained by scanning the whole bill;
    생성된 지폐이미지를 미리 정해진 일정 갯수의 단위셀로 분할하는 단계;Dividing the generated bill image into a predetermined number of unit cells;
    상기 분할된 각 단위셀별로, 얻어진 센서데이터를 이용하여 각 단위셀을 대표하는 대표값을 산출하는 단계;Calculating a representative value representing each unit cell by using the obtained sensor data for each of the divided unit cells;
    상기 산출된 각 단위셀별 대표값을 인자로 하는 지폐의 특징벡터를 추출하는 단계;Extracting a feature vector of a banknote having the calculated representative value for each unit cell as a factor;
    선형특징추출법을 이용하여 상기 추출된 지폐의 특징벡터의 차원을 감소시켜 대표 특징벡터를 추출하는 단계;Extracting a representative feature vector by reducing the dimension of the feature vector of the extracted banknote by using a linear feature extraction method;
    상기 추출된 대표 특징벡터에 GML(Gaussian Maximum Likelihood) 분류법을 적용하여 해당 지폐에 대한 권종 인식을 수행하는 단계;Applying a Gaussian Maximum Likelihood (GML) classification method to the extracted representative feature vector to perform roll paper recognition on the bill;
    를 포함하여 구성되는 것을 특징으로 하는 베이시안 접근법을 이용한 지폐 감별 방법.Banknote discrimination method using a Bayesian approach, characterized in that comprising a.
  2. 제 1항에 있어서,The method of claim 1,
    상기 생성된 지폐이미지를 단위셀로 분할함에 있어서는,In dividing the generated bill image into unit cells,
    생성된 지폐이미지를 가로 방향으로 10 내지 40개의 단위셀과, 세로방향으로 4 내지 20개의 단위셀로 각각 분할하는 것을 특징으로 하는 베이시안 접근법을 이용한 지폐 감별 방법.A banknote discrimination method using a Bayesian approach, wherein the generated banknote image is divided into 10 to 40 unit cells in a horizontal direction and 4 to 20 unit cells in a vertical direction.
  3. 제 1항에 있어서,The method of claim 1,
    상기 각 단위셀을 대표하는 대표값을 산출하는 단계는,Calculating a representative value representing each unit cell,
    각 단위셀 내에서 얻어진 센서데이터들의 평균, 분산 또는 최대값 중 하나를 대표값으로 산출하는 것을 특징으로 하는 베이시안 접근법을 이용한 지폐 감별 방법.A banknote discrimination method using a Bayesian approach, characterized in that one of the average, variance, or maximum of the sensor data obtained in each unit cell is calculated as a representative value.
  4. 제 1항에 있어서,The method of claim 1,
    상기 추출된 지폐의 특징벡터의 차원을 감소시켜 대표 특징벡터를 추출하는 단계에 적용하는 선형특징추출법은,The linear feature extraction method applied to the step of extracting the representative feature vector by reducing the dimension of the extracted feature vector,
    선형판별법(Linear Discriminant Analysis)인 것을 특징으로 하는 베이시안 접근법을 이용한 지폐 감별 방법.Banknote discrimination method using a Bayesian approach, characterized in that the linear discriminant analysis.
  5. 제 1항에 있어서,The method of claim 1,
    상기 추출된 대표 특징벡터에 GML 분류법을 적용하여 해당 지폐에 대한 권종 인식을 수행하는 단계는,Applying the GML classification method to the extracted representative feature vector to perform the paper type recognition for the bill,
    하기의 [수학식 3]을 이용하여 산출되는 계산식의 결과값 중, 가장 높은 ML값이 산출되는 권종으로 인식하는 것을 특징으로 하는 베이시안 접근법을 이용한 권종 인식 방법.A scoop recognition method using a Bayesian approach, characterized in that it recognizes as the scoop that the highest ML value is calculated among the results of the calculation calculated using Equation 3 below.
    [수학식 3][Equation 3]
    Figure PCTKR2012000548-appb-I000003
    Figure PCTKR2012000548-appb-I000003
    여기서, here,
    ML : 유사도(Maximum Likelyhood),ML: Maximum Likelyhood,
    i : 권종의 종류,i: type of winding species,
    Y : 대표 특징벡터,Y: representative feature vector,
    Figure PCTKR2012000548-appb-I000004
    : 권종별 분산행렬,
    Figure PCTKR2012000548-appb-I000004
    : Distributed matrix by volume,
    Figure PCTKR2012000548-appb-I000005
    : 권종별 평균벡터
    Figure PCTKR2012000548-appb-I000005
    : Average vector by volume type
  6. 지폐를 감별하는 방법에 있어서,In the method of discriminating a bill,
    적외선 센서를 통해 지폐 전체를 스캔하여 센서데이터를 획득하고, 적외선 패턴이 존재하는 템플릿의 영역을 특정하는 단계;Scanning the whole banknote through the infrared sensor to obtain sensor data, and specifying an area of the template in which the infrared pattern exists;
    특정된 템플릿 영역을 미리 정해진 일정 갯수의 단위셀로 분할하는 단계;Dividing the specified template region into a predetermined number of unit cells;
    상기 분할된 단위셀별로, 얻어진 센서데이터를 이용하여 각 단위셀을 대표하는 대표값을 산출하는 단계;Calculating a representative value representing each unit cell by using the obtained sensor data for each of the divided unit cells;
    상기 산출된 각 단위셀별 대표값을 인자로 하는 지폐의 특징벡터를 추출하는 단계;Extracting a feature vector of a banknote having the calculated representative value for each unit cell as a factor;
    선형특징추출법을 이용하여 상기 추출된 지폐의 특징벡터의 차원을 감소시켜 대표 특징벡터를 추출하는 단계;Extracting a representative feature vector by reducing the dimension of the feature vector of the extracted banknote by using a linear feature extraction method;
    상기 추출된 대표 특징벡터에 GML(Gaussian Maximum Likelihood)분류법을 적용하여 해당 지폐에 대한 진위 판별을 수행하는 단계;Performing authenticity determination on the banknote by applying a Gaussian Maximum Likelihood (GML) classification method to the extracted representative feature vector;
    를 포함하여 구성되는 것을 특징으로 하는 베이시안 접근법을 이용한 지폐 감별 방법.Banknote discrimination method using a Bayesian approach, characterized in that comprising a.
  7. 제 6항에 있어서,The method of claim 6,
    상기 특정된 템플릿 영역을 단위셀로 분할함에 있어서는,In dividing the specified template region into unit cells,
    특정된 템플릿 영역의 스캔 면적에 따라, 각각의 템플릿 영역별로 가로 방향으로 2 내지 20개의 단위셀과, 세로방향으로 2 내지 10개의 단위셀로 분할하는 것을 특징으로 하는 베이시안 접근법을 이용한 지폐 감별 방법.A bill discrimination method using a Bayesian approach characterized by dividing into 2 to 20 unit cells in the horizontal direction and 2 to 10 unit cells in the vertical direction according to the scan area of the specified template region. .
  8. 제 6항에 있어서,The method of claim 6,
    상기 각 단위셀을 대표하는 대표값을 산출하는 단계는,Calculating a representative value representing each unit cell,
    각 단위셀 내에서 얻어진 센서데이터들의 평균, 분산 또는 최대값 중 하나를 대표값으로 산출하는 것을 특징으로 하는 베이시안 접근법을 이용한 지폐 감별 방법.A banknote discrimination method using a Bayesian approach, characterized in that one of the average, variance, or maximum of the sensor data obtained in each unit cell is calculated as a representative value.
  9. 제 6항에 있어서,The method of claim 6,
    상기 추출된 지폐의 특징벡터의 차원을 감소시켜 대표 특징벡터를 추출하는 단계에 적용하는 선형특징추출법은,The linear feature extraction method applied to the step of extracting the representative feature vector by reducing the dimension of the extracted feature vector,
    주성분 분석법(Principal Component Analysis)인 것을 특징으로 하는 베이시안 접근 법을 이용한 지폐 감별 방법.Banknote discrimination method using a Bayesian approach, characterized in that the principal component analysis (Principal Component Analysis).
  10. 제 6항에 있어서,The method of claim 6,
    상기 추출된 대표 특징벡터에 GML 분류법을 적용하여 해당 지폐에 대한 진위 판별을 수행하는 단계는,Applying the GML classification to the extracted representative feature vector to determine the authenticity of the bill,
    하기의 [수학식 4]를 이용하여 산출되는 계산식의 결과값(ML)이 기 설정된 기준값 미만일 경우 위폐로 판별하는 것을 특징으로 하는 베이시안 접근법을 이용한 지폐 감별 방법.A banknote discrimination method using a Bayesian approach, characterized in that the counterfeit determines if the result value (ML) of the calculation formula calculated using Equation 4 below is less than the predetermined reference value.
    [수학식 4][Equation 4]
    Figure PCTKR2012000548-appb-I000006
    Figure PCTKR2012000548-appb-I000006
    여기서, here,
    ML : 유사도(Maximum Likelyhood),ML: Maximum Likelyhood,
    i : 권종의 종류,i: type of winding species,
    Y : 대표 특징벡터,Y: representative feature vector,
    Figure PCTKR2012000548-appb-I000007
    : 권종별 분산행렬,
    Figure PCTKR2012000548-appb-I000007
    : Distributed matrix by volume,
    Figure PCTKR2012000548-appb-I000008
    : 권종별 평균벡터
    Figure PCTKR2012000548-appb-I000008
    : Average vector by volume type
PCT/KR2012/000548 2011-01-21 2012-01-20 Method for discriminating banknotes using a bayesian approach WO2012099435A2 (en)

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
KR10-2011-0006275 2011-01-21
KR1020110006280A KR101232684B1 (en) 2011-01-21 2011-01-21 Method for detecting counterfeits of banknotes using Bayesian approach
KR10-2011-0006280 2011-01-21
KR1020110006275A KR101232683B1 (en) 2011-01-21 2011-01-21 Method for recognizing denomination of banknotes using Bayesian approach

Publications (3)

Publication Number Publication Date
WO2012099435A2 true WO2012099435A2 (en) 2012-07-26
WO2012099435A9 WO2012099435A9 (en) 2012-11-01
WO2012099435A3 WO2012099435A3 (en) 2012-12-20

Family

ID=46516266

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/KR2012/000548 WO2012099435A2 (en) 2011-01-21 2012-01-20 Method for discriminating banknotes using a bayesian approach

Country Status (1)

Country Link
WO (1) WO2012099435A2 (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2018060453A (en) * 2016-10-07 2018-04-12 グローリー株式会社 Currency classification device and currency classification method
WO2020003150A3 (en) * 2018-06-28 2020-04-23 3M Innovative Properties Company Image based novelty detection of material samples

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
GB2586568A (en) * 2019-04-18 2021-03-03 Gerard Rohan Jayamanne Don A technique for detecting counterfeit banknotes, travelers cheques or money orders by a method to be used with smart phone applications (apps)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH09171552A (en) * 1995-10-18 1997-06-30 Fuji Xerox Co Ltd Picture recognizing device
US20030192765A1 (en) * 1998-12-02 2003-10-16 Mars Incorporated, A Delaware Corporation Classification method and apparatus
KR100751855B1 (en) * 2006-03-13 2007-08-23 노틸러스효성 주식회사 Recognizing the denomination of a note using wavelet transform
KR20070093209A (en) * 2006-03-13 2007-09-18 노틸러스효성 주식회사 Recognizing the denomination of a note using wavelet transform

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH09171552A (en) * 1995-10-18 1997-06-30 Fuji Xerox Co Ltd Picture recognizing device
US20030192765A1 (en) * 1998-12-02 2003-10-16 Mars Incorporated, A Delaware Corporation Classification method and apparatus
KR100751855B1 (en) * 2006-03-13 2007-08-23 노틸러스효성 주식회사 Recognizing the denomination of a note using wavelet transform
KR20070093209A (en) * 2006-03-13 2007-09-18 노틸러스효성 주식회사 Recognizing the denomination of a note using wavelet transform

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2018060453A (en) * 2016-10-07 2018-04-12 グローリー株式会社 Currency classification device and currency classification method
EP3511912A4 (en) * 2016-10-07 2020-05-06 Glory Ltd. Currency classification device and currency classification method
US11017273B2 (en) 2016-10-07 2021-05-25 Glory Ltd. Currency classification device and currency classification method
WO2020003150A3 (en) * 2018-06-28 2020-04-23 3M Innovative Properties Company Image based novelty detection of material samples
US11816946B2 (en) 2018-06-28 2023-11-14 3M Innovative Properties Company Image based novelty detection of material samples

Also Published As

Publication number Publication date
WO2012099435A3 (en) 2012-12-20
WO2012099435A9 (en) 2012-11-01

Similar Documents

Publication Publication Date Title
CN107610322B (en) Banknote version identification method and device, electronic equipment and storage medium
Ghadekar et al. Handwritten digit and letter recognition using hybrid dwt-dct with knn and svm classifier
CN102667810B (en) Face recognition in digital images
Laavanya et al. Real time fake currency note detection using deep learning
CN102722708B (en) Method and device for classifying sheet media
CN108108753B (en) Identification method and device for check box selection state based on support vector machine
WO2007068867A1 (en) Banknote validation
CN114492768B (en) Twin capsule network intrusion detection method based on small sample learning
JPH09282517A (en) Identifying method for paper sheets
Pham et al. Deep learning-based fake-banknote detection for the visually impaired people using visible-light images captured by smartphone cameras
KR101232683B1 (en) Method for recognizing denomination of banknotes using Bayesian approach
CN104969268A (en) Authentication of security documents and mobile device to carry out the authentication
US8170332B2 (en) Automatic red-eye object classification in digital images using a boosting-based framework
WO2012099435A2 (en) Method for discriminating banknotes using a bayesian approach
CN102110323A (en) Method and device for examining money
Darwish et al. Building an expert system for printer forensics: A new printer identification model based on niching genetic algorithm
CN106683261B (en) A kind of false distinguishing method and apparatus of bank note
Dittimi et al. Multi-class SVM based gradient feature for banknote recognition
KR101232684B1 (en) Method for detecting counterfeits of banknotes using Bayesian approach
CN111553361A (en) Pathological section label identification method
Joshi et al. Source printer identification from document images acquired using smartphone
Gebremeskel et al. Developing a Model for Detection of Ethiopian Fake Banknote Using Deep Learning
Xu et al. Coin recognition method based on SIFT algorithm
US11100322B2 (en) Character authenticity determination
CN110415424B (en) Anti-counterfeiting identification method and device, computer equipment and storage medium

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 12736612

Country of ref document: EP

Kind code of ref document: A2

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205N DATED 27/09/2013)

122 Ep: pct application non-entry in european phase

Ref document number: 12736612

Country of ref document: EP

Kind code of ref document: A2