WO2017197884A1 - 一种纸币管理方法及其系统 - Google Patents

一种纸币管理方法及其系统 Download PDF

Info

Publication number
WO2017197884A1
WO2017197884A1 PCT/CN2016/112111 CN2016112111W WO2017197884A1 WO 2017197884 A1 WO2017197884 A1 WO 2017197884A1 CN 2016112111 W CN2016112111 W CN 2016112111W WO 2017197884 A1 WO2017197884 A1 WO 2017197884A1
Authority
WO
WIPO (PCT)
Prior art keywords
banknote
image
information
module
management method
Prior art date
Legal status (The legal status 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 status listed.)
Ceased
Application number
PCT/CN2016/112111
Other languages
English (en)
French (fr)
Inventor
柳永诠
柳伟生
孙伟忠
赵楠楠
王福艳
金彬
刘云江
卢丙峰
崔彦身
金迪
焦仁刚
戈兰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Julong Co Ltd
Original Assignee
Julong Co Ltd
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
Application filed by Julong Co Ltd filed Critical Julong Co Ltd
Priority to KR1020187037126A priority Critical patent/KR102207533B1/ko
Priority to JP2019513099A priority patent/JP6878575B2/ja
Priority to RU2018145018A priority patent/RU2708422C1/ru
Priority to EP16902263.9A priority patent/EP3460765B1/en
Priority to US16/303,355 priority patent/US10930105B2/en
Publication of WO2017197884A1 publication Critical patent/WO2017197884A1/zh
Priority to SA518400454A priority patent/SA518400454B1/ar
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • 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
    • G07D7/2016—Testing patterns thereon using feature extraction, e.g. segmentation, edge detection or Hough-transformation
    • 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
    • G—PHYSICS
    • G07—CHECKING-DEVICES
    • G07D—HANDLING OF COINS OR VALUABLE PAPERS, e.g. TESTING, SORTING BY DENOMINATIONS, COUNTING, DISPENSING, CHANGING OR DEPOSITING
    • G07D11/00—Devices accepting coins; Devices accepting, dispensing, sorting or counting valuable papers
    • G07D11/20—Controlling or monitoring the operation of devices; Data handling
    • G07D11/28—Setting of parameters; Software updates
    • 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
    • G07D7/004—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 using digital security elements, e.g. information coded on a magnetic thread or strip
    • 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
    • G07D7/2008—Testing patterns thereon using pre-processing, e.g. de-blurring, averaging, normalisation or rotation
    • 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
    • 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

Definitions

  • the invention belongs to the financial field, and particularly relates to a banknote management system and a method thereof.
  • bank system With the continuous improvement of the level of financial information application, the bank system's currency anti-counterfeiting, business process management and financial security are gradually becoming more intelligent.
  • Banknote management can maintain the security and stability of the national financial sector, realize the circulation of RMB circulation, counterfeit currency management, ATM.
  • Banknote management, asset management and cash in and out of the warehouse are of great significance.
  • Banknote management mainly deals with the comprehensive processing of information such as banknote information and business information.
  • the number of the banknotes in the banknote information plays an increasingly important role in the management of banknotes.
  • the DSP identification mode it is often limited to the network transmission efficiency and the influence of the position and orientation of the banknote in the DSP identification.
  • the recognition efficiency and the robustness of the recognition algorithm are relatively poor.
  • the patent with the application number CN201510702688.2 In the application, the edge is fitted by the gray threshold and direction search, and the edge line is filtered by the threshold to obtain the slope of the region.
  • the recognition face is backward, and the progressive neural network recognizes Crown number.
  • the prior art has the following problems: the orientation of the banknotes and the effective positioning of the characters cannot be solved efficiently, and the range of characters after recognition is large, which is easy to cause incorrect division of characters, and the amount of data for post-image processing and recognition is large. , the recognition efficiency is reduced; the rapid tilt change of the banknote image caused by the banknote cannot be well adapted, the tilt of the banknote cannot be corrected and recognized in time; the robustness of the identification of the damaged banknote is low, and the corresponding banknote damage is not provided. Identify the processing method.
  • the first technical problem to be solved by the present invention is that the banknote management system in the prior art cannot achieve high-efficiency accurate collection and identification of banknote information, thereby providing a banknote capable of efficiently and accurately collecting and identifying banknote information. Management methods and their systems.
  • the second technical problem to be solved by the present invention is to propose a method for identifying a crown number, which effectively solves the problem that the object to be identified is damaged, dirty, quickly folded, etc., while ensuring the efficiency of the crown number recognition.
  • the banknote management method of the present invention comprises the following steps:
  • step 2) transmitting the banknote characteristic information, the business information, and the information of the banknote information processing device described in step 1) to the main control server;
  • the main control server performs integration processing on the received banknote characteristic information, the business information, and the information of the banknote information processing device, and classifies the banknotes.
  • step 1) one or more of image, infrared, fluorescent, magnetic, and thickness measurement are used.
  • the banknote features are collected.
  • the sorting process of the banknotes in the step 3) is specifically: classifying the banknotes into the different coin bins according to the classified categories.
  • the barn is a container or space that houses the banknotes.
  • the banknote information includes one or more of a currency, a face value, a face, an authenticity, a newness, a stain, and a crown number; wherein the face refers to the forward and reverse orientation of the banknote.
  • the service information includes record information of payment, payment, deposit or withdrawal, business time period information, operator information, transaction card number information, identity information of the person and/or agent, two-dimensional code information, and a package number. One or more of them.
  • the identifying of the banknote feature specifically comprises the following steps:
  • Step a extracting the gray image of the region where the banknote feature is located, and performing edge detection on the gray image; the edge detection can be realized by a conventional canny detection, sobel detection, etc., and then combined with a straight line to obtain an edge line equation.
  • Step b rotating the image; correcting and mapping the image of the banknote after the edge detection to correct the image, thereby facilitating segmentation and recognition of the number image.
  • the rotation method may adopt a coordinate point conversion method, or Correction is performed according to the detected edge equation, and the transformation equation is obtained, which can also be implemented by polar coordinate rotation or the like;
  • Step c Locating a single number in the image, specifically: performing binarization processing on the image by adaptive binarization to obtain a binarized image; and then projecting the binarized image, the conventional image
  • the projection is completed by only one vertical projection and one horizontal projection.
  • the specific projection direction and number of times can be adjusted according to the specific environment and accuracy requirements of the recognition. For example, it is also possible to use a projection with a tilt angle direction, or multiple times. Projection combination; Finally, by setting the moving window, using the moving window registration method, the number is segmented, and the image of each number is obtained. Due to common problems such as damage and dirt of the banknote, there is dirt on the image of the crown number.
  • the present invention adds a moving window registration method after image projection. Precisely determine the position of the character; the way the moving window is registered, that is, by setting a fixed window, for example Similar to the window template mode, etc., the number area can be narrowed to achieve more accurate area positioning, and all the ways of sliding matching by setting a fixed window can be applied to the present application;
  • Step d tightening characters included in the image of each number, and returning each number image
  • the normalization includes size normalization and shading normalization; the tightening operation of the character is based on the step c, and the detailed positioning is performed on the character that divides the approximate position. To further reduce the amount of data to be processed by subsequent image recognition, which greatly ensures the overall operating speed of the system;
  • Step e using a neural network to identify the normalized number image to obtain a banknote feature; preferably, the banknote feature is a crown number.
  • the edge detection in the step a further comprises: setting a grayscale threshold value, performing a line search from the upper and lower directions according to the threshold value, acquiring an edge, and detecting the edge, using a straight line scanning method to obtain The pixel coordinates of the edge line; then the least squares method is used to obtain the edge line equation of the image, and at the same time obtain the horizontal length, vertical length and slope of the banknote image.
  • the rotating in the step b further comprises: obtaining a rotation matrix based on the horizontal length, the vertical length and the slope, and determining the rotated pixel point coordinates according to the rotation matrix.
  • the rotation matrix can be obtained by means of polar coordinate conversion, that is, a polar coordinate transformation matrix.
  • polar coordinate conversion that is, a polar coordinate transformation matrix.
  • the inclination angle of the banknote can be obtained by the obtained linear equation of the edge, and the pixel points are calculated according to the angle and the length of the edge.
  • Polar coordinate transformation matrix can also be calculated by ordinary coordinate transformation method, for example, according to the inclination angle and the edge length, the center point of the banknote is set as the coordinate origin, and the conversion matrix of each coordinate point in the new coordinate system is calculated.
  • other matrix transformation methods can also be used to correct the rotation of the banknote image.
  • the performing binarization processing on the image by adaptive binarization specifically includes: obtaining a histogram of the image, and setting a threshold Th, when the gray value in the histogram is from 0 to When the number of points of Th is greater than or equal to a predetermined value, Th is used as an adaptive binarization threshold at this time, and the image is binarized to obtain a binarized image.
  • the projecting of the binarized image is performed in three different directions.
  • the moving window registration in the step c specifically includes: designing a moving window for registration, the window is horizontally moved on the vertical projection view, and the position corresponding to the minimum value of the total number of black points in the window is a crown The best position for dividing the word number in the left and right direction.
  • the window is a pulse sequence with a fixed interval, and the width between the pulses is preset by an interval between the image of the crown number.
  • each of said pulses has a width of 2-10 pixels.
  • the tightening in the step d includes: separately performing binarization on the image of each number, performing regional growth on the binarized image of each acquired number, and finally, re-orienting the region.
  • the area obtained after the growth One or two areas whose area is larger than a certain preset area threshold are selected, and the rectangle in which the selected area is located is a rectangle that is tightly packed by each number image.
  • the growth of the region may employ, for example, an eight-neighbor region growth algorithm.
  • the image of each number is separately binarized, specifically comprising: extracting a histogram of the image of each number, obtaining a binarization threshold by a histogram bimodal method, and then performing binarization according to the binarization
  • the threshold binarizes the image of each of the numbers.
  • the size normalization in the step d is size normalization using a bilinear interpolation algorithm.
  • the normalized size is one of the following: 12*12, 14*14, 18*18, 28*28, and the unit is a pixel.
  • the shading normalization in the step d comprises: acquiring a histogram of the image of each number, calculating a number foreground gray average value and a background gray average value, and normalizing the brightness and darkness
  • the pixel gray value is compared with the foreground gray average and the background gray average, respectively, and the pixel gray value before normalization is set to the corresponding specific gray value according to the comparison result.
  • step b and step c further comprising a face-facing determining step of: determining a banknote size by the rotated image, determining a face value according to the size; dividing the target banknote image into n blocks, and calculating The average value of the brightness in each block is compared with the template stored in advance, and when the difference is the smallest, it is determined as the face of the template.
  • the template can be pre-set in a variety of ways, as long as it can be contrasted by banknote images, such as different denominations, different brightness values, color differences, or other features that can be converted to brightness values. Used as a comparison template.
  • the pre-stored template divides images of different faces of different denominations into n blocks, and calculates a brightness average value in each block as a template.
  • the step of judging the newness degree is further included: first extracting an image of a preset number of dpi, using the entire area of the image as a feature area of the histogram, and scanning the pixel points in the area, In the array, the histogram of each pixel is recorded, and a certain proportion of the brightest pixel is counted according to the histogram, and the average gray value of the brightest pixel is obtained as the basis for judging the newness and the oldness.
  • the preset number of dpi images may be, for example, a 25 dpi image, etc., and the certain ratio may be adjusted according to specific needs, and may be, for example, 40%, 50%, or the like.
  • a damage identification step of: obtaining a transmitted image by separately arranging a light source and a sensor on both sides of the banknote; and detecting the transmitted image after the rotation point by point, when the point is of When two adjacent pixels are simultaneously less than a predetermined threshold, the point is determined to be a damage point.
  • the detection of the damage point can be divided into broken corner damage, hole breakage and the like in more detail.
  • the step of recognizing the handwriting further comprises: in the fixed area, scanning the pixels in the area, placing them in an array, recording a histogram of each pixel point, and counting according to the histogram A preset number of the brightest pixels is obtained, and an average gray value is obtained, and a threshold is obtained according to the average gray value, and a pixel whose gray value is smaller than the threshold is determined as a character point.
  • the preset number may be, for example, 20, 30, etc., and is not understood here as a limitation of the protection range; the threshold value may be derived according to the average gray scale value, and various methods may be adopted, and the average gray value may be directly used as the threshold value. It is also possible to use the function of the average gray value as a variable to solve the threshold.
  • the neural network in the step e adopts a secondary classification convolutional neural network; the first level classification classifies all the numbers and letters involved in the crown number, and the second level classification respectively in the first level classification Some categories are reclassified.
  • the number of categories of the first-level classification may be set according to classification needs and setting habits, and may be, for example, 10 categories, 23 categories, 38 categories, etc., and is not limited thereto.
  • the second-level classification is also based on the first-level classification. For the classifications that are easy to misjudge, feature approximation or low accuracy, the second-level classification is performed again, so that the crown number is further advanced with a higher recognition rate. Differentiate and identify, and the number of specific input categories and the number of output categories of the second-level classification can be set according to the category setting of the first-level classification, classification needs and setting habits, etc., and is not limited thereto. .
  • the network model structure of the convolutional neural network is set as follows:
  • Input layer only one image as a visual input, the image being a grayscale image of a single crown number to be identified;
  • C1 layer is a convolution layer, which consists of 6 feature maps
  • S2 layer is a downsampling layer, and uses image local correlation principle to subsample the image
  • C3 layer is a convolution layer, using a preset convolution kernel to deconvolution layer S2, each feature map in the C3 layer is connected to S2 by means of incomplete connection;
  • S4 layer for the downsampling layer, sub-sampling the image by using the principle of image local correlation
  • the C5 layer is a simple stretch of the S4 layer and becomes a one-dimensional vector
  • the number of outputs of the network is the number of classifications, and the C5 layer constitutes a fully connected structure.
  • the C1 layer and the C3 layer are each convoluted by a 3x3 convolution kernel.
  • the banknote information processing device is one or more of a banknote sorting machine, a money counter, and a money detector;
  • the information of the banknote information processing device is one or more of a manufacturer, a device number, and a financial institution in which it is located.
  • the banknote information processing device is a self-service financial device; the information of the banknote information processing device is one or more of a banknote record, a banknote number, a manufacturer, a device number, and a financial institution in which it is located.
  • the banknote management method is characterized in that a plurality of the banknote processing information devices respectively collect, identify and process banknote information in their respective businesses, and transmit the banknote information to a site host or a cash center host, and then The branch host or cash center host transmits the banknote information to the master server.
  • the present invention also provides a banknote management system, the banknote management system including a banknote information processing terminal and a main control server;
  • the banknote information processing terminal comprises a banknote sending module, a detecting module and an information processing module;
  • the banknote sending module is configured to transport the banknotes to the detecting module
  • the detecting module collects and recognizes the characteristics of the banknote
  • the information processing module processes and processes the banknote features collected and recognized by the detecting module, outputs the banknote characteristic information, and transmits the same;
  • the main control server is configured to receive the banknote characteristic information, the business information, and the information of the banknote information processing terminal, process the received three types of information, and classify the banknotes.
  • the main control server processes the received information, and specifically includes processing such as aggregation, storage, sorting, querying, tracking, and exporting.
  • the detection module can also be applied to the identification system of the crown number of the DSP platform, and can be embedded or coupled to a conventional money detector, a money counter, an ATM, etc., and the like.
  • the detection module includes Image preprocessing module, processor module, CIS image sensor module;
  • the image preprocessing module further includes an edge detection module and a rotation module;
  • the processor module further includes a number positioning module, a tightening module, a normalization module, and an identification module;
  • the number positioning module performs binarization processing on the image by adaptive binarization to obtain a binarized image; then, the binary image is projected; finally, by setting a moving window, using a moving window to register By dividing the number, obtaining an image of each number, and transmitting the image of each number to the tightening module; the manner of registering the moving window, that is, by setting a fixed window, for example, a window template manner Etc., narrow the number area to achieve more accurate area positioning, and all the way to set the fixed window sliding match can be applied to this application.
  • a fixed window for example, a window template manner Etc.
  • the normalization module is configured to normalize the image processed by the tightening module; preferably, the normalization includes size normalization and light and dark normalization.
  • the number positioning module further comprises a window module, and the window module designs a registration moving window according to the crown number spacing, horizontally moves the window on the vertical projection view, and calculates black in the window.
  • the window module designs a registration moving window according to the crown number spacing, horizontally moves the window on the vertical projection view, and calculates black in the window. The sum of points;
  • the window module can also compare the sum of the black points in different windows.
  • the tightening module separately binarizes the image of each number, performs regional growth on the binarized image of each obtained number, and finally selects one of the regions obtained after the region is grown. Or two areas whose area is larger than a certain preset area threshold, and the rectangle in which the selected area is located is a rectangle that is tightly packed by each number image.
  • the growth of the region may employ, for example, an eight-neighbor region growth algorithm.
  • the image of each number is separately binarized, specifically comprising: extracting a histogram of the image of each number, obtaining a binarization threshold by a histogram bimodal method, and then performing binarization according to the binarization
  • the threshold binarizes the image of each of the numbers.
  • the detecting module further includes a compensation module for compensating an image obtained by the CIS image sensor module, and the compensation module pre-stores pure white and pure black collected luminance data, and combines the settable pixel points. Gray reference value, to obtain a compensation coefficient;
  • the compensation coefficients are stored to the processor module and a lookup table is created.
  • the identification module uses the trained neural network to implement the identification of the crown number.
  • the neural network adopts a two-class classified convolutional neural network; the first-level classification classifies all the numbers and letters involved in the crown number, and the second-level classification separately performs the partial categories in the first-level classification. classification.
  • the number of categories of the first-level classification may be set according to classification needs and setting habits, and may be, for example, 10 categories, 23 categories, 38 categories, etc., and is not limited thereto.
  • the second-level classification is also based on the first-level classification. For the classifications that are easy to misjudge, feature approximation or low accuracy, the second-level classification is performed again, so that the crown number is further advanced with a higher recognition rate. Differentiate and identify, and the number of specific input categories and the number of output categories of the second-level classification can be set according to the category setting of the first-level classification, classification needs and setting habits, etc., and is not limited thereto. .
  • the network model structure of the convolutional neural network is set as follows:
  • Input layer only one image as a visual input, the image is a grayscale image of a single crown number to be identified image;
  • C1 layer is a convolution layer, which consists of 6 feature maps
  • S2 layer is a downsampling layer, and uses image local correlation principle to subsample the image
  • C3 layer is a convolution layer, using a preset convolution kernel to deconvolution layer S2, each feature map in the C3 layer is connected to S2 by means of incomplete connection;
  • S4 layer for the downsampling layer, sub-sampling the image by using the principle of image local correlation
  • the C5 layer is a simple stretch of the S4 layer and becomes a one-dimensional vector
  • the number of outputs of the network is the number of classifications, and the C5 layer constitutes a fully connected structure.
  • the C1 layer and the C3 layer are each convoluted by a 3x3 convolution kernel.
  • the identification module further comprises a neural network training module for training the neural network.
  • the processor module can employ a chip system such as an FPGA.
  • the processor module further comprises: a face-oriented judging module, configured to determine the orientation of the banknote.
  • the processor module further includes a new and old degree judging module for judging the degree of oldness of the banknote.
  • the processor module further includes a damage identification module for identifying a broken location in the banknote.
  • the breakage includes corners, holes, and the like.
  • the processor module further includes a handwriting recognition module for identifying a handwriting on the banknote.
  • the main control server end classifies the banknotes by specifically classifying the banknotes into the different coin bins according to the classified categories.
  • the banknote characteristic information includes one or more of a currency, a face value, a face, an authenticity, a newness, a stain, and a crown number;
  • the service information includes record information of payment, payment, deposit or withdrawal, business time period information, operator information, transaction card number information, identity information of the person and/or agent, two-dimensional code information, and a package number.
  • record information of payment, payment, deposit or withdrawal business time period information, operator information, transaction card number information, identity information of the person and/or agent, two-dimensional code information, and a package number.
  • the banknote information processing terminal is one of a banknote sorting machine, a money counter, a money detector, and a self-service financial device; further preferably, the self-service financial device is an automatic teller machine (ATM) or an automatic deposit machine.
  • ATM automatic teller machine
  • the present invention also provides a banknote information processing terminal which is the banknote information processing terminal included in the banknote management system.
  • the banknote management method of the present invention can realize intelligent management of the crown number.
  • the banknote information of the bank clearing device can be traced, the counterfeit currency management, the unified management of the crown number, the business electronic log, and the data.
  • the management and operation cost of the extension equipment can also promote the good operation of equipment such as the sorting machine and the money counter;
  • the banknote management method of the invention realizes the high efficiency of collecting and identifying the banknote information while ensuring the accuracy of the identification information, especially in the identification of the crown number, ensuring the overall method and the speed of the system operation. In this case, the robustness of the method is improved, and it is well able to cope with the difficulty in identifying the crown number recognition due to banknote fouling, incompleteness, and rapid folding;
  • the method provided by the invention occupies less system resources and is faster than conventional algorithms in the prior art, and can be well combined with devices such as ATMs and money detectors.
  • FIG. 1 is a schematic diagram of a method for identifying an embodiment of the present invention
  • FIG. 2 is a schematic diagram of an edge detection method according to an embodiment of the present invention.
  • FIG. 3 is a schematic view showing a banknote image and an actual banknote in a banknote handling process according to an embodiment of the present invention
  • FIG. 4 is a schematic view showing an arbitrary point rotation of a banknote according to an embodiment of the present invention.
  • FIG. 5 is a schematic diagram of setting a mobile window according to an embodiment of the present invention.
  • FIG. 6 is a schematic structural diagram of a neural network according to an embodiment of the present invention.
  • This embodiment provides a banknote management method, which specifically includes the following steps:
  • the banknotes are The information processing device collects the characteristics of the banknote by means of image, infrared, fluorescent, magnetic, and thickness measurement.
  • the banknote characteristic information includes currency type, face value, face, authenticity, newness, degree of stain, and crown number; as a specific implementation manner of the embodiment, the banknote information processing device is a banknote sorting machine;
  • the information of the processing device is the manufacturer, the device number, and the financial institution where it is located;
  • banknote information processing devices is not unique, including but not limited to six, at least one;
  • the banknote information processing device may also be one or more of a money counter or a money detector; the information of the banknote information processing device may also be omitted from the manufacturer and the device. Number or one or more of the financial institutions in which it is located;
  • the banknote information processing device may also be a self-service financial device; specifically, the banknote information processing device may be an automatic teller machine, an automatic deposit machine, or a circulating automatic teller machine. , self-service inquiry machine, self-service payment machine.
  • the information of the banknote information processing device may be one or more of a banknote record, a banknote number, a manufacturer, a device number, and a financial institution;
  • the service information includes record information of payment, payment, deposit or withdrawal, business time period information, operator information, transaction card number information, identity information of the agent and the agent, QR code information, packet number;
  • the manner in which the banknote feature information is transmitted to the main control server is not unique, and those skilled in the art may change the banknote feature information, the service information, and the banknote information processing device according to actual conditions.
  • the transmission path of the information for example, directly transmitting the banknote characteristic information, the information of the banknote information processing device, and the business information described in step 1) to the main control server;
  • a person skilled in the art may also omit or replace part of the service information in this embodiment according to actual needs, that is, omit or replace the record information of payment, payment, deposit or withdrawal, business time period information, operator information, Transaction card number information, one or more of the identity information of the agent and the agent, the QR code information, and the package number;
  • the main control server performs integration processing on the received banknote characteristic information, the business information, and the information of the banknote information processing device, and classifies the banknotes.
  • the sorting process of the banknotes is specifically: after sorting the banknotes, the banknotes are sorted into different coin bins according to the classified categories.
  • the method for identifying the feature of the banknote is described below as an example. As shown in FIG. 1 , the method includes the following steps:
  • Step a extracting the grayscale image of the region where the crown number is located, and performing edge detection on the grayscale image; the edge detection can be implemented by conventional canny detection, sobel detection, etc., and then combined with straight line fitting to obtain an edge straight line equation. However, it is necessary to test the empirical threshold for edge detection to ensure the speed of the method.
  • the edge detection in the step a further includes: setting a grayscale threshold, performing a line search from the upper and lower directions according to the threshold, acquiring an edge, and detecting the edge, using a straight line sweep In the face mode, the pixel coordinates of the edge line are obtained; then the edge line equation of the image is obtained by the least square method, and the horizontal length, vertical length and slope of the banknote image are obtained at the same time.
  • a threshold linear regression segmentation technique can be adopted, and the calculation speed is fast, and is not limited by the image size, and is detected at other edges.
  • the threshold linear regression segmentation technique only a small number of pixel points need to be found on the upper and lower edges, and the straight line equation of the edge can be quickly determined by the straight line fitting method. No matter how big or small the image is, you can find a small number of points to calculate.
  • the straight line search method is used here to detect the edge of the banknote from the upper and lower directions.
  • Step b Rotating the image; correcting and mapping the image of the banknote after the edge detection to correct the image, thereby facilitating segmentation and recognition of the number image.
  • the rotation method may adopt a coordinate point transformation method.
  • the method, or correction according to the detected edge equation obtains the transformation equation, and can also be implemented by polar coordinate rotation or the like;
  • the rotating in step b further comprises: obtaining a rotation matrix based on the horizontal length, the vertical length, and the slope, and determining the rotated pixel point coordinates according to the rotation matrix.
  • the rotation matrix can be obtained by means of polar coordinate conversion, that is, a polar coordinate transformation matrix.
  • polar coordinate conversion that is, a polar coordinate transformation matrix.
  • the inclination angle of the banknote can be obtained by the obtained linear equation of the edge, and the pixel points are calculated according to the angle and the length of the edge.
  • Polar coordinate transformation matrix can also be calculated by ordinary coordinate transformation method, for example, according to the inclination angle and the edge length, the center point of the banknote is set as the coordinate origin, and the conversion matrix of each coordinate point in the new coordinate system is calculated.
  • other matrix transformation methods can also be used to correct the rotation of the banknote image.
  • the image can be rotated and corrected by a rectangular coordinate transformation. Since p points are collected per millimeter in the horizontal direction during image acquisition, q is collected per millimeter in the vertical direction. a point.
  • the rotation of the banknote image at any point, the whole process of rotation is a point A(x s , y s ) on the image of the banknote given arbitrarily, and the point A' corresponding to the actual banknote is found (x' s , y ' s ), the point B' is rotated by the angle ⁇ to obtain the point B'(x' d , y' d ), and finally the point B' is found to correspond to the point B (x d , y d ) on the image of the rotated banknote.
  • Step c Locating a single number in the image, specifically: performing binarization processing on the image by adaptive binarization to obtain a binarized image; and then projecting the binarized image, the conventional image
  • the projection is completed by only one vertical projection and one horizontal projection.
  • the specific projection direction and number of times can be adjusted according to the specific environment and accuracy requirements of the recognition. For example, it is also possible to use a projection with a tilt angle direction, or multiple times. Projection combination; Finally, by setting the moving window, using the moving window registration method, the number is segmented, and the image of each number is obtained. Due to common problems such as damage and dirt of the banknote, there is dirt on the image of the crown number.
  • the present invention adds a moving window registration method after image projection. Precisely determine the position of the character;
  • the performing binarization processing on the image by adaptive binarization specifically includes: obtaining a histogram of the image, setting a threshold Th, when the gray in the histogram When the number of points from 0 to Th is greater than or equal to a preset value, and Th is used as an adaptive binarization threshold at this time, the image is binarized to obtain a binarized image; The image is projected and a total of three different directions are projected.
  • the setting the moving window specifically includes: the window horizontally moving on the vertical projection view, and the sum of the black dots in the window The position corresponding to the minimum value is the optimal position for dividing the left and right direction of the crown number.
  • an overall adaptive binarization method may be employed for the binarization of the image.
  • the blacker color is the crown number area
  • the brighter white is the background area.
  • the number of points whose gradation value is 0 to Th and N are obtained on the histogram.
  • N> 2200 (empirical value)
  • the corresponding threshold Th is the threshold of adaptive binarization.
  • the binarized image is projected, and the three projections can be combined to determine the up, down, left, and right positions of each number.
  • the first horizontal projection determine the row where the number is located
  • the second projection in the vertical direction determine the position of each digit in the left and right direction
  • the third time is to horizontal projection of each small image, determine each The position of the number in the up and down direction.
  • the above-mentioned three projection method can achieve good results for single number segmentation of most banknotes, but is dirty for the image of the crown number, and the effect of sticking between characters and characters is poor. Especially for the adhesion of three or more characters, it can hardly be separated.
  • a window movement registration method can be employed. Because the resolution of the crown number and size of the clearing machine is fixed, the size of each character is fixed, and the spacing between each character is also fixed. The design of the window can be designed according to the spacing of the crown numbers on the banknotes, as shown in Figure 5.
  • the window moves horizontally on the vertical projection map, and the position corresponding to the minimum value of the total number of black points in the window is the optimal position for dividing the left and right direction of the crown number. Since the recognition algorithm is used on the banknote sorter, the accuracy and speed are satisfied, and the resolution of the original image is 200 dpi.
  • the design of the window has a pulse width of 4 pixels, and the width between the pulses is designed according to the interval between the number images. After testing, the method can fully meet the real-time and accuracy requirements of the banknote sorting machine.
  • Step d tightening characters included in the image of each number, and normalizing each number image, the normalization including size normalization and light and dark normalization;
  • the characters which are divided into the approximate positions are repositioned in detail to further reduce the amount of data to be processed by the subsequent image recognition, which greatly ensures the overall running speed of the system.
  • the cubic projection method is only a preliminary positioning of a single number, and for a lot of dirty single numbers, it cannot be really tightened.
  • the binarization method mentioned above is to binarize the entire image, and the calculated threshold does not apply to the binarization of a single character.
  • the 2005 version of RMB 100 the first four characters are red, the last six characters are black, which will result in uneven brightness of each character of the acquired grayscale image.
  • Each small image can be binarized separately.
  • the binarization uses a histogram bimodal adaptive binarization method.
  • the histogram bimodal method is an iterative method for thresholding.
  • an initialization threshold T 0 is set , and then the threshold of binarization is obtained after K iterations.
  • K is a positive integer greater than 0, where the background gray average of the kth iteration And foreground grayscale mean They are:
  • the condition for exiting the iteration when the number of iterations is sufficient (for example, 50 times), or the threshold results of the two iterations are the same, that is, the thresholds of the kth and k-1th times are the same, the iteration is exited.
  • the normalization may be as follows: the normalization here is for the next neural network identification. Considering the requirements of calculation speed and accuracy, the image size when the size is normalized cannot be too large or too small. Too big, causing too many subsequent neural network nodes, the calculation speed is too slow, too small, and the information loss is too much. Tested several normalized sizes Small, 28*28, 18*18, 14*14, 12*12, and finally selected 14*14.
  • the normalized scaling algorithm uses a bilinear interpolation algorithm.
  • the normalizing process in the step d specifically includes: performing size normalization by using a bilinear interpolation algorithm; and the shading normalization comprises: acquiring a histogram of the image of each number Figure, calculating the foreground grayscale average value and the background grayscale average value, and comparing the pixel grayscale values before the normalization of the brightness and darkness with the foreground grayscale average value and the background grayscale average value, according to the comparison result, The pixel gray value before normalization is set to the corresponding specific gray value.
  • normalization of the degree of shading must also be performed.
  • Step e using a neural network to identify the normalized number image to obtain a crown number.
  • the neural network described above can be implemented using a Convolutional Neural Network (CNN) algorithm.
  • CNN Convolutional Neural Network
  • CNN Convolutional Neural Network
  • a small portion of the image (locally perceived region) is used as the input to the lowest layer of the hierarchical structure, and the information is sequentially transmitted to different layers, each layer passing through a digital filter to obtain the most significant features of the observed data.
  • This method is capable of obtaining salient features of the invariant observation data for translation, scaling, and rotation, because the local perceptual region of the image allows the neuron or processing unit to access the most basic features.
  • the main feature on the image of the crown number is the edge and Corner points are therefore very suitable for identification using CNN methods.
  • the neural network employs a secondary classification convolutional neural network; the first fraction The class classifies all the numbers and letters involved in the prefix number, and the second level classifies the partial categories in the first level classification separately.
  • the number of categories of the first-level classification may be set according to classification needs and setting habits, and may be, for example, 10 categories, 23 categories, 38 categories, etc., and the second-level classification is the same.
  • the second-level classification is performed again, so that the crown number is further distinguished and identified by the higher recognition rate, and the second level
  • the number of specific input categories and the number of output categories of the classification can be set in detail according to the category setting of the first level classification, the classification needs and setting habits, and the like.
  • CNN convolutional neural network
  • the RMB has no letter V, and the letter O and the number 0 are printed exactly the same. Therefore, we use the secondary classification method to identify the number of the crown number. .
  • the first level classification classifies all numbers and letters into 23 categories:
  • the first category A 4
  • T J J is the 2005 version and all versions of the RMB
  • the second level classification is the classification of A 4, B 8, C 6G, O D Q, E L F, S 5, T J, Z 2 respectively.
  • the above two-level CNN classification method involves nine neural network models, which are respectively recorded as: CNN_23, CNN_A4, CNN_B8, CNN_CG6, CNN_ODQ, CNN_ELF, CNN_S5, CNN_JT, CNN_Z2.
  • FIG. 6 is a schematic structural diagram thereof.
  • the input layer of the network there is only one picture, which is equivalent to the visual input of the network, which is the gray image of the single number to be identified.
  • the grayscale image is chosen here for the information not to be lost, because if the binarized image is identified, some image edge and detail information will be lost in the process of binarization.
  • the brightness of each grayscale image is normalized, that is, the brightness and darkness normalization.
  • the C1 layer is a convolutional layer.
  • the convolutional layer has the advantage of convolution operation, which can enhance the original signal characteristics and reduce noise. It consists of six feature maps. Each neuron in the feature map is connected to a 3*3 neighborhood in the input. The size of the feature map is 14*14.
  • Both the S2 and S4 layers are downsampling layers. Sub-sampling the image using the principle of image local correlation can reduce the amount of data processing while retaining useful information.
  • the C3 layer is also a convolutional layer. It also deconvers the layer S2 through the 3x3 convolution kernel. The resulting feature map has only 4x4 neurons. For the sake of simplicity, only six different convolution kernels are designed. There are 6 feature maps. One thing to note here is that each feature map in C3 is connected to S2 and not fully connected. Why not connect each feature map in S2 to the feature map of each C3? There are two reasons. First, An incomplete connection mechanism keeps the number of connections within a reasonable range. Second, and most importantly, it undermines the symmetry of the network. Since different feature maps have different inputs, they are forced to extract different features. The composition of this non-full join result is not unique.
  • the first two feature maps of C3 are input with three adjacent feature map subsets in S2, and the next two feature maps are input with four adjacent feature map subsets in S2, and then one is not
  • the adjacent three feature map subsets are inputs, and the last one takes all feature maps in S2 as inputs.
  • the last set of S to C layers is not downsampled, but a simple stretch of the S layer, which becomes a one-dimensional vector.
  • the number of outputs of the network is the number of classifications of the neural network, and the last layer constitutes a fully connected structure.
  • the training of the neural network can be carried out in the following ways:
  • the calculation formula of the jth feature map of the first layer is as follows:
  • the * sign indicates convolution
  • the convolution kernel k performs a convolution operation on all associated feature maps of the l-1th layer, and then sums, plus an offset parameter b, taking the sigmoid function Get the ultimate incentive.
  • the residual of the jth feature map of the first layer is calculated as follows:
  • the first layer is a convolution layer
  • the l+1th layer is a downsampling layer
  • the downsampling layer and the convolution layer are in one-to-one correspondence.
  • up(x) is to expand the size of the l+1th layer to be the same as the size of the first layer.
  • Randomly select the renminbi crown number as a training sample about 100,000, the number of training is more than 1000 times, approaching The accuracy is less than 0.004.
  • step b and step c further comprising a face-facing determining step of determining a banknote size by the rotated image, determining a face value according to the size, and dividing the target banknote image into n Blocks, calculating the average value of the brightness in each block, compared with the pre-stored template, when the difference is the smallest, it is judged as the corresponding face of the template; the pre-stored template is to divide the image of different facets of different denominations It is n blocks, and the average value of the brightness in each block is calculated as a template.
  • the face value of the banknote can be determined by the banknote size detection + template matching method. First determine the face value of the banknote by the size of the banknote. Then, after determining the face of the banknote, 16*8 identical rectangular blocks are divided inside the banknote image, and the brightness average value in each rectangular block is calculated, and the 16*8 brightness average data is placed in the memory as template data. . In the same way, the average value of the brightness of the target banknote is obtained, and compared with the template data, the difference is found to be the smallest. The face of the banknote can be confirmed.
  • step b and the step c further comprising a damage identification step: obtaining a transmitted image by separately arranging a light source and a sensor on both sides of the banknote; and respectively, after the rotated transmitted image It is detected that when two adjacent pixels of the point are simultaneously less than a predetermined threshold, the point is determined to be a damage point.
  • the light source and the sensor are distributed on both sides of the banknote, that is, the transmission mode.
  • the light source encounters the banknote, only a small part of the light can penetrate the banknote and hit the sensor component, and the light that does not encounter the banknote is completely hit on the sensor component. Therefore the background is white and the banknotes are also grayscale. Damage includes notches and holes. The detection of notch and hole is applied by the damage identification technology. The difference is that the detection area is different. The corner detection detects the four corners of the banknote, and the hole is the middle area for detecting the banknote.
  • the banknotes may be divided into upper left, lower left, upper right, and lower right, respectively, four regions on the rotated transmissive banknote image. Then, the four regions are respectively detected point by point, and the adjacent two pixels are simultaneously smaller than the threshold, then the point is judged to be a damage point. If the two adjacent points do not satisfy the condition less than the threshold, it indicates that the angle corresponding to the intersection does not have Damage feature.
  • the position of the missing corner has been black Filled, if the banknote has a corner and hole feature, then the pixel is white.
  • the pixel value of the point determined to be the corner is changed to the black pixel value, thus achieving the filling. . Therefore, the entire banknote is searched for on the four sides of the banknote. If the banknote is found to have a broken feature, it indicates that the banknote has a hole, otherwise the banknote has no holes.
  • the hole area is incremented by one each time a pixel point smaller than the threshold is searched. The area of the hole will eventually be obtained after the search is completed.
  • the following manner may be adopted: in the fixed area, the pixels in the scanning area are placed in an array, and the histogram of each pixel point is recorded, and the histogram is calculated according to the histogram. The brightest pixel points, the average gray value is obtained, and the threshold is calculated. A pixel point smaller than the threshold is determined to be a handwriting +1.
  • the embodiment provides a banknote management system, and the banknote management system includes a banknote information processing terminal and a main control server end;
  • the banknote information processing terminal comprises a banknote sending module, a detecting module and an information processing module;
  • the banknote sending module is configured to transport the banknotes to the detecting module
  • the detecting module collects and recognizes the characteristics of the banknote
  • the information processing module processes and processes the characteristics of the banknotes collected and recognized by the detecting module, and outputs the characteristics of the banknotes and transmits them.
  • the banknote characteristic information specifically includes a currency, Denomination, orientation, authenticity, old and new, defacement, crown number;
  • the main control server is configured to receive the banknote characteristic information, the service information, and the information of the banknote information processing terminal, process the received three types of information, and classify the banknote; in this embodiment, As a preferred implementation manner, the main control server end sorts the banknotes by specifically classifying the banknotes into the different coin bins according to the classified categories.
  • the service information includes record information of collection, payment, deposit or withdrawal, business time period information, operator information, transaction card number information, identity information of the agent and the agent, and Dimension code information, packet number;
  • the main control server processes the received information, and specifically includes collecting, storing, sorting, querying, tracking, and exporting the received information.
  • the banknote information processing terminal described in this embodiment can be used alone.
  • the banknote information processing terminal is a banknote sorting machine; as an alternative technical solution of the embodiment, the banknote information processing terminal can also be replaced with one of a money counter, a money detector, and a self-service financial device;
  • the self-service financial device may be any one of an automatic teller machine, an automatic deposit machine, a circulation automatic teller machine, a self-service inquiry machine, and a self-service payment machine.
  • the design of the detection module is not unique. In this embodiment, a specific implementation manner is provided.
  • the detection module can also be applied to the identification system of the crown number of the DSP platform, and can be embedded or connected.
  • the utility model is used in combination with a conventional money detector, a money counter, an ATM, and the like.
  • the detection module includes: an image preprocessing module, a processor module, and a CIS image sensor module;
  • the image preprocessing module further includes an edge detection module and a rotation module;
  • the processor module further includes a number positioning module, a tightening module, a normalization module, and an identification module;
  • the number positioning module performs binarization processing on the image by adaptive binarization to obtain a binarized image; then, the binary image is projected; finally, by setting a moving window, using a moving window to register Means, dividing the number, obtaining an image of each number, and transmitting the image of each number to the tightening module;
  • the normalization module is used to normalize the image processed by the tightening module.
  • the normalization is size normalization and light and dark normalization.
  • the number positioning module further includes a window module, and the window module designs a registration moving window according to the crown number spacing, and moves the window horizontally on the vertical projection image, and calculates the The sum of the number of black points in the window; the window module can also compare the sum of the black points in different windows.
  • the specific manner of the positioning can be carried out by the method in Embodiment 1.
  • the tightening module extracts a histogram for an image of each number, obtains a binarization threshold by using a histogram bimodal method, and then, according to the binarization threshold, each number of the number The image is binarized, and the binarized image of each number obtained is subjected to regional growth. Finally, in the region obtained after the region is grown, one or two regions having an area larger than a predetermined area threshold are selected. The rectangle in which these selected regions are located is the rectangle in which each number image is nested.
  • the growth of the region may employ, for example, an eight-neighbor region growth algorithm.
  • the banknote image needs to be compensated, and the compensation module may be disposed in the detection module for the CIS.
  • the image obtained by the image sensor module is compensated, and the compensation module pre-stores pure white and pure black collection light.
  • the degree data is combined with the gray reference value of the settable pixel point to obtain a compensation coefficient; the compensation coefficient is stored in the processor module, and a lookup table is established.
  • the white paper is pressed on the CIS image sensor, the collected bright level data is stored in the CISVL[i] array, and the collected dark level data is stored in CISDK[i], and the formula is passed.
  • CVLMAX is a configurable pixel point gray reference value. According to experience, the gray value of white paper is set to 200.
  • the compensation coefficient calculated by the DSP chip is transmitted to the random memory of the FPGA (processing module) to form a lookup table. After that, the FPGA chip multiplies the collected pixel point data by the compensation coefficient of the corresponding pixel in the lookup table, and directly obtains the compensated data, and then transmits the compensated data to the DSP.
  • the identification module uses the trained neural network to implement the identification of the crown number.
  • the neural network adopts a secondary classification convolutional neural network; the first level classification classifies all the numbers and letters involved in the crown number, and the second level classification is respectively in the first level classification. Some of the categories are reclassified. It should be noted that the number of categories of the first-level classification may be set according to classification needs and setting habits, and may be, for example, 10 categories, 23 categories, 38 categories, etc., and the second-level classification is the same.
  • the second-level classification is performed again, so that the crown number is further distinguished and identified by the higher recognition rate, and the second level
  • the number of specific input categories and the number of output categories of the classification can be set in detail according to the category setting of the first level classification, the classification needs and setting habits, and the like.
  • the structure of the convolutional neural network described above can be implemented by using the neural network structure in Embodiment 1 above.
  • the processor module may further include at least one of the following: a face-to-face determination module for determining the orientation of the banknote; a new and old degree determination module for determining the oldness and the oldness of the banknote; and the damage identification a module for identifying a damaged location in the banknote; a handwriting recognition module for identifying a handwriting on the banknote.
  • a face-to-face determination module for determining the orientation of the banknote
  • a new and old degree determination module for determining the oldness and the oldness of the banknote
  • the damage identification a module for identifying a damaged location in the banknote
  • a handwriting recognition module for identifying a handwriting on the banknote.
  • the processor module can use, for example, an FPGA (Kyoto Yage M7 chipset) Chip type system such as M7A12N5L144C7).
  • the chip's main frequency is (FPGA main frequency 125M, ARM main frequency 333M), the occupied resources are (Logic 85%, EMB 98%), the recognition time is 7ms.
  • the accuracy is 99.6% or more.

Landscapes

  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computer Security & Cryptography (AREA)
  • Inspection Of Paper Currency And Valuable Securities (AREA)
  • Character Discrimination (AREA)
  • Image Analysis (AREA)

Abstract

本发明提供一种纸币管理方法，包括采用纸币信息处理装置对纸币特征进行采集、识别和处理，得到纸币特征信息；将所述的纸币特征信息、业务信息、所述纸币信息处理装置的信息一起传输至主控服务器；所述主控服务器对接收的信息进行加工，并对纸币进行分类处理。本发明还提供了相应的纸币管理系统。本发明的上述方法能够在保证运算速度的同时，提高识别的鲁棒性，保证了实际应用中的准确性和实用性。

Description

一种纸币管理方法及其系统 技术领域
本发明属于金融领域，具体涉及一种纸币管理系统及其方法。
背景技术
随着金融信息化应用水平的不断提升，银行系统的货币反假、业务流程管理和金融安全逐步趋向智能化，纸币管理对维护国家金融领域的安全和稳定实现人民币流通痕迹管理、假币管理、ATM配钞管理、残损币管理和现金出入库管理具有重大的意义。
纸币管理主要是针对纸币信息、业务信息等信息的综合处理，纸币信息中的冠字号码在纸币管理中发挥越来越重要的作用，通过将冠字号码的信息与业务等信息相关联，可大大方便纸币追踪和查询。这就使得纸币管理中对于冠字号码以及其他信息的采集和识别，尤其是对于待识别区域中的冠字号码的识别，有着较高的要求，不仅要求准确率高，识别效率和识别速度也要高。
在现有技术中，随着DSP技术的发展，通过DSP平台，配合计算机视觉技术和图像处理技术，实现对冠字号码的识别，较为常见。而在具体的识别算法中，常用的方法有模板匹配、BP神经网络、支持向量机等，也有采用多重神经网络融合的方式实现识别，例如，在申请号为CN201410258528.9的专利申请中，通过分别设计训练两个神经网络的方式，实现识别，即通过冠字号码的图像矢量特征训练一个特征提取网络，再结合一个BP神经网络来识别，通过对上述两个网络的加权融合，实现对冠字号码的识别。而在DSP识别方式中，往往局限于网络传输效率以及DSP识别中对纸币的位置、朝向等影响，其识别效率及识别算法的鲁棒性都比较差，例如在申请号为CN201510702688.2的专利申请中，通过灰度阈值与方向搜索的方式，拟合出边缘，再通过阈值对边缘线进行筛选，获得区域斜率，结合神经网络训练识别面向后，通过逐行扫描及后续的神经网络识别出冠字号码。
又如在一现有技术中，如论文《基于图像分析的人民币清分方法研究与实现》中，期采用了卷积神经网络的方式对冠字号码进行识别，但是，上述方案中仅通过最简单二值化对字符进行划分，无法实现对字符的有效套紧，而这将直接影响后续需要处理的数 据量，直接影响算法的实用价值；并且上述技术方案中仅采取了对分割字符的简单大小处理，没有有效对预处理及分割后的图像进行套紧以及图像数据的有效归一化处理，而这种简单的大小处理，将对后续的神经网络识别带来繁重的数据处理量，极大降低了后续的识别效率；并且，上述技术方案中同样没有很好地处理纸币残缺对纸币识别及图像的处理造成的影响。虽然上述技术方案理论上能够达到一定的识别准确率，但是，由于其运算识别效率低下，不能很好地转化为商业实用方法，不能适应在现实纸币识别中的速度要求。
可见，现有技术存在以下问题：不能高效率地解决对纸币的朝向及字符的有效定位，其识别后的字符范围较大，容易造成字符的错误划分，并且后期图像处理及识别的数据量大，降低了识别效率；对于走钞造成的纸币图像的快速倾斜变化不能很好地适应，不能及时对纸币的倾斜进行纠正并识别；对破损纸币识别的鲁棒性低，没有提供相应的纸币破损识别处理方式。
发明内容
为此，本发明所要解决的第一个技术问题在于现有技术中的纸币管理系统不能实现高效率的准确采集和识别纸币信息，进而提供一种可高效率、准确采集和识别纸币信息的纸币管理方法及其系统。
本发明所要解决的第二个技术问题在于提出了一种冠字号码的识别方法，在保证了冠字号码识别的效率的情况下，有效解决了待识别对象破损、脏污、快速翻折等情况下识别算法的鲁棒性问题。
本发明所述的纸币管理方法，包括以下步骤：
(1)采用纸币信息处理装置对纸币特征进行采集、识别和处理，得到纸币特征信息；
(2)将步骤1)中所述的纸币特征信息、业务信息以及所述纸币信息处理装置的信息一起传输至主控服务器；
(3)所述主控服务器对接收的所述纸币特征信息、所述业务信息、所述纸币信息处理装置的信息进行整合加工处理，并对纸币进行分类处理。
优选地，所述步骤1)中通过图像、红外、荧光、磁、测厚中的一种或多种方式对 所述纸币特征进行采集。
优选地，所述步骤3)中对纸币进行分类处理具体为：将纸币分类后，使其按分类后类别进入到不同的币仓中。所述仓币即容纳纸币的容器或空间。
优选地，所述纸币信息包括币种、面值、面向、真伪、新旧程度、污损、冠字号码中的一种或多种；其中，所述面向是指纸币的正反朝向。
优选地，所述业务信息包括收款、付款、存款或取款的记录信息，业务时间段信息，操作员信息，交易卡号信息，办理人和/或代办人身份信息，二维码信息，封包号中的一种或多种。
优选地，所述纸币特征的识别具体包括如下步骤：
步骤a、提取纸币特征所在区域的灰度图像，并对灰度图像进行边缘检测；该边缘检测，可以通过常规的canny检测、sobel检测等方式实现，再结合直线拟合，获得边缘直线方程，但需要对边缘检测时的经验阈值进行试验设定，以保证方法的运算速度。
步骤b、对图像进行旋转；即将边缘检测后的纸币的图像进行坐标点纠正和映射，以将图像摆正，从而方便号码图像的分割和识别，该旋转方法，可以采用坐标点变换方法，或者依据检测出的边缘方程进行纠正，获得变换方程，也可以以极坐标旋转等方式实现；
步骤c、对图像中的单个号码进行定位，具体包含：通过自适应二值化，对图像进行二值化处理，获得二值化图像；然后对所述二值化图像进行投影，常规的图像投影仅通过一次垂直投影和一次水平投影来完成，具体的投影方向和次数，可以依据识别的具体环境及精度要求做调整，例如还可以采用带有倾斜角度方向的投影等，或者采用多次多重投影结合；最后通过设置移动窗口，采用移动窗口配准的方式，对号码进行分割，得到每个号码的图像，由于纸币的破损、脏污等常见问题，对于冠字号码图像上有脏污，字符与字符之间存在粘连的纸币效果较差，尤其是对三个或三个以上字符的粘连，几乎分割不开，因此，本发明在图像投影之后，又加入了移动窗口配准的方式，精确确定字符的位置；该移动窗口配准的方式，即通过设置固定窗口的方式，例如类似窗口模板方式等，缩小号码区域，实现更精准的区域定位，而所有通过设置固定窗口滑动匹配的方式，均能够适用于本申请之中；
步骤d、对所述每个号码的图像中包含的字符进行套紧，并对每个号码图像进行归 一化处理；优选地，所述归一化包含尺寸归一化和明暗归一化；字符的套紧操作，是在步骤c的基础上，对分割出大致位置的字符，再次进行详细定位，以进一步减少后续图像识别要处理的数据量，这大大保证了系统的整体运行速度；
步骤e、采用神经网络对归一化后的号码图像进行识别，获得纸币特征；优选的，所述纸币特征为冠字号码。
优选地，所述步骤a中的边缘检测进一步包括：设定一灰度阈值，依据该阈值从上、下两方向进行直线搜索，获取边缘，这一边缘检测，采用直线扫面的方式，获取边缘直线的像素坐标；再通过最小二乘法，获得图像的边缘直线方程，并同时获得纸币图像的水平长度、垂直长度和斜率。
优选地，所述步骤b中的旋转，进一步包括：基于所述水平长度、垂直长度和斜率，获得旋转矩阵，依据所述旋转矩阵，求取旋转后的像素点坐标。所述旋转矩阵，可以通过极坐标转换的方式获得，即极坐标转换矩阵，例如可以通过获取到的边缘的直线方程，得到纸币的倾斜角度，依据该角度以及边缘的长度，计算各像素点的极坐标转换矩阵；也可以通过普通的坐标转换方式计算，例如依据该倾斜角度和边缘长度，将纸币的中心点设定为坐标原点，计算每个坐标点的在新坐标系中的转换矩阵等；当然，也可以采用其他的矩阵变换法方式进行纸币图像的旋转纠正。
优选地，所述步骤c中，所述通过自适应二值化对图像进行二值化处理，具体包括：求取图像的直方图，设置一阈值Th，当直方图中灰度值由0到Th的点数和大于等于一预设值时，以此时的Th作为自适应二值化阈值，对图像进行二值化，获得二值化图像。
优选地，所述对所述二值化图像进行投影，共进行三次不同方向投影。
优选地，所述步骤c中的移动窗口配准具体包括：设计配准用移动窗口，所述窗口在垂直投影图上水平移动，窗口内的黑点数总和最小值所对应的位置，即为冠字号码左右方向分割的最佳位置。
优选地，所述窗口为间隔固定的一脉冲序列，脉冲之间的宽度由冠字号码图像之间的间隔预先设置。
优选地，每个所述脉冲的宽度为2-10个像素。
优选地，所述步骤d中的套紧，具体包括：对所述每个号码的图像单独进行二值化，对获取到的每个号码的二值化图像进行区域增长，最后，再对区域增长后得到的区域里， 选取一个或两个面积大于某一预设面积阈值的区域，所述选取后的区域所在的矩形即为每个号码图像套紧后的矩形。该区域增长可以采用例如八邻域区域增长算法等。
优选地，对所述每个号码的图像单独进行二值化，具体包含：对所述每个号码的图像提取直方图，采用直方图双峰法获取二值化阈值，再依据该二值化阈值将所述每个号码的图像进行二值化。
优选地，所述步骤d中的尺寸归一化采用双线性插值算法进行尺寸归一化。
更为优选地，归一化后的尺寸为下述中的一个：12*12、14*14、18*18、28*28，单位为像素。
优选地，所述步骤d中的所述明暗归一化包括：获取所述每个号码的图像的直方图，计算号码前景灰度平均值和背景灰度平均值，并将明暗归一化之前的像素灰度值分别与前景灰度平均值和背景灰度平均值进行比较，依据该比较结果，将归一化之前的像素灰度值设置为对应的特定灰度值。
优选地，在所述步骤b、步骤c之间，进一步包括面向判断步骤：通过所述旋转后的图像确定纸币尺寸，依据所述尺寸确定面值；将目标纸币图像分割为n个区块，计算各区块中的亮度均值，与预先存储的模板比较，差值最小时，判断为模板对应的面向。该模板可以通过多种方式进行预先设置，只要能够通过纸币图像的对比，例如面额不同，朝向不同而引起的亮度值差别、颜色差别，或其他能够转换为亮度数值的其他特征等等，均能够作为比较模板使用。
优选地，所述预先存储的模板，是将不同面值纸币的不同面向的图像，分割为n个区块，并计算各区块中的亮度均值，作为模板。
优选地，在所述步骤b、步骤c之间，进一步包括新旧程度判断步骤：首先提取预设数量dpi的图像，将该图像全部区域作为直方图的特征区域，扫描区域内的像素点，放在数组里，记录各个像素点的直方图，根据直方图统计出一定比例的最亮像素点，求取所述最亮像素点的平均灰度值，作为新旧程度判断依据。这一预设数量dpi图像，可以是例如25dpi图像等，该一定比例，可以根据具体需要进行调整，可以是例如40％、50％等等。
优选地，在所述步骤b、步骤c之间，进一步包括破损识别步骤：通过在纸币两侧分别设置光源和传感器，获取透射后图像；对旋转后的透射后图像逐点检测，当该点的 相邻两像素点同时小于一预设阈值时，则判断该点为破损点。该破损点的检测，可以更详细地分为缺角破损、孔洞破损等等。
优选地，在所述步骤b、步骤c之间，进一步包括字迹识别步骤：在固定区域内，扫描区域内的像素点，放在数组里，记录各个像素点的直方图，根据直方图统计出预设数量个最亮像素点，求取平均灰度值，依据该平均灰度值得出阈值，灰度值小于阈值的像素点判定为字迹点。该预设数量可以是例如20、30等，此处并不以此为保护范围的限定理解；该依据平均灰度值得出阈值，可以采用多种方法，可以该平均灰度值直接作为阈值，也可以采用以该平均灰度值作为变量的函数，求解出阈值。
优选地，所述步骤e中的神经网络采用二级分类的卷积神经网络；第一级分类将冠字号码涉及的所有数字和字母进行分类，第二级分类分别对第一级分类中的部分类别进行再次分类。此处需要说明的是，该第一级分类的类别数量可以根据分类需要和设置习惯等进行设置，可以是例如10类、23类、38类等，此处不以此为限，而该第二级分类同样，是在第一级分类的基础上，针对部分容易误判、特征近似或准确率不高等的分类中，再次进行二级分类，从而以更高的识别率将冠字号码进一步区分识别，而该第二级分类的具体输入类别数量以及输出类别数量，则可以根据第一级分类的类别设置以及分类需要和设置习惯等，进行详细设定，此处并不以此为限。
优选地，所述卷积神经网络的网络模型结构依次设置如下：
输入层：仅以一个图像作为视觉输入，所述图像为待识别的单个冠字号码的灰度图像；
C1层：是一个卷积层，该层由6个特征图构成；
S2层：为下采样层，利用图像局部相关性原理，对图像进行子抽样；
C3层：是一个卷积层，采用预设卷积核去卷积层S2，C3层中的每个特征图采用不全连接的方式连接到S2中；
S4层：为下采样层，利用图像局部相关性原理，对图像进行子抽样；
C5层：C5层是S4层的简单拉伸，变成一维向量；
网络的输出个数为分类个数，与C5层组成全连接结构。
优选地，所述C1层、C3层均通过3x3的卷积核进行卷积。
优选地，所述纸币信息处理装置为纸币清分机、点钞机、验钞机中的一种或多种； 所述纸币信息处理装置的信息为制造厂商、设备编号、所在金融机构中的一种或多种。
或者，所述纸币信息处理装置为自助金融设备；所述纸币信息处理装置的信息为配钞记录、钞箱号、制造厂商、设备编号、所在金融机构中的一种或多种。
所述纸币管理方法是由若干个所述纸币处理信息装置分别对其相应的业务中的纸币信息进行采集、识别和处理，并将所述纸币信息传输至网点主机或现金中心主机，再由所述网点主机或现金中心主机将所述纸币信息传输至主控服务器。
此外，本发明还提供了一种纸币管理系统，所述纸币管理系统包括纸币信息处理终端和主控服务器端；
所述纸币信息处理终端包括送钞模块、检测模块、信息处理模块；
所述送钞模块用于将纸币输送至所述检测模块；
所述检测模块对纸币特征进行采集和识别；
所述信息处理模块加工处理所述检测模块采集和识别的纸币特征，输出为纸币特征信息，并将其传输；
所述主控服务器端，用于接收所述纸币特征信息、业务信息、所述纸币信息处理终端的信息，对接收的上述三类信息进行加工，并对纸币进行分类处理。
所述主控服务器端对接收的信息进行加工，具体包括汇总、存储、整理、查询、追踪、导出等处理。
所述检测模块还能够适用于DSP平台的冠字号码的识别系统，可以嵌入或联接到市面上常规的验钞机、点钞机、ATM等设备结合使用，具体而言，所述检测模块包括图像预处理模块、处理器模块、CIS图像传感器模块；
所述图像预处理模块进一步包括边缘检测模块、旋转模块；
所述处理器模块进一步包括号码定位模块、套紧模块、归一化模块、识别模块；
所述号码定位模块，通过自适应二值化，对图像进行二值化处理，获得二值化图像；然后对所述二值化图像进行投影；最后通过设置移动窗口，采用移动窗口配准的方式，对号码进行分割，得到每个号码的图像，并将所述每个号码的图像传输给套紧模块；该移动窗口配准的方式，即通过设置固定窗口的方式，例如类似窗口模板方式等，缩小号码区域，实现更精准的区域定位，而所有通过设置固定窗口滑动匹配的方式，均能够适用于本申请之中。
所述归一化模块用于对套紧模块处理后的图像进行归一化；优选地，所述归一化包括尺寸归一化及明暗归一化。
优选地，所述号码定位模块进一步包括窗口模块，所述窗口模块依据冠字号码间距，设计配准用移动窗口，将所述窗口在垂直投影图上水平移动，并计算所述窗口内的黑点数总和；
所述窗口模块还可以将不同窗口内的所述黑点数总和进行比较。
优选地，所述套紧模块对每个号码的图像单独进行二值化，对获取到的每个号码的二值化图像进行区域增长，最后，再对区域增长后得到的区域里，选取一个或两个面积大于某一预设面积阈值的区域，所述选取后的区域所在的矩形即为每个号码图像套紧后的矩形。该区域增长可以采用例如八邻域区域增长算法等。
优选地，对所述每个号码的图像单独进行二值化，具体包含：对所述每个号码的图像提取直方图，采用直方图双峰法获取二值化阈值，再依据该二值化阈值将所述每个号码的图像进行二值化。
优选地，所述检测模块还包括补偿模块，用于对CIS图像传感器模块获得的图像进行补偿，所述补偿模块预先存储纯白及纯黑的采集亮度数据，并结合可设定的像素点的灰度参考值，得到补偿系数；
所述补偿系数存储至处理器模块，并建立查找表。
优选地，所述识别模块利用训练好的神经网络实现冠字号码的识别。
优选地，所述神经网络采用二级分类的卷积神经网络；第一级分类将冠字号码涉及的所有数字和字母进行分类，第二级分类分别对第一级分类中的部分类别进行再次分类。此处需要说明的是，该第一级分类的类别数量可以根据分类需要和设置习惯等进行设置，可以是例如10类、23类、38类等，此处不以此为限，而该第二级分类同样，是在第一级分类的基础上，针对部分容易误判、特征近似或准确率不高等的分类中，再次进行二级分类，从而以更高的识别率将冠字号码进一步区分识别，而该第二级分类的具体输入类别数量以及输出类别数量，则可以根据第一级分类的类别设置以及分类需要和设置习惯等，进行详细设定，此处并不以此为限。
优选地，所述卷积神经网络的网络模型结构依次设置如下：
输入层：仅以一个图像作为视觉输入，所述图像为待识别的单个冠字号码的灰度图 像；
C1层：是一个卷积层，该层由6个特征图构成；
S2层：为下采样层，利用图像局部相关性原理，对图像进行子抽样；
C3层：是一个卷积层，采用预设卷积核去卷积层S2，C3层中的每个特征图采用不全连接的方式连接到S2中；
S4层：为下采样层，利用图像局部相关性原理，对图像进行子抽样；
C5层：C5层是S4层的简单拉伸，变成一维向量；
网络的输出个数为分类个数，与C5层组成全连接结构。
优选地，所述C1层、C3层均通过3x3的卷积核进行卷积。
优选地，所述识别模块还包括神经网络训练模块，用于训练所述神经网络。
优选地，该处理器模块可以采用例如FPGA等芯片系统。
优选地，所述处理器模块还包括：面向判断模块，用于判断纸币的朝向。
优选地，所述处理器模块还包括新旧程度判断模块，用于判断纸币的新旧程度。
优选地，所述处理器模块还包括破损识别模块，用于将纸币中的破损位置识别出来。该破损包括缺角、孔洞等等。
优选地，所述处理器模块还包括字迹识别模块，用于识别纸币上的字迹。
优选地，所述主控服务器端对纸币进行分类处理具体为：将纸币分类后，使其按分类后类别进入到不同的币仓中。
优选地，所述纸币特征信息包括币种、面值、面向、真伪、新旧程度、污损、冠字号码中的一种或多种；
优选地，所述业务信息包括收款、付款、存款或取款的记录信息，业务时间段信息，操作员信息，交易卡号信息，办理人和/或代办人身份信息，二维码信息，封包号中的一种或多种；
优选地，所述纸币信息处理终端为纸币清分机、点钞机、验钞机、自助金融设备中的一种；进一步优选地，所述自助金融设备为自动取款机(ATM)、自动存款机、循环自动柜员机(CRS)、自助查询机、自助缴费机中的一种。
本发明还提供了纸币信息处理终端，所述纸币信息处理终端为上述纸币管理系统中包含的所述纸币信息处理终端。
本发明的上述技术方案的有益效果如下：
1、本发明的纸币管理方法，可实现冠字号码的智能管理，通过本发明的方法，可以对银行清分设备的纸币信息追溯、残假币管理、冠字号码统一管理、业务电子日志、数据统计分析、设备状态监控、客户质疑币管理、配钞管理、远程管理、设备资产管理的精细化管理，实现了设备及业务“事前监控，事中跟踪，事后分析”，不仅大幅降低了银行清分机类设备的管理运行成本，还可促进清分机及点钞机等设备的良好运行；
2、本发明的纸币管理方法，实现了在高效率的采集和识别纸币信息的同时，保证识别信息的准确性，尤其是在冠字号码识别上，在保证了整体方法及系统运行的速度的情况下，提高了方法的鲁棒性，能够很好地应付实际应用中，由于纸币污损、残缺、快速翻折等对冠字号码识别带来的识别困难；
3、本发明提供的方法占用系统资源少，比现有技术中的常规算法运算速度快，能够很好地与ATM、验钞机等设备结合使用。
附图说明
图1为本发明实施例的识别方法示意图；
图2为本发明实施例的边缘检测方法示意图；
图3为本发明实施例的走钞过程中的纸币图像与实际纸币示意图；
图4为本发明实施例的纸币任意点旋转的示意图；
图5为本发明实施例的移动窗口设置示意图；
图6为本发明实施例的神经网络结构示意图。
具体实施方式
为使本发明要解决的技术问题、技术方案和优点更加清楚，下面将结合附图及具体实施例进行详细描述。本领域技术人员应当知晓，下述具体实施例或具体实施方式，是本发明为进一步解释具体的发明内容而列举的一系列优化的设置方式，而这些设置方式之间均是可以相互结合或者相互关联使用的，除非在本发明明确提出了其中某些或某一具体实施例或实施方式无法与其他的实施例或实施方式进行关联设置或共同使用。同时，下述的具体实施例或实施方式仅作为最优化的设置方式，而不作为限定本发明的保护范围的理解。
此外，本领域技术人员应当了解，一下具体实施方式及实施例中所列举出来的对于参数设定的具体数值，是作举例解释用，作为一可选的实施方式，而不应当理解为对本发明保护范围的限定；而其中涉及到的各算法及其参数的设定，也仅作为距离解释用，而对下述参数的形式变换以及对下述算法的常规数学推导，均应视为落入本发明的保护范围之内。
实施例1：
本实施例提供了一种纸币管理方法，具体包括以下步骤：
(1)由六个纸币信息处理装置分别对其相应的业务中的纸币的纸币特征进行采集、识别和处理，得到所述纸币特征信息；其中，作为本实施例优选的实现方式，所述纸币信息处理装置通过图像、红外、荧光、磁、测厚的方式对所述纸币特征进行采集。所述纸币特征信息包括币种、面值、面向、真伪、新旧程度、污损和冠字号码；作为本实施例的具体实现方式，所述纸币信息处理装置为纸币清分机；所述纸币信息处理装置的信息为制造厂商、设备编号、所在金融机构；
需要说明的是，所述纸币信息处理装置的个数并不唯一，包括但不限于六个，至少为一个；
作为本实施例可替换的实现方式，所述纸币信息处理装置还可以为点钞机或验钞机中的一种或多种；所述纸币信息处理装置的信息还可以是省略制造厂商、设备编号、所在金融机构中的一项或多项；
作为本实施例的另一则可替换的实现方式，所述纸币信息处理装置还可以为自助金融设备；具体而言，所述纸币信息处理装置可以是自动取款机、自动存款机、循环自动柜员机、自助查询机、自助缴费机中的任意一种。所述纸币信息处理装置的信息可以为配钞记录、钞箱号、制造厂商、设备编号、所在金融机构中的一种或多种；
(2)将步骤1)中所述的纸币特征信息传输至网点主机，再由所述网点主机传输至主控服务器，并且，将业务信息以及所述纸币信息处理装置的信息传输至主控服务器；其中，作为本实施例的优选实现方式，所述业务信息包括收款、付款、存款或取款的记录信息，业务时间段信息，操作员信息，交易卡号信息，办理人和代办人身份信息，二维码信息，封包号；
需要说明的是，所述纸币特征信息传输至所述主控服务器的方式并不唯一，本领域技术人员可根据实际情况更改所述纸币特征信息、所述业务信息、所述纸币信息处理装置的信息的传输路径，例如，将步骤1)中所述的纸币特征信息、所述纸币信息处理装置的信息、业务信息直接传输至主控服务器；
另外，本领域技术人员还可根据实际需要省略或者替换部分本实施例中的所述业务信息，即省略或者替换收款、付款、存款或取款的记录信息，业务时间段信息，操作员信息，交易卡号信息，办理人和代办人身份信息，二维码信息，封包号中的一项或多项；
(3)所述主控服务器对接收的所述纸币特征信息、所述业务信息、所述纸币信息处理装置的信息进行整合加工处理，并对纸币进行分类处理。作为本实施例的优选实现方式，所述对纸币进行分类处理具体为：将纸币分类后，使其按分类后类别进入到不同的币仓中。
作为本实施例的优选实现方式，下面以冠字号码的识别方法为例，对所述纸币特征的识别方法进行说明，如图1所示，具体包括如下步骤：
步骤a、提取冠字号码所在区域的灰度图像，并对灰度图像进行边缘检测；该边缘检测，可以通过常规的canny检测、sobel检测等方式实现，再结合直线拟合，获得边缘直线方程，但需要对边缘检测时的经验阈值进行试验设定，以保证方法的运算速度。
在一个具体的实施方式中，所述步骤a中的边缘检测进一步包括：设定一灰度阈值，依据该阈值从上、下两方向进行直线搜索，获取边缘，这一边缘检测，采用直线扫面的方式，获取边缘直线的像素坐标；再通过最小二乘法，获得图像的边缘直线方程，并同时获得纸币图像的水平长度、垂直长度和斜率。
在一个具体的实施方式中，如图2所示，为保证边缘检测的准确性和计算的速度，可以采用阈值线性回归分割技术，计算速度快，不受图像大小的限制，在其他的边缘检测理论中，是需要对边缘的每一个像素点都要计算的，这样的话，图像越大，计算时间越长。而采用阈值线性回归分割技术，只需要在上下边缘上找到少量的像素点，通过直线拟合的方式可以很快速的确定边缘的直线方程。无论图像大或小都可以找少量的点来计算。
具体而言，由于纸币图像的边缘亮度与背景黑色差异很大，非常容易找到一个阈值 来区分纸币和背景，因此这里采用直线搜索的方法从上、下两个方向检测纸币边缘。上、下方向我们分别沿直线X＝{xi}，(i＝1,2,…,n)搜索得到纸币上边沿Y1＝{y1i}，下边沿Y2＝{y2i}。
利用最小二乘法求出斜率k1，k2，截距b1，b2。求取上下沿中线的斜率K，截距B。已知中线必然要经过中点(x0,y0)，所以沿直线y＝K·x+B
因此可以得到如下关系式：
Figure PCTCN2016112111-appb-000001
利用最小二乘法求k1，b1：
Figure PCTCN2016112111-appb-000002
Figure PCTCN2016112111-appb-000003
Figure PCTCN2016112111-appb-000004
同理可以计算出k2，b2：
Figure PCTCN2016112111-appb-000005
因此可以得到纸币的上沿、下沿中线y＝K·x+B
Figure PCTCN2016112111-appb-000006
由于纸币的上沿、下沿中线y＝K·x+B必然经过纸币的中点(x0,y0)，所以沿直线y＝K·x+B进行搜索得到左端点(xl,yl)和右端点(xr,yr)，最后可以得到纸币图像的中点为：
Figure PCTCN2016112111-appb-000007
得到纸币中点之后，我们需要来求得纸币的水平方向长度L和垂直方向上的长度W，这样在下节就可以建立纸币的长宽模型。因此有：
Figure PCTCN2016112111-appb-000008
然后我们在直线y＝y0附近取Y＝{yi}，(i＝1,2,…,m)进行直线搜索得到纸币左边沿X1＝{x1i}和右边沿X2＝{x2i}，因此有：
Figure PCTCN2016112111-appb-000009
步骤b、对图像进行旋转；即将边缘检测后的纸币的图像进行坐标点纠正和映射，以将图像摆正，从而方便号码图像的分割和识别，该旋转方法，可以采用坐标点变换方 法，或者依据检测出的边缘方程进行纠正，获得变换方程，也可以以极坐标旋转等方式实现；
在一具体的实施方式中，所述步骤b中的旋转，进一步包括：基于所述水平长度、垂直长度和斜率，获得旋转矩阵，依据所述旋转矩阵，求取旋转后的像素点坐标。所述旋转矩阵，可以通过极坐标转换的方式获得，即极坐标转换矩阵，例如可以通过获取到的边缘的直线方程，得到纸币的倾斜角度，依据该角度以及边缘的长度，计算各像素点的极坐标转换矩阵；也可以通过普通的坐标转换方式计算，例如依据该倾斜角度和边缘长度，将纸币的中心点设定为坐标原点，计算每个坐标点的在新坐标系中的转换矩阵等；当然，也可以采用其他的矩阵变换法方式进行纸币图像的旋转纠正。
在一具体的实施方式中，如图3所示，可以采用直角坐标变换的方式对图像进行旋转纠正，由于在图像采集过程中水平方向上每毫米采集p个点，垂直方向上每毫米采集q个点。在之前的纸币图像边缘检测中我们已经计算出了纸币图像的水平长度AC＝L，垂直长度BE＝W和斜率K。因此对纸币图像的几何计算得到下边的公式：
由于
Figure PCTCN2016112111-appb-000010
因此
Figure PCTCN2016112111-appb-000011
AD＝p·AD'＝L·cos2θ       (1-11)
Figure PCTCN2016112111-appb-000012
Figure PCTCN2016112111-appb-000013
而
Figure PCTCN2016112111-appb-000014
则
Figure PCTCN2016112111-appb-000015
Figure PCTCN2016112111-appb-000016
所以
Figure PCTCN2016112111-appb-000017
同理：
Figure PCTCN2016112111-appb-000018
所以
Figure PCTCN2016112111-appb-000019
由于AB'为实际纸币的长Length，B'F'为实际纸币的宽Wide，因此有：
Figure PCTCN2016112111-appb-000020
纸币图像任意点的旋转，旋转的整个过程是对任意给出的纸币图像上的某一点A(xs,ys)，找到点A对应于实际纸币的点A'(x's,y's)，把点A'旋转θ角后得到点B'(x'd,y'd)，最后找到点B'对应于旋转后的纸币图像上的点B(xd,yd)。
结合图4，纸币上的任意点旋转时，
Figure PCTCN2016112111-appb-000021
Figure PCTCN2016112111-appb-000022
Figure PCTCN2016112111-appb-000023
Figure PCTCN2016112111-appb-000024
Figure PCTCN2016112111-appb-000025
如有旋转前的纸币图像中心为(x0,y0)，旋转后的纸币图像中心为(xc,yc)，这样可得：
Figure PCTCN2016112111-appb-000026
步骤c、对图像中的单个号码进行定位，具体包含：通过自适应二值化，对图像进行二值化处理，获得二值化图像；然后对所述二值化图像进行投影，常规的图像投影仅通过一次垂直投影和一次水平投影来完成，具体的投影方向和次数，可以依据识别的具体环境及精度要求做调整，例如还可以采用带有倾斜角度方向的投影等，或者采用多次多重投影结合；最后通过设置移动窗口，采用移动窗口配准的方式，对号码进行分割，得到每个号码的图像，由于纸币的破损、脏污等常见问题，对于冠字号码图像上有脏污，字符与字符之间存在粘连的纸币效果较差，尤其是对三个或三个以上字符的粘连，几乎分割不开，因此，本发明在图像投影之后，又加入了移动窗口配准的方式，精确确定字符的位置；
在一具体的实施方式中，所述步骤c中，所述通过自适应二值化对图像进行二值化处理，具体包括：求取图像的直方图，设置一阈值Th，当直方图中灰度值由0到Th的点数和大于等于一预设值时，以此时的Th作为自适应二值化阈值，对图像进行二值化，获得二值化图像；所述对所述二值化图像进行投影，共进行三次不同方向投影。优选地，所述设置移动窗口具体包括：所述窗口在垂直投影图上水平移动，窗口内的黑点数总和 最小值所对应的位置，即为冠字号码左右方向分割的最佳位置。
在一具体的实施方式中，对图像的二值化，可以采用整体自适应二值化的方法。首选，求图像的直方图，亮度较黑色的是冠字号码区域，亮度较为白色的是背景区域。在直方图上求灰度值为0到Th的点数和N，当N>＝2200(经验值)时，所对应的阈值Th即为自适应二值化的阈值。该方法的最大优点是计算时间短，可以满足清分机快速点钞的实时性要求，并且具有很好的自适应性。
在一具体的实施方式中，对二值化后的图像进行投影，可以采用三次投影结合的方式，确定每个号码所在的上下左右位置。其中，第一次进行水平方向投影，确定号码所在的行，第二次进行垂直方向投影，确定每个号码所在的左右方向位置，第三次是对每个小图进行水平方向投影，确定每个号码所在的上下方向位置。
在一具体的实施方式中，上述三次投影方法对于大多数纸币的单个号码分割都能取得良好的效果，但是对于冠字号码图像上有脏污，字符与字符之间存在粘连的纸币效果较差，尤其是对三个或三个以上字符的粘连，几乎分割不开。为了克服这一困难，在一个具体的实施方式中，可采用窗口移动配准法。因为清分机采集的冠字号码大小分辨率固定，每个字符大小固定，每个字符之间的间距也固定，窗口的设计可以根据纸币上冠字号码的间距设计，如图5所示。窗口在垂直投影图上水平移动，窗口内的黑点数总和最小值所对应的位置，即为冠字号码左右方向分割的最佳位置。由于该识别算法用在纸币清分机上，准确性和快速性都要满足，原始图像的分辨率为200dpi。窗口的设计每个脉冲宽度为4个像素，脉冲之间的宽度根据号码图像之间的间隔设计，经过测试，该方法完全能够满足纸币清分机实时性和准确性要求。
步骤d、对所述每个号码的图像中包含的字符进行套紧，并对每个号码图像进行归一化处理，所述归一化包含尺寸归一化和明暗归一化；字符的套紧操作，是在步骤c的基础上，对分割出大致位置的字符，再次进行详细定位，以进一步减少后续图像识别要处理的数据量，这大大保证了系统的整体运行速度。
三次投影法仅仅是对单个号码的初步定位，对于很多脏污的单个号码，都不能真正的套紧。上面提到的二值化方法是对整个图像做二值化，所计算得到的阈值并不适用于单个字符的二值化。例如2005版人民币一百元，前四个字符是红色，后六个字符是黑色，这会导致采集到的灰度图像每个字符的明暗程度不均，在一具体的实施方式中，还 可以对每个小图单独进行二值化。
在一个具体的实施方式中，该二值化采用的是基于直方图双峰的自适应二值化方法。直方图双峰法是一种迭代法求阈值的方法。特点：自适应，快速，准确。具体的，可以采用以下的一个优选的实施方式来实现：
首先设定一个初始化阈值T0，然后经过K次迭代后得到二值化分割的阈值。K为大于0的正整数，这里第k次迭代的背景灰度平均值
Figure PCTCN2016112111-appb-000027
和前景灰度平均值
Figure PCTCN2016112111-appb-000028
分别为：
Figure PCTCN2016112111-appb-000029
Figure PCTCN2016112111-appb-000030
则第k次迭代的阈值为：
Figure PCTCN2016112111-appb-000031
退出迭代的条件：当迭代次数足够多(例如50次)，或者两次迭代计算的阈值结果相同，即第k次和第k-1次的阈值相同，则退出迭代。
二值化后，对每个小图要进行八邻域区域增长算法，目的是去除面积过小的噪声点。最后，在对每个小图区域增长后得到的区域里，选取一个或两个面积大于某一个经验值的区域，这些区域所在的矩形即为每个号码图像套紧后的矩形。综上，该套紧方法的步骤为二值化，区域增长，区域选取，它的优点是抗干扰性强，计算速度快。
在二值化之后，需要对图像进一步进行归一化处理，在一个具体的实施方式中，上述归一化可以采用如下方式：这里的归一化是为了下一步的神经网络识别。考虑到计算速度和准确性的要求，尺寸归一化时的图像大小不能太大，也不能太小。太大，造成后续的神经网络节点过多，计算速度慢，太小，信息损失过多。测试了几种归一化尺寸大 小，28*28,18*18,14*14,12*12，最后选择了14*14。归一化的缩放算法采用双线性插值算法。
在一个具体的实施方式中，所述步骤d中归一化处理具体包括：采用双线性插值算法进行尺寸归一化；所述明暗归一化包括：获取所述每个号码的图像的直方图，计算号码前景灰度平均值和背景灰度平均值，并将明暗归一化之前的像素灰度值分别与前景灰度平均值和背景灰度平均值进行比较，依据该比较结果，将归一化之前的像素灰度值设置为对应的特定灰度值。
在又一具体的实施方式中，为了减少训练模板数量，还必须进行明暗程度的归一化。首先在每个小图的直方图上计算号码前景灰度平均值Gb，和背景灰度平均值Gf。设，V0ij为每个像素灰度归一化之前的值，V1ij为每个像素灰度归一化之后的值，计算方法如下。
Figure PCTCN2016112111-appb-000032
步骤e、采用神经网络对归一化后的号码图像进行识别，获得冠字号码。
在一个具体的实施方式中，上述神经网络可以采用卷积神经网络(CNN)算法来实现。
卷积神经网络(CNN)在本质上是一种输入到输出的映射，它能够学习大量的输入与输出之间的映射关系，而不需要任何输入和输出之间的精确的数学表达式，只要用已知的模式对卷积网络加以训练，网络就具有输入输出对之间的映射能力。在CNN中，图像的一小部分(局部感受区域)作为层级结构的最低层的输入，信息再依次传输到不同的层，每层通过一个数字滤波器去获得观测数据的最显著的特征。这个方法能够获取对平移、缩放和旋转不变的观测数据的显著特征，因为图像的局部感受区域允许神经元或者处理单元可以访问到最基础的特征，冠字号码图像上的主要特征是边缘和角点，因此非常适合采用CNN的方法进行识别。
在一个具体的实施方式中，所述神经网络采用二级分类的卷积神经网络；第一级分 类将冠字号码涉及的所有数字和字母进行分类，第二级分类分别对第一级分类中的部分类别进行再次分类。此处需要说明的是，该第一级分类的类别数量可以根据分类需要和设置习惯等进行设置，可以是例如10类、23类、38类等，而该第二级分类同样，是在第一级分类的基础上，针对部分容易误判、特征近似或准确率不高等的分类中，再次进行二级分类，从而以更高的识别率将冠字号码进一步区分识别，而该第二级分类的具体输入类别数量以及输出类别数量，则可以根据第一级分类的类别设置以及分类需要和设置习惯等，进行详细设定。
下面以一个优选的实施方式，例举可适用于本发明技术方案中的一个具体的卷积(CNN)神经网络的结构及训练方式：
一、CNN神经网络的结构
因为需要对数字和字母混合识别，某些数字和字母非常相似，无法区分，人民币没有字母V，字母O和数字0印刷完全一样，所以，我们对冠字号码的识别采用了二级分类的方法。第一级分类把所有数字和字母归为23类：
第一类：A 4
第二类：B 8
第三类：C G 6
第四类：O D Q
第五类：E L F
第六类：H
第七类：K
第八类：M
第九类：N
第十类：P
第十一类：R
第十二类：S 5
第十三类：T J(J为2005版及一切版本的人民币)
第十四类：U
第十五类：W
第十六类：X
第十七类：Y
第十八：Z 2
第十九：1
第二十类：3
第二十一类：7
第二十二类：9
第二十三类：J(J为2015新版人民币)
第二级分类是分别对A 4，B 8，C 6G，O D Q，E L F，S 5，T J，Z 2的分类。
以上的二级CNN分类方法涉及到9个神经网络的模型，分别记为：CNN_23，CNN_A4，CNN_B8，CNN_CG6，CNN_ODQ，CNN_ELF，CNN_S5，CNN_JT，CNN_Z2。
以第一级分类的CNN神经网络为例，图6是它的结构示意图。网络的输入层：只有一个图，相当于网络的视觉输入，即为待识别的单个号码灰度图像。这里选用灰度图像是为了信息不丢失，因为如果对二值化图像进行识别，则在二值化的过程中会损失一些图像的边缘和细节信息。为了不受图像明暗效果的影响，对每个灰度小图的亮度进行了归一化处理，即明暗归一化。
C1层是一个卷积层，卷积层存在的好处是通过卷积运算，可以使原信号特征增强，并且降低噪音，由6个特征图Feature Map构成。特征图中每个神经元与输入中3*3的邻域相连。特征图的大小为14*14。C1有156个可训练参数(每个滤波器5*5＝25个unit参数和一个bias参数，一共6个滤波器，共(3*3+1)*6＝60个参数)，共60*(12*12)＝8640个连接。
S2和S4层均为下采样层，利用图像局部相关性的原理，对图像进行子抽样，可以减少数据处理量同时保留有用信息。
C3层也是一个卷积层，它同样通过3x3的卷积核去卷积层S2，然后得到的特征map就只有4x4个神经元，为了计算简单，仅仅设计了6种不同的卷积核，所以就存在6个特征map了。这里需要注意的一点是：C3中的每个特征map是连接到S2中并不是全连接的。为什么不把S2中的每个特征图连接到每个C3的特征图呢？原因有二。其一， 不完全的连接机制将连接的数量保持在合理的范围内。其二，也是最重要的，其破坏了网络的对称性。由于不同的特征图有不同的输入，所以迫使他们抽取不同的特征。这种非全连接结果的组成方式并不唯一。例如，C3的前2个特征图以S2中3个相邻的特征图子集为输入，接下来2个特征图以S2中4个相邻特征图子集为输入，然后的1个以不相邻的3个特征图子集为输入，最后1个将S2中所有特征图为输入。
最后一组S层到C层不是下采样，而是S层的简单拉伸，变成一维向量。网络的输出个数为该神经网络的分类个数，与最后一层组成全连接结构。这里的CNN_23共有23个类别，所以有23个输出。
二、神经网络的训练可以通过以下方式进行：
假设第l层为卷积层，第l+1层为下采样层，则第l层第j个特征图的计算公式如下：
Figure PCTCN2016112111-appb-000033
其中，*号表示卷积，是卷积核k在第l-1层所有关联的特征图上做卷积运算，然后求和，再加上一个偏置参数b，取sigmoid函数
Figure PCTCN2016112111-appb-000034
得到最终的激励。
第l层的第j个特征图的残差计算公式如下：
Figure PCTCN2016112111-appb-000035
其中第l层为卷积层，第l+1层为下采样层，下采样层与卷积层是一一对应的。其中up(x)是将第l+1层的大小扩展为和第l层大小一样。
误差对b的偏导数公式为：
Figure PCTCN2016112111-appb-000036
误差对k的偏导数公式为：
Figure PCTCN2016112111-appb-000037
随机选择人民币冠字号码作为训练样本，约10万个，训练次数1000回以上，逼近 的精度小于0.004。
在一具体的实施方式中，在所述步骤b、步骤c之间，进一步包括面向判断步骤：通过所述旋转后的图像确定纸币尺寸，依据所述尺寸确定面值；将目标纸币图像分割为n个区块，计算各区块中的亮度均值，与预先存储的模板比较，差值最小时，判断为模板对应的面向；所述预先存储的模板，是将不同面值纸币的不同面向的图像，分割为n个区块，并计算各区块中的亮度均值，作为模板。
具体而言，可通过纸币尺寸检测+模板匹配方式来确定纸币的面向值。先通过纸币尺寸确定纸币的面值。然后在确定纸币的面向，在纸币图像内部分割了16*8个相同的矩形块，并计算出每个矩形块内的亮度均值，将这16*8个亮度均值数据置于存储器中作为模板数据。同理取得目标纸币的亮度均值，与模板数据做比较，找到差值最小的。可确认纸币的面向。
此外，在一具体的实施方式中，还可以加入纸币新旧程度的判断，首先提取25dpi图像，将25dpi图像全部区域作为直方图的特征区域，扫描区域内的像素点，放在数组里，记录各个像素点的直方图，根据直方图统计出50％最亮像素点，求取平均灰度值，以该灰度值作为新旧程度判断的依据。
在一个具体的实施方式中，在所述步骤b、步骤c之间，进一步包括破损识别步骤：通过在纸币两侧分别设置光源和传感器，获取透射后图像；对旋转后的透射后图像逐点检测，当该点的相邻两像素点同时小于一预设阈值时，则判断该点为破损点。
在具体实施方式中，纸币破损识别时采用的是发光源和传感器分布在纸币的两侧，即透射方式。发光源遇到纸币仅有少部分光线能够穿透纸币打到传感器件上，而没有遇到纸币的光线完全打到了传感器件上。因此背景为白色，纸币也为灰度图。破损包含缺角和孔洞。缺角和孔洞的检测都是应用破损识别技术的，不同的是检测的区域不同，缺角检测的是纸币的四个角，孔洞是检测纸币的中间区域。
在又一具体的实施方式中，对于纸币缺角，可分别在旋转完的透射纸币图像上分成左上、左下、右上、右下，四个区域。然后分别对这四个区域逐点检测，相邻两个像素点同时小于阈值，则判断此点为破损点，如果相邻两点不满足小于阈值的条件，则表明该交点对应的角不具有破损特征。
对于纸币上的孔洞检测，在搜索完了纸币的缺角之后，由于缺角的位置已经被黑色 填充了，如果纸币上有缺角和孔洞特征，那么这个像素点是白色的，在搜索纸币的过程中，把确定是缺角的点的像素值改为黑色的像素值，这样就实现了填充。所以再以纸币的四边为边界搜索整张纸币。如果搜索到纸币具有破损特征，则表明纸币具有孔洞，否则此纸币没有孔洞。当每搜索到一个小于阈值的像素点时，孔洞面积将加1。搜索结束后最终将得到孔洞的面积。
在又一具体的实施方式中，对于字迹的检测，可采用以下方式：在固定区域内，扫描区域内的像素点，放在数组里，记录各个像素点的直方图，根据直方图统计出20个最亮像素点，求取平均灰度值，计算得出阈值。小于阈值的像素点判定为字迹+1。
实施例2：
本实施例提供一种纸币管理系统，所述纸币管理系统包括纸币信息处理终端和主控服务器端；
所述纸币信息处理终端包括送钞模块、检测模块、信息处理模块；
所述送钞模块用于将纸币输送至所述检测模块；
所述检测模块对纸币特征进行采集和识别；
所述信息处理模块加工处理所述检测模块采集和识别的纸币特征，输出为纸币特征信息，并将其传输；本实施例中，作为具体的实现方式，所述纸币特征信息具体包括币种、面值、面向、真伪、新旧程度、污损、冠字号码；
所述主控服务器端，用于接收所述纸币特征信息、业务信息、所述纸币信息处理终端的信息，对接收的上述三类信息进行加工，并对纸币进行分类处理；本实施例中，作为优选的实现方式，所述主控服务器端对纸币进行分类处理具体为：将纸币分类后，使其按分类后类别进入到不同的币仓中。
本实施例中，作为具体的实现方式，所述业务信息包括收款、付款、存款或取款的记录信息，业务时间段信息，操作员信息，交易卡号信息，办理人和代办人身份信息，二维码信息，封包号；
作为本实施例的优选实现方式，所述主控服务器端，对接收的信息进行加工，具体包括对接收的信息汇总、存储、整理、查询、追踪、导出处理。
需要说明的是，本实施例中所述的纸币信息处理终端可以单独使用，本实施例中， 所述纸币信息处理终端为纸币清分机；作为本实施例可替换的技术方案，所述纸币信息处理终端还可替换为点钞机、验钞机、自助金融设备中的一种；其中，所述自助金融设备可以是自动取款机、自动存款机、循环自动柜员机、自助查询机、自助缴费机中的任意一种。
需要说明的是，所述检测模块的设计方式并不唯一，本实施例中提供一种具体的实现方式，所述检测模块还能够适用于DSP平台的冠字号码的识别系统，可以嵌入或联接到市面上常规的验钞机、点钞机、ATM等设备结合使用，具体而言，所述检测模块包括：图像预处理模块、处理器模块、CIS图像传感器模块；
所述图像预处理模块进一步包括边缘检测模块、旋转模块；
所述处理器模块进一步包括号码定位模块、套紧模块、归一化模块、识别模块；
所述号码定位模块，通过自适应二值化，对图像进行二值化处理，获得二值化图像；然后对所述二值化图像进行投影；最后通过设置移动窗口，采用移动窗口配准的方式，对号码进行分割，得到每个号码的图像，并将所述每个号码的图像传输给套紧模块；
所述归一化模块用于对套紧模块处理后的图像进行归一化，本实施例中，所述归一化为尺寸归一化及明暗归一化。
在一个具体的实施方式中，所述号码定位模块进一步包括窗口模块，所述窗口模块依据冠字号码间距，设计配准用移动窗口，将所述窗口在垂直投影图上水平移动，并计算所述窗口内的黑点数总和；所述窗口模块还可以将不同窗口内的所述黑点数总和进行比较。该定位的具体方式，可以采用实施例1中的方法进行。
在又一具体的实施方式中，所述套紧模块对每个号码的图像提取直方图，采用直方图双峰法获取二值化阈值，再依据该二值化阈值将所述每个号码的图像进行二值化，对获取到的每个号码的二值化图像进行区域增长，最后，再对区域增长后得到的区域里，选取一个或两个面积大于某一预设面积阈值的区域，这些选取后的区域所在的矩形即为每个号码图像套紧后的矩形。该区域增长可以采用例如八邻域区域增长算法等。
在一个具体的实施方式中，由于常规的纸币图像获取中，纸币的新旧、残损等状况不一，所以需要对纸币图像进行补偿，则可以在所述检测模块中设置补偿模块，用于对CIS图像传感器模块获得的图像进行补偿，所述补偿模块预先存储纯白及纯黑的采集亮 度数据，并结合可设定的像素点的灰度参考值，得到补偿系数；所述补偿系数存储至处理器模块，并建立查找表。
具体而言，将白纸压在CIS图像传感器上，采集亮电平数据存储在CISVL[i]数组里，在采集暗电平数据存储在CISDK[i]里，通过公式
CVLMAX/(CISVL[i]-CISDK[i])
取得补偿系数。其中CVLMAX为可设定的像素点灰度参考值，按照经验，白纸的灰度值设置为200。
将DSP芯片计算得出的补偿系数，传送到FPGA(处理模块)的随机存储器里，形成一个查找表。之后FPGA芯片对采集到的像素点数据乘以查找表中对应像素点的补偿系数，直接得到补偿后的数据，再传送给DSP。
在一具体的实施方式中，所述识别模块利用训练好的神经网络实现冠字号码的识别。
在一个具体的实施方式中，所述神经网络采用二级分类的卷积神经网络；第一级分类将冠字号码涉及的所有数字和字母进行分类，第二级分类分别对第一级分类中的部分类别进行再次分类。此处需要说明的是，该第一级分类的类别数量可以根据分类需要和设置习惯等进行设置，可以是例如10类、23类、38类等，而该第二级分类同样，是在第一级分类的基础上，针对部分容易误判、特征近似或准确率不高等的分类中，再次进行二级分类，从而以更高的识别率将冠字号码进一步区分识别，而该第二级分类的具体输入类别数量以及输出类别数量，则可以根据第一级分类的类别设置以及分类需要和设置习惯等，进行详细设定。
在一个更为具体的实施方式中，上述的卷积神经网络的结构可以采用上述实施例1中的神经网络结构实现。
在一个更为具体的实施方式中，上述的处理器模块还可以包括以下至少一种模块：面向判断模块，用于判断纸币的朝向；新旧程度判断模块，用于判断纸币的新旧程度；破损识别模块，用于将纸币中的破损位置识别出来；字迹识别模块，用于识别纸币上的字迹。这些模块所采用的功能实现方法，可以采用实施例1中所例举的方法。
在一具体的实施方式中，该处理器模块可以采用例如FPGA(京微雅格M7芯片 具 体型号M7A12N5L144C7)等芯片系统。芯片的主频为(FPGA主频125M，ARM主频333M)，占用的资源是(Logic 85％，EMB 98％)，识别时间为7ms。准确度为99.6％以上。
显然，上述实施例仅仅是为清楚地说明所作的举例，而并非对实施方式的限定。对于所属领域的普通技术人员来说，在上述说明的基础上还可以做出其它不同形式的变化或变动。这里无需也无法对所有的实施方式予以穷举。而由此所引伸出的显而易见的变化或变动仍处于本发明创造的保护范围之中。

Claims (25)

  1. 一种纸币管理方法，其特征在于，包括以下步骤：
    (1)采用纸币信息处理装置对纸币特征进行采集、识别和处理，得到纸币特征信息；
    (2)将步骤1)中所述的纸币特征信息、业务信息、所述纸币信息处理装置的信息一起传输至主控服务器；
    (3)所述主控服务器对接收的所述纸币特征信息、所述业务信息、所述纸币信息处理装置的信息进行整合加工处理，并对纸币进行分类处理。
  2. 根据权利要求1中所述的纸币管理方法，其特征在于，所述纸币特征的识别具体包括如下步骤：
    步骤a、提取纸币特征所在区域的灰度图像，并对灰度图像进行边缘检测；
    步骤b、对图像进行旋转；
    步骤c、对图像中的单个号码进行定位，具体包含：通过自适应二值化，对图像进行二值化处理，获得二值化图像；然后对所述二值化图像进行投影；最后通过设置移动窗口，采用移动窗口配准的方式，对号码进行分割，得到每个号码的图像；
    步骤d、对所述每个号码的图像中包含的字符进行套紧，并对每个号码图像进行归一化处理；优选地，所述归一化包含尺寸归一化和明暗归一化；
    步骤e、采用神经网络对归一化后的号码图像进行识别，获得纸币特征；优选的，所述纸币特征为冠字号码。
  3. 根据权利要求2中所述的纸币管理方法，其特征在于，所述步骤a中的边缘检测进一步包括：设定一灰度阈值，依据该阈值从上、下两方向进行直线搜索，获取边缘；再通过最小二乘法，获得图像的边缘直线方程，并同时获得纸币图像的水平长度、垂直长度和斜率。
  4. 根据权利要求2或3中所述的纸币管理方法，其特征在于，所述步骤b中的旋转，进一步包括：基于所述水平长度、垂直长度和斜率，获得旋转矩阵，依据所述旋转矩阵，求取旋转后的像素点坐标。
  5. 根据权利要求2中所述的纸币管理方法，其特征在于，所述步骤c中，所述通过 自适应二值化对图像进行二值化处理，具体包括：求取图像的直方图，设置一阈值Th，当直方图中灰度值由0到Th的点数和大于等于一预设值时，以此时的Th作为自适应二值化阈值，对图像进行二值化，获得二值化图像。
  6. 根据权利要求2中所述的纸币管理方法，其特征在于，所述步骤c中的移动窗口配准具体包括：设计配准用移动窗口，所述窗口在垂直投影图上水平移动，窗口内的黑点数总和最小值所对应的位置，即为冠字号码左右方向分割的最佳位置。
  7. 根据权利要求2所述的纸币管理方法，其特征在于，所述步骤d中的套紧，具体包括：对所述每个号码的图像单独进行二值化，对获取到的每个号码的二值化图像进行区域增长，再对区域增长后得到的区域里，选取一个或两个面积大于某一预设面积阈值的区域，选取后的区域所在的矩形即为每个号码图像套紧后的矩形。
  8. 根据权利要求7所述的纸币管理方法，其特征在于，对所述每个号码的图像单独进行二值化，具体包含：对所述每个号码的图像提取直方图，采用直方图双峰法获取二值化阈值，再依据该二值化阈值将所述每个号码的图像进行二值化。
  9. 根据权利要求2所述的纸币管理方法，其特征在于，所述步骤d中的明暗归一化包括：获取所述每个号码的图像的直方图，计算号码前景灰度平均值和背景灰度平均值，并将明暗归一化之前的像素灰度值分别与前景灰度平均值和背景灰度平均值进行比较，依据该比较结果，将归一化之前的像素灰度值设置为对应的特定灰度值。
  10. 根据权利要求2所述的纸币管理方法，其特征在于，在所述步骤b、步骤c之间，进一步包括面向判断步骤：通过所述旋转后的图像确定纸币尺寸，依据所述尺寸确定面值；将目标纸币图像分割为n个区块，计算各区块中的亮度均值，与预先存储的模板比较，差值最小时，判断为模板对应的面向；
    和/或，在所述步骤b、步骤c之间，进一步包括新旧程度判断步骤：首先提取预设数量dpi的图像，将该图像全部区域作为直方图的特征区域，扫描区域内的像素点，放在数组里，记录各个像素点的直方图，根据直方图统计出一定比例的最亮像素点，求取所述最亮像素点的平均灰度值，作为新旧程度判断依据；
    和/或，在所述步骤b、步骤c之间，进一步包括破损识别步骤：通过在纸币两侧分别设置光源和传感器，获取透射后图像；对旋转后的透射后图像逐点检测，当该点的相邻两像素点同时小于一预设阈值时，则判断该点为破损点；
    和/或，在所述步骤b、步骤c之间，进一步包括字迹识别步骤：在固定区域内，扫描区域内的像素点，放在数组里，记录各个像素点的直方图，根据直方图统计出预设数量个最亮像素点，求取平均灰度值，依据该平均灰度值得出阈值，灰度值小于阈值的像素点判定为字迹点。
  11. 根据权利要求2所述的纸币管理方法，其特征在于，所述步骤e中的神经网络采用二级分类的卷积神经网络；第一级分类将冠字号码涉及的所有数字和字母进行分类，第二级分类分别对第一级分类中的部分类别进行再次分类。
  12. 根据权利要求1中所述的纸币管理方法，其特征在于，所述步骤1)中通过图像、红外、荧光、磁、测厚中的一种或多种方式对所述纸币特征进行采集。
  13. 根据权利要求1中所述的纸币管理方法，其特征在于，所述步骤3)中对纸币进行分类处理具体为：将纸币分类后，使其按分类后类别进入到不同的币仓中。
  14. 根据权利要求1-13中任意一项所述的纸币管理方法，其特征在于，
    所述纸币特征信息包括币种、面值、面向、真伪、新旧程度、污损、冠字号码中的一种或多种；
    和/或，所述业务信息包括收款、付款、存款或取款的记录信息，业务时间段信息，操作员信息，交易卡号信息，办理人和/或代办人身份信息，二维码信息，封包号中的一种或多种。
  15. 根据权利要求1-14中任意一项所述的纸币管理方法，其特征在于，所述纸币信息处理装置为纸币清分机、点钞机、验钞机中的一种或多种；所述纸币信息处理装置的信息为制造厂商、设备编号、所在金融机构中的一种或多种。
  16. 根据权利要求1-14中任意一项所述的纸币管理方法，其特征在于，所述纸币信息处理装置为自助金融设备；所述纸币信息处理装置的信息为配钞记录、钞箱号、制造厂商、设备编号、所在金融机构中的一种或多种。
  17. 根据权利要求15或16所述纸币管理方法，其特征在于，所述纸币管理方法是由若干个所述纸币处理信息装置分别对其相应的业务中的纸币信息进行采集、识别和处理，并将所述纸币信息传输至网点主机或现金中心主机，再由所述网点主机或现金中心主机将所述纸币信息传输至主控服务器。
  18. 一种纸币管理系统，其特征在于，所述纸币管理系统包括纸币信息处理终端和 主控服务器端；
    所述纸币信息处理终端包括送钞模块、检测模块、信息处理模块；
    所述送钞模块用于将纸币输送至所述检测模块；
    所述检测模块对纸币特征进行采集和识别；
    所述信息处理模块加工处理所述检测模块采集和识别的纸币特征，输出为纸币特征信息，并将其传输；
    所述主控服务器端，用于接收所述纸币特征信息、业务信息、所述纸币信息处理终端的信息，对接收的上述三类信息进行加工，并对纸币进行分类处理。
  19. 根据权利要求18中所述的纸币管理系统，其特征在于，所述检测模块包括图像预处理模块、处理器模块、CIS图像传感器模块；
    所述图像预处理模块进一步包括边缘检测模块、旋转模块；
    所述处理器模块进一步包括号码定位模块、套紧模块、归一化模块、识别模块；
    所述号码定位模块，通过自适应二值化，对图像进行二值化处理，获得二值化图像；然后对所述二值化图像进行投影；最后通过设置移动窗口，采用移动窗口配准的方式，对号码进行分割，得到每个号码的图像，并将所述每个号码的图像传输给套紧模块；
    所述归一化模块用于对套紧模块处理后的图像进行归一化；优选地，所述归一化包括尺寸归一化及明暗归一化。
  20. 根据权利要求19所述的纸币管理系统，其特征在于，所述号码定位模块进一步包括窗口模块，所述窗口模块依据冠字号码间距，设计配准用移动窗口，将所述窗口在垂直投影图上水平移动，并计算所述窗口内的黑点数总和；所述窗口模块还可以将不同窗口内的所述黑点数总和进行比较。
  21. 根据权利要求19所述的纸币管理系统，其特征在于，所述套紧模块对每个号码的图像单独进行二值化，对获取到的每个号码的二值化图像进行区域增长，再对区域增长后得到的区域里，选取一个或两个面积大于某一预设面积阈值的区域，所述选取后的区域所在的矩形即为每个号码图像套紧后的矩形。
  22. 根据权利要求19所述的纸币管理系统，其特征在于，所述检测模块还包括补偿模块，用于对CIS图像传感器模块获得的图像进行补偿，所述补偿模块预先存储纯白及纯黑的采集亮度数据，并结合可设定的像素点的灰度参考值，得到补偿系数；所述补 偿系数存储至处理器模块，并建立查找表。
  23. 根据权利要求18中所述的纸币管理系统，其特征在于，所述主控服务器端对纸币进行分类处理具体为：将纸币分类后，使其按分类后类别进入到不同的币仓中。
  24. 根据权利要求18-23中任意一项所述的纸币管理系统，其特征在于，所述纸币特征信息包括币种、面值、面向、真伪、新旧程度、污损、冠字号码中的一种或多种；
    和/或，所述业务信息包括收款、付款、存款或取款的记录信息，业务时间段信息，操作员信息，交易卡号信息，办理人和/或代办人身份信息，二维码信息，封包号中的一种或多种；
    和/或，所述纸币信息处理终端为纸币清分机、点钞机、验钞机、自助金融设备中的一种；优选地，所述自助金融设备为自动取款机、自动存款机、循环自动柜员机、自助查询机、自助缴费机中的一种。
  25. 一种纸币信息处理终端，其特征在于，所述纸币信息处理终端为权利要求18-24中任意一项所述的纸币管理系统中包含的所述纸币信息处理终端。
PCT/CN2016/112111 2016-05-20 2016-12-26 一种纸币管理方法及其系统 Ceased WO2017197884A1 (zh)

Priority Applications (6)

Application Number Priority Date Filing Date Title
KR1020187037126A KR102207533B1 (ko) 2016-05-20 2016-12-26 지폐 관리 방법 및 시스템
JP2019513099A JP6878575B2 (ja) 2016-05-20 2016-12-26 紙幣管理方法、システム、プログラム及び記録媒体
RU2018145018A RU2708422C1 (ru) 2016-05-20 2016-12-26 Способ и система управления банкоматами
EP16902263.9A EP3460765B1 (en) 2016-05-20 2016-12-26 Banknote management method and system
US16/303,355 US10930105B2 (en) 2016-05-20 2016-12-26 Banknote management method and system
SA518400454A SA518400454B1 (ar) 2016-05-20 2018-11-18 طريقة إدارة أوراق مالية ونظامها

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201610341020.4A CN105957238B (zh) 2016-05-20 2016-05-20 一种纸币管理方法及其系统
CN2016103410204 2016-05-20

Publications (1)

Publication Number Publication Date
WO2017197884A1 true WO2017197884A1 (zh) 2017-11-23

Family

ID=56910314

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2016/112111 Ceased WO2017197884A1 (zh) 2016-05-20 2016-12-26 一种纸币管理方法及其系统

Country Status (8)

Country Link
US (1) US10930105B2 (zh)
EP (1) EP3460765B1 (zh)
JP (1) JP6878575B2 (zh)
KR (1) KR102207533B1 (zh)
CN (1) CN105957238B (zh)
RU (1) RU2708422C1 (zh)
SA (1) SA518400454B1 (zh)
WO (1) WO2017197884A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111967690A (zh) * 2020-09-07 2020-11-20 中国银行股份有限公司 一种外币配送方法及系统
CN114120518A (zh) * 2021-11-26 2022-03-01 深圳怡化电脑股份有限公司 纸币连张检测方法、装置、电子设备及存储介质

Families Citing this family (40)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105957238B (zh) 2016-05-20 2019-02-19 聚龙股份有限公司 一种纸币管理方法及其系统
CN106548558B (zh) * 2016-11-07 2019-07-23 广州广电运通金融电子股份有限公司 一种票据一维信号的检测方法及装置
CN108074321B (zh) * 2016-11-14 2020-06-09 深圳怡化电脑股份有限公司 一种纸币的图像边界提取方法及装置
CN106683257A (zh) * 2016-12-30 2017-05-17 深圳怡化电脑股份有限公司 冠字号定位方法及装置
CN106933948B (zh) * 2017-01-19 2021-03-09 浙江维融电子科技股份有限公司 一种金融数据的统一入库方法
CN106910276B (zh) * 2017-02-24 2019-04-26 深圳怡化电脑股份有限公司 检测纸币新旧的方法及装置
CN106952391B (zh) * 2017-02-27 2019-06-07 深圳怡化电脑股份有限公司 一种污损纸币识别方法及装置
CN107484429B (zh) * 2017-04-18 2020-04-07 深圳怡化电脑股份有限公司 一种金融终端的出钞控制方法、系统及金融终端
CN107085882A (zh) * 2017-06-02 2017-08-22 深圳怡化电脑股份有限公司 一种确定假钞的方法及装置
CN107481394B (zh) * 2017-07-03 2019-10-11 深圳怡化电脑股份有限公司 纸币的识别方法、识别装置及终端设备
CN108022243A (zh) * 2017-11-23 2018-05-11 浙江清华长三角研究院 一种基于深度学习的图像中纸张检测方法
CN108492445A (zh) * 2018-02-06 2018-09-04 深圳怡化电脑股份有限公司 纸币分类的方法及装置
CN108492446B (zh) * 2018-02-07 2020-09-15 深圳怡化电脑股份有限公司 一种纸币边缘查找方法以及系统
KR102095511B1 (ko) * 2018-02-23 2020-04-01 동국대학교 산학협력단 딥 러닝 기반의 지폐 적합 판단 장치 및 방법
CN108717708B (zh) * 2018-03-30 2021-04-13 深圳怡化电脑股份有限公司 规则图像的边沿斜率的计算方法及装置
JP6842177B2 (ja) * 2018-04-06 2021-03-17 旭精工株式会社 硬貨識別方法、硬貨識別システム及び硬貨識別プログラム
CN109448219A (zh) * 2018-10-25 2019-03-08 深圳怡化电脑股份有限公司 图像匹配方法、装置、票据鉴别仪及计算机可读存储介质
CN109685968A (zh) * 2018-12-15 2019-04-26 西安建筑科技大学 一种基于卷积神经网络的纸币图像缺陷的识别模型构建以及识别方法
GB2581803B (en) * 2019-02-26 2021-10-06 Glory Global Solutions International Ltd Cash-handling machine
CN111724335A (zh) * 2019-03-21 2020-09-29 深圳中科飞测科技有限公司 检测方法及检测系统
CN110415425B (zh) * 2019-07-16 2021-09-10 广州广电运通金融电子股份有限公司 基于图像的硬币检测及识别方法、系统及存储介质
KR102331078B1 (ko) * 2019-12-30 2021-11-25 주식회사 포스코아이씨티 딥러닝 기반의 철강제품 이미지 인식 시스템 및 인식 방법
CN111292463A (zh) * 2020-01-17 2020-06-16 深圳怡化电脑股份有限公司 一种纸币识别方法、装置、服务器及存储介质
US11367254B2 (en) * 2020-04-21 2022-06-21 Electronic Arts Inc. Systems and methods for generating a model of a character from one or more images
CN111583502B (zh) * 2020-05-08 2022-06-03 辽宁科技大学 基于深度卷积神经网络的人民币冠字号多标签识别方法
CN111627145B (zh) * 2020-05-19 2022-06-21 武汉卓目科技有限公司 一种图像精细镂空图文的识别方法及装置
CN112651289B (zh) * 2020-10-19 2023-10-13 广东工业大学 一种增值税普通发票智能识别与校验系统及其方法
CN113298812B (zh) * 2021-04-22 2023-11-03 中国电子产品可靠性与环境试验研究所（（工业和信息化部电子第五研究所）（中国赛宝实验室）） 图像分割方法、装置、系统、电子设备及可读存储介质
CN112990150A (zh) * 2021-05-10 2021-06-18 恒银金融科技股份有限公司 一种基于投影双向累和测定冠字号上下边界的方法
CN113129143B (zh) * 2021-05-20 2024-12-06 中国银行股份有限公司 数据处理方法、装置、设备及可读存储介质
CN114140928B (zh) * 2021-11-19 2023-08-22 苏州益多多信息科技有限公司 一种高精准度的数字彩统一化查票方法、系统及介质
CN114677599B (zh) * 2021-12-15 2024-08-02 南京数维测绘有限公司 一种基于无人机摄影测量的黄土梯田损毁部位识别方法
CN114626999A (zh) * 2022-01-26 2022-06-14 上海昊博影像科技有限公司 一种平板探测器坏点快速校正方法及装置
CN115131910B (zh) * 2022-05-30 2024-02-13 华中科技大学同济医学院附属协和医院 一种基于大数据的票据检验系统
CN115100786B (zh) * 2022-06-16 2024-06-04 中国银行股份有限公司 一种现钞管控方法、系统、设备及存储介质
TWI826155B (zh) * 2022-11-30 2023-12-11 元赫數位雲股份有限公司 識別隨機多合一帳務憑證影像以自動獲取多組帳務關聯資訊之帳務管理系統
CN117291209B (zh) * 2023-02-02 2024-05-17 深圳牛图科技有限公司 一种基于多核异构架构的条码识别智能终端
CN117237966B (zh) * 2023-11-13 2024-01-30 恒银金融科技股份有限公司 基于面额数字字符内轮廓的纸币识别方法和装置
CN117746107B (zh) * 2023-12-05 2024-07-09 青岛希尔信息科技有限公司 一种基于数据分析的财务实体报表综合管理系统
CN119091534B (zh) * 2024-08-01 2025-11-28 天津大学 非接触式快速点钞方法、系统、设备、储存介质及程序

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101536047A (zh) * 2006-11-06 2009-09-16 光荣株式会社 纸张识别装置以及纸张识别方法
CN102136167A (zh) * 2010-11-29 2011-07-27 东北大学 一种纸币清分鉴伪装置及方法
WO2012127741A1 (ja) * 2011-03-18 2012-09-27 日立オムロンターミナルソリューションズ株式会社 紙葉類処理装置、紙葉類仕分け装置及び紙葉類仕分けシステム
CN105303676A (zh) * 2015-10-27 2016-02-03 深圳怡化电脑股份有限公司 一种纸币的版本识别方法和系统
CN105957238A (zh) * 2016-05-20 2016-09-21 聚龙股份有限公司 一种纸币管理方法及其系统

Family Cites Families (22)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2894375B2 (ja) * 1991-03-20 1999-05-24 富士電機株式会社 パターン判別方法
JP2002015317A (ja) * 2000-06-29 2002-01-18 Toyo Commun Equip Co Ltd 紙片の画像データ変換方法
CN1213592C (zh) * 2001-07-31 2005-08-03 佳能株式会社 采用自适应二值化的图象处理方法和设备
US6970236B1 (en) 2002-08-19 2005-11-29 Jds Uniphase Corporation Methods and systems for verification of interference devices
DE102004013903A1 (de) 2004-03-22 2005-10-20 Giesecke & Devrient Gmbh System zur Bearbeitung von Wertdokumenten
JP2006280499A (ja) 2005-03-31 2006-10-19 Omron Corp 真券判定システムおよびその動作方法、価値媒体処理装置およびその動作方法、動線管理サーバおよび動線管理方法、監視管理サーバおよび監視管理方法、ホール管理サーバおよびホール管理方法、データセンタサーバおよびその動作方法、並びにプログラム
US7724957B2 (en) * 2006-07-31 2010-05-25 Microsoft Corporation Two tiered text recognition
JP5184824B2 (ja) * 2007-06-15 2013-04-17 キヤノン株式会社 演算処理装置及び方法
CN101359373B (zh) * 2007-08-03 2011-01-12 富士通株式会社 退化字符的识别方法和装置
JP5229874B2 (ja) 2008-02-13 2013-07-03 株式会社ユニバーサルエンターテインメント 紙幣管理システム
US20100125515A1 (en) * 2008-11-14 2010-05-20 Glory Ltd., A Corporation Of Japan Fund management system
CN102142168A (zh) * 2011-01-14 2011-08-03 哈尔滨工业大学 纸币清分机高速高分辨率号码采集装置及其识别方法
CN102509091B (zh) * 2011-11-29 2013-12-25 北京航空航天大学 一种飞机尾号识别方法
JP5900195B2 (ja) 2012-07-03 2016-04-06 沖電気工業株式会社 自動取引装置
CN102800148B (zh) * 2012-07-10 2014-03-26 中山大学 一种人民币序列号识别方法
JP5954038B2 (ja) 2012-08-09 2016-07-20 沖電気工業株式会社 紙幣処理装置、及び紙幣処理方法
JP5914687B2 (ja) 2012-10-24 2016-05-11 日立オムロンターミナルソリューションズ株式会社 紙葉類処理装置、紙葉類仕分け装置及び紙葉類仕分けシステム
JP6342739B2 (ja) * 2014-07-28 2018-06-13 日立オムロンターミナルソリューションズ株式会社 紙葉類識別装置、紙葉類処理装置、および紙葉類識別方法
CN104866867B (zh) * 2015-05-15 2017-12-05 浙江大学 一种基于清分机的多国纸币序列号字符识别方法
CN105354568A (zh) * 2015-08-24 2016-02-24 西安电子科技大学 基于卷积神经网络的车标识别方法
CN105335710A (zh) * 2015-10-22 2016-02-17 合肥工业大学 一种基于多级分类器的精细车辆型号识别方法
CN105261110B (zh) * 2015-10-26 2018-04-06 江苏国光信息产业股份有限公司 一种高效dsp纸币冠字号识别方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101536047A (zh) * 2006-11-06 2009-09-16 光荣株式会社 纸张识别装置以及纸张识别方法
CN102136167A (zh) * 2010-11-29 2011-07-27 东北大学 一种纸币清分鉴伪装置及方法
WO2012127741A1 (ja) * 2011-03-18 2012-09-27 日立オムロンターミナルソリューションズ株式会社 紙葉類処理装置、紙葉類仕分け装置及び紙葉類仕分けシステム
CN105303676A (zh) * 2015-10-27 2016-02-03 深圳怡化电脑股份有限公司 一种纸币的版本识别方法和系统
CN105957238A (zh) * 2016-05-20 2016-09-21 聚龙股份有限公司 一种纸币管理方法及其系统

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP3460765A4 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111967690A (zh) * 2020-09-07 2020-11-20 中国银行股份有限公司 一种外币配送方法及系统
CN111967690B (zh) * 2020-09-07 2023-09-08 中国银行股份有限公司 一种外币配送方法及系统
CN114120518A (zh) * 2021-11-26 2022-03-01 深圳怡化电脑股份有限公司 纸币连张检测方法、装置、电子设备及存储介质
CN114120518B (zh) * 2021-11-26 2024-02-02 深圳怡化电脑股份有限公司 纸币连张检测方法、装置、电子设备及存储介质

Also Published As

Publication number Publication date
EP3460765A1 (en) 2019-03-27
JP6878575B2 (ja) 2021-05-26
EP3460765A4 (en) 2020-01-15
CN105957238A (zh) 2016-09-21
EP3460765B1 (en) 2023-02-01
US20200320817A1 (en) 2020-10-08
KR102207533B1 (ko) 2021-01-26
SA518400454B1 (ar) 2021-09-27
JP2019523954A (ja) 2019-08-29
CN105957238B (zh) 2019-02-19
RU2708422C1 (ru) 2019-12-06
US10930105B2 (en) 2021-02-23
KR20190004807A (ko) 2019-01-14

Similar Documents

Publication Publication Date Title
WO2017197884A1 (zh) 一种纸币管理方法及其系统
CN106056751B (zh) 冠字号码的识别方法及系统
CN110598699B (zh) 一种基于多光谱图像的防伪票据鉴伪系统和方法
CN102034108B (zh) 基于多分辨网格特征配准的清分机纸币面值面向分类方法
US9396404B2 (en) Robust industrial optical character recognition
CN103279735B (zh) 一种金融票据识别模块中积灰检测方法及系统
WO2017190574A1 (zh) 一种基于聚合通道特征的快速行人检测方法
CN105894656A (zh) 一种纸币图像识别方法
CN111583502B (zh) 基于深度卷积神经网络的人民币冠字号多标签识别方法
CN106203237A (zh) 集装箱拖车编号的识别方法和装置
CN107103683B (zh) 纸币识别方法和装置、电子设备和存储介质
CN104537364A (zh) 一种基于纹理分析的美元纸币面额及版本识别方法
Forczmański et al. Stamps detection and classification using simple features ensemble
CN112308141A (zh) 一种扫描票据分类方法、系统及可读存储介质
CN109543554B (zh) 票据检测方法、装置、终端及计算机可读存储介质
Devi et al. Fake currency detection using Inception V3
Hmood et al. Statistical edge-based feature selection for counterfeit coin detection
CN116051864A (zh) 一种基于包装图像的智能识别系统
CN108320372A (zh) 一种折叠纸币识别方法
Vishnu et al. Currency detection using similarity indices method
Das et al. Deep Learning Approach for fake currency recognition
Sooruth et al. Automatic South African Coin Recognition through Visual Template Matching
Yan et al. Research on the counting algorithm of bundled steel bars based on the features matching of connected regions
Kanojia et al. Recognition and verification of Indian currency notes using digital image processing
Heikkonen et al. A computer vision approach to digit recognition on pulp bales

Legal Events

Date Code Title Description
ENP Entry into the national phase

Ref document number: 2019513099

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

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

Ref document number: 16902263

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 20187037126

Country of ref document: KR

Kind code of ref document: A

ENP Entry into the national phase

Ref document number: 2016902263

Country of ref document: EP

Effective date: 20181220