KR101149843B1 - Method for recognizing serial number of security paper money - Google Patents

Abstract

The present invention relates to a method for recognizing a serial number of a security that recognizes the serial number of the security, and (a) acquiring an image by scanning an entire image of the security that passes through the money counter or the bill counting machine using an image sensor. Wow; (b) generating an image of a serial number region by preprocessing the entire image of the security obtained in step (a); (c) extracting the respective characters in the image of the serial number region generated in step (b), correcting the extracted respective characters, and reconstructing the characters by obtaining the least adjacent rectangles, and then distributing the pixels. Generating a pattern vector for each letter consisting of the number of cells in a row M × column N according to the < RTI ID = 0.0 > and (d) recognizing a serial number by using a decision function of the pattern vector for each letter generated in step (c).

According to the present invention, it is possible to clarify the result value and the result process by constructing the feature vector by the pixel distribution through the decision function. In addition, by recognizing letters according to the structural characteristics of each letter, they can be applied only by modification according to the structural features of the letter when issuing new securities. After extracting the image of the serial number area from the image, there is an advantage that the speed can be improved by tilting the small-capacity image.

Serial Number, Decision Tree, H Beam

Description

Method for recognizing serial number of security paper money}

The present invention relates to a method for recognizing serial numbers of securities capable of recognizing serial numbers in securities. More specifically, the result values and process are clearly defined by constructing feature vectors by pixel distribution through a decision function. As it can be identified, it has excellent variable selectivity, small memory occupying capacity, low production cost, fast recognition speed, and character recognition according to the structural characteristics of each character. It can be applied only by modification according to structural features, which can shorten the development period. In the pre-processing process, after extracting the image of the serial number area from the image of the whole securities, the oil price can be improved by tilting the small capacity image. It relates to a method for recognizing serial numbers of securities.

In general, a bill counter or bill counting machine is used for counting the number of securities in banknotes or by classifying two kinds of bonds. It is becoming.

Such a bill counter or bill counter equipped with a conventional serial number recognition function has a problem in that it processes the recognized serial number because it outputs the scanned image as it is. The method had a problem in that the recognition rate was lowered. In particular, in the case of damaged or old banknotes, there was a problem in that the false recognition rate was increased due to the deformation of the image state.

In order to solve this problem, Korean Patent No. 0893613 has been proposed.

1 is a flowchart illustrating a conventional method of recognizing securities, extracting an image of a serial number portion of a securities (S180), and performing the predetermined preprocessing process (S190). Remove the upper and lower blank image from the image (S191), remove the space between each character of the serial number (S192), separate each character of the serial number (S193), and for the separated characters Image processing is performed on binary images of 0 and 1 (S194).

The images which have undergone the preprocessing process (S190) are substituted into the input layer 101 of the neural network function 100 prepared in advance. The probability values for each character are output as the output data.

Here, the process of preparing the neural network function will be briefly described. First, the type of the training example is set, the training set is formed, and the input feature expression of the learning function is determined. The standard data is generated for the prepared images by combining the front and back sides of each bill, and the forward and reverse directions of each bill, and the neural network function for the recognition of the bill is completed by inserting the standard data into the neural network function. .

The backpropagation algorithm is mainly used to learn the neural network, in which the error propagates back from the output node to the internal node.

In other words, by comparing the output of the neural network and the output required in the sample, the error in the output is calculated, the error, the actual output, and the magnification are calculated to apply the weight through the hidden layer 102 to reduce the error. .

However, the conventional serial number recognition method as described above has the following problems.

First, the neural network has a long time required for data conversion / manipulation, which requires a fast CPU, which acts as a factor in the increase in unit price, and above all, a huge storage space is required to accommodate the data. .

Second, since the neural network only provides classification or prediction results and does not explain why such results are extruded, the process of the result is not clear, and there is a problem that cannot be solved when an error occurs.

Third, the process of inputting the standard data into the neural network function is complicated, and it takes a long time to build the model. There was a problem that it takes a longer time required to build the model because it is necessary to repeat the number of times.

Fourth, because the re-learning is required from the beginning when issuing new securities, the development period was significantly delayed.

Fifth, it was difficult for non-experts to use because designing model building and setting various parameter values required expertise.

The serial number recognition method of the securities of the present invention for solving the above problems includes the steps of: (a) acquiring an image by scanning an entire image of the securities passing through a banknote or banknote scanner using an image sensor; (b) generating an image of a serial number region by preprocessing the entire image of the security obtained in step (a); (c) extracting the respective characters in the image of the serial number region generated in step (b), correcting the extracted respective characters, and reconstructing the characters by obtaining the least adjacent rectangles, and then distributing the pixels. Generating a pattern vector for each letter consisting of the number of cells in a row M × column N according to the < RTI ID = 0.0 > (d) As shown in Figs. 5 and 6, the pattern vector for each letter generated in step (c) is a binary tree form, as indicated by the classification model already made to predict the value of an arbitrary pattern. Recognizing a serial number by using a decision function that predicts a classification value to which the pattern belongs by repeatedly performing the task of asking the attribute value of the pattern.

In the method of recognizing the serial number of the securities of the present invention, step (b) includes extracting an outline of the securities (b-1), calculating a slope of the securities from which the outline is extracted (b-2), and (b-3) extracting the reference point of the security through the slope of the security calculated in (b-2) and extracting an image of the serial number region from the security that has undergone the step (b-3) ( b-4).

In the step (c) of the method of recognizing the serial number of the securities of the present invention, the generation of the pattern vector for each character is performed by scanning the M and N times in the row (M) and column (N) directions of the minimum adjacent rectangle for each character. It is characterized by generating a pattern vector according to the matrix form.

In step (d) of the method of recognizing the serial number of the securities of the present invention, the decision function is a letter-by-letter pattern composed of horizontal (row (M)) and vertical (column (N)) cell units generated in step (c). After the reference value is set according to the number of cells of the vector, the method recognizes the serial number using a decision function through the H-beam, and uses the H-beam-shaped mask to classify the generated MxN pattern vector. As shown in Fig. 6, the regions determined as having the pixel of the pattern vector are filled with both vertical regions at the left and right ends, such as 'H', are filled with only the left and right sides, or both the top and bottom of the central vertical region. It is characterized by predicting the value of the generated pattern through recursive partitioning which repeatedly asks questions such as whether it is empty, the top and bottom sides are empty, the middle horizontal line is filled, or is empty. All.

In the securities serial number recognition method of the present invention, the input character image is composed of a specific vector based on pixel distribution, and the result vector and process of recognition can be clearly identified through the decision function of the feature vector. Its excellent and small memory occupies a low manufacturing cost and a fast recognition speed.

In addition, by recognizing the character according to the structural characteristics of each character formally, it is possible to shorten the development period by applying only a modification according to the structural characteristics of the character when issuing new securities.

In addition, it is a useful invention that can improve the speed by extracting the image of the serial number area from the image of the entire security in the preprocessing process by tilting the image of the small capacity.

First, prior to describing the present invention, descriptions and drawings of general contents will be omitted.

As shown in Figure 2 the present invention is configured to extract the serial number (SN) of the security (10) through (a) ~ (d) step.

First, the entire image I1 of the security that passes through the banknote counter or the banknote sorter is obtained by using a conventional image sensor (not shown) formed in the banknote counting machine or the bill counting machine (not shown). (step a)

Thereafter, as shown in FIG. 3, the serial number region included in the securities is extracted from the entire image I1 of the acquired securities through a preprocessing process to extract the image I2 of the serial number region. step)

Here, the pre-processing step is a step (b-1) to extract the outline (boundary line) of the security by using the mask to find the portion of the maximum tilt value in the vertical and horizontal direction as a boundary region and the outline is extracted Calculating a reference point for setting a serial number area using the slope and candidate points of the securities calculated in the step (b-2) and the (b-2); (B-3) extracting a reference point which is a reference for extracting an image, and extracting an image (I2) of a serial number area from a security by using the reference point in the step (b-3) (b-4). It is composed of

At this time, in the present invention, in the state in which the tilt correction for the entire security image I1 is not made, the image I2 of the serial number region is first extracted.

Meanwhile, when the extraction of the image I2 of the serial number region is completed as described above, as shown in FIG. 4, each letter C is extracted from the image I2 of the serial number region, and the extracted character C is obtained. ), And after reconstructing the letter (C) to be recognized by a conventional microcomputer (not shown) by obtaining a minimum adjacent rectangle, the row according to the pixel distribution as shown in FIG. Generate a letter-by-letter pattern vector 10 consisting of the number of cell units of column N (step c).

The generation of the pattern vector 10 divides an image of a binarized character region into an M × N region and generates a number of pixels of the image in each region in order to generate the pattern vector 10 having an M × N size. If the number of pixels is greater than or equal to a predetermined reference value according to the calculated reference value, the pattern vector 10 corresponding to the area is displayed to generate a binary character image to be recognized as the pattern vector 10.

The pattern vector 10 generated as described above forms M × N cells in the row M and column N directions, respectively, as shown in FIG. C) 2 is marked on the marked part and 0 is marked on the blank part, so that it is reconstructed so that it can be displayed by coordinate points such as (0,0) and (0,1).
Then, the pattern vector 10 can recognize the serial number SN by using a decision function (step d).

Then, the pattern vector 10 can recognize the serial number SN by using a decision function (step d).
Here, the decision function is a binary tree form as shown in Figs. 5 and 6, and iteratively asks the attribute value of the pattern as indicated by the classification model already made to predict the value of the arbitrary pattern. This function predicts the classification value to which the pattern belongs.

Further, in the step (d), the reference value is set in the pattern vector 10 generated in the step (c) according to the number of cells in the row (M) × column (N), and then a decision is made using an H-beam mask. Recognize a serial number (SN) using a function.

The pattern vector 10 includes a plurality of rows and columns, and the number of the pattern vectors 10 may be further extended for more accurate recognition rate.

In particular, in step (d), the reference value is changed according to the number of rows and columns so that the algorithm C can be recognized to recognize each letter C of the serial number SN.

For example, as shown in (a) of FIG. 6, the number of rows M of the pattern vector 10 is 5 and the number of columns N is 3, and X and K are determined using a decision function. Assuming,

In step (c), a pattern vector 10 having a 3 × 5 matrix form is created according to the pixel distribution of the character image.

In addition, the decision function in step (d) selects cells in which the marking portion may exist according to the shape characteristics of the letter C in the pattern vector 10 according to the number of rows and columns of the pattern vector 10. To set the reference value.

Of course, if the number of rows and columns in the step (c) is incorrect, the position of the cell in which the marking portion of the letter (C) is located is incorrect, so the reference value should be reset again.

On the other hand, when the setting of the reference value is completed as described above, the program is input to the decision function using the set reference value and the H beam.

In the present embodiment, the coordinate points (0, 2), (0, 1), (4, 2), and (4, 2) are not 0 using the H beam, or the coordinate points (0, 1), ( If 4,1) is 0, it is programmed to recognize as A, H, K, M, N, R, U, V, W, X, Y characters.

Characters selected as described above are configured to be recognized as H, K, N X characters when the coordinate points (0, 1), (4, 1) is 0 again.

Subsequently, if the coordinate point (2,0) is 0 or (2,2) is 0, the character is recognized as K and X characters. If the coordinate point (2,0) is 0, the character is recognized as X character.

Of course, such a reference value may vary according to the set value of the reference value since the reference value becomes different depending on the number of characters, numbers, alphabets, Korean characters, and the like, and the number of rows and columns of the pattern vector 10. It becomes possible to deform.

That is, according to the present invention, a new securities may be obtained by setting a rule according to a predetermined pattern by varying a reference value according to the number of rows and columns of the pattern vector 10 and the type of letters (alphabet, numbers, etc.) to be sorted. When the font, magnification, position, etc. are changed at the time of introduction, the recognition process is simple, and the processing speed of the securities recognition process is improved, and the recognition rate is not generated. The question-and-answer decision function, rather than a method, requires only a space to store basic data, thereby reducing the cost of the product.

Although the above-described embodiments are described in order to describe the most preferred embodiment of the present invention, the present invention is not limited thereto and may be modified in various forms without departing from the technical spirit of the present invention.

1 is a flowchart illustrating a method of recognizing a serial number of a conventional security.

Figure 2 is a flow chart showing the recognition sequence of the securities serial number in accordance with the present invention.

Figure 3 is a state diagram showing the step b-1 ~ b-4 in accordance with the present invention.

Figure 4 is a state diagram showing a state of reconstructing the image of the serial number area according to the present invention.

5 is a state diagram showing a pattern vector according to the present invention;

6A is a state diagram showing the binarization distribution of a pattern vector according to the present invention, and FIG. 6B is a block diagram illustrating an example of programming a decision function.

◈Description of code for main parts of drawing

C: letter SN: serial number I1: full image of securities

I2: Image of serial number area 10: Pattern vector

Claims (4)

(a) acquiring an image by scanning an entire image of the securities passing through a banknote counter or bill counting machine using an image sensor; (b) generating an image of a serial number area by preprocessing the entire image of the security obtained in step (a); (c) extracting the respective characters in the image of the serial number region generated in step (b), correcting the extracted respective characters, and reconstructing the characters by obtaining the least adjacent rectangles, and then distributing the pixels. Generating a pattern vector for each letter consisting of the number of cells in a row M × column N according to the < RTI ID = 0.0 > (d) Iteratively repeats the operation of querying the attribute values of the pattern as directed by the classification model that has already been created to predict the value of the arbitrary pattern by using the pattern vector for each character generated in the step (c) in the form of a binary tree. Recognizing the serial number by using a decision function to predict the classification value to belong to the pattern; Serial number recognition method of a security characterized in that consisting of. The method of claim 1, wherein step (b) comprises: extracting an outline of the security (b-1), calculating a slope of the security from which the outline is extracted (b-2), and (b-2) (B-4) extracting the reference point of the security through the slope of the security calculated in the step (b-3) and extracting an image of the serial number area from the security that went through the step (b-3) (b-4) Serial number recognition method of the securities characterized by consisting of. The method of claim 1, wherein the generating of the pattern vector for each character in the step (c) is performed by scanning the M and N times in the row (M) and column (N) directions of the smallest adjacent rectangle for each character. A method for recognizing serial numbers of securities characterized by generating a pattern vector. The method of claim 1, wherein the decision function in step (d) is performed using a decision function composed of horizontal (row (M)) and vertical (column (N)) cell units generated in step (c). The method of using the H-beam type mask to recognize the number but classify the generated MxN pattern vector is determined that there are pixels of the pattern vector. Through recursive partitioning that repeatedly asks questions such as whether it is filled in on the left, right, left or right, or the top and bottom of the central vertical area, the top and bottom are empty, or the center horizontal line is full or empty. A method for recognizing serial numbers of securities characterized by predicting the value of the generated pattern.

Images