KR100888674B1 - Method for measuring similarity using frenquency range - Google Patents

Abstract

Rotationally correcting the banknote image obtained from the image sensor; Converting the rotationally corrected banknote image into a banknote image suitable for fast Fourier transform for optimal comparison of frequency domains; Extracting frequency information of the transformed image using a fast Fourier transform algorithm; And comparing the extracted frequency information with frequency information of a standard bill to calculate a similarity. The method of measuring similarity of a bill using a frequency domain is disclosed.

Frequency Information, Similarity, Compression, Pseudo Banknote Area

Description

Method for measuring similarity using frenquency range

1 is a database of banknote images.

Figure 2 shows the procedure for identifying the denominations of the dollar using the target country as the United States.

3 shows a bill image obtained from the CIS.

4 illustrates a binarized banknote image and a center point.

5 shows an edge extracted banknote image.

6 shows a method of extracting a rotation angle.

7 shows a banknote image rotated.

8 illustrates a binarized banknote image.

9 shows the extracted pseudo banknote area and its center point.

10 shows a banknote image with a custom transformation on the X-axis.

11 shows a method of setting an actual processing area by X, Y projection.

12 shows a compressed banknote image.

13 shows a frequency domain of a banknote image.

The present invention relates to a method for measuring bills of similarity using a frequency domain, and more particularly, to a method for measuring bills of similarity using an image processing algorithm by frequency domain comparison that exhibits strong characteristics against damaged and wrapped damaged banknotes during distribution.

In the banknote processing system, various sensors have been used for the classification of winding papers and identification of counterfeit. Among them, it is very active to derive the identification result based on the printed contents of the paper.

As a device for this purpose, inexpensive devices have used a method of processing one-dimensional image data obtained by continuously reading a point position of bills transferred using a one-point optical sensor. On the other hand, relatively expensive devices use CIS (Contact Image Sensor) -type one-dimensional sensors to continuously obtain images of bills transported to obtain images of two-dimensional planes and to process brightness information for each element of the image. Dimensional image analysis is performed.

In general, an image processing process for identifying the type and counterfeit of a banknote through a banknote image is as follows.

FIG. 1 is a database of banknote images. As shown in FIG. 1, a standard image serving as a reference for identifying a banknote type is created and stored together with an index key. For example, "1f" represents "one dollar bill front image", "kr10r" represents "South Korea 10,000 won banknote back image", and "jp5f" represents "Japanese five thousand won banknote front image".

In the case of using such a bill image database, since there are many types of standard images registered for each country, the calculation of the similarity takes too long and the distance of the decision space for determining the similarity between the paper types is shortened, and thus the accuracy of the judgment is lowered. Therefore, by selecting the target country to be identified from the standard image database of various countries, the number of standard images is limited and the types of bills are selected.

When a query image is generated, the corresponding species is selected by selecting an image having the highest similarity compared to a standard image stored in a pre-stored bill image database.

Figure 2 shows the procedure for identifying the denominations of the dollar using the target country as the United States.

Specifically, when the tilted query image of the $ 10 bill is input (FIG. 1A), the tilt and the center coordinates of the query image are calculated and the portion occupied by the real image of the banknote is extracted from the entire image (FIG. 1B). Rotation and position correction are performed using the calculated slope and the center coordinate (FIG. 1C).

Subsequently, the similarity is calculated by comparing each of the US dollar standard images stored in the standard image database, and then the corresponding species is found from the standard image showing the highest similarity (FIG. 1D).

In general, an image recognition target area may be largely formed in a spatial domain and a frequency domain. In the case of the recognition method by measuring the similarity of the spatial domain, the pre-image processing procedure for this is complicated because stable recognition is achieved only when the reference pixel and the similarity measure reference pixel of the target bank coincide. On the other hand, in the recognition method based on the similarity measurement of the frequency domain, the one-to-one comparison can be made for a given frequency domain without having to obtain an accurate reference pixel as in the spatial domain comparison. have.

However, in the case of using the frequency domain, a fast Fourier transform algorithm is used to speed up the extraction of the spatial frequency components of the banknote image. In addition, this method changes the placement of spatial frequencies as the rotation and the size change, but does not change the transition of the image. Therefore, the area corresponding to the power of 2 in the banknote image should be acquired, and the study of the preprocessing process and effective method for minimizing the change of rotation and size has a decisive influence on the recognition performance of the system using the same.

Accordingly, it is an object of the present invention to provide a banknote image processing method that exhibits strong characteristics in the processing of banknotes damaged in the distribution process, enables high-speed processing and is suitable for real-time processing.

Specific features for realizing this object of the present invention will be clearly understood from the following description.

The present invention for achieving the above object is a step of rotationally correcting the bill image obtained from the image sensor; Converting the rotationally corrected banknote image into a banknote image suitable for fast Fourier transform for optimal comparison of frequency domains; Extracting frequency information of the transformed image using a fast Fourier transform algorithm; And calculating similarity by comparing the extracted frequency information with frequency information of a standard banknote.

Preferably, the rotation correction step, the step of calculating the center point of the obtained banknote image; Extracting an edge component for the banknote image; Calculating a rotation angle of the obtained banknote image; And rotating the obtained banknote image by using the calculated rotation angle.

In the converting step, binarization is performed using a dynamic reference value of the luminance of the banknote region to exclude the damaged region, a pseudo banknote region is obtained from the X and Y projections of the binarized banknote image, and a center point of the pseudo banknote region. Calculating; Determining an actual processing area including the feature of the bill from the X and Y projective information; And compressing the actual processing area.

Preferably, after obtaining the pseudo-banknote area, the method may further include a step of custom transforming the binary banknote image with respect to the pseudo-banknote area with respect to the X-axis.

Preferably, frequency extraction may be performed using the following equation.

는 지폐 영상의 공간영역의 각 픽셀값이며,

는 푸리에 변환된 공간 주파수 영역의 주파수 값이다) (here,

Is the value of each pixel in the space area of the banknote image,

Is the frequency value of the Fourier transform space frequency domain)

Next, a bill image processing method according to an embodiment of the present invention will be described with reference to the drawings.

Find the center of a bill

As shown in FIG. 3, the size of the bill image obtained from the CIS (Contact Image Sensor) is almost constant, but various rotations are made, and thus a process for correcting it is required. The following steps are for this purpose, first finding the center of the bill.

Since the background of the original image obtained from the CIS has a value close to 0 (BLACK) at most gray levels, after binarization based on a threshold value near 0, as shown in FIG. The center of the x and y coordinate values of is determined as the center of the banknote image.

는 흰색영역의 픽셀 수이다. here,

Is the number of pixels in the white area.

Edge ( edge ) extraction

Next, as shown in FIG. 5, the edge component of the binarized banknote image is extracted. The size of the mask used for convolution for edge extraction is 3 × 3, and the mask value is {1, 1, 1, 1, -8, 1, 1, 1, 1}.

Rotation Angle Extraction

In order to correct the rotated bill, it is necessary to know the rotation angle of the acquired bill image.

Specifically, as shown in FIG. 6, the x point x _{c -α} , x _{c + α} separated from the x value x _c of the center point of the banknote image by the y axis from the top (or bottom) to the y-axis individually until an edge appears. Find two points (x _{c -α} , y _{c -α} ) and (x _{c + α} , y _{c + α} ) by searching with, and find the angle of inclination based on this information. At this time, since there is a possibility of incorrectly extracting the rotation angle due to damaged banknotes, by applying a number of α values, the mode of the tilt is selected as the rotation angle.

Rotation

As shown in Fig. 7, the original image is rotated using the calculated rotation angle.

After correcting the banknote image through the above process, in the following steps, the image is compressed to a size suitable for fast Fourier transform for optimal comparison of the frequency domain. These steps consist of finding a pseudo region and an actual processing region.

Binarization

First, as shown in FIG. 8, the binarization is performed through the dynamic reference value of the brightness of the banknote area in order to extract the pseudo banknote area to be processed. Here, "binarization" means conversion to a black and white image based on a predetermined threshold value. In addition, the dynamic reference value means that a certain boundary value necessary in this process is dynamically changed according to the brightness of the image. That is, it refers to a threshold value that changes in conjunction with an average brightness value of a binarization target region, not a threshold value of an absolute value that is generally designated as a fixed value.

Reextract Pseudo Banknote Area and Center Point

As shown in Fig. 9, in the X and Y projections of the binarized image, the pseudo-bill area for the width and length of the bill is obtained and the center of the rectangular area is obtained. That is, since banknotes in circulation are likely to be damaged, it is necessary to obtain a pseudo area and extract an accurate bill area excluding the damaged area before extracting the correct bill area.

Fit transformation for X axis

An image in which the banknote image obtained according to the performance of the CIS is shifted with respect to the X axis may be obtained. Therefore, in order to fit the tilted banknote image in the pseudo-banknote area, as shown in FIG. 10, a custom transformation is performed on the X-axis.

Actual processing area setting by X, Y projection

As shown in Fig. 11, the actual processing area is determined from the X, Y projection information. Here, the actual processing area refers to an area that includes the characteristics of the banknote. For example, the edges are white spaces, which are common features of all banknotes, and thus need to be excluded. For this purpose, the actual processing area is determined as the part where the projection value is minimum between the center point and the left end, the center point and the right end, the center point and the top, and the center point and the bottom.

compression

Since the image required for the fast Fourier transform should be an image having a square size of 2, as shown in FIG. The compression method compresses pixels in a processing area by applying an average value based on the size of a compressed image.

Frequency extraction

As shown in FIG. 13, frequency information of a compressed image is extracted using a fast Fourier transform algorithm.

는 지폐 영상의 공간영역의 각 픽셀값이며,

는 푸리에 변환된 공간 주파수 영역의 주파수 값에 해당한다. 주파수 추출은 일반적인 2차원 푸리에 변환식에서 중첩된 합 부분의 분리성질을 이용한다. 먼저 열이나 행 중의 하나의 방향으로 고속 푸리에 변환을 수행하고 그 결과를 이용하여 나머지 한 방향으로 고속 푸리에 변환을 수행하면 2차원 이산 푸리에 변환이 완성된다. 결국 1차원 푸리에 변환을 2번 수행하는 것이다. 이렇게 변환된 결과는 효율적인 유사도 산출 도 13과 같이 중심이 저주파가 되도록 재배치하게 된다.here,

Is the value of each pixel in the space area of the banknote image,

Is the frequency value of the Fourier transformed spatial frequency domain. Frequency extraction uses the separation property of the sum of overlapping parts in the general two-dimensional Fourier transform. First, the fast Fourier transform is performed in one direction of a column or a row, and the fast Fourier transform is performed in the other direction using the result to complete a two-dimensional discrete Fourier transform. In the end, one-dimensional Fourier transform is performed twice. Efficient Similarity Calculation As shown in FIG.

Similarity calculation

The similarity calculation uses a frequency table of standard banknotes prepared in the same process as the recognition process and a cumulative distance of one-to-one Euclidean distances by spatial frequencies of the frequency table for the banknotes recognized through the above process. Preferably, in this calculation, a comparison is made only for the 1/2 quadrant except for the third quarter where the spatial frequency domain is the same.

In other words, the similarity calculation is performed by comparing the spatial frequency domain unlike the spatial domain. Since the one-to-one correspondence is established in the spatial frequency domain without having to set the comparison criteria as in the spatial domain, the general cumulative Euclidean value is obtained. Is calculated for each reference bill and the category corresponding to the highest value is determined as the recognition result.

This method is achieved by, for example, a mutual correlation method. This expression

값을 산출한다.Here, I (x, y) and Q (xk, y + l) represent the magnitude of each frequency domain value to be compared. That is, the frequency transform value of the query image is shifted by the coordinate values k and l with respect to the frequency transform value of the comparison target image.

Calculate the value.

Bill classification

Banknotes are classified by calculating mutual similarity of spatial frequency features based on the spatial frequency feature database previously created and stored from the extracted spatial frequencies.

The image processing method including this process is Banknote Discriminator, Banknote Sorting Machine, Banknote Denominator, Banknote Recognizer, High Performance Banknote Counter, ATM (Automated) Teller Machine), or KIOSK image processing can be applied directly.

Each of the steps described in the present invention may be programmed to allow a computer to perform the above-described processing, and may also implement an integrated circuit including part or all of a functional module that performs the processing of each of the steps.

Thus, as long as each of the above steps according to the present invention is performed, the implementation manner is not limited, and various modifications are possible.

Edge detection by the edge search proposed in the present invention, unlike the detection of four sides by the conventional four-way search, can easily exclude the detection error of the uneven portion due to damage.

In the method according to the present invention, since the arrangement of spatial frequencies changes according to rotation and change of magnitude, the Fourier transform does not change in the transition of the image, and thus is more powerful than the processing for the spatial domain.

In addition, an improvement in processing time can be expected without a change in recognition rate.

Compared with the cross-correlation of the existing spatial domain, the setting of the reference pixel is not necessary.

Claims (5)

delete Rotationally correcting the banknote image obtained from the image sensor; Converting the rotationally corrected banknote image into a banknote image suitable for fast Fourier transform for optimal comparison of frequency domains; Extracting frequency information of the transformed image using a fast Fourier transform algorithm; And Comparing the extracted frequency information and the frequency information of the standard bill to calculate the similarity, The rotation correction step, Calculating a center point of the acquired banknote image; Extracting an edge component for the banknote image; Calculating a rotation angle of the obtained banknote image; And And using the calculated rotation angle to rotate the obtained banknote image in a similarity degree of the banknote using the frequency domain. The method according to claim 2, The conversion step, Performing a binarization using the dynamic reference value of the luminance of the bill area to exclude the damaged area, obtaining a pseudo bill area from the X and Y projections of the binarized bill image, and calculating a center point of the pseudo bill area; Determining an actual processing area including the feature of the bill from the X and Y projective information; And And measuring the similarity of bills using the frequency domain, comprising compressing the actual processing region. The method according to claim 3, And obtaining a pseudo-bill region, and further, performing a custom transformation on the X-axis to fit the binarized banknote image to the pseudo-bill region. Rotationally correcting the banknote image obtained from the image sensor; Converting the rotationally corrected banknote image into a banknote image suitable for fast Fourier transform for optimal comparison of frequency domains; Extracting frequency information of the transformed image using a fast Fourier transform algorithm; And Comparing the extracted frequency information and the frequency information of the standard bill to calculate the similarity, Frequency extraction is a method of measuring the similarity of bills using the frequency domain, characterized in that performed using the following equation.

(here,

Is the value of each pixel in the space area of the banknote image,

Is the frequency value of the Fourier transform space frequency domain)

Images