JP2007156764A - Figure detection processing program and paper sheets processing device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To detect a seal stamped on an optional position of a voucher at high speed. <P>SOLUTION: A pattern detection processing program comprises a step of extracting a multi-value contour at high speed, and a step of voting contour sorting points. Thus, an amount of calculation necessary for Hough voting can be reduced, and a stamped position can be detected at high speed. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a predetermined graphic detection method, program, and apparatus. In particular, it relates to detection of a seal.

  A slip reader (OCR, also referred to as an optical reader) is used to efficiently process a large amount of paper slips. One of the functions required for the slip reading apparatus is a seal detection function for determining whether or not a stamp such as a receipt stamp is pressed on the slip. For example, the slip processing business can be divided into a counter business and a back business, and the back business may verify whether or not the slip processing at the counter is appropriate depending on the presence or absence of a receipt stamp related to cash accounting. A slip reading device having a seal detection function is used in business processing such as separating slips without a seal with a sorting machine and checking again only the slip by a human.

  There is a technique using the Hough technique as a technique for detecting a seal (see, for example, Patent Documents 1 and 2). Non-Patent Document 1 has a detailed description of Hough. This method detects the vertex coordinates of a circle by heuristic processing, and uses the same vertex as a reference to reduce the number of dimensions of Hough voting to one dimension. Disclosed is a method for extracting circular figures by the Hough voting method.

Patent Document 1 is a patent relating to a method for detecting postmarks of postal items by performing Hough voting using a one-dimensional feature vector sequence extracted as a peripheral feature (for example, a feature obtained by projecting an image on a vertical axis) as a key. . Patent document 2 is a patent relating to a method for detecting a specific seal (copy prohibition, confidential information, etc.). A red-marked area such as “copy prohibited” stamps is extracted by color image processing, and a hough is used to determine the type. Therefore, it is difficult to apply to the detection of a stamp on a slip image picked up by multi-value gray gradation.

Japanese Patent Laid-Open No. 10-334240 JP 2000-123113 A Koji Ichikawa (others), “High-speed circle detection by Hough transform with one-dimensional voting array”, IEICE J80-D-2, no. 11, pp. 2949-2957 (1997)

  An object of the present invention is to provide a recording medium on which an OCR device that reads a paper slip detects a seal or determines a seal type and sorts paper slips based on the result, and a recording medium recording the seal detection program. There is.

  In the conventional seal detection method using the Hough voting, the Hough voting using the peripheral feature projected in one direction or the entire range of parameters representing the shape and size of the target figure by scanning all points on the image. It was detected by performing parameter space Hough voting. However, the parameter space Hough voting requires a huge amount of calculation for voting. In addition, when using peripheral features, a one-dimensional feature vector sequence is input, so Hough voting can be calculated at high speed. However, since peripheral features obtained from projected images in one direction are used as features, peripherals are used. The presence of a stamped portion within a predetermined area for calculating the feature is a precondition for applying the process. Therefore, it is not suitable for detection of a stamp pressed at an arbitrary place on the slip.

In addition, in the method of Non-Patent Document 1, since a circle vertex coordinate candidate cannot be detected for a missing circle lacking at the upper end, a circle is extracted by a Hough vote from a two-dimensional plane, so that the processing speed is greatly reduced. There is.

  The outline of the representative invention disclosed in the present application is as follows.

A contour point sequence is detected from the captured image of the paper, a plurality of segment points are detected from the contour point sequence, and the number of voting target points determined by the figure to be detected is determined from the segment points. A graphic detection processing program for detecting a detection target by calculating graphic information from voting target points and voting on likelihood as a graphic as a detection target.

In the conventional method, there is a restriction that the amount of calculation is increased or the area is limited. On the other hand, in the present invention, it is possible to reduce the amount of calculation by eliminating the limitation of the stamp position and performing Hough voting from the contour division points.

  In order to realize the object of the present invention, contour segment approximate voting processing based on high-speed multi-valued contour extraction is performed. As a result, high speed and robust detection of the seal is achieved.

  This process will be described in detail below with reference to the drawings.

  First, a hardware configuration as a premise of the present invention will be described. FIG. 1 is a block diagram relating to the hardware configuration of the high-speed slip reader. In the high-speed slip reader, the paper document is converted into electronic data by the image pickup device (0101), stored in the external storage device (0105) and the memory (0106), and read by the central processing unit (0107). The stamp detection program is stored in the external storage device (0105) or (0106), and the central processing unit (0107) detects the stamp on the captured electronic data image. When specifying the type of stamp, the type of stamp is determined with reference to the name dictionary stored in the external storage device (0104) or the memory (0106). These processes can be operated by a human through the operation terminal device (0102), and a processing result log or the like is displayed through the display terminal device (0103). Information such as processing results is stored in the external storage device (0105) as necessary, or data is sent to the external device through the communication device (0107) and the external communication line (0109). The above devices are connected by an internal bus (0108). Reflecting the result of the stamp detection, the entered slips are distributed and accumulated in the defined box by the sorter device (0104).

  FIG. 2 shows a functional block diagram of the processing. These may be realized by reading a program into a computer, or may be based on cooperative processing with hardware. The input paper slip is picked up in processing 0201 and becomes an electronic image. Subsequent processing is performed based on this electronic image. Since an electronic image is generally gray or color, binarization and contour extraction are performed in process 0202. In particular, if these are performed simultaneously, the speed can be increased. Further, the contour is scanned, and a candidate for a stamped portion is detected in process 0203. For the stamped part found, if necessary, the type of the stamp is checked by the process 0204 according to the signature definition defined in the dictionary 0206. The process 0204 may be selectively applied. Depending on the presence / absence of the stamp or the type of the stamp, the input paper slip is transferred to the stack of the sorter unit by processing 0205 according to the distribution definition defined in the dictionary 0207. The main point of the present invention lies in the stamp detection unit 0203. Therefore, the details of the stamp determination process 0204 and the signature dictionary 0206 are omitted. The seal detection process will be described in detail with reference to FIGS.

  FIG. 3 is a flowchart of processing for extracting a contour. Start with process 0301. In process 0302, a threshold value is selected. There can be multiple thresholds. After processing 0303, scanning of each line of the image is repeated. The line is scanned and the left and right ends of the run are determined according to a given threshold. In process 0304, link information of the connection relation of the run is calculated using the relation between the left and right ends of the obtained run, the run on the right side, and the run on the right. The process proceeds to process 0305, and if the image edge has not yet been reached on the currently scanned line, the process returns to process 0303, and processes 0303 to 0304 are repeated. When the image end is reached, it is determined in process 0306 whether there is something to be operated on the next line, and if there is a next line, processes 0303 to 0305 are repeated. Since the connection relation information of the run is obtained by the above scanning, the link can be scanned and the contour line can be formed by the process 0307.

  FIG. 4 shows a graphic voting flow according to the segmented N-point voting method. For example, if the voting graphic is a circle, the center coordinates and size of the circle can be estimated by voting three points. Processing is as follows.

  First, it is assumed that the contour run has already been created at the start stage (0401). In process 0402, one of a plurality of contour runs is selected. Contour runs may be selected by selecting all the contours one by one or by ignoring contour runs that are smaller than the size of the figure to be voted and processing only large runs. and so on. In the case where a limitation condition may be set for the size or position of a graphic drawn on a sheet, the number of votes can be limited by performing the contour selection process (0402), and the processing speed can be increased.

  Next, in process 0403, a segment point is obtained from the contour point group, and an end point serving as a starting point for N-point voting is determined. The division point mentioned here means a bending point when the outline is approximated by a straight line. There are two advantages of voting to find a figure not only from all the points on the contour, but only from the dividing points. One is to increase the processing speed by reducing the number of calculations of graphic information and the number of votes. The other is that the calculation result of the graphic information is stabilized. The stability of the calculation result will be described in the explanation of FIG. As a method for calculating the dividing point, for example, there is a linear approximation method. In the linear approximation method, a next segment point C is obtained between two points A and B that have already been determined as segment points, and the same processing is repeated for two sets of points A to C and C-B. It is. As the end point that is the starting point of the N-point voting, for example, the upper left point on the contour is set as the start point and the end point. In order to obtain the segment point C between the two points A and B, there is a method such as selecting a point farthest from the line segment passing through the two points A and B.

  Next, in process 0404, N points to be voted on are determined from the extracted segment points. The voting N points are determined by the figure to vote. The determination of N points is automatically set by the apparatus according to an algorithm for calculating voting information. When obtaining a circle figure, since the center coordinates and size of the circle can be estimated from the three points, the three points including the end points and one division point between them are the voting points. As the simplest example, three consecutive division points are voting points. However, when the distance between consecutive segment points is shorter than a threshold value determined according to the size of the figure, there is a case where processing is performed such as selecting another segment point instead of the next segment point. This is because if the graphic information is calculated from adjacent segment points, the calculation error may increase. The calculation error will be described with reference to FIG. After determining the voting N points, in processing 0405, graphic information is obtained from these N points using the end points and the segment points.

  Graphic information can generally be expressed as a vector representing the posture of a graphic. That is, the coordinates of the center of gravity of the figure, the size of the figure, and the inclination of the figure. In the case of a circular figure, since the inclination of the figure can be ignored, the vector representing the graphic information is a four-dimensional vector (a circle center coordinate X, a circle center coordinate Y, a circle radius R, an ID indicating a circle) ). However, in the following description, if the shape of the figure to be calculated is clear in context, the ID representing the figure shape is omitted. In this case, the graphic information representing the shape of a circle is a three-dimensional vector (circle center coordinate X, circle center coordinate Y, circle radius R). In general Hough transform, a multidimensional array representing a parameter space is prepared, and voting is performed by increasing the value of the array closest to the detected figure by one. In this method, voting is performed by registering vector information of detected figures. This eliminates the need to secure a multidimensional array for the parameter space in advance, so that the memory capacity for calculation can be reduced.

  By the processing from 0403 to 0405, the calculation and voting of graphic information is performed only on the division points, not on all the points arranged on the contour. As a result, the processing speed can be increased by reducing the number of graphic information calculations / voting times, and the stability of graphic information calculation can be ensured.

  A process 0406 is a process for recording the obtained graphic information in a table. In process 0407, it is determined whether or not to continue scanning for the contour currently being scanned. In the case of continuing scanning, the process returns to the end point determination process again, and the processes from 0404 to 0406 described above are performed. For example, in the case of circle voting, three new votes can be formed by combining one of the end points of the previous stage and the intermediate point to make an end point in a new step, and further selecting one point between the end points. In this case, two end point set candidates can be created, but in order to cover all classification methods, the end point candidates are stored in a queue or stack, and the end points are determined by popping them, and the intermediate points are selected. Just go. Since the concept of a queue or stack is common in information processing, the description is omitted here. The selection of a queue or stack is a story that depends on whether the depth-first search is good or the breadth-first search is good depending on the nature of the target problem, so it cannot be determined uniquely. If the above scanning is continued and there are no more end point candidates in the queue or stack, or if the scanning is not continued from the arrangement information such as the end points / partition points, the process 0407 is ended and the next process is started. . For example, in the case where the scanning is not continued when obtaining a circular figure, it can be concluded that if the end points and the division points are arranged almost on a straight line, even if further calculations are performed, only three points on the straight line can be obtained. In that case, since the estimation accuracy of the circle deteriorates, the calculation can be aborted. In process 0408, it is determined whether to scan another contour. If there is still a contour to be scanned, the processes 0402 to 0407 are repeated. If no further processing is performed, the process proceeds to the next processing. In process 0409, based on the graphic information registered as much as possible, a vote of which graphic is likely is performed. Voting can be performed by calculating the degree of proximity between vectors. That is, an inter-vector distance measure is introduced, and when the vectors are similar, the likelihoods of both vectors are added. Eventually, a vector having many similar vectors is considered to have a high figure likelihood and becomes a candidate. For example, the Euclidean distance can be considered simply as a distance measure between vectors. A Mahalanobis distance in which the weight is changed for each vector element is also possible.

  The graphic likelihood calculation process 0409 will be described in a little more detail using mathematical expressions, taking a circle vote as an example. Graphic information can be represented as vector data. When a plurality of votes (i = 1 to N) are voted, the i-th graphic information zi (i = 1 to N) is described as follows.

At this time, the likelihood LT of the vector data is determined based on the following equation.

The function LB is a function for calculating the likelihood according to the proximity of the distance between the two vectors. That is, as the number of similar votes increases, the overall likelihood LT increases. For example, as the function LB, the Euclidean distance is converted by an exponential function.

Functions with different tolerances for each function or variable

And so on.
In the graphic likelihood calculation process 0409, since the distance between vectors is calculated, the calculation amount at this time is generally M × (M−1) with respect to the number M of voted vectors. However, the number of M can be reduced because unnecessary voting can be omitted by performing the segment line segment approximation. Furthermore, by creating an index table on the two-dimensional coordinates based on the center of gravity of the vector, it is also possible to reduce N of objects for which distance calculation is performed. The amount of calculation using a two-dimensional index is reduced as follows. Suppose that the i-th graphic information is a vector (Xi, Yi, Ri). Assume that the possible value of X is from 0 to MAX, and similarly the possible value of Y is from 0 to MAX. In addition, a two-dimensional array IDX [10] [10] used for registering graphic information is prepared in advance. At this time, the i-th graphic information is registered in the 10 * Xi / MAX, 10 * Yi / MAX-th arrangement of the two-dimensional array IDX, that is, IDX [10 * Xi / MAX] [10 * Yi / MAX]. . When calculating the distance between vectors in the process 0409, the distance between those registered in the same arrangement of the two-dimensional array IDX [] [] is calculated. That is, the graphic information (Xi, Yi, Ri) and the distance calculation are performed only for other graphic information registered in IDX [10 * Xi / MAX] [10 * Yi / MAX]. Although the comparison of the calculation amount with the conventional method will be described later, the calculation amount can be greatly reduced by the above processing as compared with the conventional method.

Finally, there is a process 0410 for determining the maximum likelihood figure candidate based on the figure likelihood obtained by the above process. Here, the graphic information is sorted by the voting likelihood, and the acceptance / rejection is determined by the threshold.

FIG. 5 shows a schematic diagram in the case of obtaining the center coordinates from the three-point vote. The outline of the stamped portion is indicated by a thick line 0501. The outline is scanned, and three points (0502, 0503, 0504) are selected on the outline. The center coordinates of the arc formed by these three points can be obtained as the intersection of two bisectors (0507, 0508) of two line segments (0505, 0506) obtained from the three points. Expressed with specific mathematical formulas. Assume that the coordinates of three points (0502, 0503, 0504) are given as follows.

At this time, the center of the circle intersects at the points multiplied by k and l in the normal direction from the centers of the line segments X12 and X23. Therefore, the following simultaneous equations may be solved.

The values of k and l obtained by solving the above simultaneous equations are as follows.

The center coordinates are obtained using this k or l. If the central coordinate (0509) is (cx, cy), the solution is

It is obtained.
From Equation 8, it can be explained that the calculation result of the graphic information using the segment points is stable. The coordinates of the three points X1, X2, and X3 arranged on the contour are actually on the true circumference due to factors such as paper twist, distortion during scanning, and discretization error when digitizing the image. Observed at a position deviated from the coordinates. Assume that the coordinates of X3 are observed with a shift of (0, Δy). At this time, the center coordinates of the circle are

It is obtained. Here, if Δy in the numerator is ignored and an approximate expression of 1 / (1 + δ) ≈1−δ is taken into consideration, the fractional expression part can be developed as follows.

This is due to the estimated center coordinates of the circle.

It is shown that an error proportional to is generated. The smaller the value of Equation 11, the smaller the estimation error of the center coordinates of the circle. If it is assumed that the distance between the points X1-X2 and the distance between the points X2-X3 are substantially equal, the equation 11 can be further simplified, and the estimation error of the center coordinates of the circle can be interpreted as being proportional to the following equation. .

When calculating graphic information from adjacent points on the contour line instead of segment points, y21 + y32, which is the denominator of Equation 12, takes a value close to 0, resulting in a large error in the estimation of the center coordinates of the circle. However, this problem can be avoided when the graphic information is calculated from the dividing points because the value of y21 + y32, which is the denominator of Equation 12, can be expected to be larger. The above is the reason why the calculation of graphic information using segment points is stable.

  Here, we compare the amount of calculation between the conventional method and this method. FIG. 9 shows a conventional parameter space voting type Hough voting pseudo code for extracting a circle or an arc, and FIG. 10 shows a Hough voting pseudo code based on a segmented contour voting method.

  When the calculation amount of the conventional method is calculated based on this pseudo code, the number W × H of all points on the image, the number R of variations of the assumed circle size, the number of votes M for a circle of a certain radius, It can be seen that the calculation amount of the Hough vote is a value proportional to W × H × R × M. For this reason, the larger the size of the slip, the longer it takes to detect a circle stamp from the entire slip surface.

On the other hand, in the Hough voting by the segmented contour voting method, the calculation amount of voting is proportional to the number of contour segment points L on the image, and N (N− The calculation amount proportional to 1) is applied. Accordingly, the calculation amount is on the order of L + N (N−1) at the maximum. When a two-dimensional index is used for the voting vector, the eigennumber for obtaining the maximum likelihood voting result can be suppressed to α times N, so that the calculation amount is L + N × α. Thus, it can be seen that this method can greatly reduce the amount of calculation compared to the conventional method.

FIG. 6 shows the results of contour extraction and stamp detection using an actual slip image. As a result of investigating the outline of the stamped portion (0601), the center coordinate is indicated by a large cross (0602).

An operation mode of the slip reading apparatus when the seal detection function is used will be described with reference to FIG. Reference numeral 0701 denotes a main body of the slip reading apparatus, and reference numeral 0702 denotes a sorter unit attached to the right. The slip reading apparatus can operate with only the 0701 main body or a plurality of 0702 sorter units connected to the main body. Reference numeral 0703 denotes a paper feeding device of the slip reading device. Enter a paper slip here. The paper slip is distributed to the stacker unit 0704 of the slip reading apparatus main body, or to the stacker unit 0705 and the stacker 0707 of the sorter unit. At this time, the paper slip is conveyed in the direction of the arrow at 0706.

FIG. 8 is a schematic diagram showing how the slips are distributed to the stacker when the seal detection function is used. The most basic function is the flow shown in the upper part of FIG. In other words, the seal detection process (0801) is performed, a slip without a seal is distributed to the stacker 1 (0802), and a slip with a seal is distributed to the stacker 2 (0803). The lower part of FIG. 8 is a flow in the case of performing distribution by adding various slip determination processes to the stamp detection process. For example, slips without a seal are manually distributed for checking by hand, and slips with a seal are classified according to the type of business such as electricity and gas, or classified according to the store where the slip is received. In this case, by combining the seal detection function and the slip determination function, it is possible to store the stackers sorted for each type of slip. In this case, a seal detection process 0804 is performed, and a slip without a seal is distributed to the stacker 1 (0805). For a slip with a seal, a slip discrimination process (0806) is subsequently performed. In the slip determination process, for example, the type of the slip is determined by reading a bar code, a mark, or a numerical string written on the slip, or the reception store is determined by reading a stamped character. There is processing of. The slip determination process is not a single process, but may be a multi-stage process such as barcode reading or stamp reading. As a result of the slip discrimination process, if any discrimination result does not exist, it is assigned to stacker 2 (0807). If there is any discrimination result, it is assigned to the stacker 3 or a separate stacker according to the discrimination result (0808). It should be noted that the same processing can be performed even if the above-described conditions of no stamping and stamping are reversed. That is, it is possible to perform a slip determination process for a case without a seal.

In the configuration of the slip processing apparatus in FIG. 7, how to assign the stacker in FIG. 8 is determined depending on the operation policy. When only the main body (0701) of FIG. 7 performs the processing in the upper part of FIG. 8, stacker 1 (0802) and stacker 2 (0803) are allocated from the stacker group (0704) of the main body. When using the sorter unit (0702) in addition to the main body (0701), it may be possible to distribute more slips to the side closer to the worker (main body side), or to work by bending the waist. In some cases, the slips are distributed to the upper stage of the stacker group (0705) of the sorter unit. In addition, an operation may be considered in which distribution is performed by dividing into several stacker groups according to reading results such as slip types, stamp types, presence / absence of stamps, and the like. The stacker group can be divided into the side closer to the main body (0701) and the side farther from the main body (0701), the upper stack and the lower stacker group, or the stacker group (0704) of the main body and the stacker group of the sorter unit (0702) There is something to divide.
If the history during operation is used, the stacker assignment can be automatically changed to the optimum assignment by stacker placement automatic learning. For example, in order to reduce the amount of movement due to a person's slip extraction operation located on the main body (0701) side, papers having a large handling amount are distributed to a stacker (0705) close to the main body, and the paper feeding device decreases as the handling amount decreases. A stacker assignment to distribute to a stacker (0707) that is far from the stacker is considered. At this time, a function for automatically determining the number of stackers assigned can be realized by automatically counting the frequency of presence / absence of stamps during operation. For example, when there are many slips with stamps, it is possible to calculate how many stacks are allocated from the side closer to the main body using the following calculation formula. Assume that a maximum of A sheets can be stored in one stacker. The number of slips processed per hour is combined with the time for paper feeding and the time of automatic sorting, the rate C of slips with stamps by automatic learning, and BC slips with stamps are input per hour. If the generation interval of the slip removal work due to the stacker becoming full is designated as D hours or more, the number N of stackers for the slip with a stamp can be obtained by the equation N ≧ BCD / A. The N stackers are allocated to the stacker group (0704) of the main body and the stacker group (0705) of the sorter unit close to the main body, and the remaining stackers are allocated to the stacker group (0706) away from the main body. Do it automatically. The stacker layout learning can be performed using a slip type log, a stamp store log, and the like in addition to the stamp presence / absence log.

The hardware block diagram of a high-speed slip reading apparatus. The functional block diagram of a high-speed slip reader. Contour extraction processing flow. Seal detection processing flow. The outline figure of a circle stamp detection process. Sealed part detection result. The stacker structural example of a high-speed slip reader. Imprint detection function and stack distribution strategy. Pseudo code for parameter space voting. Pseudo code for section contour voting.

Explanation of symbols

010 ... Image capturing device 0102 constituting high-speed slip reading device. Operation terminal device 0103 constituting high-speed slip reading device. Display terminal device 0104 constituting high-speed slip reading device. Sorter device 0105 constituting high-speed slip reading device. External storage device 0106 constituting the slip reading device ... Memory 0107 constituting the high speed slip reading device ... Central processing unit 0108 constituting the high speed slip reading device ... Internal bus 0109 constituting the high speed slip reading device ... Constructing the high speed slip reading device Communication device 0201 Image capturing unit 0202 in stamping process Outline extracting unit 0203 in stamping process Stamp detecting unit 0204 in stamping process Stamp determining unit 0205 in stamping process Stamp processor 0206 used in stamping determination in stamping process Sorter processing unit 0207 in Sorting dictionary 0301 used in sorter processing in the image ... Start of contour extraction processing 0302 ... Threshold selection in contour extraction processing 0303 ... Image line operation in contour extraction processing 0304 ... Left and right run link processing in contour extraction processing 0305 ... Same line in contour extraction processing Scan continuation determination 0306... Another line scanning transition determination in contour extraction processing 0307 ... contour link scanning 0308 in contour extraction processing ... end of contour extraction processing 0401 ... start of specific graphic extraction processing 0402 ... contour selection 0403 in specific graphic extraction processing ... specific Division point calculation in the graphic extraction process 0404 ... Vote N point determination in the specific graphic extraction process 0405 ... Graphic information calculation in the specific graphic extraction process 0406 ... Graphic information registration in the specific graphic extraction process 0407 ... Same contour scanning in the specific graphic extraction process Next determination 0408 ... Determination of transition to another contour scanning in specific graphic extraction processing 0409 ... Graphic likelihood calculation in specific graphic extraction processing 0410 ... Maximum likelihood graphic candidate determination in specific graphic extraction processing 0411 ... End of specific graphic extraction processing 0501 ... the contour 0502 ... the feature point 0503 on the contour contour 0503 ... the feature point 0504 on the contour contour 0505 ... the line segment 0506 between the feature points on the contour contour 0506 ... between the feature points on the contour contour Line segment 0507 ... A bisector line 0508 connecting line feature points on the contour of the discriminator 0508 ... A bisector line 0509 connecting the feature points on the pendulum contour ... 0602 ... Display of the detection result of the stamp 0701 ... Main body of the slip reader 0702 ... Sorter unit connected to the slip reader 0703 ... Paper input section of the slip reader 0 04 ... slip reader stacker stacker 0705 ... sorter units of the main body (the side closer to the paper input section)
0706: Sheet transport direction in the sorter unit 0707: Stacker unit stacker (the side farther from the paper input unit)

Claims (5)

A contour point sequence is detected from the captured image of the paper,
A plurality of segment points are detected from the contour point sequence, the number of voting target points determined by the figure to be detected is determined from the segment points, and graphic information is calculated from the voting target points,
A graphic detection processing program for detecting the detection target by voting on likelihood as a graphic as the detection target based on the graphic information.   2. The graphic detection processing program according to claim 1, wherein the detection target is a circular stamp, and three division points are determined as the voting target points.   The graphic detection processing program according to claim 1 or 2, wherein the inflection point when the contour point sequence is linearly approximated is set as the division point.   When calculating the above contour point sequence, the same contour line between the end points of the runs can be calculated by calculating the link information between the end points of the already scanned upper run and the end points of the currently scanned run. The figure detection processing program according to any one of claims 1 to 3, wherein a contour point sequence is formed by calculating whether the image belongs. An imaging device that captures an input paper sheet, a central processing unit that extracts presence or absence of a seal from the captured image,
A sorter device that sorts the paper sheets according to the presence or absence of the seal,
The central processing unit detects a contour point sequence from the captured image,
A plurality of segment points are detected from the contour point sequence,
The number of voting points determined by the figure to be detected is determined from the above division points,
Calculate graphic information from the above voting points,
A paper sheet processing apparatus, wherein the stamp is detected by voting on the likelihood as a graphic as the detection target based on the graphic information.

Images