JPS60120485A - Pattern recognizing system - Google Patents

Abstract

PURPOSE:To process a preprinted part in the same way as that applied to a part printed in dropout colors, by subtracting a reference slip image from an inputted slip image in which a pattern to be read is recorded. CONSTITUTION:By using a slip (reference slip) on which no patterns to be read, such as sharacters, graphic forms, etc., are recoded, information (reference slip image) on the reference slip is stored in a memory 51. The slip image of an inputted slip is taken into the memory 51 in the same way and stored in another memory 52. A subtraction process between the normalizing image (memory 5) of the inputted slip and the thicker image (memory 56) of the reference slip is performed by a control section 2 and the part of the reference slip image contained in the inputted slip is erased and the remaining image is stored in a memory 56. The content of the memory 56 is the characters, symbols, or graphic forms recorded on the inputted slip and a recognition processing section 6 recognizes the content.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Application of the Invention] The present invention relates to a pattern recognition method, and particularly to a pattern recognition method suitable for recognizing characters and figures written in a complex background.

[Background of the invention]

Most of the conventional pattern recognition methods, such as character recognition methods, print a character entry frame in dropout color and read the characters written in the frame. However, due to the increase in printing costs, there is a growing demand for using forms that are arbitrarily created without being aware of the character recognition method. The problem when using such a form is that already printed circular characters, tables, lines, etc. (hereinafter referred to as "preprints") become interfering information, making it impossible to extract the characters to be read. There is an idea that it would be possible to recognize the preprint including the part, understand its content, and then edit only the characters to be read, but it is difficult to fully understand the content, and it is necessary to read parts that do not originally need to be read. There are problems such as a decrease in reading speed due to the reading of .

[Purpose of the invention]

An object of the present invention is to provide a pattern recognition method that can easily read only a pattern to be read written on a preprinted form, such as a character, a symbol, or a single figure, even if an arbitrary form is used. It's about doing.

[Summary of the invention]

In order to achieve this purpose, in the present invention, a digital image C (reference form image) of a form in which the target pattern to be read is not filled out is stored in advance, and a reference form image is created from the input form image in which the target pattern to be read is written. By subtracting , only the pattern to be read, for example, the written character pattern, is left, and the preprint portion can be processed in the same manner as if it were printed in dropout color.

[Embodiments of the invention]

Hereinafter, the present invention will be explained in detail using the drawings.

FIG. 1 is a block diagram showing the configuration of an apparatus that implements the recognition method of the present invention. Each part of the device is connected to a bus 1, and the overall operation is controlled by a control unit 2. The information (form image) on the form 3 is scanned multiple times by the photoelectric conversion device 4, photoelectrically converted, further digitized, and stored in the memory 51 via the bus 1. At the time of digitization, a known highly efficient encoding process may be performed, which can save the storage capacity of the memory that stores the form image, which may be advantageous in terms of cost. Further, in the following explanation, it is assumed that each pixel is binarized into one bit, but one pixel may be expressed with multiple values, or color information may be added using a color scanner.

First, use a form (standard form) in which no pattern to be read, such as characters or figures, is written, and the information on this standard form (
A reference form image) is stored in the memory 51. A normalized image obtained by performing tilt detection processing and tilt correction processing, which will be described later, by the control unit 2 on this reference form image is stored in the memory 52. Line extraction processing, which will be described later, is performed on this normalized image under program control of the control unit 2, and the extracted line pattern is stored in the memory 53. After that, the control unit 2
A thick image obtained by performing a known thickening process is stored in the memory 54.

The processing for the standard form is thus completed, and the processing moves on to the input form, that is, the form on which characters, symbols, figures, etc., which are patterns to be read, have been written.

The form image of the input form is stored in the memory 51 in the same way as the standard form.
, and its normalized image is stored in memory 52. The line pattern extracted from the normalized image of the input form is stored in the memory 55. The control unit 2 performs alignment processing, which will be described later, on the line pattern of the reference form (memory 53) and the line pattern of the input form (memory 55), and the position correction amounts (g, h) are calculated in the control unit 2. Stored in temporary memory. The control unit 2 creates a normalized image of the input form (memory 52) and a thicker image of the reference form (
Alignment processing and subtraction processing with the memory 56) are performed,
The portion of the reference form image that is omitted from the input form is erased, and the remaining image is stored in the memory 56. Details of the alignment process and subtraction process will be described later.

Thus, the contents of the memory 56 are recognized characters, symbols, or figures written on the input form, and these are extracted and recognized by the recognition processing section 6, but the contents of the processing are well known and will therefore be omitted.

As mentioned above, the entire apparatus is controlled by the control unit 2, but instructions from the user, such as whether the form 3 supplied to the apparatus is a reference form or an input form, are determined by the control unit 2.
Input from the keyboard 7 and instructions and responses from the control unit 2 are displayed on the display 8.

Next, various processes performed by the control section 2 will be explained. First, FIG. 2 shows a flowchart of the tilt detection process P AD (Pr
ogiam 77nalysis])iagram
) format. In Figure 2, each pixel on the form is P(
i, j), where j is the scanning line number (5 lines from 0 to J-1) and i is the pixel number in each scanning line (one pixel per scanning line). Reference numeral 201 is an entrance to the inclination detection process, and a form image P stored in the memory 51 is provided. In step 202, the variables S, . . . are initialized. In step 203, the processes of steps 204 to 214 are repeated for the angle θ in steps of Δθ from 1θl to θλ. Processes 204 to 214 are processes for determining the sharpness Sa in the θ direction and determining the direction θ that gives the maximum value Sm. Reference numeral 204 calculates the horizontal peripheral distribution H(j) about the scanning line j, and at this time, the direction of integration is tilted by an angle θ. 205 is the initialization of H(j), and 206 is 207 to 209-! for i! This is a process of repeatedly calculating the integral value H(j). 207 calculates the coordinates (ie l Jll) in the original image that are mapped to the coordinates (i, j) when rotated by the angle θ, and is defined by, but it can also be approximately substituted by good. 208 is a coordinate limit check, and 209 is an integration. 210-211 is sharpness S
, 210 is initialization, 211 to 212
This is the process of calculating the sharpness S defined by . The sharpness S, which is calculated based on the fact that characters preprinted on forms are generally lines that extend horizontally, and also contains many horizontal lines such as ruled lines, It has the property of becoming larger when the integration directions match. It would be better if not only the horizontal peripheral distribution but also the vertical peripheral distribution were evaluated at the same time (9), but this is omitted as this can be easily inferred. 213 and 214 are the parts that determine the maximum value S of 8θ and its direction. 215 is an exit from the tilt detection process, and the parameter θ1 is given to the output.

FIG. 3 shows a flowchart of the tilt correction process in PAD format. 3
01 is the entrance, the form image P and the rotation angle θ. is given. According to the double loop of 302 and 303, the following process is repeated for the scanning line number j and the pixel number i within the scanning line. 304 is a coordinate limit check where the coordinate m is within the limit, and when it is within the limit, the pixel P(i
,.

je) to pixel Q(i,j) in the new image.

When (i, +j,,) jumps out of the limit, Q with 307
Let (i, j) be 0. 308 is an exit of the tilt correction process, which outputs the rotated normalized image Q(i, j) and stores it in the memory 52.

Next, the process of extracting the straight line portion (10) from the preprint pattern will be explained. FIG. 11 is a diagram showing the principle. Only line segments of a certain length or more are extracted from a preprint pattern like (5) to create a pattern like Q3). This is projected horizontally to create a 0-like pattern (this is called a line pattern). Actually, there is no need to create the pattern (3), and (3) is directly converted to (Q).

FIG. 4 shows a flowchart of this line extraction process in PAD format.

Line extraction includes horizontal line extraction and vertical line extraction, but horizontal line extraction will be described here, and vertical line extraction will be omitted since it is the same. In FIG. 4, 401 is the entrance to line extraction, and the normalized image Q stored in the memory 52
is given.

In step 402, the processes in steps 403 to 409 are repeated for scanning line number j to obtain the long black continuous integral value AU). 4
03 is initialization. 404 is the pixel number i in the scanning line
This is a repetition of

In 405, it is determined whether the pixel Q (i, j) is 1 or 0, and if it is 1, the black run length B is counted in 406.

If Q(i,j) is O, then (11
) When the run length B to the previous pixel is greater than the threshold ε,
Integration processing is performed in 408, and the run length B is reset in 409. The reason why B is added to the integral value A(j) at 410 after the loop ends is that in loops below 404, the final point (i=I-1
) is not integrated. By adding the judgment of 407, integration is performed on A(j) only when there is a sufficiently long black horizontal line segment, so it is less affected by noise, and characters and symbols written on the form are separated from short line segments. Therefore, it can be ignored for the most part. For documents that do not have long line segments, the threshold value ε may be set appropriately small.

Reference numeral 411 is an exit of this process, and the long continuous integral value A is given to the output and stored in the memory 53.

As can be seen from the above explanation, the long continuous integral value A has a pattern that reflects the existence of line segments, and this is called a line pattern. Note 1. If peaks are extracted from line pattern A, the presence of long line segments themselves can be detected.

FIG. 5 shows a flowchart of the alignment process in PAD.

For vertical alignment, horizontal line pattern A (12
), and a vertical line pattern is used for the horizontal direction, but here only for the vertical direction. 50
1 is the entrance, and two line patterns A, +, A2 are given. A! A2 is extracted from the reference form image (memory 53), and A2 is extracted from the input form image (memory 55). 5 (12 is initialization. 503 is the process of repeating the processes 504 to 508 for the positional deviation amount d from (d+) to dl. 504f
d initialization, 505 is a repetition for scanning line j, and 506 is a process of calculating the difference between two line patterns by shifting the position by d, and integrating the absolute value.

If written as a formula, it is. 507 and 508, the minimum value of Cd and d that gives the minimum value can be seen. Reference numeral 509 denotes an exit, and the output is given dyu, which is stored in the temporary memory of the control unit 2. d
, is the vertical positional shift correction amount, and indicates that if the reference form image (the normalized image (13) of it) is shifted by d in the vertical direction, it will best match the input form image.

FIG. 6 shows a flowchart of the subtraction process in PAD format.

601 is an entrance, and a reference form image Q1, an input form image Q2, a horizontal positional deviation correction amount g1, and a vertical positional deviation correction scale are given. Here, the standard form image Ql
is a thick image of the standard form stored in the memory 56, and the input form image Q2 is the thick image of the standard form stored in the memory 56.
Assume that the normalized image of the input form stored in 2 is used. Moreover, h is written as d6 in FIG. 5, and g is obtained by performing the same processing as in FIG. 5 in the horizontal direction. Scan the entire image in a double loop of 602.603. 604 is a limit check to prevent the subscript from jumping out of the image area. 6
In 05, the reference form image Ql(i,j) is positionally corrected by g1 in the horizontal direction and h in the vertical direction, and the pixel Ql(i+g
, j ten h), its negation and input form image Q
2(i,j). Therefore, by the above processing (14), the image R created is one in which the portion of the reference form image included in the input form image Q2 is deleted. 607 is an exit, and the image R is given as an output and stored in the memory 56.

The operation of the embodiment shown in FIG. 1 can be understood from the above.

In addition, in FIG. 1, memories 51, 52, 53, 54°, 55.
56 is assumed to be different, but different areas within the same memory may be allocated. Alternatively, all or part of the processing performed by the recognition processing section 6 may be executed under program control of the control section 2.

Furthermore, one of the peripheral distribution pattern H(j) used for tilt detection and the line pattern (black long continuous integral value) AU) used for position correction can be substituted for the other, and even if the roles are exchanged. good.

Further, if a guide or the like is attached to the paper feeding device (not shown) to prevent the occurrence of skew of the form, the tilt detection process and the tilt correction process become unnecessary. It is also possible to perform tilt correction and position correction by detecting document edges or specific marks on the document and aligning these portions. moreover,
The tilt correction of the reference form (15) may be omitted, and the angle correction and alignment of the input form may be performed by rotating the input form at a plurality of angles and searching for the angle and position that best matches the input form. It is also effective to smooth line edges and remove noise such as voids and spots on the reference form. If the thinning method is still used as the recognition method, the input form image may be thinned instead of thickening the reference form image. If a contour analysis method is employed as the recognition method, contours may be extracted from the input form image and portions included in the reference form image may be deleted. When there is a part of the standard form image that you want to read (for example, a title part), by erasing that part with white paint or the like and registering the standard form image, the erased part can be included in the pattern to be read. Further, in the above description, processing is performed on a binary pattern, but multivalued images can also be processed, and in that case, the negation/logical product part in FIG. 6 605 can be replaced with a magnitude comparison operation. Equation (16) becomes as follows. A thick operation can also be defined using a magnitude comparison operation for a multivalued image. It can also be applied to color images.

Next, another embodiment of the present invention will be described with reference to FIG. In this embodiment, a plurality of types of normalized and thickened standard form images are registered in the memory 5, for example, the memory 54, and when processing an input form, the user responds to the input form by, for example, an instruction from the keyboard. This embodiment differs from the first embodiment in that a reference form image (and its line pattern) is selected and alignment processing and subtraction processing are performed, but the rest is the same. FIG. 7 shows a flowchart in PAD format of the process of calculating the address and extracting the image when the number of the reference form image is specified. 701
is given the reference form image number k at the entrance. 7
In 02, the size of the th standard form image Ih-J
Take out k.

(17) 703 sets the initial value of K to O. 704 performs the sum of products of 705, and the total number of pixels of the reference form image up to the (k-1)th is calculated as =Σ×mJ1. Add l. The reference form images are sequentially stored in the memory area G, and K indicates the start address of the second reference form image among them. A double loop of 706 and 707 allows G wholesale to G (K
The process 708 of transferring the contents up to +I, , J, -1) to the output buffer area Q is repeated. 710 is the exit,
Jk is given as output to the reference form image Q and its size specified from the keyboard. Note that the process of registering a plurality of reference form images in the memory area G can be performed by combining the processes reversed to those shown in FIG. 7, so a description thereof will be omitted.

In the second embodiment, it is also effective to display the reference form image instructed from the keyboard or the like on the display 8 for confirmation.

According to this embodiment, a plurality of forms can be registered in advance and one of the same type as the input form can be selected from among them. There is no need to register a standard form for each form, allowing high-speed processing.

Next, a third embodiment will be described with reference to FIGS. 8 and 9. This embodiment differs from the first embodiment in that the form image is divided and alignment is performed in each divided area.

FIG. 8 shows the case of four divisions. Generally speaking, it is divided into U equal parts in the horizontal direction and 7 equal parts in the vertical direction, and the (u, v
) By applying the processing of FIGS. 4 and 5 to the divided area, the position correction amount (g, ,
, hu, ) are observed. This correction amount is used to perform subtraction processing in place of the one shown in Fig. 6, and its flowchart (PAD
(format) is shown in Figure 9. 901 is an entrance, which includes a reference form image Q1, an input form image Q2, position correction tg,
h is given. Here, g and h are arrays of size U×■. 902 is a loop that repeats 903 to 905, and 903 to 905 are horizontal position corrections t g w, v
is given as a two-dimensional array, and is averaged with respect to the subscript V to obtain a -dimensional array G(u). G(
u) gives the horizontal position correction amount in each divided area when the form is divided into TJs in the horizontal (19) direction. 906-909 are the above 902-
Processing similar to step 905 is performed in the vertical direction to provide a vertical position correction amount H(v).

910 to 912 are position correction amounts E per pixel in the horizontal direction
This is a process that gives (i). In step 910, it is repeated for the sections ((U-1)) between the midpoints of the divided regions, and in step 911, it is repeated for the number of pixels α=I/U in each section. 9
12 is a linear interpolation formula. 913-914 are 2 at both ends
E(i) for half intervals is given here, and here the value is set as a value that is slid to the middle value of the region. 915-9
19 is a process in which the same processing as 910 to 914 described above is performed in the vertical direction to give a position correction amount FO per pixel in the vertical direction, and β in 916 is determined by β=J/V.

A double loop of 920.921 scans the entire image. 922 is a subtraction process, and in the horizontal direction E(i),
The pixel Qs (E(i) 2. F(j) ) on the standard form image corrected by F(j) in the vertical direction is extracted, and its negation and input form image Qh(''-j)(2
0). Therefore, the image R created by the above processing is obtained by erasing the portion of the reference form image included in the input form image Q2.

923 is an exit, and image R is given to the output.

In addition, in this example, the position correction amount g determined as a two-dimensional array
Although wv and h□ are averaged in the horizontal and vertical directions, they may be treated as two-dimensional arrays without being converted into -dimensional arrays, and the position correction amount at each pixel may be determined by two-dimensional interpolation. Although the image is divided into equal parts in this embodiment, it is easy to modify the image to be divided into unequal parts. It's a moat, like in this example, line pattern AI * A2 for each divided area
Instead of adjusting the position correction amount for the entire area, line pattern A for the entire image! (j) and A2(J), find the FO) (and E(i) in the horizontal direction) that best matches Ar(FG)) and A2(j), and perform the subtraction process. You may go. Such correspondence can be efficiently determined by dynamic programming matching.

According to this embodiment, the standard form image can be reliably erased even if the form image expands or contracts due to changes in the restrictive scanning pinch of the form due to the expansion and contraction of paper and the effects of the photoelectric conversion device (21). effective.

Next, an embodiment will be described in which the type of a form having a preprint is automatically identified, and it is possible to read a mixture of different types of forms without specifically assigning a form identification number. This embodiment is characterized in that the type of the form is automatically identified by actively utilizing the preprint information of the form.

This embodiment will be explained using FIG. 10. FIG. 10 is a flowchart showing the process of identifying the type of form in PAD format.

Reference numeral 1001 indicates various initializations, including initialization of the inside of the device and processing for reading various information of registered forms, which will be described later, from the external storage device 9. 10o2 selects registration mode or reading mode. 1003 to 1 in registration mode
007 processing is performed. At step 1003, the identification number n of the form to be registered is keyed in from the keyboard 7. 10
In step 04, the reference form to be used is converted into a digital pattern (22)-n by the photoelectric conversion device 4, and the reference form image P(n) is stored in the memory 5. This reference form image P (n) is subjected to normalization and thickening processing as necessary, as described above.

At step 1005, characteristic patterns E (n
). The example of extracting the feature pattern is as described above, but the reference form image P (n) may be used as it is as the feature pattern E(n). 1006
Now, input the format information F (n) from the keyboard 7, and input the standard form image P (n) or characteristic pattern E (n).
It is also stored in the memory 5. In 1007, P(n),
Information such as E(n) and F(n) is written to the external storage device 9, but if a nonvolatile memory is used as the memory 5, it is not necessarily necessary to use the external storage device 9. In this case, at least a memory that stores the feature pattern for the reference form (or the reference form image if the reference form image itself is used as the feature pattern), for example,
In the configuration of FIG. 1, the memories 53 and 54 may be nonvolatile memories.

(23) If the reading mode is selected in 1002, 100
8 to 1018 are performed. At 1008, an input form image P is obtained from the form on which the target pattern to be read is written. Samurai stores it in memory 5.

At step 1009, a characteristic pattern Ex such as a peripheral distribution pattern or a line pattern is extracted from P, but since this process has been described in detail in the first embodiment, it will be omitted here. 10
10 is feature matching, and feature pattern E! of the input form is used. This is the process of calculating the form identification number m corresponding to , and its contents consist of 1011 to 1015. Here, it is assumed that the line pattern stored in the memory 55 is used as the feature pattern E of the input form, and the line pattern stored in the memory 53 is used as the feature pattern E (n) of the reference form. 1011 is a loop related to already registered standard forms, n is the form identification number, and N is the type of the registered standard forms. In 1o12, the degree of similarity, for example, the distance, between Ex and E (n) is calculated.

Similarity calculations may be those used in normal pattern recognition, and distance and likelihood correlation (24) may be used as similarity measures. In 1013, the form identification number m that gives the minimum value of the distance D11 is determined.

After the loop of 1011 ends, DIIl reaches the threshold θ at 1014.
,.

Determine whether it is smaller. If it becomes smaller, it is because the registered form corresponding to the input form was not found.
1015 rejection processing is performed. In the reject process, a list of registered forms is displayed on the display 8, and the user is asked to select from the list, or if there is no matching form among the registered forms, the registration mode is executed to add a standard form. Perform processing such as When the form identification number m is determined as described above, the registered form image P (e) is subtracted from the input form image P at 1016 to obtain an image R in which only the patterns to be read, such as written characters and symbols, remain. In step 1017, characters and symbols are cut out and recognized using the format information F from the image R. 1
At 018, the recognition result is displayed on the display 8 or an output device such as a printer.

Note that if the preprinted part of the form to be used is not close to the part where characters and symbols are written and the cutout can be easily made (25), then 1017 may be omitted.

1019 is a termination process, which includes outputting various information to the external storage device 9, etc.

In this embodiment, since a line pattern is used as a characteristic pattern, the characters and symbols written on the form can be almost ignored, and there is an advantage that the identity of the form can be determined more accurately.

Note that when a line pattern is used as a characteristic pattern, it is not possible to distinguish between different forms if they have the same straight line part, but in such a case, specify a partial area on the form and It is a good idea to take various characteristics of Since the straight line part has already been detected, the partial area can be specified based on the straight line part, and the adverse effects of positional deviation can be avoided.

In this way, according to this embodiment, even if existing forms are used as they are, it is possible to read a mixture of different types of forms, which has the effect of significantly reducing the effort for designing new forms and printing costs.

〔Effect of the invention〕

(26) As explained above, according to the present invention, it is possible to read only the pattern to be read regardless of the preprint part, and there is no need to redesign and print forms, and normal forms can be used as they are. I can do it.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of an apparatus implementing a pattern recognition method according to the present invention, and FIGS. 2, 3, 4, 5, 6, 7, and 9. 10 is a flowchart explaining the processing in the control section of FIG. 1, FIG. 8 is a diagram showing an example of a form division state, and FIGS. 11(a), (b), and (c) are diagrams showing line patterns. This is a diagram for illustration purposes only. 1... Bus, 2... Control unit, 3... Report, 5...
・Memory, 6... Recognition processing unit, 7... Keyboard,
8... Disp (27) Live 11 Figure (Δ) (Bn521- (C)

Claims (1)

[Claims] 1. A standard form image converted into a digital image is stored, and a portion included in the standard form image is deleted from the input form image obtained by photoelectrically converting and digitizing the input form. The pattern recognition method is characterized in that the pattern to be read written in the input form is cut out from the erased image and recognized. 2. In the pattern recognition method according to claim 1, a direction with respect to the reference form image. A pattern recognition method characterized in that after normalizing the position, thickness, etc., and also normalizing the input form image, the input form image and the reference form image are aligned. 3. In the pattern recognition method according to claim 1 or 2, a desired reference form image is selected from among a plurality of reference form images, and the selected reference form image is extracted from the input form image. A pattern recognition method that is characterized by erasing parts included in . 4. In the pattern recognition method according to claim 1, 2, or 3, a plurality of partial regions are extracted from each of the input form image and the reference form image, and each of the plurality of regions is , calculate the alignment amount at each pixel on the form image from the alignment amount in each of the multiple areas obtained as a result, perform alignment using the alignment amount at each pixel, and then A pattern recognition method characterized by erasing a portion included in the reference form image from an input form image. 5. In the pattern recognition method according to claim 3, a feature pattern extracted in advance for each of the plurality of reference form images is stored, and the feature pattern is extracted from the input form image. , a pattern recognition method characterized in that the desired reference form image is selected by comparing it with a characteristic pattern of the reference form image. 6. In the pattern recognition method according to claim 5, predetermined format information is stored for each part of the plurality of reference form images, and the format information for the selected reference form is used to identify the target to be read. A pattern recognition method characterized by recognizing patterns.