JP2550012B2 - Pattern cutting and recognition method - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern cutout and recognition method, for example, a method for cutting out a pattern of a form written under natural writing conditions and a method for recognizing the pattern in an optical character reading device. It is about.

[Background of the Invention]

Characters to be read by a conventional character reading device (hereinafter abbreviated as OCR) must be correctly written in a character frame 11 set for each character, as shown in FIG. 1 (a). In that case, the protrusion from the frame is allowed,
As shown in FIG. 1 (b), the degree is about 1.0 to 1.5 mm in the up-down direction, and it is not in the adjacent frame in the left-right direction.

By the way, in order to spread OCR more and more, we do not have to write the characters inside the OCR unique character frame as described above, and as shown in Figs. It is necessary to enable the writing conditions as written.

The conventional character frame has a large size and the gap 5 between character frames is 1.0 to 1.5 mm, while the character frame with relaxed conditions has a smaller size as shown by 12 in FIG. , Gears between character boxes 6, 7 are 0 mm. As a result, there is a problem in that the characters protrude from the frame 12, and the characters overlap each other in the vertical direction, or the characters easily come into contact with each other. Further, in the case where the character pattern components are separated, for example, in the case of the character "ka" or the like in FIG. 2, most of the components may be in the adjacent frame, which is a problem in character reading.

In the prior art, for example, in Japanese Unexamined Patent Publication No. 59-98283, when there is ambiguity in the determination of the boundary separating one character pattern when cutting out into one character pattern, the boundary between patterns is determined. Create a plurality of hypotheses and send the unit pattern of each hypothesis to the recognition unit, and the recognition unit comprehensively judges the recognition result of the unit pattern of each hypothesis and selects a single hypothesis from each hypothesis. The method is disclosed, and letters, numbers, and
It is known to be effective in the range of katakana. However, if the recognition target is expanded to kanji, this method will cause the cutout / recognition part to cut when the deviation of the character "ka" in the kanji "chemistry" and the stroke are written far apart, as shown in FIG. It is judged that the hypothesis that Katakana “Ihi” is cut out and recognized by the cutout unit and recognition unit as Katakana “Ihi” is valid, and the recognition result “Kana” as Kanji is selected as a single hypothesis. There was a problem that I could not do.

[Object of the Invention]

An object of the present invention is to largely protrude from a character frame in order to solve such a conventional problem, or to cut out a pattern including kanji on a form written under natural writing conditions such that it is in contact with an adjacent character, It is to provide a means for recognizing those patterns.

[Outline of Invention]

The present invention inputs a form that does not have a character frame for each character, cuts out a pattern including characters from the two-dimensional video pattern converted into an electric signal by the cutout unit, and cuts out the pattern in the recognition unit into a pattern dictionary. A pattern segmentation and recognition method that is performed by comparing each pattern in FIG. 1 with each pattern in FIG. The components are arranged in order, a component list having the pattern component as an element is created, the attribute of the pattern component is calculated, and the attribute of the calculated pattern component is referred to to construct a character 1 or Make multiple hypotheses about combinations of multiple pattern components, and based on each of the multiple hypotheses, A combination of pattern components to be recognized is constructed from the min list and matched against each pattern in the pattern dictionary for recognition.For a recognition result based on multiple hypotheses, a single hypothesis is accepted as a predetermined regulation. If a plurality of hypotheses are accepted, each recognition result based on the plurality of hypotheses is obtained, and the language dictionary is obtained for each recognition result. It is characterized by outputting a recognition result based on a certain hypothesis as a final recognition result by collating with the content of.

Example of Invention

Hereinafter, the principle and embodiments of the present invention will be described with reference to the drawings.

The principle of the present invention lies in the following three points. That is,
(1) When ambiguity occurs in pattern cutting, a plurality of hypotheses are first set and a unit pattern based on each hypothesis is sent to the recognition unit without making a forcible judgment.
(2) The recognition unit identifies partial patterns and contact patterns and recognizes them based on comprehensive judgment. If a single candidate category cannot be determined, multiple candidate categories for each unit pattern are sent to the language processing unit. send. (3) The language processing unit makes a comprehensive judgment from a sequence of a plurality of hypotheses of a plurality of candidate categories by using language knowledge, and cuts out the result and checks the validity of recognition to eliminate the ambiguity.

FIG. 4 is a diagram showing various states of adjacent character patterns.

First, in FIG. 4A, the patterns 31 and 32 overlap each other in the vertical direction. In this case, a correct pattern can be cut out by extracting a continuous black area as a pattern component. A method of extracting a continuous black area as a pattern component has been known in the past, and not only when the characters are correctly written in the frame, but also when the text is simply overlapped, the area along the black area is outside the frame. Since even the portion that protrudes is extracted, it is possible to correctly cut out into a unit pattern.

Next, in FIG. 4 (b), the pattern is separated into portions 33 and 34, and most of the separated components 34 are contained in the adjacent frames. When it is unclear whether the pattern 34 belongs to the frame 21 or the frame 22, a multiple hypothesis is created by considering both cases as possible cases. Then, both cases are separately sent to the recognition unit and the language processing unit, and which hypothesis is correct is determined from the result.

Next, FIG. 4 (c) shows a case where the separated character pattern 36 contacts.

FIG. 4D shows a case where the separation patterns are in contact with each other. That is, in the case of FIG. 4B, the separation pattern is only the character “non”, whereas in the case of FIG. 4D, both the characters “non” and “normal” are separation patterns. , Their separation patterns are in contact.

In FIG. 4, complete patterns are in contact with each other. That is, the patterns are not separated, but the adjacent patterns are in contact with each other.

In FIG. 4 (f), the same as in (b), but the separation patterns 40, 41 themselves exist as the correct character patterns "a" and "hi", and the pattern obtained by integrating these also exists as the character "ka". This is the case.

A method of recognizing the cases shown in FIGS. 4B to 4F will be described below.

5 to 9 are diagrams showing the recognition principle of the present invention. FIG. 5 is a diagram for explaining the operation when the cutout unit makes a plurality of hypotheses, and FIG. 6 is a separated pattern. FIG. 7 is a diagram showing a recognition result when a component touches an adjacent character, FIG. 7 is an operation explanatory diagram when the separated pattern components touch each other, and FIG. 8 shows a recognition result when the complete pattern touches each other. FIG. FIG. 9 is an operation explanatory diagram when the separated pattern and the integrated pattern exist as characters.

First, as shown in FIG. 4 (b), the separated pattern components
A recognition method when 34 is in the adjacent frame will be described with reference to FIG.

In FIG. 5, 51 and 53 are two hypotheses output by the clipping unit. 200 is a recognition unit, 100 is a pattern dictionary, 101
˜104 are partial dictionaries in the pattern dictionary. The recognition unit 200 inputs the two hypotheses 51 and 53, performs character recognition, and outputs recognition results 52 and 54 for each. That is, the results for the first hypothesis are (SP.non) and (RJ.?). Here, (SP.non) means "non" sub-pattern, which is recognized by referring to the partial pattern dictionary 101, and (RJ.?) Is a reject (unread). is there. Furthermore, the results for the second hypothesis are (AC. Non) and (AC.
It means that it has been accepted (accepted) as the character "non" or "normal". Therefore, the second hypothesis is valid, and the recognition result is the character "non, ordinary". Note that the four partial dictionaries 101 to 104 provided in the pattern dictionary 100 are newly provided, and conventionally only the dictionary 104 having a normal pattern was provided. The partial dictionary 101 is a dictionary of partial patterns, the partial dictionary 102 is a dictionary of patterns in which a partial pattern contacts another character, and the partial dictionary 103 is a dictionary of contact character patterns.

Next, a recognition method when the separated character pattern component of FIG. 4 (c) is in contact with an adjacent character will be described.

In this case, as shown in FIG. 6, the cutout result is 55, and the recognition result 56 is (SP. Non), (SC. Non, general). Here, (SC.non, ordinary) means that the partial pattern of the character "non" and the character "normal" are in contact,
It is recognized by referring to the partial pattern dictionary 102.
From this result, it can be determined that the read character is the character “non, ordinary”.

Next, a recognition method when the separated pattern components shown in FIG. 4D are in contact with each other will be described.

In this case, as shown in FIG. 7 (a), two hypotheses 57 and 59 stand and recognition results 58 and 60 are obtained. Further, in this case, in particular, the sub-pattern 61, that is, in FIG.
Recognize 38 alone, resulting in 62 (SS.non, mediocre). Hypothesis 57 is the case where sub pattern 38 is assumed to be added on the right side, and hypothesis 59 is the case where sub pattern 38 is assumed to be added to the left side. (SP. Non)
(RJ.?) Is the "non" sub-pattern and Rigid (totally unknown), and (RJ.?) (SP. Ron) is the Rigid (totally unknown) and "Robo" sub-pattern. Also, (SS.
(Non, ordinary) means that the sub pattern of “non” and the sub pattern of the character “normal” are in contact with each other, and the dictionary 102 of patterns in which the partial pattern and other characters are in contact with each other.
Is referred to. By summing up these results, the answer is judged to be the letters "non" and "normal".

Next, a recognition method when the complete patterns of FIG. 4 (e) contact each other will be described.

In this case, as shown in FIG. 8, the whole is sent to the recognition unit without being forcibly divided, and the same pattern is searched for and recognized by referring to the partial pattern dictionary 103. The result (CC. Dai, Yama) was obtained, but it can be judged that this is the contact between the letters "Large" and "Mountain."

Next, a recognition method in the case where the partial pattern itself and the pattern obtained by integrating the partial patterns in FIG. 4 (f) exist as character patterns will be described.

In this case, as shown in FIG. 9, two hypotheses of 63 and 64 stand, and the recognition results of 63 are (AC · a), (AC · hi),
(AC / gaku) has 64 recognition results (AC / a), (AC / g)
Science) is obtained. Both of them are equal in recognition result, and it is not possible to judge which one should be adopted at this stage. Therefore, referring to the linguistic dictionary, although the word “Ichaku” does not exist, but the word “chemistry” exists, it can be determined that the recognition target character (group) is “chemistry”. As the language dictionary, it is possible to use not only words but also data of various knowledge such as grammar, rhetoric, and vocabulary.

As described above, as the operating principle of the present invention, the outline of the method of comprehensively judging the recognition result and the language processing result and giving the final answer has been described. It is realized by processing.

First, with respect to FIGS. 4A to 4F, FIGS.
The following is a summary of what was processed in the figure.

(A) (AC. Large) (AC. Mountain) → (AC. Large) (AC. Mountain) (C) (SP. Non) (SC. Non, general) → (AC. Non) (AC. General) (E) (CC. Large, Mountain) → (AC. Large) (AC. Mountain) The recognition result code for each hypothesis on the left side is rewritten to the recognition result code on the left side.

The generalized rules of these are called rewriting rules.

In the cutout method according to the present invention, the rewriting rule is as follows.

R1: (▲ ▼ .a) (▲ ▼ .b) + (AC.a) (AC.b)
→ (AC.a) (AC.b) R2: (SP.a) (SC.a, b) → (AC.a) (AC.b) R3: (SP.a) (▲ ▼. *) + (▲ ▼. *) (SP.b) → (RCG) R4: (SP.a) (▲ ▼. *) + (▲ ▼. *) (SP.b) + (SS.ab) → (AC. a)
(AC.b) R5: (CC.a, b) → (AC.a) (AC.b) R6: (AC.p) (AC.q) (SP.b) + (AC.a) (AC .b) → (DCT) where ▲ ▼ means not AC and shows all except AC. Also, * can take any value. (RC
G) means to recognize only the sub-pattern that gives the ambiguity of cutout. (DCT) means that if there is ambiguity in recognition, make a decision by referring to the language rules. Rule R1 corresponds to the above expressions (a) and (b). Even if there is a place where a and b are not accepted (recognized), if there is another accepted place, it is recognized. I will do it. The rule R2 corresponds to the above equation (c), and when the sub pattern of a is recognized and the contact between the sub pattern of a and the pattern of b is recognized, a and b are accepted ( Be recognized). Rule R3 corresponds to the above equation (d), and
When the sub-pattern of b is recognized and a candidate of any value other than accept is given, it is separated. Instruct to recognize only the sub-pattern that has Also, rule R4
Also corresponds to (d), and shows a case where only the sub-pattern processed according to the rule of R3 is comprehensively recognized including the recognition result. That is, the sub of a
When three patterns, that is, a pattern that cannot be recognized as a pattern, a pattern that cannot be recognized as a sub pattern of b, and a pattern in which the sub pattern of a and the pattern of b are in contact, are recognized, a acceptance, b Become an accept. Rule R5 corresponds to (e), and shows that the pattern in which a and b are in contact is a accept or b accept. The rule R6 corresponds to (f), and two hypotheses are affirmed, that is, a recognition result of p accept, q accept, and b accept is given, and a recognition result of a accept and b accept is given. In this case, the linguistic knowledge is referred to for a accept and b accept.

FIG. 10 is a block diagram of a character reading device showing an embodiment of the present invention.

This character reading device consists of a pattern observation unit 800, a cutout unit 900, a form format dictionary 950, and a pattern recognition unit.
200, pattern dictionary 100, recognition result final determination unit 400, recognition result rewriting rule dictionary 300, language dictionary 500, language processing unit 600
Consists of

Characters are written on the form 75 under natural writing conditions as shown in FIG. When the form 75 is input to the pattern observation unit 800 and subjected to photoelectric conversion and preprocessing (binarization, form skew correction), the two-dimensional image pattern is sent to the pattern cutout unit 900 as an electric signal. The pattern cutout unit 900 refers to the frame position parameter from the form format dictionary 950 and cuts out a set of patterns determined to correspond to one character from the image of one form and sends them to the pattern recognition unit 200. In the pattern recognition unit 200,
Pattern for one character that was input (may be a sub-pattern or a pattern for two characters that touched as described above)
Is compared with each dictionary pattern stored in the pattern dictionary 100 shown in FIG.
Send to 0. The final determination unit 400 sequentially applies each rewriting rule in the rewriting rule dictionary 300 to the recognition result until there are no more rewriting rules, and performs processing according to the rewriting result. That is, from the conditions of the rules R1 to R5, the symbolized recognition result that matches this is selected and applied, and the result is adopted. The language processing unit 600 refers to the language dictionary 500 for the recognition result, and when it remains undetermined, that is, performs the process for the result corresponding to the rule R6.

In FIG. 10, the pattern observing section 800 can be realized by a known technique, and therefore its explanation is omitted.

The process after the pattern cutout unit 900 will be described in more detail.

FIG. 11 is a diagram showing the content of data corresponding to the cutout processing and recognition processing flowchart of FIG.

In step 701, one line of image pattern 712 is cut out from one form of image pattern 711. Next, step 7
In 02, the pattern component for each black background is extracted by utilizing the continuity of the black background pattern, and after ordering in the horizontal direction, the component list 713 is created. Furthermore, the attribute of each component is calculated, and the component attribute list 714 is created. The component attributes include the upper and lower ends of each component, the coordinates of the left and right ends, the total length of the contour, and the like.

Next, in step 703, a hypothesis of a boundary between characters is established from the component attribute list 714 and the information in the form format dictionary 950 to create a character list 715. This list 715 shows which component each character pattern is composed of. In FIG. 11, the first hypothesis is that the orders 1, 2, and 5 are one character, 3 and 4, respectively. Becomes one character,
The second hypothesis is the order 5 only, one letter, 1 and 2, 3 and 4
Is assumed to be one character in total.

Next, in step 704 of the recognition unit, the component list 713, the component attribute list 714, and the character list 715 are input, the components included in the character list 715 are collected and pattern matching is performed, and the result is written in the result list 716. The result code indicating the matching result is (SP.a), (SC.a, b), (SS.a, b), (C
It takes a symbolic form such as (Ca, b), (AC.a), (RJ.a). As described above, these meanings are the sub-pattern of category a, the sub-pattern of category a and the category b, respectively.
Pattern, contact between sub-patterns of category a and category b, contact between category a and category b, pattern of category a. Candidate category is category a but has the meaning of "reject". There is.

In step 705, all rules in the rewriting rule dictionary 300 are referenced to the result list 716, sequentially applied until there are no applicable rules, and the processing according to the finally obtained result is performed. Into the result list 717.

In step 706, the result list 717 is examined and
It is determined whether or not the determination (DCT) has been made in 705,
When (DCT) is present, the matching with the language dictionary 500 is performed and the matching result is written in the result list 718. The result list 718 is the final output.

In step 707, it is judged whether or not all the lines on the form are completed, and if not completed, the process returns to step 701 and the above processing is repeated until it is completed.

〔The invention's effect〕

As described above, according to the present invention, the writing condition of the form is relaxed, and when the adjacent patterns overlap each other, when a part of the patterns exists with a large deviation from the original position, the adjacent patterns are further separated from each other. Even in the case of contact, the character reading device can cut out and recognize a proper pattern, which is extremely convenient for the user.
It becomes possible to spread OCR more and more.

In addition, in the embodiment, the cutting and recognition of the character in the character reading device is described, but the present invention is not limited to the character,
Of course, it can be applied to general patterns such as voice.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 shows a character frame of a conventional OCR form, FIG. 2 shows a character frame when restrictions are relaxed, and FIG. 3 shows a state of a kanji pattern. FIG. 4 is a diagram showing various states of adjacent character patterns, FIGS. 5, 6, 7, 8, and 9 are explanatory diagrams showing the recognition principle of the present invention, and FIG. 10 is an embodiment of the present invention. FIG. 11 is a functional block diagram of the character reading device, and FIG. 11 is a diagram showing a flow of the cutting and recognition processing of FIG. 75 …… Form, 100 …… Pattern dictionary, 200 …… Pattern recognition section, 300 …… Recognition result rewriting rule dictionary, 400 …… Recognition result final judgment section, 500 …… Language dictionary, 600 …… Language processing section,
800 …… Pattern observation section, 900 …… Pattern cutting section, 95
0 …… Form format dictionary.

Front page continued (72) Inventor Osamu Kunisaki 292 Yoshida-cho, Totsuka-ku, Yokohama City Hitachi Ltd., Microelectronics Device Development Laboratory (72) Inventor Kiyomichi Kurino 2880 Kozu, Odawara City Hitachi Ltd. Odawara Plant ( 72) Inventor Shozo Kadota 2880, Kozu, Odawara City, Odawara Plant, Hitachi Ltd. (56) References JP 59-91582 (JP, A) JP 59-112367 (JP, A) JP 58- 48182 (JP, A) JP-A-57-209574 (JP, A)

Claims (3)

(57) [Claims] 1. A form which does not have a character frame for each character is input, a pattern including characters is cut out from a two-dimensional video pattern converted into an electric signal by a cutout unit, and the pattern cut out by the recognition unit is used as a pattern. A pattern cutout and recognition method that is performed by comparing with each pattern in a dictionary, wherein a region of continuous black pixels is extracted as a pattern component for a region corresponding to one character line of the two-dimensional video pattern, The order of the arrangement of the pattern components is performed, a component list having the pattern components as elements is created, the attributes of the pattern components are calculated, and characters are formed by referring to the attributes of the calculated pattern components. A plurality of hypotheses about a combination of one or more pattern components On the basis of the above, a combination of pattern components to be recognized is constructed from the component list and is recognized by collating with each pattern in the pattern dictionary. If one hypothesis is accepted, the recognition result is output as the final recognition result, and if multiple hypotheses are accepted, each recognition result based on multiple hypotheses is obtained and A pattern segmentation and recognition method characterized by outputting a recognition result based on one probable hypothesis as a final recognition result by comparing the recognition result with the contents of a language dictionary. 2. The pattern cutout and recognition according to claim 1, wherein the attributes of the pattern component to be calculated include the upper and lower ends of the pattern component, the coordinates of the left and right ends, and the total contour length. Method. 3. A partial pattern of a pattern according to claim 1, wherein in forming a combination of pattern components and matching with each pattern in a pattern dictionary, (1) the pattern component represents a one-character category. If it is determined that the pattern component is a partial pattern, and the recognition result is a signal indicating the category name of the character category, (2) the pattern component is a complete pattern representing one character category. If it is judged that there is a pattern, a display indicating a complete pattern and a signal indicating the category name of the character category that is the recognition result, and (3) a pattern in which the pattern component represents a plurality of character categories When it is determined that the touched pattern is, the display indicating the touched pattern and the Pattern extraction and recognition characterized by outputting signals indicating category names and applying a predetermined regulation to the output signals for each pattern component that is the target of the above combination to test the validity of the hypothesis Method.