JP2000207491A - Reading method and device for character string - Google Patents

Abstract

PROBLEM TO BE SOLVED: To correctly segment and recognize the characters even against a character string where the Kanji (Chinese characters) touches each other by deciding the possibility of touching not for the coupling components included in a character string image but for the character pattern candidates which are generated by combining those coupling components. SOLUTION: A line segmenting part 602 extracts a line including a reading object character from a character line image which is inputted to a character string reading part 601 and outputs the line as a character line image. A character segmenting part 603 segments the character pattern candidates from the character line image. A character identification processing part 604 compares every character pattern candidate with a reference pattern that is previously registered and outputs the character category having the least difference as a character recognition result. Thus, a correct character segmentation candidate is outputted based on the over-segmentation and by a multiple hypothesis verification method. A character string collation processing part 605 collates the lattice of character identification result with a prepared reading object word and outputs the character string reading result while correcting the error of a character identifying part.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and an apparatus for reading and recognizing a handwritten character string, and more particularly to an automatic character string reading method for correctly recognizing characters in contact with each other.

[0002]

2. Description of the Related Art Reading a character string described on a document mainly requires the following four processing procedures.

(1) Character line extraction: extraction of character line image from document image (2) Character extraction: extraction of character pattern from character line image (3) Character identification: character type (character code) of each character pattern (4) Character string collation: Determining the result of character string reading using language knowledge that has been obtained in advance The above four processes have a relationship as shown in FIG. As a conventional technique, as a technique for extracting a character line image from a document image in (1), a method of extracting a character line in a region after extracting a region in a top-down manner (Akiyama et al. Iterative Document Image Segmentation Using Features "IEICE Technical Report PRU85-41).
With regard to the character extraction of (2), a continuous pattern of black pixels in a character line image (hereinafter referred to as a connected component) is used as a unit to create a character pattern candidate, and a combination of a plurality of connected components is generated. The method of verifying these hypotheses and outputting the correct character segmentation results is described in the literature (Fujisawa et al., "Character segmentation method for automatic separation of contacted handwritten characters", 1984 IEICE General Conference, 1984) No. 1588). In order to obtain a correct character segmentation result, the character pattern candidate is given a certainty factor using the result of character identification, and the character segmentation of (2) and the character identification of (3) are made to act complementarily. Others "Character Extraction and Recognition from Handwritten Character Strings Using Linguistic Information" IEICE Technical Report PRU-92-
40) etc. As a technique relating to the character string collation of (4), a method of generating a finite automaton from a lattice of character identification results and extracting candidate words by inputting words in a dictionary (see Marukawa et al. Error Correction Algorithm for IPSJ Transactions on Information Processing, Vol. 35, No. 6, etc.).

However, since the above-described character segmentation method uses a connected component as a unit, there is a problem that when characters are in contact with each other, a correct character pattern cannot be obtained.

In order to solve this problem, there is known a method of forcibly cutting a connected component including a contact to obtain a character pattern.
For example, there is a method of classifying strokes between touching characters between touching characters and identifying and cutting contact points by stroke shape analysis (Fujisawa et al., “Pattern Extraction and Recognition Method”, Patent No. 2550012). . As one of the contact separation methods, for example, as shown in FIG.
There is a method in which the number of black pixels is counted in the vertical direction of the character string to create a histogram, and a point at which this shape changes is cut and separated as a contact position candidate.

[0006]

SUMMARY OF THE INVENTION The problem to be solved by the present invention is to cut a contact character for obtaining a correct character pattern candidate for a kanji character or a kana character string including a contact character.
An object of the present invention is to realize a separating function so as to reduce the generation of erroneous character pattern candidates as much as possible, thereby speeding up the processing time of the entire character string reading process.

The conventional contact character extraction method described above focuses on a large connected component in a character line image, and
It is a technique to separate. The determination as to whether or not a contact is included is basically made by comparing the size of the connected component with a threshold value. Unlike alphanumeric characters, kanji are composed of partial and adjacent characters, so a single character pattern often consists of multiple connected components. The size is often the same as the size of a character that is not touching. For example, in the character string shown in FIG.
Although there are three contacts of 01, 302, and 303, the first two contacts indicate that the connected component including this is not particularly large compared to other connected components, and based on only its size. One character can be determined.

[0008] Therefore, when reading a kanji character string, if an attempt is made to apply a conventional contact character cutting / separating method, it is not determined that there is a contact based on the size of the connected component, and character extraction may fail. there were. On the other hand, if the threshold value relating to the size of the connected component in the contact determination is loosened, most of the character patterns are determined to be in contact, and as a result of the cutting / separation processing, the number of character patterns as a cut-out hypothesis increases significantly, The processing time for pattern generation and the time for character identification processing for these character patterns are increased, and the efficiency of the entire character string reading processing is greatly reduced.

[0009] Even if the connection component including the contact is successfully cut / separated, correct character pattern candidates may exist separately in different character pattern candidates. For example, if the newly generated character pattern candidate is "day" or "month"
In the case of kanji biased or near, such as, it is not only possible to obtain a correct character segmentation result by simply using a connected component newly generated by cutting / separation as a character pattern, Recognition as another character such as "month" leads to erroneous recognition of a character string.

[0010]

In order to solve the above-mentioned problems, according to the present invention, first, the determination of the possibility of contact is made by combining connected components, not connected components in a character string image. Is performed for the character pattern candidates that have been set. By judging the possibility of contact with a character pattern candidate, even if the contact portion exists only in a small connected component, if the character pattern including the connected component becomes large, it is judged that there is a possibility of contact. By performing the cutting / separating process, a correct character extraction result can be obtained. In the character string of FIG. 5 used as an example earlier, the image of the character line is composed of connected components surrounded by a rectangle in FIG.
In the method of the present invention, the possibility of contact is determined for a character pattern candidate obtained by combining the above. Therefore, it is determined that there is a possibility of contact for a character pattern candidate as shown in FIG.

Second, for a character pattern candidate determined to have a possibility of contact, a connected component including a contact is selected from the character pattern candidates and cut / separated. In this case, since the same connected component may be the target of cutting in a plurality of character pattern candidates determined to be in contact with each other, by storing the connected component of the target of cutting / separating processing, The separation process should be activated only once. This prevents the same connected component from being cut a plurality of times and prevents an increase in character patterns. This also prevents an increase in the processing time of the entire character string reading process.

Third, the process of determining the contact character pattern for the character cut-out hypothesis and the process of separating (cutting) the contact character pattern are recursively repeated until there is no character pattern determined to be a contact character. Thus, even if three or more characters are continuously touched, character extraction can be performed.

Fourth, it is determined whether or not a character pattern candidate newly generated by the cutting / separating process is integrated with an adjacent character pattern. If this criterion is satisfied, these components are integrated to form a new one. Generate a character pattern. In this way, when a part of a character such as unbalanced or adjacent is in contact, after separating the contact part, the remaining part of the character can be integrated to obtain a correct character pattern.

Fifth, the validity of the cutting is determined based on the relative positional relationship between the connected component newly generated by the cutting / separating process and the connected component not to be cut in the character pattern candidate. If determined to be inappropriate, cancel the disconnection. As a result, generation of non-character patterns can be prevented, and an increase in character pattern candidates can be suppressed.

[0015]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the present invention will be described with reference to the drawings. First, the flow of a character string reading process including a contact character will be described with reference to a data flow diagram and FIG.

This embodiment is a character string reading process in which an image including a character string to be read is input and a character string as a result of character string recognition is output. First, the line cutout 602 extracts a line including a character to be read from the character line image input to the character string reading 601 and outputs it as a character line image. The character line image is a black and white binary image. Character extraction 603 extracts character pattern candidates from the character line image. A character pattern candidate is created using a continuous portion (connected component) of black pixels in the character line image as a unit. The character pattern candidate corresponds to one arc of the cut-out hypothesis network described later. A character pattern candidate is composed of only one connected component or a combination of connected components. At the stage of just creating character pattern candidates, it is common that the correct character segmentation method is not uniquely determined, so all possible character pattern candidates are enumerated, and the order relation between the character pattern candidates is stored in memory in a network format. Remember. This network is called a cut-out hypothesis network. FIG. 13 shows a cut-out hypothesis network for the character line image in FIG. Network nodes correspond to character pattern boundaries, character line breaks, and arcs correspond to character patterns.

In order to select a correct character extraction candidate from the character pattern candidates in the extraction hypothesis network, the degree of certainty of character extraction is weighted based on the character identification result for the character pattern candidate, or the interval between character patterns and the character pattern candidate are selected. There is known a method of weighting the certainty factor using a geometric feature such as height and width. The latter is described in the literature (Koga et al. Segmentation of Japanese Handw
ritten Characters Using Peripheral Feature Analysi
s International Conference for Pattern Recognitio
n, 1998, pp. 1137-1141).

The character identification processing 604 compares each character pattern candidate with a standard pattern registered in advance, and outputs a character category having the smallest difference as a character recognition result. As described above, a correct character extraction candidate, that is, a candidate character string is output in the character extraction 603 by the multiple hypothesis testing method based on over-segmentation.

The character string collation processing 605 compares the lattice of the character identification result with a word to be read prepared in advance, and outputs a character string read result while correcting an error in the character identification unit.

FIG. 1 created in the character extracting process 603
In the cut-out hypothesis network as shown in FIG. 3, character pattern candidates are configured using connected components as the minimum unit. In the portion indicated by each arc, only one connected component or a combination of a plurality of connected components is shown.
One character pattern candidate is shown. When (a part of) a plurality of characters constitute the same connected component due to contact between characters, it is necessary to separate them separately and create a new character candidate pattern. In this embodiment, the separation of the contact and the process of creating a new character pattern candidate are realized by operating a character pattern candidate on the network after generating a cut-out hypothesis network.

The procedure of the contact separation and the process of creating a new character pattern candidate will be described with reference to FIG. Hereinafter, FIG. 13 is used as an example of the cut-out hypothesis network.

First, in step 801 of FIG. 15, a cut-out hypothesis network is created from an input character line image.

In step 802, contact separation and new character candidate pattern addition processing are performed on all character candidate patterns on the cut-out hypothesis network.

In the first step 803 in the contact separation and new character candidate pattern addition processing, the character candidate pattern on the cut-out hypothesis network shown in FIG. Make a decision. The process of determining whether or not a contact character is included will be described later.
Among the character pattern candidates on the cut-out hypothesis network in FIG. 13, character pattern candidates 901, 902, and 903 shown by bold lines in FIG.
Is determined to have a possibility of contact.

If it is determined in step 803 that the contact character is not included, the process proceeds to step 804, and the process proceeds to the next character pattern candidate in the loop of step 802.

If it is determined in the determination process of step 803 that a contact character is included, the process proceeds to step 805. In step 805, for a character pattern candidate determined to include a contact character, a connected component including a contact is estimated from among the connected components constituting the character pattern candidate, and this is cut and separated, and a new connected component image is formed. Generate. The processing for cutting and separating the connected component will be described later.

The connected component newly generated by the connected component cutting / separating process is a connection component that has not been subjected to the cutting / separating process among the character pattern candidates determined to have a possibility of contact in step 806. New character pattern candidates are generated by selecting and combining appropriate components from the components. As a result, from the character pattern candidates 901, 902, and 903 determined to include the contact in FIG. 17, the character pattern candidates 1001, 1002, 1003, 1004, 1005, and 1006 in FIG.
Is generated and added to the cut-out hypothesis network.

Even if a new character pattern candidate is generated by cutting / separating a contact character pattern, there is a case where a connected component constituting a correct character pattern candidate is divided into different character pattern candidates on a cut-out hypothesis network. is there.
Therefore, after performing the character pattern candidate addition processing to the cut-out hypothesis network after the cutting / separation processing, in step 807, the character pattern candidate newly added by the cutting / separation processing and the adjacent character pattern on the cut-out hypothesis network The candidates are integrated to generate a new character pattern candidate.
As a result, character pattern candidates 1101, 1102, and 1103 are newly generated in FIG.

Steps 803 to 807 are repeatedly executed until no new character pattern candidate is generated. Through the above processing, a character cut-out hypothesis network including a correct character pattern candidate can be generated for a character string image including a contact character.

Next, the details of the processing from processing steps 803 to 807 in FIG. 15 will be described for each step.

First, the contact character determination in step 803 will be described. Japanese characters and kana characters differ from alphanumeric characters in that one character pattern, such as one-sided or side-by-side, often consists of multiple connected components, and the size of the connected components in the character is smaller than the size of the character pattern There are many. In such a case, even if the characters are in contact with each other, if the connected component including the contact is taken out, the size of the connected component is not different from that of the non-contacting character. Therefore, for Japanese kanji character strings,
When trying to apply the above-described contact character cutting / separating method, it is not determined that there is a contact based on the size of the connected component, and there is a possibility that a correct character extraction result cannot be obtained. Therefore, instead of determining whether or not the character is a contact character based on the size of the connected component in the input character line, the contact character is cut by noting a character pattern candidate on the cut-out hypothesis network.

The flow of the contact character determination process for a character pattern candidate will be described with reference to FIG.

For each character pattern candidate in the cut-out hypothesis network of FIG.
It is determined whether or not a contact is included by comparing the geometric feature of the character pattern candidate with a threshold value. Examples of the geometric feature include the length of the character pattern candidate in the line direction and the aspect ratio. As the threshold value for the length of the character pattern candidate in the line direction, a value 1302 obtained by multiplying the width 1301 of the character line image in FIG. 13 by a coefficient can be used.

If the character pattern candidate does not satisfy the condition in step 1201, the process proceeds to step 1202, and the contact character determination processing ends. The character pattern candidate 701 in FIG. 13 does not satisfy the above conditions, and is determined not to be a contact character. Character pattern candidate 702 satisfies the upper limit.

In step 1203, for the character pattern candidate satisfying the condition of step 1201, a connected component that is actually cut / separated is selected. As a selection criterion, the length of the connected component in the row direction is adopted, and the connected component having the longest length is selected as a target of cutting / separation processing. By comparing the size of the character and the size of the connected component including the contact with the connected component that satisfies this criterion, by providing a process of excluding from the target of cutting if the connected component is smaller than the threshold value Thus, the accuracy of the contact determination can be improved. This prevents useless cutting processing from being performed on character pattern candidates consisting only of small-width connected components, such as the character "kawa", or character pattern candidates consisting only of small connected components due to blurring. The purpose is to: As shown in FIG. 14, the width 1401 of the character line image is regarded as the estimated width of the character pattern, and is multiplied by a coefficient to obtain a threshold value 1402 relating to the lower limit of the width of the connected component. Figure
The connected component 1403 having the largest width among the connected components constituting the 13 character pattern candidates 702 is selected. This connected component 1403
Is larger than the threshold value 1402 relating to the lower limit of the connected component width, the connected component 1403 is selected as a target of the cutting / separating process.

The contact determination process for a character pattern candidate is as follows.
Since this is performed for all character pattern candidates on the cut-out hypothesis network, for example, the character candidate pattern 702 in FIG.
And the character candidate pattern 703 are both determined to include contact, and in step 1203, the same connected component 1403 (FIG. 14) is cut off redundantly, which is inefficient. Therefore,
The connected components to be cut are registered in a cut-target connected component management table as shown in FIG. 15, and when the same connected component is to be cut repeatedly, a character pattern candidate including this connected component is Then, among the candidates determined to be in contact with each other, only the connected component in the smallest pattern is set as a disconnection target. In the cut target connected component management table, the identification number of the connected component, or the relative coordinates of the left end 1501, the right end 1502, the upper end 1503, and the lower end 1504 of the connected component in the character line image are stored, so that the connected component is stored. Can be identified. By having a pointer 1505 to a character pattern candidate, it is possible to associate with a character pattern candidate including the connected component.

In the contact character pattern determination process for the character candidate pattern 702, in step 1204, a search is performed in the cut target connected component management table to determine whether the connected component 1403 determined to be cut is already registered as a cut target connected component. Find out by doing. If it has already been registered, the process proceeds to step 1205, and the contact character pattern determination processing ends.

If not registered, the following processing is performed for each character pattern candidate on the cut-out hypothesis network. First, in step 1206, the coordinates 1501 to 1504 of the connected component are stored in the cut target connected component management table of FIG. Thereafter, a new character pattern candidate is generated based on the connected component after cutting, but if the character pattern candidate selected as the contact character pattern candidate includes an extra connected component, after cutting, The connected component is also included in the generated new character pattern candidate, and as a result, a correct character pattern candidate cannot be obtained. So step 120
In step 7, as the character pattern candidate for the connected component to be cut, the smallest character pattern candidate among the contact character pattern candidates including the connected component is selected and registered in the cut target connected component management table. In this way, instead of the character candidate pattern 703 in FIG. 13, the character candidate pattern 702 is registered in the character pattern candidate pointer field 1505 of the cut target connected component management table as a contact character pattern candidate for the cut target connected component 1403.

In this contact character pattern determination, if it is determined that the character is not a contact character, or if the connected component to be cut cannot be identified, the process proceeds to step 804 in FIG. 8, and the next character pattern candidate on the cut-out hypothesis network is contact character determined. Perform processing.

If it is determined that the character is a contact character, and if the connected component to be cut can be identified, the process proceeds to step 805, where the connected component is cut / separated. The above is the contact character determination processing in step 803.

Next, the processing for cutting / separating connected components in step 805 in FIG. 8 will be described.

FIG. 16 shows an example of kanji contact. In the case of vertical writing, the upper character "Harai" often contacts the lower character, such as 1601 and 1602. On the other hand, for horizontal writing, 1603
There is a case where “Harai” of the left character touches the right character as in “1” and a case where “Export” of the right character touches the left character as in 1604. Therefore, when cutting the connected component, (1) In the case of vertical writing, the place where the vertical straight line is in contact with the horizontal straight line at the bottom (2) In the case of horizontal writing, the horizontal straight line is the right vertical direction Location where it touches the straight line The connected component is cut using the location where the vertical straight line touches the horizontal straight line on the right side as the cut point. If the contact point is 1
There are cases where there are no other pixels of the same character pattern candidate in the horizontal direction from the contact point as in 601 and cases where there are other pixels in the same character pattern as in 1602. , And two methods of cutting point identification based on the variation of the thickness of the linear component and the cutting point identification using the peripheral distribution of black pixels in the row direction.

FIG. 17 shows the overall flow of the cutting / separating process of the connected component. First, in step 1701, the connected component is cut by tracking the change in the stroke width, and then in step 1702, the connected component is cut based on the marginal distribution shape. Hereinafter, cutting / separating processing of each connected component for a vertically written character string will be described.

First, the cutting process based on the stroke width change tracking will be described.

By extracting the contour from the character image,
A method of identifying the place where a stroke touches and cutting the contact is described in the literature (H. Fujisawa et al., “Segmentation Methods for
Character Recognition: From Segmentation to Docume
nt Structure Analysis '' Proc. of IEEE Vol.80, No.7,
July 1992). In the first connected component cutting method in this embodiment, a method of preparing a run-length image perpendicular to the character line direction and examining a change in the length of an adjacent run to extract a stroke in the image is adopted. . Figure 18 shows the flow of the connected component cutting / separating process by tracking the change in the stroke width.

First, in step 1801,
Create a run-length image perpendicular to the row direction for the connected component to be separated. Next, in step 1802, a stroke is extracted from the run-length image.
As shown in FIG. 19, when the average value of the lengths of runs determined to belong to the same linear component at that time is W ₁ 1901, a value in a certain range from W ₁ is set as a threshold 1902, and adjacent runs are determined. If the difference in length is within the threshold value 1902, the strokes are regarded as the same, and if the difference exceeds the threshold value, the strokes are regarded as different. Such a stroke extraction method is described in the literature (D. Nishiwaki “AN
ew Recognition Method of Black Ruled Line Form '' Pr
oc. of Second IAPR Workshop on GraphicsRecognition
1997).

The following processing is performed in step 1803 for each extracted stroke.

First, in step 1804, a contact point between strokes satisfying the following conditions is regarded as a cutting point.

(A) The stroke has a certain length and is a (vertical) stroke parallel to the row direction. (B) The length where the stroke in the row direction and the stroke in contact with the end of the row are perpendicular to the row. (C) The length of the stroke in the line direction is longer than a certain percentage of the stroke perpendicular to the line. (D) The stroke in the line direction under the above conditions is the beginning of the line relative to the line and the stroke in the vertical direction. For the stroke 2001 in Fig. 20, the width 2003 can be calculated from the average of the lengths of the runs included in the stroke, and the length 2004 can be calculated from the number of runs that make up the stroke. Therefore, it can be used to determine whether the stroke is parallel to the row direction (condition (a)). Since the length of the stroke 2002 that is in contact with the stroke 2001 at the end of the line can be estimated from the length of the run 2005, it can be determined that the stroke is a stroke perpendicular to the line (condition (b)). Linewise stroke length 2
Condition (c) can also be determined from 004 and a stroke 2005 perpendicular to the line. Furthermore, if the connection relation between strokes is held using a stroke management table as shown in FIG. 21, it is possible to compare the length of the stroke that contacts the line at the beginning of the vertical stroke (condition ( d)). In FIG. 21, reference numeral 2101 denotes an area for holding information on one stroke, and the inside thereof includes a stroke length 2102, a width 2103, the number of strokes at the beginning of the line 2104, and a table relating to the strokes in contact. Pointer to area 21
05, the number of strokes in contact with the end of the line 2106, a pointer 2107 to an area in the table relating to the touching stroke, and a pointer 2108 to run length data 2109 of the stroke.

If all of these conditions are not satisfied, the flow advances to step 1805 to shift to the processing related to the next stroke in step 1803.

If all the conditions are satisfied, step 180
Proceed to 6 to perform cutting processing of the boundary between the strokes. The disconnection of the connected component is performed by deleting one run from the boundary between two strokes in the stroke in the row direction.

In the following step 1807, it is checked whether or not the connected component has been divided into a plurality by the process of deleting the run at the stroke boundary portion in the step 1806. Perform labeling processing from the run-length data as described in the literature (Mori "Pattern Recognition" IEICE)
It is sufficient to check whether a plurality of connected components are extracted.

Here, if the connected component to be cut is ring-shaped, and one run data is not divided into a plurality of connected components even if one run data is deleted, the connected component newly created by cutting is rejected. In step 1808, the processing shifts to the processing related to the next stroke in step 1803.

If it has been divided into a plurality of connected components, the flow advances to step 1809 to receive the newly generated connected components by cutting, and the process proceeds to step 1803 for the next stroke.

When the processing for all the strokes is completed in step 1803, the flow advances to step 1810 to rank the cutting point candidates. By cutting and selecting cutting candidates based on this order and limiting the number of candidates, excessive cutting can be prevented, errors in character string reading processing can be reduced, and processing speed can be improved. In the present embodiment, the product of the lengths of the intersecting strokes, that is, the product of l1 and l2 in FIG. 20, is adopted as a standard for ranking the cutting point candidates, and the candidate with a larger value is positioned higher. .

Finally, the width of the stroke selected as the top cut candidate in step 1811 is output as the reference value of the stroke width of the character including the connected component, and the connected component cutting process based on the stroke width change analysis ends. .

Next, the connected component cutting process by tracking the change of the projected shape in step 1702 in FIG. 17 will be described.

A method of calculating a marginal distribution in the line direction of a character line image and cutting out characters at the boundary between the “valley” and the “mountain” is described in the literature (Y. Lu “On the Segmentation o”).
f Touching Characters '' International Conference on
Document Analysis and Recognition 1993)
Widely known. The flow of the connected component cutting / separating process by the peripheral distribution shape tracking in the present embodiment is based on the same concept. The processing flow is described according to FIG.

First, in step 2201, the peripheral distribution of black pixels in the row direction of the input connected component is calculated. A marginal distribution such as 2302 is obtained from the connected component 2301 in FIG.

Subsequently, in step 2202, an estimated value of the width of the stroke constituting the connected component is calculated. If the stroke width can be obtained in the connected component cutting process (step 1701) by tracking the stroke width change in FIG. 17, this value is used. If it cannot be obtained, the center (for example, 3/4) excluding both ends from the peripheral distribution in the row direction of the connected component image obtained in step 2201
Range) 2303 is the minimum value of the stroke width. The reason why both ends of the marginal distribution are excluded from the target of the minimum value calculation is to exclude from the stroke width estimation a state in which the line width becomes narrower due to, for example, “spreading” when writing a character.

The selection of the cut point of the connected component based on the marginal distribution shape uses two threshold values, large and small, and in the case of vertical writing, a boundary where the shape changes from the saddle portion to the ridge is obtained. In step 2203, these thresholds are obtained by multiplying the constant of the estimated line width obtained in step 2202. When the line width is Lw, the larger threshold value h is given by h in the equation α × Lw.
Α> β as shown in 2304 and w 2305, h assumes the minimum value of the width when multiple strokes overlap, and w assumes the maximum value of the line width. Of the marginal distribution of the connected component,
Among the areas having a value equal to or less than a plurality of stroke widths, a minimum point is found among the points equal to or smaller than the maximum value of one stroke width, and the minimum point closest to the end of the line is set as a candidate cutting point of the connected component.

In step 2204, using the two threshold values, the following processing is performed for each pixel by writing from the head of the line to the end of the line in the search range of the central portion of the connected component image described above.

In step 2205, it is determined whether or not the position is a local minimum value in the saddle portion of the peripheral distribution shape. That is,
If the value of the marginal distribution at that position is smaller than the threshold value h,
It is determined that the peripheral distribution shape is located in the saddle portion. In this case, if it is smaller than the threshold value w, a position that can be placed in the search range as a minimum value is stored in step 2207. If these conditions are not satisfied, the process proceeds to step 2206, and the process for the next position on the search range is continued.

When the processing within the search range is completed, step 220
Proceeding to 8, the minimum value is obtained from the values stored as the minimum values of the peripheral distribution in the search range, and the position in the search range at that time is output as a cutting point candidate. If there are a plurality of locations having the same local minimum value, the location closest to the end of the line is output as a cutting point candidate.

In step 2209, a blank of a connected component is inserted by blanking one pixel in the direction perpendicular to the row at the position selected.

The processing for cutting / separating connected components in step 805 in FIG. 8 has been described above. For horizontally written character strings, by applying conditions where the order relation between strokes perpendicular to the line direction and strokes parallel to the line direction is replaced, judgment of cutting conditions is performed, and cutting / separation processing of connected components is performed. The connected component can be cut for both contacts classified in (2).

Next, the validity verification of the disconnection processing in step 806 in FIG. 8 and the addition processing to the character extraction hypothesis network will be described.

When the cutting / separating process of the connected component including the contact is completed in the character pattern candidate determined to have the touch, the character pattern candidate is next created from the newly generated connected component by the cutting / separating process. Must be added to the cut-out hypothesis network. In this embodiment, two character pattern candidates are newly generated by one cutting / separating process for the contact character pattern candidate. for that reason,
In the contact character pattern candidate, it is necessary to create a character pattern candidate by combining a connected component that has not been subjected to the cutting / separating process with a connected component newly generated by the cutting / separating process. In the present embodiment, it is determined whether each connected component is included in the first character pattern candidate or the second character pattern candidate based on the relative positional relationship of the connected components. The character pattern candidate 2401 in FIG. 24 is a character pattern candidate 901 determined to include a contact character on the cut-out hypothesis network in FIG.
02 is a connected component to be cut / separated. In the connected component cutting / separating process, the component is separated into connected components 2403 and 2404. Of these connected components, the connected component 2403 at the beginning of the line is used as a character pattern candidate at the beginning of the line, and the connected component 2404 at the end of the line is used as a character pattern candidate at the end of the line. It is determined which character pattern includes the component 2405. The criterion for determination is (i) if the end 2407 at the end of the line of the connected component 2405 that is not subject to cutting / separation processing is closer to the beginning of the line than the end 2408 of the connected component 2403 at the beginning of the line, the character is a candidate for a character pattern at the beginning of the line. Include (ii) If the end 2406 at the beginning of the line of the connected component 2405 that is not the target of cutting / separation processing is closer to the end of the line than the end 2409 of the connected component 2404 at the end of the line, it is included in the character pattern candidate at the end of the line (iii In other cases, it is assumed that the same character pattern as that of the connected component having a larger portion overlapping with the connected component 2405 that is not to be cut is used. In FIG. 24, since the connected component 2405 is included in the character pattern candidates at the beginning of the line, character pattern candidates 2410 and 2411 are generated, and the character pattern candidates 1001 and 1002 in FIG.
Is added to the extraction hypothesis network.

At this time, if the connected components that have not been subjected to the cutting / separating process have an inclusive relationship in the row direction with respect to the connected components subject to the cutting / separating process, the result of the cutting / separating process is obtained. And no new character pattern candidates are created from the generated connected components.

The above is the verification of the validity of the disconnection process in step 806 in FIG. 8 and the process of adding the character extraction hypothesis network.

Next, the process of adding integrated character pattern candidates in step 807 of FIG. 8 will be described.

A new character pattern candidate is added by integrating the character pattern candidate newly added to the cut-out hypothesis network by the cutting / separation processing with the adjacent character pattern candidates. This is because even if only one of the kanji that was in contact with the cut / separation process was cut out and the other side was already on the cut-out hypothesis network as another character pattern candidate, the correct character pattern This is to generate candidates. In step 807 in FIG. 8, all character candidate patterns newly generated by the cutting / separation processing are integrated with adjacent arcs on the cut-out hypothesis network to form new character pattern candidates. If the length in the line direction is smaller than the threshold value corresponding to the maximum value of the width of one character, the character pattern candidates are integrated and a new character pattern candidate is cut out and added to the hypothesis network.
A character pattern candidate 1101 and a character pattern candidate 1102 in FIG. 11 are generated and added from the character pattern candidate 1002 in FIG.

The above is the process of adding integrated character pattern candidates in step 807 of FIG.

Thus, step 802 in FIG.
Is performed on all character patterns. The processing is also performed for a character pattern candidate newly generated by the cutting / separating process. As a result, even when two or more contact characters are included in one character pattern candidate, as described above, it is determined that there is a possibility of a contact character from the size and aspect ratio for each character pattern. At times, correct character pattern candidates can be obtained by applying a cutting / separating process recursively.

FIG. 10 shows the hardware configuration in this embodiment.
See Figure 25. The present invention described so far is applied to, for example, a character string reading device 2501 as shown in FIG. A scanner 2502 controls image input. The scanner 2502 is connected to a pre-processing device 2503 that controls extraction of a character line image in which a character string to be read is described from an input page. The preprocessing device 2503 is connected to a bus 2505 of the recognition device via an input / output interface 2504. Reference numeral 2506 denotes a character string reading result display terminal for displaying a character string reading result. The character string reading result display terminal 2506 is connected to the bus 2505 of the recognition device via the input / output interface 2507. Reference numeral 2508 denotes an external storage device for storing programs required for the present invention, standard patterns for character identification, and the like. A hard disk or the like is used as the external storage device. The external storage device is connected to the bus 2505 of the recognition device via the input / output interface 2509. The recognition device 2501 has a bus 2505 for connecting each part inside the recognition device, control of the whole recognition device, and character string reading processing. Processing unit 2510, which controls input / output, and input / output interfaces 2504, 2507, 250, which control input / output with external devices
9. It has a memory 2511 for storing programs, tables, dictionaries, etc. necessary for reading character strings. When starting the device,
First, a program and a standard pattern dictionary are loaded from the external storage device 2508 to the memory 2511. After the loading is completed, when the character line image is supplied from the preprocessing device 2503 to the recognition device 2501,
The character line image is copied to the memory 2511. Further, the arithmetic processing unit 2510 recognizes the character string while generating the cut-out hypothesis network in the memory 2511 according to the method described in the present invention, and transmits the read result to the character string read result display terminal 2506. The character string reading result display terminal 2506 displays the recognition result on its screen.

The embodiments of the character reading method and apparatus according to the present invention have been described above. The characters read in the present invention are handwritten character strings. The handwritten characters can be used not only for reading a sentence containing kana and kanji characters written in a document but also for reading characters such as numbers and symbols on a form and an address name written on a mail. It is necessary to quickly and surely read the contents of such a large number of forms, postcards and sealed letters. For example, in the case of a postal matter, the entered address, name, postal code, and the like can be accurately and quickly read, and can be used for sorting for delivery. Existing mail sorting machines optically read mail addresses and recognize the read black-and-white grayscale images as characters, and classify them by address according to the recognition contents. It is clear that the application of the invention improves the accuracy of the sorting and also improves the sorting speed, except that it depends on a mechanical mechanism.

[0077]

According to the present invention, correct character extraction and recognition can be performed even if the kanji is in contact with a character string. In particular, the present invention is also effective in a case where a character portion such as a bias or a side touches and the connected connected component does not become particularly large as compared with others. Further, the present invention can be applied to any apparatus or method for recognizing handwritten characters.

[Brief description of the drawings]

FIG. 1 is a general processing flow of character string recognition.

FIG. 2 is a diagram showing an example of a method of extracting a character from a contacted number.

FIG. 3 is a diagram illustrating an example of a touched kanji character string image.

FIG. 4 is a diagram illustrating an example of a connected component in a touched kanji character string image.

FIG. 5 is a diagram showing an example of a contacted Kanji character pattern candidate.

FIG. 6 is a diagram showing a processing flow of one embodiment of character string recognition in the present invention.

FIG. 7 is a diagram illustrating an example of a cut-out hypothesis network.

FIG. 8 is a diagram showing a flow of a character extracting process for a character string including a contact kanji character.

FIG. 9 is a diagram illustrating an example of an expression on a cut-out hypothesis network of a result of contact character pattern candidate determination.

FIG. 10 is a diagram illustrating an example of a result of cutting / separating processing of a contact character pattern candidate.

FIG. 11 is a diagram illustrating an example of integration of a character pattern candidate generated as a result of the cutting / separation processing with an existing character pattern candidate.

FIG. 12 is a diagram showing a contact character determination processing flow for a character pattern candidate.

FIG. 13 is a diagram illustrating an example of contact character determination for a character pattern candidate.

FIG. 14 is a diagram illustrating an example of selection of a cutting target connected component from character pattern candidates.

FIG. 15 is a diagram illustrating an example of a cutting target connected component management table.

FIG. 16 is a diagram showing an example of how to contact Chinese characters.

FIG. 17 is a diagram showing an overall flow of a connected component cutting process.

FIG. 18 is a diagram illustrating the connection component cutting / analysis by the stroke width change analysis.
Flow of separation processing.

FIG. 19 is a diagram illustrating a method of extracting a straight-line component from a run-length image.

FIG. 20 is a diagram showing a cutting determination method based on a stroke width change analysis.

FIG. 21 is a diagram illustrating an example of a stroke management table.

FIG. 22 is a diagram illustrating a flow of a connected component cutting / separating process by projection shape change tracking.

FIG. 23 is a diagram illustrating a cutting determination method based on marginal distribution shape analysis.

FIG. 24 is a diagram showing how to combine connected components of a new character pattern candidate after cutting / separation processing.

FIG. 25 is a diagram illustrating a hardware configuration of a character string reading device.

 ────────────────────────────────────────────────── ─── Continuing on the front page (72) Inventor Masafumi Koga 1-280 Higashi-Koigakubo, Kokubunji-shi, Tokyo Inside the Central Research Laboratory, Hitachi, Ltd. (72) Inventor Yutaka Sakami 1-280 Higashi-Koigakubo, Kokubunji-shi, Tokyo Hitachi, Ltd. Central Research Laboratory (72) Inventor Hiromichi Fujisawa 1-280 Higashi-Koigakubo, Kokubunji-shi, Tokyo F-term in Central Research Laboratory, Hitachi, Ltd. 5B029 AA01 CC27 DD10 EE08 5B064 AA04 AB03 AB14 AB16 CA08 CA10 EA34

Claims (11)

[Claims] 1. A character line extracting process for extracting a character line from an image on which a character string is described, a character pattern extracting process for extracting a character pattern candidate from the character line, and a character pattern extracting process for extracting the character pattern candidate In a character string reading method having character identification processing for identification and character string collation processing for extracting a sequence of words from the character identification result, there is a possibility that a character pattern candidate generated from a character line image may include a contact. A contact character determination process for determining whether or not a character pattern is present,
A contact point estimation process for estimating a contact position, a character pattern cutting / separation process for cutting / separating a character pattern at the estimated contact point, and a new character pattern candidate generated from the partial pattern generated in the above cutting / separation process And a character string reading method. 2. A character string reading method according to claim 1, wherein the character pattern candidate including a contact is determined based on the size of the character pattern candidate including the partial character pattern regardless of the size of the partial character pattern including the contact part. A character string reading method comprising a contact character determination process for determining whether a character string is a character string. 3. The character string reading method according to claim 1, wherein a connected component including a contact is selected from the connected components as constituent elements of a character pattern candidate determined to include a contact in the contact character determination processing. A character string reading process comprising: a process of selecting a character string; and a process of storing the selected connected component and controlling the cutting / separating process so that the cutting / separating process of the same connected component is not performed repeatedly. Method. 4. The character string reading method according to claim 1, wherein the processing of cutting / separating a contact character pattern image and the processing of generating a character pattern candidate are performed recursively as long as it is determined that there is a contact character. Character string reading method. 5. A character string reading method according to claim 1, wherein a character pattern candidate newly generated by a cutting / separating process of a connected component determined to have a possibility of contact and an existing character pattern adjacent to the character pattern candidate are generated. For the pattern that is integrated with the character pattern candidate, it is determined whether or not the character pattern candidate is a character pattern candidate based on the positional relationship with the preceding and following character pattern candidates and its size. A process of generating and adding candidates. 6. The character string reading method according to claim 1, wherein a character pattern candidate newly generated by the cutting / separating process of the connected component is compared with another connected component in the same character pattern candidate. Rejecting the cutting / separating process from the positional relationship. 7. A character string described in a document is read as a grayscale image, b) a continuous portion of black pixels of the grayscale image is cut out as a connected component, and c) a character pattern candidate including the connected component is If it is determined that the character is a contact character, d) cut the connected component having the maximum length in the row direction of the connected component in the character pattern to generate two character patterns; A character string reading method that performs character recognition on character patterns obtained by repeating steps d through d. 8. A character string reading method according to claim 7, wherein in step c, it is determined whether the character is a contact character based on the size or aspect ratio of the character pattern candidate. 9. A scanner for converting a handwritten character string written on a document into grayscale image data, and a memory for holding data which is a continuous portion of black pixels of the grayscale image and is cut out as a connected component. In the case where the character pattern candidate including the connected component is a contact character, a new character pattern generated by cutting the connected component having the maximum length in the line direction of the connected component in the character pattern is A character string reading device, comprising: processing means for performing character recognition. 10. A scanner for converting a handwritten character string described in a postal matter into grayscale image data, and a memory for holding a continuous portion of black pixels of the grayscale image cut out as a connected component. In the case where the character pattern candidate including the connected component is a contact character, a new character pattern generated by cutting the connected component having the maximum length in the line direction of the connected component in the character pattern is A mail sorting machine characterized by processing means for executing character recognition and mail sorting based on the recognized result. 11. A scanner for converting a handwritten character string described in a form into gray-scale image data, a memory for holding a continuous portion of black pixels of the gray-scale image and cut out as a connected component, When the character pattern candidate including the connected component is a contact character, a character corresponding to a new character pattern generated by cutting the connected component having the maximum length in the line direction of the connected component in the character pattern And a processing unit for performing recognition.