JPH0528305A - Device and method for image recognition - Google Patents

Abstract

PURPOSE:To improve recognition ratio by extracting character strings as candidate character strings, evaluating the recognition result of the constituting characters of respective candidate character strings and feeding the evaluation result back to determine whether or not the character string constitution is proper. CONSTITUTION:An input means 1 reading an image 5, and a character string extraction part 2 extracting plural character strings as the candidate character strings when respective character elements are integrated in accordance with environmental conditions and when the character elements can be integrated into plural character strings, are provided. A character recognition part 3 segmenting the constituting characters of the candidate character strings, preparing a normalized image of the constituting characters corresponding to their tilt, extracting the features of the normalized image, checking the features with the features of the standard pattern of a dictionary and recognizing the constituting characters, is provided. A character string evaluation part 4 outputting the candidate character strings in which the recognition inavailable ratio of the plural candidate character strings is minimized, as the recognition result, is also provided.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention mixes characters and figures such as various design drawings and maps according to a format in which the direction of a character string and the bases of the characters constituting the character string are basically parallel. The present invention relates to an image recognition device that reads a document (hereinafter referred to as an image).

With the spread of CAD / CAM and computer mapping, there is automatic image input using an optical reader as a means for digitizing images such as design drawings and maps created in the past.

Along with this, it is important to read an image and correctly recognize the read image.

[0004]

2. Description of the Related Art FIG. 7 is a diagram showing a character string extraction unit of an image recognition apparatus. FIG. 8 is a diagram showing a character recognition unit of the image recognition device.

FIG. 11 is a diagram showing a conventional example. 12,
FIG. 13 is a flowchart of a conventional example. FIG. 14 is a diagram showing grouping of vectors.

FIG. 15 is a diagram showing an example of creating a character string.
In FIG. 11, an image in which characters and drawings are mixed is read by the scanner 201, converted into raster data, and stored in the frame memory 209 (S1 in FIG. 12).

At this time, the scanner controller 202 controls the scanner 201 and stores the raster data in the frame memory 209. Frame memory 20
For the raster data of No. 9, the raster = vector conversion unit (R = V conversion unit) 203 generates vector data by means of thinning the raster data, approximation of broken lines, etc., and outputs the result to the graphic information table 211. Stored in (Fig. 1
2 S2).

For each vector in the graphic information table 211, the character / graphic separation unit 204 groups mutually continuous vectors (S3 in FIG. 12), and the character has a relatively small area, and the other graphic / character Until the measurement of all groups is completed using the feature of being isolated from (Fig. 12
S4) Check the size of each group (S5 in FIG. 12),
If the size of the group matches the character size (S6 in FIG. 12), it is extracted as a character element (S7 in FIG. 12).
Extract groups other than character elements as graphic elements.
(S8 of FIG. 12).

The separated character elements are the character information table 2
Stored in 12. Therefore, as shown in FIG. 14, the groups 101 and 102 are character elements, group 10
3 is a graphic element.

As shown in FIG. 7, a character information table 212
With respect to the character element stored in (S10 of FIG. 13), the surrounding search unit 401 of the character string extraction unit 205 searches for other character elements around the focused character element (S11 of FIG. 13). ).

If a character element exists as a result of the search (S12 in FIG. 13), the distance calculation unit 40 calculates the distance between the center of gravity of the character element of interest and the center of gravity of any other searched character element.
4 calculates, the angle calculation unit 405 calculates the difference between the inclination of any other character element and the inclination when the character element of interest to any other character element is integrated, and the obtained distance and angle are calculated. The possibility of being integrated as a character string in the integrated evaluation unit 402 is represented as an evaluation value, and the smaller the distance and the angle, the higher the evaluation value.

However, the weight of the evaluation of the angle and the distance is higher than the weight of the distance. When a plurality of character elements are searched by the peripheral search unit 401 (S in FIG. 13).
13), the searched arbitrary character element is changed to another searched character element, and the above-mentioned processing is performed in the same manner.

The above processing is performed for all the character elements searched for the noted character element to obtain respective evaluation values. The determination unit 403 determines that the integration of character elements is 1
If there is only one, the character element is integrated (S14 in FIG. 13). If there are multiple character element integrations, the character element having the highest evaluation value from the obtained evaluation values is selected. The integration is basically extracted as a character string (Fig. 1
3 S15).

However, even if the evaluation value is the highest, if the evaluation value is lower than the predetermined level value, the character element of interest is extracted as independent, that is, it is not integrated anywhere.

That is, conventionally, the evaluation value that is equal to or higher than the level value and is maximum is extracted. This character string extraction processing is performed on all the character elements stored in the character information table 212 (S9 in FIG. 13).

For example, when there are a plurality of character strings such as a character string 105 and a character string 106 in the integrable range for the character element 104 of interest as shown in FIG. The distance calculation unit 404 measures the distances L1 and L2 from the center of gravity of the nearest characters of both character strings.

Further, the inclination of the character string 105 and the character string 105
The angle calculation unit 405 measures the difference between the inclination when the character element 104 is integrated with the character string 104 and the difference between the inclination of the character string 106 and the inclination when the character element 104 is integrated with the character string 106.

Then, in the integrated evaluation unit 402, the distance calculated by the distance calculation unit 404 and the angle calculation unit 405,
An evaluation value is obtained from each angle, and the determination unit 403 determines whether to integrate the character element 104 into the character string 105, the character string 106, or neither of them by the two evaluation values.

The result of the character string extraction is the character information table 2
Stored in 12. As shown in FIG. 8, the image cutout unit 50 of the character recognition unit 206 is based on the inclination information of the character string and the position information of each character stored in the character information table 212.
1 cuts out an image for one character from the frame memory 209 (S17 in FIG. 13).

In order to recognize the cut out character image, the characteristics of the character image are compared with the characteristics of the standard pattern in the dictionary 505. At that time, in order to match the direction of collation between the character feature and the feature of the standard pattern of the dictionary 505, a rotation process according to the inclination of the character string is applied to the clipped character image by the image rotation unit 502 shown in FIG. After the normalized image is created by the feature extracting unit 503, the feature extracting unit 503 extracts the feature, and the feature matching unit 504 matches the feature of the normalized image with the feature of the standard pattern of the dictionary 505 to obtain a character recognition result (FIG. 13). S18).

The recognition processing of the characters constituting the above character string is performed for all character strings (S16 in FIG. 13), and the obtained character recognition result is stored in the character information table 212 (S19 in FIG. 13).

On the other hand, with respect to the graphic elements of the graphic information table 212, graphic primitive recognition of straight lines, circles, arcs, etc. is performed in the vector operation unit 207 as follows. First, a vector that is a candidate for a circular arc is searched from the length of the vector and the angle between two consecutive vectors.

Next, from the searched vectors, candidates for the radius and center of a possible arc are obtained, and among the obtained candidates, the one consisting of the largest number of vectors within a given threshold is circled, This is the extraction result of the arc.

Vectors other than circles and arcs are set as straight line candidates, and among them, an array of vectors whose error is less than or equal to a threshold value and which has a certain length or more is set as a straight line. The result of the graphic primitive recognition is stored in the graphic information table 211 again.

The image recognition is completed as described above, and the recognition results of the graphic information table 211 and the character information table 212 are transferred to the host computer 21 via the communication control unit 208.
4 is output.

In the host computer 214, after the recognition result is displayed and confirmation / correction work is performed by the operator, the recognition result is stored in the CAD drawing database or the mapping database, and the information is created when the CAD system creates a drawing, for example. Used as.

[0027]

However, in the prior art, when the character elements are integrated and extracted as a character string, the character strings are integrated only from the surrounding situation even if they can be integrated into a plurality of character strings. There was a problem that the character string was erroneously extracted and it became unrecognizable.

That is, in an image such as a design drawing in which characters and figures are mixed, according to the format that the direction of the character string and the bases of the characters forming the character string are basically parallel,
In addition, the character string is written with an arbitrary inclination.

When recognizing each character forming the character string, first, when the character forming the character string is cut out and collated with the characteristics of the standard pattern of the dictionary, the inclination of the character string is aligned to match the direction of the collation. It is necessary to create a normalized image in which characters are rotated according to the above, and to extract the features of the normalized image.

Therefore, if an incorrect character string is extracted,
A normalization image is created by performing an incorrect rotation process on each character that makes up the character string, and the characteristics of the normalization image are compared with the characteristics of the standard pattern in the dictionary. Is not normally performed, and the extracted character string is likely to be unrecognizable.

Therefore, according to the present invention, when an image is read according to a format in which the direction of the character string and the bases of the characters forming the character string are basically parallel, and the character string is extracted,
When a character element can be integrated into multiple character strings, multiple candidate character strings are listed, the recognition result of each character in each candidate character string is evaluated, and feedback is given to the propriety of the character string configuration. An object of the present invention is to provide an image recognition device capable of reducing the extraction of character strings and improving the degree of recognition.

[0032]

FIG. 1 is a diagram showing the first principle of the present invention. In the figure, reference numeral 1 is an input means for reading an image 5 having at least characters written according to a predetermined format.

Reference numeral 2 denotes a character string extraction unit that, when individual characters are integrated and extracted as a character string, and if they can be integrated into a plurality of character strings, the plurality of integrable character strings are extracted as candidate character strings. To do.

A character recognition unit 3 cuts out and recognizes each character according to the inclination and the character position of the character string which are the constituent requirements of each candidate character string extracted by the character string extraction unit 2.

A character string evaluation unit 4 counts the number of unrecognizable characters of the character elements in each candidate character string recognized by the character recognition unit 3, and recognizes the unrecognizable rate (the total number of characters forming the candidate character string and its candidates). The candidate character string having the smallest ratio of the number of unrecognizable characters of the character string) is output as the recognition result.

FIG. 2 is a second principle diagram of the present invention. In the figure, 1 is an input unit, 2 is a character string extraction unit, and 3 is a character recognition unit, which are the same as those in FIG. 1 described above, so a detailed description thereof will be omitted.

A character string selecting unit 6 counts the number of unrecognizable characters of the character elements in each candidate character string recognized by the character recognizing unit 3 and recognizes all candidate character strings whose unrecognizable rate satisfies the setting criterion. Is output, and any one of them is selected.

[0038]

According to the first aspect of the present invention, when the character element can be integrated into a plurality of character strings at the time of extracting the character string in the character string extracting unit 2, the plurality of the character strings that can be integrated are candidate characters. Extract all as columns.

In order to make the constituent characters of the plurality of candidate character strings extracted by the character string extraction unit 2 cut out according to the inclination and character position of the character string, and to allow the character recognition unit 3 to collate with the characteristics of the standard pattern of the dictionary. In addition, after creating a normalized image obtained by performing rotation processing on the characters cut out according to the inclination of the character string and performing feature extraction, the features of the normalized image and the features of the standard pattern of the dictionary are collated, Perform recognition processing.

When a character string in an oblique or vertical direction is written according to a predetermined format, the individual characters forming the character string are inclined so that the direction of the character string is parallel to the base of each character (parallel to the base of the character. And the direction of the character is).

Therefore, in order to recognize each character forming the character string, the direction of the character string is regarded as the direction of the character, and the character is cut out in an inclined state. On the other hand, when matching the characteristics of the character with the characteristics of the standard pattern of the dictionary, it is necessary to match the directionality of the matching. Therefore, the direction of the character cut out in an inclined state matches the direction of the standard pattern of the dictionary. Normalization must be done to rotate the cut out characters.

Therefore, if the wrong character string is extracted, the direction of the character string does not match the direction of each character forming the character string. Therefore, the constituent characters of the extracted character string may be cut out, and the cut out characters may be rotated in order to match the direction of the character with the direction of the standard pattern. Is big.

However, in the present invention, when a character element can be integrated into a plurality of character strings, all of the plurality of character strings that can be integrated are extracted as candidate character strings, and the characters constituting the extracted candidate character strings are extracted. After recognition, the character string evaluation unit 4 counts the number of unrecognizable characters for each candidate character string,
Since the candidate character string having the smallest unrecognizable rate is output as the recognition result, there is a high possibility that the character string to be extracted is output.

Due to the above, it is possible to reduce erroneous recognition and unrecognizable character due to erroneous character string extraction,
It is possible to obtain a highly accurate recognition result. According to claim 2 of the present invention, after recognizing each character forming the candidate character string, the character string selecting unit 6 counts the number of unrecognizable characters of each candidate character string, and the unrecognizable rate satisfies the setting criterion. All columns are output as recognition results, and any one of them is selected.

Therefore, when the means of claim 2 is used, the recognition result can be selected from a plurality of candidate character strings including the recognition result output by claim 1, so that the recognition result with higher accuracy than that of claim 1 is obtained. Can be obtained.

[0046]

EXAMPLE First, a first example will be described. Figure 3
FIG. 3 is a diagram showing a first embodiment of the present invention.

The flowcharts are shown in FIGS. 4, 5 and 12. FIG. 6 is a diagram showing an example of extracting candidate character strings in this embodiment. FIG. 7 is a diagram showing a character string extraction unit of the image recognition device.

FIG. 8 is a diagram showing a character recognition unit of the image recognition apparatus. Since the image reading, vector data generation, vector grouping, and character element / graphic element separation are performed as in the conventional example, detailed description thereof is omitted (see FIG. 12).

As shown in FIG. 7, the character information table 212
For the character element stored in (S21 in FIG. 4), the peripheral search unit 401 of the character string extraction unit 205 searches for another character element around the target character element (S22 in FIG. 4). ).

If a character element exists as a result of the search (S23 in FIG. 4), the distance calculation unit 404 calculates the distance between the center of gravity of the character element of interest and the center of gravity of any other searched character element.
Is calculated by the angle calculation unit 405, and the angle calculation unit 405 calculates the difference between the inclination of any other character element and the inclination when the character element of interest to the other arbitrary character element is integrated, and from the obtained distance and angle. The possibility of being integrated as a character string in the integrated evaluation unit 402 is represented as an evaluation value, and the smaller the distance and the angle, the higher the evaluation value.

However, the weight of the evaluation of the angle and the distance is higher than the weight of the distance. When a plurality of character elements are searched by the peripheral search unit 401, the searched arbitrary character element is changed to another searched character element, and the above-described processing is similarly performed.

The above processing is performed for all the character elements searched for the noted character element to obtain respective evaluation values. The determination unit 403 sets a threshold value, and from the evaluation values obtained by the integrated evaluation unit 402, the integration of all character elements having an evaluation value larger than the threshold value is extracted as a candidate character string.

Therefore, there is a case where a certain character element can be integrated into a plurality of character strings. This character string extraction processing is performed for all the character elements of the character information table 212 (S20 of FIG. 4).

When the character elements can be integrated into a plurality of character strings as shown in FIG. 6 (S24 in FIG. 4), the plurality of character strings are regarded as candidate character strings (T41, T42, T43 in FIG. 6A). (B) TTT, 444, 123) and extracted (FIG. 4).
S25), the constituent characters of each candidate character string and other candidate character strings sharing the character are added to each candidate character string as link information.

When a character element can be integrated into only one character string, it is extracted as an independent character string (a character string in which the constituent characters of one character string do not become the constituent characters of another character string) (see FIG. 4). S26).

The result of the character string extraction and the link information are stored in the character information table 212. Character information table 2
The image cutout unit 501 of the character recognition unit 206 cuts out an image of one character from the frame memory 209 based on the inclination information of the independent character string or the candidate character string stored in 12 and the position information of each character (S28 in FIG. 4). .

In order to recognize the cut out character image, the characteristics of the character image are compared with the characteristics of the standard pattern in the dictionary 505. At that time, in order to match the direction of collation between the character feature and the feature of the standard pattern of the dictionary 505, the image rotation unit 502 shown in FIG. 8 cuts out rotation processing according to the inclination of the independent character string or the candidate character string. The normalized image is created by applying it to each character image, the features are extracted by the feature extraction unit 503, and then the features of the normalized image and the dictionary 50 are extracted.
The characteristics of the standard pattern of No. 5 are collated by the characteristic collating unit 504 to obtain a character recognition result (S29 in FIG. 4).

This recognition processing is performed for all the independent character strings and candidate character strings, and the obtained character recognition result is stored in the character information table 212 (S27 in FIG. 4). When a character group having a plurality of candidate character strings (FIGS. 6 (a) and 6 (b), the arrow indicates the direction of the character string) is present, the character string evaluation unit 301 determines whether or not a plurality of candidate character strings of the character group are present. For each, the recognition result is evaluated.

Specifically, the number of unrecognizable characters of the characters forming the candidate character string is counted (S31 of FIG. 4), and the number of unrecognizable characters of all candidate character strings based on the link information is counted (see FIG. 4). S30), the candidate character string having the smallest unrecognizable rate is output as a correct answer, and the other candidate character strings in the character information table 212 are deleted (S32 in FIG. 5).

Since the character string direction of the candidate character string in FIG. 6A is horizontal, it is not necessary to perform rotation processing on each character forming the candidate character string for normalization. Then, the feature of each character that constitutes the candidate character string is extracted, and the dictionary 505 is extracted.
When compared with the features of the standard pattern of, each character is correctly recognized because it is in the positive direction.

On the other hand, since the character string direction of the candidate character string in FIG. 6B is vertical, a normalized image is created by rotating each character forming the candidate character string by 90 degrees. However, the features are extracted from the normalized image, and the dictionary 5
Even if the characters of the standard pattern of No. 05 are matched, each character forming the candidate character string is not in the normal direction, and thus cannot be recognized or is erroneously recognized.

Therefore, comparing the unrecognizable rates of the candidate character strings shown in FIGS. 6A and 6B, the unrecognizable rate of the candidate character strings shown in FIG. 6A is lower. The candidate character string shown in is output as a recognition result, and (b)
The candidate character string shown in is deleted.

Next, the second embodiment will be described. Figure 9
FIG. 6 is a diagram showing a second embodiment of the present invention. 4 and 1
0, the flowchart is shown in FIG.

The second embodiment differs from the first embodiment in that the character string evaluation unit 301 is replaced by a character string selection unit 302. Incidentally, the description of the same processing as that of the first embodiment will be omitted (see FIGS. 12 and 4).

In the character string selection unit 302, the number of unrecognizable characters of each candidate character string is counted (S31 in FIG. 4), and after counting (S30 in FIG. 4), among the candidate character strings in the character information table 212, Setting criteria (for example, the unrecognizable rate is 1
/ 3 or less) is deleted, and the candidate character string satisfying the setting criterion is sent to the host computer 214 via the communication control unit 208, and the display (DP) is controlled under the control of the host computer 214. 21
5 is displayed (S33 in FIG. 10).

The man-machine interface (MMI) 216 connected to the host computer 214 allows the operator to select one of the candidate character strings displayed on the display 215, but the displayed candidate character string should be recognized. If the operator determines that the image is different from the image, another candidate character string is displayed and selected based on the link information (S34 in FIG. 10).

On the other hand, with respect to the graphic elements of the graphic information table 211 in the first and second embodiments, in the vector operation unit 207, graphic primitives such as straight lines, circles and arcs are recognized as in the conventional case, and the result is obtained. Are again stored in the graphic information table 211.

The image recognition is completed as described above, and the recognition results of the graphic information table 211 and the character information table 212 are passed to the host computer 214 such as a workstation having a CAD program or a mapping program via the communication control unit 208. Is output.

In the host computer 214, after the recognition result is displayed and confirmation / correction work is performed by the operator, the recognition result is stored in the CAD drawing database or the mapping database, and information is created when the CAD system creates a drawing, for example. Used as.

[0070]

As described above, according to the present invention,
It is possible to provide a highly accurate character recognition function in an image recognition device that reads an image according to a format in which the direction of a character string and the bases of the characters that make up the character string are basically parallel. It greatly contributes to the efficiency of data base creation such as drawings and maps.

[Brief description of drawings]

FIG. 1 is a first principle diagram of the present invention.

FIG. 2 is a second principle diagram of the present invention.

FIG. 3 is a first embodiment of the present invention.

FIG. 4 is a flowchart of the first and second embodiments of the present invention.

FIG. 5 is a flowchart of the first embodiment of the present invention.

FIG. 6 is a diagram showing an example of candidate character string extraction according to the present embodiment.

FIG. 7 is a diagram showing a character string extraction unit of the image recognition device.

FIG. 8 is a diagram showing a character recognition unit of the image recognition device.

FIG. 9 is a second embodiment of the present invention.

FIG. 10 is a flow chart of a second embodiment of the present invention.

FIG. 11 is a diagram showing a conventional example.

FIG. 12 is a flowchart (part 1) of a conventional example.

FIG. 13 is a flowchart (part 2) of a conventional example.

FIG. 14 is a diagram showing grouping of vectors.

FIG. 15 is a diagram showing an example of creating a character string.

[Explanation of symbols]

1 Input means 2 Character string extractor 3 Character recognition part 4 Character string evaluation section 5 images 6 Character string selection section

Claims (4)

[Claims] 1. An image recognition device for recognizing a character by reading an image (5) having a character written according to at least a predetermined format and comparing the feature extracted from the character with the feature of a standard pattern in a dictionary. In the case where the characters can be integrated into a plurality of character strings when the character is integrated by the input means (1) for reading at least the character in the image (5) and extracted as a character string according to ambient conditions. A character string extraction unit (2) for extracting the plurality of character strings as a candidate character string, cutting out the constituent characters of the candidate character string, and creating a normalized image of the constituent characters according to the inclination of the candidate character string, A character recognition unit (3) for recognizing the constituent characters by matching the characteristics of the normalized image with the characteristics of the standard pattern of the dictionary, counting the number of unrecognizable characters of the plurality of candidate character strings, and calculating the plurality of candidates. String of Image recognition apparatus characterized by having a character string evaluation unit 識不 efficiency is output as the recognition result candidate character string is a minimum (4). 2. An image recognition apparatus for recognizing a character by reading an image (5) having a character written according to at least a predetermined format, comparing the feature extracted from the character with the feature of a standard pattern of a dictionary. In the case where the characters can be integrated into a plurality of character strings when the character is integrated by the input means (1) for reading at least the character in the image (5) and extracted as a character string according to ambient conditions. A character string extraction unit (2) for extracting the plurality of character strings as a candidate character string, cutting out the constituent characters of the candidate character string, and creating a normalized image of the constituent characters according to the inclination of the candidate character string, A character recognition unit (3) for recognizing the constituent characters by matching the characteristics of the normalized image with the characteristics of the standard pattern of the dictionary, counting the number of unrecognizable characters of the plurality of candidate character strings, and calculating the plurality of candidates. String of Output as all recognition result candidate character string 識不 efficiency meets the set criteria, the image recognition apparatus characterized by having character string selection unit for selecting one of said recognition result (6). 3. A recognition method for recognizing a character by reading an image having at least a character written according to a predetermined format, comparing the feature extracted from the character with the feature of a standard pattern of a dictionary, and recognizing the character. Of at least the above characters are read, and when the individual characters are integrated according to ambient conditions and extracted as a character string, when the characters can be integrated into a plurality of character strings, the plurality of character strings are extracted as candidate character strings, The constituent characters of the candidate character string are cut out, a normalized image of the constituent characters is created according to the inclination of the candidate character string, and the characteristics of the normalized image and the characteristics of the standard pattern of the dictionary are compared to perform the constituent characters. Is recognized, the number of unrecognizable characters of the plurality of candidate character strings is counted, and the candidate character string having the smallest unrecognizable rate of the plurality of candidate character strings is output as a recognition result. Method. 4. A recognition method for recognizing a character by reading an image having at least a character written according to a predetermined format, comparing the feature extracted from the character with the feature of a standard pattern in a dictionary, and recognizing the character. Of at least the above characters are read, and when the individual characters are integrated according to ambient conditions and extracted as a character string, when the characters can be integrated into a plurality of character strings, the plurality of character strings are extracted as candidate character strings, The constituent characters of the candidate character string are cut out, a normalized image of the constituent characters is created according to the inclination of the candidate character string, and the characteristics of the normalized image and the characteristics of the standard pattern of the dictionary are compared to perform the constituent characters. Is recognized, the number of unrecognizable characters of the plurality of candidate character strings is counted, all candidate character strings whose unrecognizable rate of the plurality of candidate character strings satisfy a setting criterion are output as a recognition result, and the recognition result is obtained. Recognition method for causing selected one of.