JPH06243294A - Character recognition postprocessing device - Google Patents

Abstract

PURPOSE:To properly perform the post processing of character recognition. CONSTITUTION:A document image 2 is segmented into individual character images by a character segmenting part 3, and a character code is outputted from a character recognition part 4 by pattern recognition. Meanwhile, a standard character space between characters in the document image 2 is calculated in a standard character space calculating part 51 by segmenting information obtained by the character segmenting part 3. A character space calculating part 52 divides each character space by this standard character space to calculate a normalized character space. A character space collating part 54 compares this normalized character space with a prescribed value stored in character segmenting knowledge data 53; and if it is smaller than the prescribed value, adjacent characters are regarded as one character. A language knowledge collating part 6 substitutes a part, which is discriminated as segmenting error by a character segmenting error detection part 5, out of the character code inputted from the character recognition part 4 with a character code in character segmenting knowledge data 53 inputted through the character space collating part 54. This substituted character code is used to perform collation with a word dictionary 7.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition post-processing device for automatically correcting a character recognition result.

[0002]

2. Description of the Related Art Conventionally, there is a character recognition device which cuts out a character from an optically read document image and recognizes the cut out character. In such a character recognition device, the recognized character is taken out word by word and compared with a word in a word dictionary to detect and correct an error in character recognition. Such processing is called character recognition post-processing. FIG. 2 is a block diagram of an example of a conventional character recognition post-processing device.

The character recognition post-processing device 21 shown in FIG.
The character code cut out by the character cutout unit 23 from the optically read document image 22 and recognized by the character recognition unit 24 is input. The character cutout unit 23 cuts out, for example, a character line as shown in FIG. 3 from the document image 22, and then the characters “′”, “a”,
Cut out "b" etc. The character recognition unit 24 obtains the degree of coincidence of each of the cut out characters “′”, “a”, “b”, etc. with the candidate character by pattern recognition, and determines the degree of coincidence from the first candidate to the nth candidate. , Output the character code.

The character recognition post-processing device 21 comprises a language knowledge collating unit 25 and a word dictionary 26. Linguistic knowledge collation unit 2
Reference numeral 5 divides the recognized character code array into words and compares them with candidate words in the word dictionary 26 for comparison. As for the delimiter for each word, for example, in the case of an English document, a method of detecting a space between words is known. As a specific example of a method of matching with a candidate word, the word having the largest number of matching characters with the reference word in which the characters output as the first candidate from the character recognition unit 24 are arranged is extracted from the word dictionary 26, and each character of both words is extracted. There is a method of outputting a candidate word having the minimum cost value, which is the total value of the degree of coincidence of (for example, Japanese Patent Publication No. 61-2003
No. 8).

[0005]

However, the above-mentioned conventional technique has the following problems. That is, the characters cut out by the character cutting unit 23 are recognized by the character recognition unit 24 as a character code, and the arrangement of the character code is used as a reference word to be matched with the candidate character in the word dictionary. If there was an error in the character segmentation of 23, the appropriate post-processing could not be performed. For example, as shown in FIG. 4, due to an error in cutting out a character "",
If it was cut out as "'", the error could not be detected.

Further, as shown in FIGS. 7 and 8, “m” is cut out as “r” or “n” due to faint characters, or the like.
In some cases, "w" may be cut out as "v" or "v", and in this case, the language knowledge collation unit 25 cannot perform correct word collation. For example, when the word “memory” is present in the document image 22 and is cut out by the character cutout unit 23 as “rnememory”, the language knowledge collation unit 2
At 5 the closest candidate word is retrieved from the word dictionary 26 with this as the reference word. Therefore, when trying to correct an error in character extraction in post-processing, it is necessary to perform processing such as matching with the word dictionary assuming all characters are in error in character extraction, and the time required for post-processing is extremely long. Sometimes it became very long, and it was sometimes corrected to a word completely different from "memory." The present invention has been made in view of the above points, and an object of the present invention is to provide a character recognition post-processing device capable of detecting a character cutting error and performing appropriate word matching.

[0007]

A character recognition post-processing apparatus of the present invention detects a character cutout error based on each character cut out from a document image and the coordinates of each character and selects a candidate character. It is characterized by comprising a character cutout error detection unit for outputting and a language knowledge collation unit for collating with the language knowledge by including the candidate characters output by the character cutout error detection unit in the character recognition result. .

[0008]

In the character recognition post-processing apparatus of the present invention, the optically read document image is cut out into each character image by the character cutting section. In addition, the character recognition section outputs a character code by pattern recognition. On the other hand, the standard character spacing calculation section of the character segmentation error detection section calculates the standard character spacing between the characters in the document image based on the segmentation information obtained by the character segmentation section. In addition, the character spacing calculation unit divides each character spacing by this standard character spacing to calculate a normalized character spacing. Then, the character interval collating unit compares the normalized character interval with a predetermined value stored in the character cut-out knowledge data, and if the value is smaller than the predetermined value, the adjacent character is regarded as one character. After that, in the linguistic knowledge collating unit, of the character codes input from the character recognizing unit, the portion determined to be the cutting error by the character cutting error detecting unit is input from the character interval checking unit. In this case, the portion which is determined to be the cutout error may be replaced with the character code in the character cutout knowledge data. After that, the word dictionary is collated.

[0009]

Embodiments of the present invention will now be described in detail with reference to the drawings. FIG. 1 is a block diagram of an embodiment of the character recognition post-processing device of the present invention. In the character recognition post-processing device 1 shown in FIG. 1, the characters cut out from the document image 2 by the character cutout unit 3 are converted into character codes by the character recognition unit 4 and input. The character cutout information of is input. In the character cutout unit 3, a character line as shown in FIG. 3 is cut out into each character “′”, “a”, “b” and the like. To cut out a character, for example, a method of detecting a position where all pixels in the Y direction are white in each X coordinate of a character line can be used. For example, in FIG. 3, at each X coordinate of X1 to X10, coordinates Y1 and Y2
Since the pixels between are all white, these coordinates X1 ...
The character line is cut out into characters "'", "a", "b", etc. at X10.

The cutout result of the portion of FIG. 3 is shown by the broken line in FIG. That is, "'" is cut out between the coordinates X5 and X6. Further, although "" is one character, since all the pixels in the Y direction are white at the coordinate X7, "'" is cut out between the coordinates X6 and X7. Then, "'" is cut out even between the coordinates X7 and X8. Therefore, the recognition result of the portion of FIG. 3 in the character recognition unit 4 becomes "'", "'", "'". The character recognition post-processing device 1 includes a character cutout error detection unit 5 and a language knowledge collation unit 6
Etc. The character cutout error detection unit 5 detects the character cutout information from the character cutout unit 3, for example, the number indicating the number of the cutout character from the beginning of the document, the coordinates of the cutout position, and the character recognition unit 4. The character code and are input, and based on these, the character cutting error is detected and the candidate character is output. On the other hand, the language knowledge collation unit 6 includes the candidate characters output by the character cutout error detection unit 5 in the character recognition result of the character recognition unit 4 and verifies the language knowledge, for example, a candidate word in the word dictionary 7.

The character cutout error detection section 5 comprises a standard character spacing calculation section 51, a character spacing calculation section 52, character cutout knowledge data 53, and a character spacing verification section 54. The standard character interval calculation unit 51 obtains the number of pixels between each character based on the character cutout information from the character cutout unit 3, for example, the X coordinate and the Y coordinate of the cut out character, and determines this as the character interval. To do. For example, as shown in FIG. 4, the character spacing between the “′” cut out at the coordinate X6 and the “′” cut out at the coordinate X7 is
There are 75 pixels. Also, "'" cut out at coordinate X7,
The character interval of "'" cut out at the coordinate X8 is 6 pixels. Such character spacing is calculated by the standard character spacing calculation unit 51 when each character is cut out. When all the characters have been cut out and all the character intervals have been calculated, all the calculated character intervals are summed, and the total value is divided by the number of character intervals (= number of characters-1) to obtain the average character. Find the interval. The standard character spacing calculator 51 uses this average character spacing as the standard character spacing in the document image 2, for example. In addition, for example, the average of the character spacing in the first several lines of the document may be set as the standard character spacing.

The character interval collating unit 54 detects an error in character recognition from the character code which is the character recognition result of the character recognizing unit 4 and the character interval calculated from the character cutting information of the character cutting unit 3. First, the average character spacing determined as described above is used as a standard character spacing, and this is compared with each character spacing. For this reason, the character spacing calculation unit 52 divides each character spacing by the standard character spacing to obtain a normalized character spacing for convenience.
For example, assuming that the standard character spacing is 15 pixels,
As shown in, the normalized character interval is 5 when the character interval is 75 pixels, and the normalized character interval is 0.4 when the character interval is 6 pixels. Then, by comparing this normalized character spacing with a predetermined value (for example, 0.5) shown in FIG. 6, it is determined whether or not the character cut out as two characters is one character. As a result, for example, what is cut out as two characters of "'" and "'" in FIG. 3 is determined as an error of one character of "". In this way, it is determined whether there is a cutout error based on the normalized character spacing, which is the ratio to the standard character spacing, instead of determining whether there is one character based on the character spacing itself. It may be narrow or wide.

Further, in order to perform more accurate error detection,
As shown in FIG. 6, the predetermined value used to determine the normalized character spacing is changed according to the adjacent character. Therefore, the following character cutting error knowledge data is configured. As shown in FIG. 6, the determination value of the normalized character spacing when the adjacent characters are “′” and “′” is 0.5. That is, when the adjacent characters are "'" and "'", the normalized character spacing is 0.
When it is less than 5, these are not two characters but one character """as a candidate character. In addition, the following data is stored in the character cut-out knowledge dictionary shown in FIG. For example, the “m” in the part of FIG. 3, the “w” in the part, and the “x” in the part are each one character, but may be recognized as two characters. Such erroneous recognition is caused by generation of a white line in the Y direction on a character due to a blur of ink or the like on a document or an abnormality in the reading unit such as a defect in one pixel of the line sensor.

For example, as shown in FIG. 7, when a white line is entered in the central part of "m", the character cutting section 3 separates "r" and "n" into two characters. Also, FIG.
As shown in, when there is a white line in the center of "w",
It is cut out as two characters divided into "v" and "v". Then, as shown in FIG. 9, when a white line is entered at the center of "x", it is divided into ">" and "<" and cut out as two characters. Therefore, in the character cut-out error knowledge data, if the adjacent characters are "r" and "n", the judgment value of the normalized character interval is 0.3.
"M" is mentioned as a candidate character when it is smaller. In addition, when adjacent characters are “v” and “v” and the determination value of the normalized character spacing is smaller than 0.25, “w” is cited as a candidate character. Then, when adjacent characters are ">" and "<" and the determination value of the normalized character spacing is smaller than 0.3, "x" is cited as a candidate character.

When the character cutout error detection unit 5 detects an error in the character cutout by the character cutout unit 3, the two characters cut out according to the character cutout error knowledge data shown in FIG. 6 are corrected into one character candidate character. And passed to the language knowledge collation unit 6. The language knowledge collating unit 6 collates the character code received from the character recognizing unit 4 with a word in the word dictionary 7, and if a word with a sufficiently high degree of coincidence cannot be obtained, the character cutting error detecting unit 5 The obtained corrected candidate characters are used for matching with the words in the word dictionary 7 for post-processing, and the result is output from the output unit 8.

Next, the operation of the above character recognition post-processing device will be described. It is assumed that the document image 2 shown in FIG. 1 has the character lines shown in FIG. In this case, when the character cutout unit 3 cuts out the character "'" at the coordinate X1, the character cutout unit 3 inputs the coordinates at this time to the character cutout error detection unit 5. At the same time, the character cutout unit 3 inputs the cutout character image of “′” to the character recognition unit 4. The character recognition unit 4 finds the closest matching pattern from the character image of "'" by pattern matching and outputs the character code corresponding to this. In this case, the character code "'" is usually output.

On the other hand, the standard character spacing calculator 51 of the character clipping error detector 5 detects and stores the spacing between the character "'" and the character "a" to be clipped next by the character clipping unit 3. To do. Also, the number of characters is counted. At this time, the count value is "1". Then, when the next character "b" is cut out, the interval between the character "a" and the character "b" is detected and stored, and the count value is counted up to "2".
After detecting and storing all the character intervals, all the stored character intervals are summed, and the total value is divided by "the number of characters-1" to calculate the standard character interval.

It is assumed that "m" shown in the part of FIG. 3 is cut at the coordinate X8 'as shown in FIG. Then, the character recognition unit 4 recognizes that one character “m” is two characters “r” and “n”. On the other hand, in the character cutout error detection unit 5, the normalized character spacing is calculated from the standard character spacing “15” calculated by the standard character spacing calculation unit 51 and the distance “4 pixels” between the characters “r” and “n”. calculate. That is, the character spacing calculation unit 52 calculates “4 ÷ 15 = 0.26” to calculate the normalized character spacing “0.26”. The result is shown in FIG. Then, the character spacing collating unit 54 refers to the character cut-out knowledge data 53. That is, when the character spacing verification unit 54 determines the normalized character spacing “0.26” calculated by the character spacing calculation unit 52 as the adjacent characters in the character cut-out knowledge data 53 shown in FIG. 6 being “r” and “n”. And the normalized character spacing "0.3". Since the former is smaller than the latter, the character cutout by the character cutout unit 3 is an error. The result is shown in FIG. Then, “m” is input to the language knowledge collation unit 6 as a candidate character, and the language knowledge collation unit 6 sends “r” and “n” sent from the character recognition unit 4.
Is replaced with “m” sent from the character cutout error detection unit 5.

The replacement character in this case does not have to be sent from the character segmentation error detection unit 5. That is, the linguistic knowledge collation unit 6 may receive only the information indicating that the number and the number of the character are cut out and the linguistic knowledge collation unit 6 may perform the process corresponding to this information. For example, the linguistic knowledge collation unit 6 may have the knowledge of “r” + “n” = “m”, and the knowledge may be replaced by this knowledge, or a candidate character does not appear for a character with a clipping error. You may make it perform a word collation as a thing. As a result, for example, the word “memory” becomes the document image 22.
It is inside, and this is the "rnemory" in the character cutting part 3
Even if the language knowledge collation unit 6 cannot extract a word with a sufficiently high degree of matching of 7 characters from the word dictionary 7, the character segmentation error detection unit 5 detects the segmentation error and outputs “m”.
The language knowledge collation unit 6 can retrieve the closest candidate word “memory” of 6 characters from the word dictionary 7 by using the corrected character code “memory”.

Further, it is assumed that the "w" shown in the part of FIG. 3 is cut at the coordinate X9 'as shown in FIG. Then, the character recognition unit 4 converts one character “w” into two characters “v”,
Recognized as "v". On the other hand, the character segmentation error detection unit 5 determines the normalized character spacing from the standard character spacing “15” calculated by the standard character spacing calculation unit 51 and the distance “3 pixels” between the characters “v” and “v”. calculate. That is,
The character spacing calculator 52 calculates "3/15 = 0.2" to calculate the normalized character spacing "0.2". The result is shown in FIG. Then, the character spacing collating unit 54 refers to the character cut-out knowledge data 53. That is, when the character spacing verification unit 54 determines the normalized character spacing “0.2” calculated by the character spacing calculation unit 52 to be “v” and “v” as the adjacent characters in the character cutting knowledge data 53 shown in FIG. It is compared with the normalized character spacing "0.25". Since the former is smaller than the latter, the character cutout by the character cutout unit 3 is an error. The result is shown in FIG. Then, “w” is input to the language knowledge collation unit 6 as a candidate character, and the language knowledge collation unit 6 sends “v”, which is sent from the character recognition unit 4.
“V” is replaced with “w” sent from the character cutting error detection unit 5.
Replace with.

Further, it is assumed that the "x" shown in the part of FIG. 3 is cut at the coordinate X10 'as shown in FIG. Then, the character recognizing unit 4 converts one character “x” into two characters “>”,
Recognized as "<". On the other hand, in the character cutout error detection unit 5, the normalized character spacing is calculated from the standard character spacing “15” calculated by the standard character spacing calculation unit 51 and the distance “4 pixels” between the characters “>” and “<”. calculate. That is,
The character spacing calculator 52 calculates "3/15 = 0.2" to calculate the normalized character spacing "0.2". The result is shown in FIG. Then, the character spacing collating unit 54 refers to the character cut-out knowledge data 53. That is, when the character spacing verification unit 54 determines the normalized character spacing “0.26” calculated by the character spacing calculation unit 52 as the adjacent characters in the character cut-out knowledge data 53 shown in FIG. 6, “>” and “<”. And the normalized character spacing "0.3". Since the former is smaller than the latter, the character cutout by the character cutout unit 3 is an error. The result is shown in FIG. Then, “x” is input to the language knowledge collation unit 6 as a candidate character, and the language knowledge collation unit 6 sends “>”, which is sent from the character recognition unit 4.
The “<” is replaced with the “x” sent from the character cutout error detection unit 5.
Replace with.

As described above, "'" and "'",
When the character spacings such as "r" and "n", "v" and "v", and ">" and "<" are closer than the standard character spacing, the character segmentation of the character segmentation unit 3 is performed. Is detected and replaced with "", "m", "w", "x", etc., respectively. Therefore, the language knowledge collation unit 6 can collate with the candidate word in the word dictionary 7 by using an appropriate candidate character, and while including the function of correcting the character segmentation error,
Post-processing of character recognition can be appropriately performed.

In the above embodiment, the character spacing between two adjacent characters is compared with a predetermined value, but the present invention is not limited to this.
For example, a clipping error may be detected by making a relative comparison with the character spacing before and after. That is, as shown in FIG. 4, when the character interval at the coordinate X7 is 6 pixels, this is changed to the character interval "7" at the coordinate X6 immediately before.
Since the character spacing at the coordinate X7 is relatively narrower than that of "5 pixels", a clipping error of """may be detected. In addition, in the above-described embodiment, one character becomes two.
Although the case of detecting a clipping error of a character has been described, the present invention is not limited to this and can be applied to a case of detecting a clipping error of one character being three or more characters. Further, in the above-described embodiment, the case of recognizing a European sentence has been described, but the present invention is not limited to this, and can be applied to a case of recognizing a Japanese sentence.

[0024]

As described above, according to the character recognition post-processing device of the present invention, from the cutout information such as the coordinates of the cutout position output by the character cutout unit and the character code output by the character recognition unit, Since the character spacing is calculated and a plurality of characters that are close to each other are regarded as one character to be collated with the candidate character in the word dictionary, it is possible to perform an appropriate character recognition post-processing including a correction function of a character cutting error. it can. Further, by detecting the cut-out error by comparing each character interval with a predetermined value, a simple and accurate error detection can be performed. Furthermore, by calculating the standard character spacing in the document and calculating the normalized character spacing, which is the ratio, and preparing a predetermined value corresponding to this and comparing it with the calculated normalized character spacing. Even when the standard character spacing differs depending on the document, the same character segmentation error knowledge data can be used to enable the same error detection.

[Brief description of drawings]

FIG. 1 is a block diagram of an embodiment of a character recognition post-processing device of the present invention.

FIG. 2 is a block diagram of an example of a conventional character recognition post-processing device.

FIG. 3 is an explanatory diagram of a character cutout procedure.

FIG. 4 is an explanatory diagram of an example (part 1) of character cutting processing.

FIG. 5 is an explanatory diagram of a character cutout error detection example (No. 1).

FIG. 6 is an explanatory diagram of an example of character cutting error knowledge data.

FIG. 7 is an explanatory diagram of a character cutout processing example (No. 2).

FIG. 8 is an explanatory diagram of a character cutting process example (3).

FIG. 9 is an explanatory diagram of a character cutout processing example (No. 4).

FIG. 10 is an explanatory diagram of a character cutting error detection example (No. 2).

FIG. 11 is an explanatory diagram of a character cutting error detection example (No. 3).

FIG. 12 is an explanatory diagram of a character segmentation error detection example (No. 4).

[Explanation of symbols]

 1 Character recognition post-processing device 5 Character cutout error detection unit 6 Language knowledge collation unit 7 Word dictionary 51 Standard character spacing calculation unit 52 Character spacing calculation unit 53 Character cutout knowledge data 54 Character spacing verification unit

Claims (3)

[Claims] 1. A character cutout error detection unit that detects a character cutout error based on the coordinates of each character cut out from a document image and outputs the detection position, and a character cutout error detection unit that outputs the detected position. A character recognition post-processing device comprising a linguistic knowledge collating unit that corrects a character at a detection position and collates it with linguistic knowledge. 2. The character cut-out error detection unit outputs the adjacent plural characters as one character when the interval between the adjacent characters cut out from the document image is smaller than a predetermined value. Character recognition post-processing device. 3. A standard character interval is calculated from an interval between adjacent characters cut out from a document image, the interval between adjacent characters is normalized by the standard character interval, and the normalized character interval is a predetermined value. The character recognition post-processing device according to claim 1, wherein when the difference is smaller than the character extraction error detection unit, the adjacent plural characters are output as one character.