JPH07262320A - Address recognition device - Google Patents

Abstract

PURPOSE:To improve a recognition rate by estimating an error part from a properly recognized character even when there is any character segmentation error or character recognition error concerning the address recognizing device for recognizing an address from an address image. CONSTITUTION:Plural segmenting candidates are outputted by a character segmenting part 11, a key character extracting part 13 extracts a key character from candidate characters which are provided by recognizing characters concerning the respective candidates with a character recognition part 12, and a place name area candidate retrieving part 14 infers a place name area based on the key character. A place name candidate evaluated value arithmetic part 16 calculates a place name candidate evaluated value corresponding to the number of matched characters between the place name candidates and a place name dictionary 18 or the like, and a place name candidate selecting part 17 selects a correct place name.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an address recognition device for reading and recognizing an address shown on a delivery or the like.

[0002]

2. Description of the Related Art Conventionally, in a character recognition device for recognizing an address character string, knowledge processing using a place name dictionary has been introduced to improve the recognition accuracy. This knowledge processing is a method of correcting the recognition result into the most probable character by collating the recognition result for each character with the place name dictionary.

A conventional address recognition device is disclosed, for example, in Japanese Patent Laid-Open No. 4-111186. FIG. 6 shows a conventional address recognition device. The character cutout unit 61 cuts out an image area for one character from the input image, and the character recognition unit 62 outputs N candidate characters for each character from the cutout image. The place name search unit 63 uses the place name dictionary 64,
A combination of characters forming a place name is obtained from the set of candidate character strings, and a place name candidate evaluation value calculation unit 65 obtains a place name candidate evaluation value based on the similarity of the recognition result. The place name candidate selection unit 66 uses the place name hierarchy dictionary 67 to select place name candidates having higher place name candidate evaluation values in order from the higher level place names, and outputs the place name candidates as a recognition result.

By recognizing the address as described above, it is possible to correct a character that the character recognition unit 62 has erroneously recognized, and it is possible to improve the recognition.

[0005]

However, in the above-mentioned conventional address recognition device, the character slicing section 61 may erroneously clip a character when recognizing a handwritten character or the like. In this case, the character recognition unit 62 cannot recognize the correct character, and the place name search unit 63 cannot search the correct place name candidate. Also, if the character position is misaligned due to a clipping error, the place name dictionary will be
Since it cannot be compared with 64, the place name search unit 63 cannot search for a correct place name candidate.

Further, since the place name candidate selecting section 66 selects place name candidates in order from the higher level place names, if the upper level place names are erroneously recognized, all lower level place names are wrong. Had the problem of recognizing it.

The present invention solves such a conventional problem, and an object thereof is to increase the character recognition rate by selecting from the most likely correct place name candidates.

[0008]

In order to achieve the above object, the present invention cuts out an input image into regions for each character,
A character cutout unit that outputs a plurality of cutout candidates as a cutout region, a character recognition unit that outputs N candidate characters for each cutout candidate, and a key character extraction that extracts a predetermined key character from the set of candidate characters Section, a place name area candidate search section for searching a place name area candidate corresponding to a place name hierarchy based on the key character, and a place name for searching a place name candidate using a place name dictionary from a candidate character set included in the place name area candidate A configuration including a search unit, a place name candidate evaluation value calculation unit that calculates the probability of the place name candidate and outputs a place name candidate evaluation value, and a place name candidate selection unit that selects a place name candidate based on the place name candidate evaluation value Is.

[0009]

According to the present invention, with the above configuration, a plurality of cutout candidates are obtained in the character cutout unit, character recognition is performed on all cutout candidates, and a candidate character set of character recognition results using all cutout candidates is set. Select the most correct place name from. Further, with respect to the place name candidates searched by the place name search unit, the place name is selected from the place name candidates of the most reliable place name hierarchy level.

Therefore, even if the character cutout unit cannot determine the result of the character cutout, by searching for the correct answer from the cutout candidates, false recognition is reduced, and further,
By selecting the place name from the place name candidates of the most reliable place name hierarchy level, the search error is eliminated.

[0011]

EXAMPLES Examples of the first invention of the present invention will be described below. FIG. 1 shows the overall configuration of the address recognition device of this embodiment.

The character cutout unit 11 cuts out an area of each character from the input image and outputs cutout candidates. The character recognition unit 12 performs character recognition from a character image and outputs a candidate character set having n candidate characters from the first candidate character to the nth candidate character for each character. Key character extraction unit 13
Is a key character indicating the place name hierarchy, for example, "prefecture",
Characters such as "city" and "town" are extracted from the candidate character set.

The place name area candidate search unit 14 searches for a place name area candidate which is a place name area of each layer based on the key character. The place name search unit 15 searches the place name dictionary 18 for each place name area candidate, and selects a combination that is a place name candidate from the combinations of candidate character sets.
The place name candidate evaluation value calculation unit 16 calculates the place name candidate evaluation value based on the degree of coincidence between the place name candidate searched by the place name search unit 15 and the place name dictionary 18 and the recognition similarity of the character recognition unit 12. The place name candidate selecting unit 17 selects the place name candidate having the largest place name candidate evaluation value among the place name candidates, and outputs the address of the recognition result.

Address recognition is performed as follows in the address recognition device having the above configuration. First, the address image to be recognized is processed by the character cutout unit 11, and the image is cut out into areas for each character. When a plurality of candidates for the cutout position are considered, a plurality of possible cutout areas are output as cutout candidates. FIG. 2 shows an example of a plurality of cutout candidates. FIG. 2 shows the result of clipping the input image of the address "Matsuoka, Fuji-shi". a to k are cutout candidates of the processing result, and the range of the vertical line of each symbol is the cutout area.

Next, the character recognition unit 12 performs character recognition on the area image for each cutout candidate, and a candidate having n candidate characters from the first candidate character to the nth candidate character for each cutout candidate. Output the character set. FIG. 2 shows the result of character recognition from the first candidate character to the tenth candidate character for each cut-out candidate.

Further, in the key character extraction unit 13, a character (to be referred to as a key character) that serves as a delimiter of the place name hierarchy from the candidate character set, for example, "prefecture", "city", "road",
"Prefecture", "city", "county", "town", "ward", "village",
Extract "large letters", "letters", etc. At this time, the key character is extracted from all the characters from the first candidate character to the nth candidate character of the candidate character set.

In the place name area candidate search unit 14, the key character extraction unit
Key character and key character extracted in 13, character string start character,
The area sandwiched between the character string end character is searched as a place name area candidate. However, when searching the place name area, the order of the key characters sandwiching the place name area is restricted by the order of the hierarchical structure of the place name. For example, after the key character "gun", only the key characters "town" and "village" are connected. Further, the length of the character string in each place name area is set to a length determined in accordance with the type of key character. For example, the place name area having the key character "ken" as the last character has a length of 3 or 4 characters.

FIG. 3 shows a search example of the place name area. Figure 3
As a key character, "fu", "gun", at each character position
This is a search for a place name area when “city”, “town”, and “ward” are extracted. At this time, the searched place name area candidate is
"○○ Prefecture", "○○ City", "○○ Ward", "○○", "○"
"County", "○○○ town", "○○○○○ city", "○○ town", and "○○○ ward"("ward" is not connected to "county") or "○○○" ○ "County"("County" with a length of 5 characters or more)
Not searched. However, this example shows an example in which there are not a plurality of cutout candidates.

The place name search unit 15 matches the characters from the first candidate character to the mth candidate character (m ≦ n) of the candidate character set with the place name dictionary 18, and finds a place name that matches even one character with the place name dictionary 18. Search as a place name candidate. Place Name Dictionary 18
The number m of the candidate characters used for the matching with and must be determined in accordance with the calculation time because the larger the number m, the more the place name candidates and the larger the calculation amount. On the other hand, the number n of the candidate characters used for extracting the key character is important for determining the place name area candidate, and thus it is preferable to set a large value to some extent.

As described above, the character cutout unit 11 and the character recognition unit
12, by using the key character extraction unit 13, the place name area candidate search unit 14, and the place name search unit 15, even if the character cutout unit 11 cannot uniquely determine the cutout of the character, it is possible to perform the address recognition. Becomes Further, by inferring the place name area using the key characters, the candidates for the place name search can be narrowed down.

The place name candidate evaluation value calculating unit 16 calculates the most probable place name among the place name candidates by calculating the matching character number calculating unit 21, the adjacent character matching number calculating unit 22, the mismatching character position correspondence number calculating unit 23, and the mismatching character position. Based on the output of the character number calculation unit 24, the evaluation value of the place name candidate is calculated by (Equation 1).

(Equation 1) Evaluation value = a × (number of matching characters) + b ×
(Number of matching adjacent characters) + c x (number of corresponding non-matching characters) -d
× (number of non-matching characters) where a, b, c, d are constants, the number of matching characters is the number of matching characters in the place name candidate and the place name dictionary 18, and the number of adjacent character matches is that both adjacent characters of the place name candidate are in both characters. Place Name Dictionary 18
And the corresponding number of non-matching characters are the totals of the place name candidates and the characters that are not matched in the place name dictionary 18 at corresponding positions. The number of unmatched characters is the number of characters that are not matched in the place name dictionary 18.

By using the place name candidate evaluation value calculation unit 16 having the above configuration, even if there is a recognition error in the character recognition unit 12 or a cutout error in the character cutout unit 11, the whole is inferred from the correct answer part. It becomes possible to do.

The place name candidate selection unit 17 searches the place name candidate evaluation value comparison unit 25 for a candidate with the highest place name candidate evaluation value, and the place name validity evaluation unit 26 searches for the place name candidate before and after. Using the place name hierarchy dictionary 19, it is checked whether there is a contradiction with the place name candidates connected to. If there is a contradiction, the last searched place name candidate is rejected, and the place name candidate evaluation value comparison unit 25 searches again for the candidate with the highest place name candidate evaluation value. When the place names in all the character areas are selected by repeating this, the selected place name candidates are output as the recognized addresses.

By using the place name candidate selecting unit 17 having the above-mentioned configuration, the candidate that seems to be the most correct answer is decided, and the uncertain candidate is decided last, so that there is less room for error and the recognition rate is improved. To do.

The place name candidate evaluation value calculating unit 16 and the place name candidate selecting unit 17 may have a conventional structure.

Next, a second embodiment of the present invention will be described. FIG. 4 shows the overall configuration of the address recognition device of this embodiment. Character cutout unit 11, character recognition unit 12, key character extraction unit 13, place name area candidate search unit 14, place name search unit 1 in the figure
5. The place name candidate evaluation unit 16 is the same as that of the first embodiment of the invention.

The place name candidate selecting section 17 selects the place name candidate having the largest place name candidate evaluation value among the place name candidates and outputs the address of the recognition result. If there is no place name that seems to be the correct answer among the place name candidates, the place name area candidate is reset and the place name search unit 15 performs the process again.

Address recognition is performed as follows in the address recognition device having the above configuration. First, the address image to be recognized is processed by the character cutout unit 11, and the image is cut out into areas for each character. When a plurality of candidates for the cutout position are considered, a plurality of possible cutout areas are output as cutout candidates.

Next, the character recognition unit 12 performs character recognition on the area image for each cutout candidate, and a candidate having n candidate characters from the first candidate character to the nth candidate character for each cutout candidate. Output the character set.

Further, the key character extraction unit 13 extracts a key character serving as a delimiter of the place name hierarchy from the candidate character set. At this time, the key character is extracted from all the characters from the first candidate character to the nth candidate character of the candidate character set.

The place name area candidate retrieving unit 14 retrieves an area sandwiched between the key characters extracted by the key character extracting unit 13 and the key characters, the character string start character, and the character string end character as place name area candidates. However, when searching the place name area, the order of the key characters sandwiching the place name area is restricted by the order of the hierarchical structure of the place name. Further, the length of the character string in each place name area is set to a length determined in accordance with the type of key character.

In the place name search unit 15, the characters from the first candidate character to the mth candidate character (m ≦ n) of the candidate character set are matched with the place name dictionary 18, and a place name that matches even one character with the place name dictionary 18 is searched. Search as a place name candidate. Place Name Dictionary 18
The number m of the candidate characters used for the matching with and must be determined in accordance with the calculation time because the larger the number m, the more the place name candidates and the larger the calculation amount. On the other hand, the number n of the candidate characters used for extracting the key character is important for determining the place name area candidate, and thus it is preferable to set a large value to some extent.

The place name candidate evaluation value calculation unit 16 calculates the place name candidate evaluation value in the same manner as in the first embodiment in order to find the most probable place name among the place name candidates.

The place name candidate selection unit 17 searches the place name candidate evaluation value comparison unit 25 for a candidate having the highest place name candidate evaluation value from the place name candidates, and the place name validity evaluation unit 26 searches the place name candidate evaluation values forward and backward. Using the place name hierarchy dictionary 19, it is checked whether there is a contradiction with the place name candidates connected to. If there is a contradiction, the last searched place name candidate is rejected, and the place name candidate evaluation value comparison unit 25 searches again for the candidate with the highest place name candidate evaluation value. Repeating this, the place names of all the character areas are selected. Further, the reject determination unit 27 determines the rejection based on the certainty of the selected place name based on the place name candidate evaluation value.

If not rejected, the selected place name is output as the recognized address. On the other hand, when rejected, the key character inference unit 28 infers a key character that could not be recognized, sets a place name area candidate according to the key character inferred by the place name area candidate setting unit 29, and selects the place name area candidate. The place name search unit 15 is used again to reprocess.

FIG. 5 shows a processing example of the place name area candidate setting unit 29. In Figure 5, the first key character extraction, "city",
“Town” and “ward” are extracted, and “○” is selected as the place name area candidate.
In this example, "○○○ city", "○○ ward", "○○ town", and "○○" are searched. When the place name determined from these place name area candidates is rejected by the reject determination unit 27, the key character inference unit 28 infers the key character that cannot be extracted from the knowledge of the place name hierarchy or the already extracted key character. To do. In FIG. 5, it is inferred that "prefecture" is in the third letter. In the place name area candidate setting unit 29, based on the key characters extracted by the key character inference unit 28, new place name area candidates, "○○ prefecture", "○"
○ "City" is set.

[0038]

As described above, according to the present invention, even if the character segmentation unit cannot determine the result of character segmentation and a segmentation error occurs, a plurality of segmentation candidates are searched for the correct answer, thereby making it It is possible to determine the correct character from the cut-out candidate and reduce misrecognition.

Further, by estimating the place name area by the key character and selecting the place name from the place name candidates of the most reliable place name hierarchy level, the place name search error can be prevented, thus improving the character recognition rate. .

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of an address recognition device according to a first embodiment of the present invention.

FIG. 2 is a diagram showing outputs of a character cutout unit and a character recognition unit according to the first embodiment of this invention.

FIG. 3 is a diagram showing an output of a place name area candidate search unit according to the first embodiment of this invention.

FIG. 4 is a block diagram showing a configuration of an address recognition device according to a second embodiment of the present invention.

FIG. 5 is a diagram showing an output of a place name area candidate setting unit according to a second embodiment of the present invention.

FIG. 6 is a block diagram showing a configuration of a conventional address recognition device.

[Explanation of symbols]

 11 Character cutout unit 12 Character recognition unit 13 Key character extraction unit 14 Place name area candidate search unit 15 Place name search unit 16 Place name candidate evaluation value calculation unit 17 Place name candidate selection unit 18 Place name dictionary 19 Place name hierarchical dictionary 21 Matched character number calculation unit 22 Adjacent characters Matching number calculation unit 23 Non-matching character position correspondence number calculation unit 24 Non-matching character number calculation unit 25 Place name candidate evaluation value comparison unit 26 Place name validity evaluation unit 27 Rejection judgment unit 28 Key character reasoning unit 29 Place name area candidate setting unit

Front page continuation (72) Inventor Hidetsugu Maekawa 1006 Kadoma, Kadoma City, Osaka Prefecture Matsushita Electric Industrial Co., Ltd. (72) Inventor, Kazuhiro Kayashima 1006 Kadoma, Kadoma City, Osaka Matsushita Electric Industrial Co., Ltd.

Claims (4)

[Claims] 1. An input image is cut out into an area for each character,
A character cutout unit that outputs a plurality of cutout candidates as a cutout region, a character recognition unit that outputs N candidate characters for each cutout candidate, and a key character extraction that extracts a predetermined key character from the set of candidate characters Section, a place name area candidate search section for searching a place name area candidate corresponding to a place name hierarchy based on the key character, and a place name for searching a place name candidate using a place name dictionary from a candidate character set included in the place name area candidate A place name candidate evaluation value calculation unit that calculates a probability of the place name candidate and outputs a place name candidate evaluation value, and a place name candidate selection unit that selects a place name candidate based on the place name candidate evaluation value are provided. Address recognition device characterized by. 2. A place name candidate evaluation value calculation unit calculates the number of matching characters between the place name candidate and the place name dictionary, and both adjacent characters of the place name candidate match the place name dictionary. An adjacent character matching number calculation unit for calculating a number, a non-matching character position correspondence number calculation unit for calculating the number of characters that are not matched in the place name candidate and the place name dictionary at corresponding positions, and the place name dictionary The address recognition device according to claim 1, further comprising a non-matching character number calculation unit that calculates the number of unmatched characters. 3. A place name candidate selection unit selects a place name candidate evaluation value comparison unit that selects a place name candidate having the largest place name candidate evaluation value, and a place name that checks whether or not the retrieved place name candidate is consistent with a place name candidate connected before and after. The address recognition device according to claim 1, further comprising a legitimacy evaluation unit. 4. A place name candidate selection unit selects a place name candidate evaluation value comparison unit that selects a place name candidate having the largest place name candidate evaluation value, and a place name that checks whether or not the retrieved place name candidate is consistent with a place name candidate connected before and after. A legitimacy evaluator, a reject evaluator that determines rejects based on the evaluation value of the selected place name, a key character inferor that infers unrecognized key characters, and a place name region that resets place name region candidates. The address recognition device according to claim 1, further comprising a candidate setting unit.