JP2998054B2 - Character recognition method and character recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition technique, and more particularly to a technique for correcting a recognition result. More specifically, when performing character recognition using an optical character reading device (OCR), etc., the recognition error was corrected for the character recognition results of the character strings written in forms such as various application documents, sales slips, and delivery slips. Related to technology.

[0002]

2. Description of the Related Art A character recognition technique is required in many fields, and in particular, a technique for correcting a character recognition result in order to improve recognition accuracy has been studied. For example, JP
JP-A-3-103393 (word recognition device)
Japanese Patent Application Laid-Open No. -121989 (word reading method) discloses a method of correcting a character recognition result for a document image in which the character pitch is not constant.

In these systems, characters are first cut out in character units from the shape characteristics of a circumscribed rectangle of characters, and the most probable character pattern is selected to recognize the character.
Next, a permissible word is searched from the word dictionary by allowing a character position shift and a character error in the candidate character string obtained as a result of the character recognition, and each searched word is compared with the character recognition result candidate character. I do. Furthermore, the part where the character is cut out by mistake is detected, the character is cut out and the character is recognized again, the word in which the candidate character is matched is extracted, and finally, the value of the value obtained by accumulating the evaluation value of the character recognition is obtained. Is determined according to.

[0004]

According to the above-mentioned prior art, a certain degree of recognition accuracy can be obtained. However, in such a method, in the first character segmentation, only one character pattern is created, character recognition and word search are performed, and the optimal word is selected from the results obtained by the word search. Therefore, if the first character segmentation was incorrect and the correct word was not included in the words obtained as a result of the word search, it was a correction error.

[0005] The character string to be read is:
When all the character strings can be covered in a word dictionary like address place names, by registering them in the word dictionary, word matching could be performed. On the other hand, when reading an unspecified number of company names or building names, it is not possible to register all the character strings that appear. There remains a problem that the entire character string cannot be verified or a correction error occurs.

The present invention has been made under the above background, and has a high recognition rate and a character recognition method capable of correctly recognizing a character that is not registered in a dictionary, and a character recognition method realizing the method. It is intended to provide a device.

A character recognition method according to a first invention provided by the present invention includes a step of dividing a character string image into unit areas, and at least one character that can be recognized as at least one character from a combination of the divided unit areas or adjacent areas. Detecting a character combination corresponding to the detected region combination and a character pattern corresponding to each of the unit regions included in the region combination, and recognizing each of the detected character patterns. results and comprising the steps of words within each character pattern to detect whether each may take the position of the character number consisting the steps of identifying a candidate character capable of recognizing each character pattern, for each candidate character, based on Identifying a word in which a character matching the candidate character exists at a position where the character pattern of the candidate character can take as a candidate word; and determining the sentence from each candidate word according to a predetermined evaluation criterion. A determining a word represented by the string image, a.

According to a second aspect of the present invention, there is provided the character recognition method according to the first aspect, wherein a character evaluation value representing a degree of similarity between each character pattern and the corresponding candidate character is generated. And comparing a character constituting the candidate word with the candidate character to detect a candidate character matching the position of the character in the candidate word, and from the character evaluation value corresponding to the detected candidate character Generating each word evaluation value; and evaluating each candidate word based on the generated word evaluation value to determine a word represented by the character string image. When evaluating candidate words, it is preferable to determine an optimum word based on the number of matching characters with respect to the character string length and the cumulative value of the evaluation values of the matching characters.

In the character recognition method according to a third aspect of the present invention, the character string image is divided into unit areas, and a recognition target area at a predetermined position is cut out from the character string image using the unit area as a minimum unit. Performing a character recognition process for each of the cut-out recognition target regions to determine candidate words, and, for each of the determined candidate words, detecting a corresponding region in an original character string image; Combining candidate words, selecting a combination that minimizes the overlap between the detected corresponding regions and the gap between the corresponding regions in the character string image, from among the combinations, corresponding to the selected combination Determining a combination of candidate words as a character represented by the character string image.

In a preferred embodiment of the character recognition method according to the third aspect of the present invention, the step of cutting out a recognition target area at a predetermined position includes setting each of the plurality of divided unit areas as a head. It is characterized in that a combined region of a plurality of adjacent unit regions is cut out as a recognition target region.

Further, a character recognition device according to a fifth invention provided by the present invention includes a first memory storing a character string image to be recognized, a second memory storing a recognition candidate character group,
A third memory storing a group of recognition candidate words, further comprising an image area dividing means for dividing the character string image stored in the first memory into unit areas, and a divided unit area or adjacent area Area feature detection means for detecting an area combination recognizable as at least one character from the combination of the characters, a character pattern corresponding to the detected area combination, and a character pattern corresponding to a unit area included in the area combination. a character pattern generating means for generating on the basis of each character pattern generated is recognized binding
A character position range detecting means for detecting whether may take what character position of the word to be a result, the candidate character selection means for selecting a candidate character corresponding to each character pattern from said second memory, the selected First candidate word selecting means for selecting, from the third memory, all candidate words having a matching character within the position range of the selected candidate character, and selecting the character from the selected candidate word according to a predetermined criterion. Word determination means for determining a word represented by the column image.

A character recognition apparatus according to a sixth aspect of the present invention includes at least a first memory storing a character string image to be recognized and a third memory storing a recognition candidate word group. An image area dividing unit that divides a character string image in one memory into unit areas, an image area extracting unit that extracts a predetermined recognition target area from the character string image using the unit area as a minimum unit, Second candidate word selecting means for performing character recognition processing for each recognition target area and selecting all candidate words from the second memory;
A region where the corresponding region in the original character string image is detected for each of the selected candidate words and a plurality of region combinations are generated, and an overlapping portion between the detected corresponding regions and a gap portion between the corresponding regions are minimized. An area combination specifying unit for specifying a combination and a compound word determining unit for determining a combination of candidate words corresponding to the specified combination as a combination of words represented by the character string image are provided.

An area combination in which the overlap between the corresponding areas and the gap between the corresponding areas is minimized is specified as follows, for example. First, the coordinates of the original character string image are set, and the distance between the words is derived by referring to the coordinate positions of the character patterns corresponding to the first character and the last character of each candidate word. Next, the distance of the overlapping portion and the distance of the gap portion of the character pattern corresponding to each word are checked, and if the value of this distance is within the allowable range, the connection of the word is recognized. Words that can be connected by this method are combined to form a compound word. In particular, when performing recognition processing using a predetermined word evaluation value, an optimal compound word is determined based on the number of matching characters with respect to the character string length of the compound word and the cumulative value of the evaluation values of the matching characters.

[0014]

According to the character recognition method of the first invention, when a character pattern is cut out from a character string image, a character pattern corresponding to each unit area and a character pattern corresponding to a combination of unit areas that can be regarded as at least one character. Is generated, all the character patterns that can be generated from the original character string are generated, and the correct character pattern to be recognized is also included in the generated character pattern. Since a correct character pattern is not missed in this way, the probability that a character to be recognized as a candidate character to be selected is selected is increased, and thus the character recognition accuracy is increased. In particular, as in the second invention, generating a character evaluation value indicating how close each character pattern is to a corresponding candidate character, and using this character evaluation value to generate an evaluation value for each candidate word Thus, the similarity between each candidate word and the original character string image is expressed as an objective numerical value.

In the character recognition method of the third invention, a plurality of recognition target areas are generated from the original character string image, and the candidate words selected for each of the recognition target areas are combined. It is recognized as a combination of multiple words. Thus, the character string image can be recognized as a compound word including a plurality of words, whereas the character string image can only be recognized as a single word in the related art.

In particular, as in the fourth invention, this character string image is constituted by cutting out, as recognition target areas, a combined area of a plurality of adjacent unit areas each of which has a plurality of divided unit areas as a leading area. All possible words are detected. In addition, the distance of the overlapping portion of the character pattern and the distance of the gap portion are checked, and if the value of this distance is within the allowable range, the connection of the word is permitted, so that the same character can be used repeatedly in two or more words. However, logically inconsistent combinations are excluded.

[0017] In the character recognition device of the fifth invention, in order to implement the character recognition method of the first invention, the first memory string image to be recognized, the recognition candidate character group second
Is stored in the third memory, and the character string image in the first memory is divided for each unit area by the image dividing means, and a combination of the divided unit area or the adjacent area is used. An area combination that can be recognized as at least one character is detected by an area feature detection unit. The generated by the character pattern generating means based on the character pattern corresponding to the unit region included in the character pattern and region combinations corresponding to each region combination to each region feature, each character pattern is the recognition result to become words The character position range detecting means detects what character position can be taken. afterwards,
The candidate character selection means selects a candidate character corresponding to each character pattern from the second memory, and selects all candidate words having a matching character within a possible position range of the selected candidate character as a first candidate word. The means selects from the third memory. Then, the word determination means determines a word represented by the character string image from the candidate words according to a predetermined evaluation criterion.

In order to implement the character recognition method of the third invention, the character recognition apparatus of the sixth invention stores a character string image to be recognized in a first memory and a group of recognition candidate words in a second memory. In advance, the character string image in the first memory is divided for each unit area by the image area dividing means. Then, a predetermined recognition target area is cut out from the character string image with the unit area as a minimum unit, and the second candidate word selecting means performs a character recognition process for each recognition target area and retrieves all candidate words from the second memory. select. The area specifying means detects a corresponding area in the original character string image for each of the selected candidate words and generates a plurality of area combinations. Then, an area combination that minimizes the overlap between the corresponding areas and the gap between the corresponding areas is specified. The compound word determining means determines a combination of candidate words corresponding to the specified combination as a combination of words represented by the character string image.

[0019]

Embodiments of the present invention will be described below in detail with reference to the drawings.

(First Embodiment) FIG. 1 is a functional block diagram of a character recognition device according to one embodiment of the present invention. In the figure, 10
Reference numeral 1 denotes a character string image storage unit (first memory) that stores a character string image (document line image) input from an image scanner or the like, and 102 denotes a character string division unit that divides a character string image into minimum unit areas. The unit (image area dividing unit) 103 detects all combinations that can be recognized as one character among combinations of adjacent unit areas, and uses an image feature to set a character pattern corresponding to the detected combination of unit areas. And a character pattern generation unit (region characteristic detection means, character pattern generation means) 104 for generating a character pattern corresponding to each unit area, from what character to what character in a word in which each character pattern is a recognition result. Position range detecting section (character position detecting means) for detecting all possible positions of
Reference numeral 05 denotes a candidate character storage (second memory) for storing recognition candidate characters and their standard patterns.

Reference numeral 106 denotes a candidate character for comparing each character pattern in the position range detected by the position range detection unit 104 with a standard pattern in the candidate character storage unit 105 and selecting a candidate character for each character pattern. A selection unit (candidate character selection means) generates a character evaluation value as an index of how close the character pattern and the standard pattern are for each selected candidate character. The candidate character selected in this way, its possible position range, and the corresponding character evaluation value are output. In this embodiment, the value of the character evaluation value increases as the character pattern and the standard pattern become closer.

A candidate word storage unit (third memory) 107 for storing a word table in which character codes are associated with words registered in a word dictionary (not shown) for each group of recognized candidate words and character positions; Reference numeral 108 denotes a candidate word selection unit (candidate word selection unit) that detects a character code corresponding to a candidate character obtained from the candidate character selection unit 106. Each candidate word in the candidate word selection unit 107 is From words, all words having a matching character within the position range of the candidate characters are selected as candidate words.

A step 109 detects, for each character in each selected candidate word, a character that matches the above-mentioned candidate character, and checks whether the detected character is within the position range of the original candidate character. Is a word evaluation value calculation unit that detects a character that satisfies these conditions for each candidate word, adds the evaluation value of the corresponding candidate character as a score of the word, and obtains a total sum value. Is the word evaluation value. Reference numeral 110 denotes a suitable word determining unit that selects a suitable word from the selected candidate words based on the word evaluation value (the above-described word determining unit).

Here, the word evaluation value calculation unit 109 only detects whether or not the detected character is within the position range of the original candidate character for the word candidate.
A plurality of character candidates may be assigned to one character position. Further, the character order of the candidate words may be reversed from the character order of the original cutout pattern. As described above, since the evaluation value for the logically inconsistent candidate character may be added to the word evaluation value, the character position of the searched word candidate is again determined. A combination of cutout patterns that maximizes the number of matching characters without duplication or inversion is determined. Then, the evaluation values of the matching candidate characters are summed to obtain a true word evaluation value, and a suitable word is selected based on the true word evaluation value. In this embodiment, one candidate word having the highest score is selected as a suitable word.

At this time, true word evaluation values may be obtained for all candidate words. However, in this embodiment, in order to simplify the processing, the words having higher scores of the original word evaluation values are used. A predetermined number was selected in order, and a true word evaluation value was determined for the selected candidate word.

Next, an example of a recognition process of a character string image of "Kawaguchi City Shiba-No-Naw" will be specifically described. Character string division unit 102
FIG. 2 shows a state in which the character string image is divided into unit areas. In this example, a minimum circumscribed rectangular area is used as a unit area. The character pattern generation unit 103 generates a character pattern from the unit area shown in FIG. As described above, the position range of each character pattern is detected, and candidate characters and character evaluation values are detected. FIG. 3 is an example of each character pattern generated in this way, the corresponding candidate character, and its character evaluation value.

As shown in FIG. 3, each unit area overlaps with each character pattern. For example, the “tree” (tree bias) in the “gutter” is the cutout patterns 13, 14, 15 in FIG.
Is used repeatedly. Then, the candidate word and the word evaluation value are obtained by the candidate word selection unit 108 and the word evaluation value calculation unit 109 according to the candidate character and its position range.
FIG. 4 shows examples of the obtained words and word evaluation values.

FIG. 5 is an explanatory diagram of a process of generating a true word evaluation value in the preferred word determination section 110. In this figure,
This shows a state in which a predetermined number of candidate words having a high word evaluation value are selected, and character patterns having candidate characters that match each character of each candidate word are indicated by numbers.

With regard to the candidate word “Kawaguchi City Shiba No Niu”,
A combination using the pattern 18 and the pattern 21 is obtained. However, as is clear from FIG. 3, since a unit area corresponding to the pattern 24 is used for the pattern 21,
Both patterns overlap. Therefore, as a combination of cutout patterns to avoid this, “3, 11, 12, 1”
5, 18, 24 ". Next, the evaluation values of the candidate characters corresponding to each cutout pattern are summed.

Referring to FIG. 3, the character evaluation value of “river” in pattern 3 is 676, the character evaluation value of “city” in pattern 11 is 757, and the character evaluation value of “shiba” corresponding to pattern 12 is 639. The character evaluation value of “gutter” in pattern 15 is 628, the character evaluation value of “no” in pattern 18 is 629, and the character evaluation value of “claw” corresponding to pattern 24 is 622. In total, a word evaluation value 3951 is obtained.

In this way, for each candidate word, an optimum combination of character patterns is obtained, and the score of the word is obtained again. FIG. 6 shows the true word evaluation of each candidate word. Finally, the candidate word with the highest score among them, “Shibainoclaw, Kawaguchi City”
Is selected as the preferred word.

As described above, according to the present embodiment, since a plurality of character patterns are generated, recognition accuracy is improved. Further, when the true word evaluation value is obtained for all the candidate words, the processing data amount becomes very large, and a large memory may be required, or the calculation time may be long. In this example, a predetermined number of words are selected in descending order of the score of the original word evaluation value, and a true word evaluation value is obtained for the selected candidate word. Can be performed, and substantially the same result as in the case where true word evaluation values are obtained for all candidate words can be obtained.

When the word evaluation value and the true word evaluation value are obtained as described above, a candidate word having a large number of characters tends to have a high word evaluation value even if the character evaluation value in the candidate word is low. . Therefore, a value obtained by dividing the total of the character evaluation values by the number of characters of the candidate word may be used as the word evaluation value or the true word evaluation value. When such an evaluation is performed, among the characters in the candidate word, the evaluation value of a word having a small ratio of characters matching the candidate character is low.

(Second Embodiment) As in the first embodiment,
A word representing a place name or the like can be registered in a dictionary in advance, and a word corresponding to the entire character string image can be easily selected. In some cases, only words corresponding to several characters are obtained. For example, if the words "second", "kawashima", and "building" of "second Kawashima building" are separately registered in the dictionary, if the character string image of the compound word "second Kawashima building" is recognized, the character Even if the word “second” can be recognized in the column image, the character image corresponding to “Kawashima Building” thereafter cannot be recognized. Second
In the embodiment, a character recognition device capable of recognizing such compound words will be described.

FIG. 7 is a functional block diagram of a character recognition device according to the second embodiment. 7, the same reference numerals are given to the same functional blocks or corresponding parts as in FIG. A character string image storage unit 101, a character string division unit 102, and a character pattern generation unit 103 are blocks that perform the same processing as in the first embodiment.

Reference numeral 701 denotes a collation start position setting unit that sets a recognition target area in a character string image with each unit area as a start position, and performs a character recognition process on the original character string image and each recognition target area. (Image area extracting means).

More specifically, the processing after the position range detection unit 104 is first performed with the head unit area as the start position.
Therefore, the original character string image is recognized as it is. Next, the start position is set to the second cutout pattern from the top, and all cutout patterns other than the first cutout pattern are output to the position range detection unit 104. Therefore, an image obtained by removing the area corresponding to the head unit area from the original character string image is the recognition target area. Hereinafter, the collation start positions are sequentially shifted in the same manner, and all cutout positions are set as collation start positions.

Thereafter, in the processing of the word range evaluation unit 109 from the position range detection unit 104, the same processing as in the first embodiment is performed for each of the recognition target areas. The preferred word determination unit 110 performs the same processing as in the first embodiment to generate a true word evaluation value. However, in the first embodiment, only one suitable word is selected. However, in this embodiment, when a predetermined evaluation criterion is satisfied, a plurality of words can be selected as a suitable word. Therefore, one word may be selected as a preferred word, or a plurality of words may be selected. In this way, a suitable word is determined for each collation start position.

Reference numeral 702 denotes a compound word determining unit which determines the distance between the candidate words by referring to the coordinate positions of the character patterns corresponding to the leading character and the ending character of each of the selected suitable words. Then, based on the result, the duplication and the interval of the character pattern corresponding to each word are checked. If the value of this distance is within the allowable range, the connection of the word is recognized. Words that can be connected by this method are combined to form a compound word, and a suitable compound word is determined based on the number of matching characters with respect to the character string length of the compound word and the cumulative value of the evaluation values of the matching characters (compound word determining means).

Next, an example of a recognition process for a character string image "2nd Kawashima Building" will be described. FIG. 8 shows a state where the character string image is divided into unit areas by the character string dividing unit 102. In this example, a minimum circumscribed rectangular area is used as a unit area.

The character pattern generation unit 103 generates a character pattern from the unit area shown in FIG. 8 as in the first embodiment. FIG. 8 shows the generated character patterns, corresponding candidate characters, and their character evaluation values. For each character pattern, the position range is detected as described above, and candidate characters and character evaluation values are detected.

In the collation start position setting unit 701, the position range detection unit 10 first sets the first cutout pattern 801 as a start position, that is, for the entire cutout pattern.
4 to the word evaluation value calculation section 109 and the processing in the preferred word determination section 110. As a result, the preferred words “second” and “yoshikawa” are obtained. Next, the collation start position is set to the second cutout pattern 802 from the beginning, and the same processing is performed to obtain “Kawashima”, “Kawacho”, and “Kohashi”. Less than,
Similarly, the matching start position is advanced to other cutout patterns 803 to 806 to obtain a plurality of candidate words.

The compound word determination unit 702 obtains the distance between words by referring to the coordinate positions of the character patterns corresponding to the first character and the last character of the candidate word, checks the connectivity of the words, and connects the words. Determine whether or not. In this example, a condition under which a certain word A and a certain word B can be connected is obtained as follows. First, words A and B are selected such that x coordinate of the left end of A ≦ x coordinate of the left end of B, and Gap is determined as follows. Gap = x coordinate of the left end of B−x coordinate of the right end of A Then, when the following expression is established, the word A and the word B are connected. − (Limit of overlap of cutout pattern) <Gap <limit of remainder of cutout pattern

For example, in FIG. 10, the "second" and "Yoshikawa" cannot be connected because the overlapping portion is too large. Also, the connection between "second" and "building" is too large to be connected, but the connection between "second" and "Kawashima" is possible. In the "Kawashima" part, there are two preferred words, but "Kawashima" is selected according to the score. As a result, "second", "Kawashima", and "building" are selected. In this way, a compound word is created by combining the connectable words, and a suitable compound word is obtained from the number of matching characters and the cumulative value of the evaluation values.

[0045]

As is apparent from the above description, according to the present invention, since all the character patterns that can be generated from the original character string are generated, the character patterns to be recognized are included therein. The probability increases. Therefore, there is an effect that the character recognition accuracy is improved.

In particular, by generating an evaluation value of each candidate word based on a character evaluation value indicating how close each character pattern and the corresponding candidate character are, each candidate word and the original character string image Can be expressed as an objective numerical value. Furthermore, by selecting the word having the highest similarity, it is possible to increase the recognition accuracy.

Further, since the character string image can be recognized as a compound word formed by combining a plurality of words,
Even if the entire compound word is not registered in the dictionary, if the individual constituent words are registered in the dictionary, the entire compound word can be recognized. All the words that may form the character string image are detected. Furthermore, the entered compound word can be recognized even if the boundaries between the words constituting the compound word are unknown and the cutout position of each character is not clear.

[Brief description of the drawings]

FIG. 1 is a functional block diagram of a character recognition device according to a first embodiment.

FIG. 2 is an explanatory diagram of a state in which a character string image is divided into unit areas.

FIG. 3 is an explanatory diagram of a generated character pattern.

FIG. 4 is an explanatory diagram of candidate words and their evaluation values.

FIG. 5 is an explanatory diagram of a method for calculating a true word evaluation value.

FIG. 6 is an explanatory diagram of a true evaluation value as a candidate word.

FIG. 7 is a functional block diagram of a character recognition device according to a second embodiment.

FIG. 8 is an explanatory diagram of a state in which a character string image is divided into unit areas.

FIG. 9 is an explanatory diagram of a character pattern.

FIG. 10 is an explanatory diagram of processing in a compound word determination unit.

[Explanation of symbols]

 101 Character string image storage unit 102 Character string division unit 103 Character pattern generation unit 104 Position range detection unit 105 Candidate character storage unit 106 Candidate character selection unit 107 Candidate word storage unit 108 Candidate word selection unit 109 Word evaluation value calculation unit 110 Suitable word Determination unit 701 collation start position setting unit 702 compound word determination unit

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G06K 9/00-9/82

Claims (6)

(57) [Claims] 1. A step of dividing a character string image into unit areas, detecting one or more area combinations that can be recognized as at least one character from a combination of the divided unit areas or adjacent areas, and further detecting the detected combination. Detecting a character pattern corresponding to the region combination and a character pattern corresponding to a unit region included in the region combination; and determining, for each detected character pattern, the character pattern in the word as a recognition result. and detecting whether may take the position of the character in the steps of identifying a candidate character capable of recognizing each character pattern, for each candidate character, and the candidate characters at positions which can be taken the original character pattern Identifying a word having a matching character as a candidate word; and determining a word represented by the character string image from each candidate word according to a predetermined evaluation criterion. Character recognition method which is characterized in that. 2. The character recognition method according to claim 1, further comprising: generating a character evaluation value indicating a degree of approximation of each character pattern and the corresponding candidate character; Comparing the candidate character with the candidate character to detect a candidate character that matches the position of the character in the candidate word, and generating a respective word evaluation value from a character evaluation value corresponding to the detected candidate character; Evaluating each candidate word based on the obtained word evaluation value to determine a word represented by the character string image. 3. A step of dividing a character string image into unit areas; a step of cutting out a recognition target area at a predetermined position from the character string image using the unit area as a minimum unit; Performing a character recognition process to determine each candidate word; for each of the determined candidate words, detecting a corresponding region in an original character string image; combining the determined candidate words; Selecting a combination that minimizes the overlap between the detected corresponding areas and the gap between the corresponding areas in the character string image; and sets the combination of candidate words corresponding to the selected combination to the character string image. And determining the character represented by the character. 4. The character recognizing method according to claim 3, wherein the step of cutting out the recognition target area at the predetermined position includes a combined area of a plurality of adjacent unit areas each of which is a leading area. A character recognition method characterized in that a character is cut out as a recognition target area. 5. A first memory storing a character string image to be recognized, a second memory storing a recognition candidate character group, and a third memory storing a recognition candidate word group,
Further, an image area dividing means for dividing the character string image stored in the first memory into unit areas, and detecting an area combination that can be recognized as at least one character from a combination of the divided unit areas or adjacent areas. Area feature detection means, a character pattern corresponding to the detected area combination, and a character pattern generation means for generating a character pattern corresponding to a unit area included in the area combination based on each area characteristic; Character position range detection means for detecting the position of a character in a word which is a recognition result in a character pattern, and a candidate character selection for selecting a candidate character corresponding to each character pattern from the second memory Means for selecting, from the third memory, all candidate words in which a matching character exists within a possible position range of the selected candidate character. That the first candidate word selection means, a character recognition apparatus characterized by having a word determining means for determining a word represented by the character string image from the candidate words said selected according to a predetermined criterion. 6. At least a first memory storing a character string image to be recognized and a third memory storing a recognition candidate word group, further comprising a character string image in the first memory. Image area dividing means for dividing each unit area; image area cutting means for cutting out a predetermined recognition target area from the character string image using the unit area as a minimum unit; and character recognition processing for each of the cut recognition target areas. Go to select all candidate words from the third memory
A candidate word selecting means, detecting a corresponding region in the original character string image for each of the selected candidate words, generating a plurality of region combinations, and determining an overlapping portion between the detected corresponding regions and a corresponding region. Region combination specifying means for specifying a region combination with a minimum gap portion; and compound word determining means for determining a combination of candidate words corresponding to the specified combination as a combination of words represented by the character string image. A character recognition device characterized in that: