JPH09245121A - Character recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To efficiently recognize a character by segmenting a series of strokes by the unit of a single character, recognizing them by each segmenting pattern to output a candidate character and comparing the recognizing result after character segmentation and each newest recognizing result with each other to detect a difference. SOLUTION: A character segmenting means 4 decides a character segmenting pattern by collecting inputted stroke data by each strokes of one character constitution. A comparing means 6 compares a past character segmenting pattern and a new character segmenting pattern to detect the difference and a character recognition means 7 recognizes stroke data for one character. A comparing means 9 compares a past candidate character and a new candidate character to detect a difference. A word recognizing means 10 decides a proper word candidate by collating a word obtained by combining the candidate characters with a word dictionary 11. A comparing means 13 compares a past candidate word and a new candidate word to detect a difference. A grammar recognition means 14 decides a suited word string and a deciding means 17 decides a proper recognizing result and output it.

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 for reading handwritten characters, and more particularly to a character recognition technique for effectively performing online handwritten character recognition processing in which strokes are successively input.

[0002]

2. Description of the Related Art A conventional on-line handwritten character recognition device is, for example, as shown in FIG. The contents of the processing are as follows. First, as the first step of the process, the handwriting stroke is stroke input means 1201.
Is entered by As a second step, a character cutout unit 1202, for example, a method as disclosed in Japanese Patent Publication No. 6-90741, cuts out a stroke set for one character as a character. Next, in the third step, the character set strokes are input to the individual character recognition unit 1203 and recognized as characters.

At this time, a plurality of candidate characters and a recognition evaluation value representing the certainty of the candidate characters are obtained as a result of the individual character recognition processing. Using this result, it is possible to narrow down the candidate characters.

Thereafter, in order to further improve the recognition rate, post-processing for determining a candidate character using linguistic restrictions such as word recognition 1204 as a fourth step and grammar recognition 1205 as a fifth step and background knowledge. Processing was performed in a hierarchical manner such as.

[0005]

However, the above-mentioned conventional online handwritten character recognition device has the following problems. After cutting out the input handwriting stroke into one character by a predetermined method, it is impossible to change the character cutout by additionally inputting a new stroke. Moreover, it is difficult to dynamically update or change the recognition result due to the change in the character segmentation. Further, a series of post-processing such as word recognition and grammar recognition reflects a partial change in the recognition result. For this reason, it was not efficient because it was necessary to input all candidate characters again and redo the processing.

In view of the above problems, the present invention provides a character recognition device which can dynamically reflect a newly input stroke and efficiently perform recognition including post-processing. It is an object.

[0007]

The present invention has the following features to attain the object mentioned above. Stroke input means, character cutout means for cutting out a series of input strokes into units that form one character, individual character recognition means for recognizing each cut out character cutout pattern and outputting a candidate character, character cutout, character A means for temporarily storing the recognition result after completion of each processing of post-processing such as recognition, word recognition, grammar recognition, a recognition result output means for obtaining and outputting an optimum recognition result from each temporarily saved processing result, and a temporary storage The present invention is characterized by comprising means for comparing the processed result of each stage and the latest recognition result of each stage to detect a difference.

[0008]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a schematic configuration diagram showing an embodiment of a character recognition device of the present invention. The device shown in FIG.
It consists of the configuration requirements of.

First, the stroke input means 1 inputs handwritten character stroke data. Stroke storage means 3
Stores sequentially the handwritten character stroke data newly input from the stroke input means 1 and stores it together with the strokes input in the past. The stroke comparison means 2 compares the stroke newly input by the stroke input means 1 with the stroke stored in the stroke storage means. The character cutout unit 4 collects a series of stroke data input by the stroke input unit 1 for each stroke forming one character, and determines a character cutout pattern. The character cutout pattern storage means 5 sequentially stores the character cutout patterns determined by the character cutout means. The cut-out pattern comparison means 6 compares the past character cut-out pattern obtained from the character cut-out pattern storage means 5 with the new character cut-out pattern obtained from the character cut-out means 4 to detect a difference. The individual character recognition unit 7 recognizes the stroke data for one character cut out as a character cutout pattern and outputs a candidate character.

The candidate character storage means 8 sequentially stores the candidate characters output by the individual character recognition means 7. The candidate character comparison means 9 compares the past candidate character obtained from the candidate character storage means 8 with the new candidate character obtained from the individual character recognition means 7 to detect a difference. The word recognizing means 10 determines a word candidate considered to be appropriate by collating a word formed by combining the candidate characters with the word dictionary 11 based on the output from the candidate character comparing means. The candidate words determined by the word recognition means 10 are sequentially stored. Candidate word comparison means 13
Detects the difference by comparing the past candidate word obtained from the candidate word storage means 12 and the new candidate word obtained from the word recognition means 10. The grammar recognition means 14 determines a grammatical matching word string in which the connection is considered to be grammatically appropriate by collating with the grammar dictionary 15 describing the connection rules of words. The grammatical matching word string storage unit 16 sequentially stores the grammar matching word string.
The recognition result determining means 17 determines the most appropriate recognition result from the contents of the candidate character storage means, the candidate word storage means 8 and the grammatical matching word string storage means 16 and outputs it as the recognition result.

Next, the details of each of the above means will be described. The stroke input means 1 is a tablet capable of detecting coordinates indicated by a pen, and detects handwriting by the pen as character stroke data (hereinafter referred to as stroke data).

The stroke comparison means 3 has the latest stroke data newly input from the stroke input means 1, and
The stroke data input in the past in the stroke storage means 3 is referred to, and strokes existing within a distance L from the latest stroke are extracted. Then, the latest stroke and the extracted stroke are output to the character cutting means 4. Although L is determined in consideration of the character spacing, the character size, etc., it can take a value of 0 to infinity. This example is shown in FIG. It is assumed that the characters “today” are manually input and strokes 201 to 211 are sequentially input. Here, when the newest stroke 211 is newly input, there are 20 strokes existing in the range of the distance L.
6, 207, 208, 209, and 210, and strokes 206 to 210 and the latest stroke 211 are sent to the character cutting means. Also, after the stroke extraction is completed, the latest stroke is stored in the stroke storage means 3 in addition to the past stroke.

The method of extracting strokes is not limited to the distance between strokes, but may be the stroke input time difference or the number of strokes.

In the example of FIG. 2, a certain period of time, for example, 100 m, is traced back from the time when the latest stroke is input.
If the strokes input within the time of sec are to be extracted, 206 to 210 are applicable, and if a predetermined number of strokes input immediately before, for example, 7 strokes are extracted, The corresponding stroke is from 204 to 210. Further, if a certain number of strokes close to the latest stroke, for example, three strokes are extracted, 210, 209, and 208 are extracted in the order of closeness.

The character cut-out means 4 cuts out a series of stroke data into a stroke set expected to constitute one character by the procedure shown in FIG. For example, it is assumed that the stroke sequence of the handwritten character "Kawasaki" (stroke data 301 to 314) shown in FIG. 3 is input. Let us consider character cutout processing by taking this cutout of stroke data as an example. First,
As shown by 301a to 314a in FIG. 5, after the stroke sequence is input (401), the circumscribed rectangle of each stroke is obtained (402).

Next, circumscribing rectangles that overlap each other are integrated to obtain an integrated rectangle (403). In the example, circumscribed rectangle 30
4a to 306a, 307a to 309a, 310a to 31
4a are integrated respectively in order of 304b, 305b, 30
6b is generated.

Next, the rectangles having overlapping portions in the vertical direction are integrated (404). However, this is the case of horizontal writing,
When writing vertically, verify the horizontal overlap. 30 in the example
5b and 306b are integrated to generate 304c.

Next, a rectangle 301 separated in the horizontal direction
If the intervals of a, 302a, 303a, 304b, and 305c are less than or equal to the threshold, they are integrated (405).
The threshold is determined as shown in FIG. Consider integration 2
Left rectangle 1 (60 in height, H1 in width, W1 in width)
1), with respect to the right side rectangle 2 (602) having the height H2 and the width W2, the distance D12 is set to the condition (D12 + W1) <(a · H
1), will be integrated when satisfying. However, a is an arbitrary constant determined from the character spacing or the like. This process is executed in the order of input and rectangle integration is performed. For example, when a = 1, the rectangle 1 (601) and the rectangle 2 (602) are (D
Since 12 + W1) <H1, the rectangle 4 (604) is integrated and generated. Next, rectangle 4 (604) and rectangle 3 (603)
However, since (D23 + W4)> H2, it is not integrated. Finally, a cutout pattern is created so that all strokes included in the integrated rectangle can be recognized as data for one character.

FIG. 7 shows an example of cutout pattern data. The cut-out pattern data has the ID of the pattern cut out as a character, the start stroke ID, and the end stroke ID as information. In the example of FIG. 7, the cutout pattern given as the character cutout ID number 1 is composed of three strokes 301, 302, and 303. Similarly, the character cutout pattern 2 is stroke 304.
It shows that it is comprised from 11 strokes of -314. Finally, the generated cutout pattern is output to the cutout pattern comparison unit 6 (406).

The cut-out pattern data in this embodiment has information on the start stroke and end stroke IDs on the assumption that the stroke IDs contained in one character are continuous, but all stroke IDs that form one character.
It may be data having information. The cutout method is not limited to this embodiment, and stroke integration may be determined by stroke contour tracking, for example.

The cut-out pattern comparison means 6 compares the latest character cut-out pattern newly cut out by the character cut-out means 4 with the character cut-out pattern cut out in the past stored in the cut-out pattern storage means 5 to determine the past. If a new character cutout pattern that is not in the character cutout pattern exists in the latest character cutout pattern, the cutout pattern is output. FIG. 8 shows an example thereof. Character cutout pattern 80 cut out in the past
It is assumed that 1, 802, 803, and 804 exist. At this time, if 805 composed of strokes 5 to 8 and 806 composed of strokes 9 to 14 are cut out as the latest character cut-out pattern, the cut-out pattern comparison means 6 selects the cut-out pattern 806 as the candidate character recognition means. Output to 7. The cutout pattern 805 is the same as the cutout pattern 803 in the past and is not output. The character cutout pattern output by the cutout pattern comparison unit 6 is stored in addition to the cutout pattern already stored in the character cutout pattern storage unit 5.

The individual character recognition means 7 recognizes the character cutout pattern output by the cutout pattern comparison means 6 and outputs a candidate character. As a method of recognizing handwritten characters, for example, a dictionary stroke pattern stored in a recognition dictionary is compared with an input stroke pattern by a pattern matching method. In the pattern matching method, the inter-pattern distance between the dictionary stroke pattern and the input stroke pattern is obtained as the sum of the corresponding inter-stroke distances. Therefore, the smaller the pattern-to-pattern distance, the more similar the character is evaluated. As a recognition result, the first to n-th candidates are output in the order of smaller inter-pattern distance. Here, n is an integer of 1 or more. The distance between patterns is output together with the candidate character as a recognition evaluation value. The recognition method in this embodiment is not limited to the pattern matching method, and any recognition method may be used as long as a plurality of ranked recognition candidate characters are output.

The candidate character is output to the candidate character comparing means 9 as the latest recognition candidate character data. FIG. 9 shows an example of recognition candidate character data. As shown in the figure, the recognition candidate character data is a candidate ID number 901 that is a unique number for each candidate.
And a start stroke ID 902, an end stroke ID 903, a candidate character code 904, and a candidate character evaluation value 905 of the character cutout pattern of the candidate character.
Therefore, when outputting up to the nth candidate, there are n pieces of recognition candidate character data having the same start stroke ID and end stroke ID. In FIG. 9, n = 2
Therefore, there are two candidates each having the same stroke range.

The candidate character comparing means 9 refers to the latest recognition candidate character data output from the individual character recognizing means 7 and the recognition candidate character data recognized in the past stored in the candidate character storing means 8 to refer to the latest recognition. M located just before the candidate character
The character recognition candidate characters are selected and output to the word recognition means 10 together with the latest recognition candidate character data.

The immediately preceding m characters can be searched by referring to the start stroke ID and end stroke ID of the recognition candidate character data. Note that the number of selected characters m is the word dictionary 11
1 is subtracted from the longest number of characters of the headword registered in. After that, the latest recognition candidate character data is added to the recognition candidate character data already recognized and stored in the candidate character storage means 8.

For example, in the example of FIG. 9, the candidate characters "knowledge" and ID 26442 with ID number 26441 are the latest recognition candidate characters.
It is assumed that the candidate character "Ori" exists and the remaining candidate words are stored in the candidate character storage means. At this time, if the longest heading word length of the word dictionary is 2, m = 1, and the candidate character connected to the latest recognition candidate character is the end stroke ID.
"Word" of candidate ID number 10251 whose number is 25, 1
0252 "Accepted", 18251 "Shinobi", 18252 "Thought"
The four candidate characters are output to the word recognition means 10 together with the latest candidate character.

The word recognition means 10 is a candidate character comparison means 9
By comparing the candidate characters output by the above with the headwords stored in the word dictionary 11, a candidate word that includes at least the latest recognition candidate character and that matches the remaining combination of recognition candidate characters is searched.

In the example of FIG. 9, "word knowledge", "recognition", and "intellect"
"Recognition" is present as a candidate word in the word dictionary 11 from among "thought", "word weave", "recognition weave", "ninori", and "shiori".

The candidate word is output to the candidate word comparison means 13 as the latest recognition candidate word data. FIG. 10 shows an example of recognition candidate word data. As shown in the figure, the recognition candidate word data has a candidate ID number 100, which is a unique number for each candidate.
1. Start stroke ID number 100 that composes the word
2, end stroke ID number 1003, word character string 1004, and evaluation value 1005. The evaluation value of the candidate word is adjusted so that a better evaluation can be obtained than the sum of the individual evaluation values of each of the constituent candidate characters. For example, when the smaller the evaluation value is, the better the evaluation is, an adjustment method of multiplying by a constant α (0 <α <1) can be considered. For example, assuming that “recognition” of the candidate ID number 10441 in FIG. 10 is α = 0.8, the evaluation value is 440. This is the word candidates 10252 "Kan" and 26441 "Kin" in FIG.
It is a better evaluation value than the value 550 obtained by individually adding the evaluation values of, and reflects that the evaluation of the character is improved by being recognized as a word.

The candidate word comparison means 13 is the word recognition means 1
The latest recognized candidate word data output by 0 and the previously recognized already recognized candidate word data stored in the candidate word storage unit 12 are compared and referenced to search for the already recognized word connected to the latest recognized word. . The latest recognition candidate word data and the already recognized candidate word data found to be connected to the latest recognition candidate word data are output to the grammar recognition means 14.

The grammar recognition means 14 is the candidate word comparison means 1
Of the latest recognition candidate word and the already recognized candidate word output by 3 which correspond to the connection rules of the words stored in the grammar dictionary 15 are recognized. The word string that matches the connection rule is
As a grammar matching word candidate string, a grammar matching word string storage unit 16
Is stored. Similar to the recognized word candidate data, the grammatical matching word candidate string has a candidate ID number, a start stroke ID number and an end stroke ID number that form the word string,
It consists of word string character string and evaluation value. The evaluation value of the candidate word string is adjusted so that a better evaluation can be obtained than the sum of the individual evaluation values of each of the constituent candidate words. For example, if a smaller evaluation value gives a better evaluation, the constant β
An adjustment method of multiplying (0 <β <1) can be considered.

The recognition result determining means 17 refers to the candidate character storage means 8, the candidate word storage means 12, and the grammar matching word string storage means 16 to cover all the input strokes and to obtain the recognition result based on the evaluation value. Output. For example, a recognition result having the smallest sum of evaluation values is selected from the candidates and is output as the recognition result. Alternatively, the product of evaluation values may be used.

The actual operation of the apparatus of this embodiment when the stroke sequences shown in FIG. 3 are sequentially input will be described with reference to FIG. The already-input stroke column 1101 illustrates strokes that have already been input. The new stroke column 1102 shows a newly input stroke and its ID number (common to FIG. 3).
In the new cut-out pattern field 1103, the start stroke ID and end stroke ID of the newly cut out cut-out pattern, and in the candidate character field 1104, the recognition result (up to the second-ranked candidate) of the new cut-out pattern and its evaluation value. In the candidate word column 1105, the word recognized as a candidate word and its evaluation value are described, and in the recognition result column 1106, the final result of recognizing the stroke input up to that point and its evaluation value are described. . The description will be made in the following order.

1: Input of Stroke 301 At this time, since there is no input stroke, a cutout pattern composed only of the stroke 301 is newly cut out and recognized. The recognition result is "No".

2: Stroke 302 input The strokes 301 and 302 are cut out as one character to form a new cutout pattern. As the recognition result, “i” having the best evaluation value is selected from the candidates that cover the strokes input at the present time.

3: Input of Stroke 303 Strokes 301 to 303 are cut out as one character to form a new cutout pattern. The recognition result at this stage is “river”.

4: Input of stroke 304 Here, two cutout patterns composed of strokes 301 to 303 and 304 to 304 are cut out. Since the former is the same as the cutout pattern already cut out when the stroke 303 is input, the new cutout pattern is only the pattern constituted by the stroke 304. Therefore, the latter is the only cut-out pattern that is subjected to character recognition processing at this stage. As a result, the candidate characters became "1" and "ku". Candidate character "Kawa" with continuous cutout patterns
There is no candidate word in the combination with “N”, and the recognition result at this stage is “kawa 1”.

5: Input of Stroke 305 The strokes 301 to 303 and 304 to 305 are cut out, and the new cutout pattern is composed of the strokes 304 to 305 as in the previous step. The recognition candidate characters of the cutout pattern are "to" and "person".
There is no candidate word, and the recognition result at this stage is "Kawato".

6: Stroke 306 input Although the strokes are cut out by the strokes 301 to 303 and 304 to 306, the new cutout pattern is composed of the strokes 304 to 306 as in the previous stage. The recognition candidate characters of the cutout pattern are "mountain" and "N".
There is no candidate word, and the recognition result at this stage is "Kawayama".

7: Stroke 307 input The strokes 301 to 303 and 304 to 306 and 307 are cut out. The new cutout pattern is stroke 3.
07 are only configured. The recognition candidate characters for the cutout pattern are "I (kanji)" and "-(long-sound symbol)."
It became. There is no candidate word, and the recognition result at this stage is “Kawayama Hajime”.

8: Stroke 308 input Strokes 301 to 303 and 304 to 306 and 307 to
It is cut out at 308, but the new cutout pattern is only the one composed of strokes 307 to 308. The recognition candidate characters of the cutout pattern are "na" and "+". There is no candidate word, and the recognition result at this stage is "Kawayama Na".

9: Input of stroke 309 Strokes 301 to 303 and 304 to 306 and 307 to
Although it is cut out at 309, the new cut-out pattern is only the one composed of strokes 307 to 309. The recognition candidate characters of the cutout pattern are "large" and "length". There is no candidate word, and the recognition result at this stage is "Kawayama Univ."

10: Stroke 310 input The strokes 301 to 303 and 304 to 310 are cut out, and the new cutout pattern is strokes 304 to 3
Only those configured with 10. The recognition candidate characters of the cutout pattern are "sex" and "enclosure." There are no candidate words, and the recognition result at this stage is "river."

11: Input of stroke 311 The strokes 301 to 303 and 304 to 311 are cut out, but the new cutout pattern is strokes 304 to 3
11 is only configured. The recognition candidate characters of the cutout pattern are "frame" and "break". There are no candidate words, and the recognition result at this stage is "river frame."

12: Stroke 312 input The strokes 301 to 303 and 304 to 312 are cut out, but the new cutout pattern is strokes 304 to 3
It is only composed of 12. The recognition candidate characters for the cutout pattern are "pass" and "special". There is no candidate word, and the recognition result at this stage is "Kawatoge".

13: Stroke 313 input: The strokes 301 to 303 and 304 to 313 are cut out, but the new cutout pattern is strokes 304 to 3
13 is the only one. The recognition candidate characters of the cutout pattern are "Mine" and "Saki". The candidate word “Kawasaki” exists, and the evaluation value is a value 210 (= 60 + 150) obtained by individually adding the evaluation values of the candidate words “kawa” and “saki” multiplied by a constant α. In the example of the figure, the evaluation value is 168 because the calculation is performed with α = 0.8. “Kawasaki” is selected as the recognition result at this stage. By the way, if there is no word dictionary "Kawasaki", it will be "Kawamine".

14: Stroke 314 input The strokes 301 to 303 and 304 to 314 are cut out, and the new cutout pattern is strokes 304 to 3
14 is the only one. The recognition candidate characters of the cutout pattern are "saki" and "saki". As in the previous step, the candidate word "Kawasaki" exists. The recognition result is "Kawasaki"
Becomes

As described above, a character that changes when a new handwritten stroke is input, a change in the recognition result accompanying it, and a dynamic reflection of the changed recognition result in post-processing such as word recognition are realized. did. In addition, efficient processing was realized by extracting only the changed recognition result and the part expected to be related to it and reprocessing.

The structure according to the present invention can be applied not only to an on-line handwritten character recognition device but also to all devices for recognizing intermittent input. For example, it can be applied to a continuous voice recognition device.

[0050]

As is apparent from the above embodiments, by using the character recognition device having the configuration of the present invention, when a new handwriting stroke is input in the online handwritten character recognition, the character segmentation and the individual character are input. At each stage of recognition, word recognition, and grammar recognition, only the difference from the past recognition result can be processed. Therefore, the recognition result can be updated with the minimum necessary processing, which is effective in performing high-speed and efficient recognition.

[Brief description of drawings]

FIG. 1 is a configuration diagram of an embodiment of a character recognition device of the present invention.

FIG. 2 is a diagram for explaining an example of a character cutout method of a character cutout unit of the embodiment of the character recognition device of the present invention.

FIG. 3 is a diagram for explaining an example of a character cutout method of a character cutout unit of the embodiment of the character recognition device of the present invention.

FIG. 4 is a diagram for explaining an example of a character cutout method of a character cutout unit of the embodiment of the character recognition device of the present invention.

FIG. 5 is a diagram for explaining an example of a character cutout method of the character cutout means of the embodiment of the character recognition device of the present invention.

FIG. 6 is a diagram for explaining an example of a character cutout method of a character cutout unit of the embodiment of the character recognition device of the present invention.

FIG. 7 is a diagram showing an example of a character cutout pattern of an embodiment of the character recognition device of the present invention.

FIG. 8 is a diagram for explaining the processing of the character cutout pattern comparison means of the embodiment of the character recognition device of the present invention.

FIG. 9 is a diagram showing an example of recognition candidate character data of an embodiment of the character recognition device of the present invention.

FIG. 10 is a diagram showing an example of candidate word data in the embodiment of the character recognition device of the present invention.

FIG. 11 is a diagram for explaining the operation of the embodiment of the character recognition device of the present invention.

FIG. 12 is a configuration diagram showing an example of a conventional character recognition device.

[Explanation of symbols]

1 ... Stroke input means, 2 ... Stroke comparison means, 3
... stroke storage means, 4 ... character cutout means, 5 ... character cutout pattern storage means, 6 ... character cutout pattern comparison means, 7 ... individual character recognition means, 8 ... candidate character storage means, 9 ... candidate character comparison means, 10 ... Word recognition means, 1
DESCRIPTION OF SYMBOLS 1 ... Word dictionary, 12 ... Candidate word storage means, 13 ... Candidate word comparison means, 14 ... Grammar recognition means, 15 ... Grammar dictionary, 1
6 ... Grammar compatible word string storage means, 17 ... Recognition result determination means

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Yasuharu Namba 1099, Ozenji, Aso-ku, Kawasaki-shi, Kanagawa Inside the Hitachi, Ltd. Systems Development Laboratory (72) Inventor Hirokazu Aoshima 1099, Ozen-ji, Aso-ku, Kawasaki, Kanagawa Hitachi Systems Development Laboratory

Claims (5)

[Claims] 1. An input unit for inputting handwriting stroke data of handwritten characters, a first storage unit for storing the input handwriting stroke data, and a unit for storing the input handwriting stroke data and the storage unit in the past. Compare the stored handwriting stroke data,
First extracting means for extracting difference stroke data, cutout means for cutting out the input handwriting stroke data as a stroke set for each character using the extracted difference stroke data, and the cutout stroke set. A character recognition device comprising a character recognition means for recognizing a character. 2. The character recognition device according to claim 1, wherein the storage unit stores the stroke set cut out by the cutting unit, the stroke set cut out by the cutting unit, and the second set in the past. Second extraction means for comparing the stroke set stored in the storage means and extracting the difference, the character recognition means performing character recognition using the extraction result of the second extraction means. Character recognition device. 3. The character recognition device according to claim 1, wherein the character recognition unit stores the character recognized by the character recognition unit, and the character recognized by the character recognition unit and the character recognized in the past. A third recognizing means for comparing the characters stored in the second storing means and extracting a difference, and a word recognizing means for recognizing a word using the third extracting means and a word dictionary storing word information. A character recognition device having. 4. The character recognition device according to claim 3, wherein the word recognition means outputs one or more candidate words. 5. The character recognition device according to claim 1, wherein the character recognition means outputs one or more candidate characters.