JPH08180138A - Character recognizing device - Google Patents

Abstract

PURPOSE: To speed up the character recognition of a Japanese document by decreasing the number of arithmetic processes for a degree of difference without altering the array order of components of multi-dimensional feature vectors of a character recognition dictionary. CONSTITUTION: Characters to be recognized are divided into plural character groups consisting of characters which are small in degree of difference, and the character groups that the characters belong to and information showing whether or not they are representative characters of the character groups are stored in a character recognition dictionary 1 together with feature vectors. The character pattern of a character to be recognized which is inputted by a character pattern input part 2 is normalized by a character pattern normalization part 3 and then passed to a feature vector extraction part 4, which extracts multidimensional feature vectors consisting of plural feature elements. An initial candidate selection part 5 compares them with the representative characters of the character groups to select a character group whose degree of difference is less than a selection reference, and a final candidate selection part 6 while sequentially totalizing and adding the degrees of differences by feature elements about all character kinds of the selected character group decides a decision at each time to replace candidate characters. Lastly, a recognition result output part 7 outputs the candidate character of the highest order as a recognition result.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition device, and more particularly to a character recognition device for recognizing Japanese characters using a feature vector having a plurality of feature element components for each divided small area.

[0002]

2. Description of the Related Art In character recognition for Japanese sentences, many kanji, hiragana, katakana, alphanumeric characters and symbols are used, and many complicated kanji are included. Therefore, the character area is divided into a large number of small areas that have overlapping parts, and a vector with multiple feature elements is set for each small area. The resulting "number of feature elements x number of small areas" component A method of comparing multidimensional feature vectors with is used.

Typical feature vectors having a plurality of feature elements as components include one defined by a line drawing pattern of a character image and one defined by a contour line pattern of the character image. In order to improve the recognition rate without increasing the total number of vector dimensions, various methods have been proposed, each of which has a different method of dividing a small area and a method of selecting characteristic elements. Below, each typical example is demonstrated.

An example of the former is a direction line element feature amount. This is to thin a character pattern displayed in a character area of 64 × 64 dots to obtain a line drawing pattern representing a skeleton line forming a character, and determine the orientation of each line element (corresponding to one pixel) forming the skeleton line. It is a 196-dimensional feature vector that is quantized in four directions of vertical (|), horizontal (-), and ± 45 degrees (/, \), and that is quantized. First, a character area of 64 × 64 dots is divided into 8 × 8 dot units, and 16 × 16 dots, which is a combination of adjacent 4 areas, are set as small areas, and each small area is divided by half in both vertical and horizontal directions. 49 overlapping small regions are formed. By creating a histogram for each direction by associating a corridor-shaped filter having a Gaussian filter-like weight that is heavy in the central part and light in the peripheral part for each small region configured in this way, 49 small regions × 4 directions = A 196-dimensional feature vector is obtained. For details, see “High-precision character recognition using directional line element features” (IEICE Transactions D-II, vo).
l. J74-D-II, No3, pp330-339, 1
March 991).

On the other hand, in the weighted direction index histogram, which is an example of the latter, the character contours of the binarized character pattern after normalization are tracked with eight connections, and the direction of the contour line at each contour point is 45.
Quantize in 4 directions. First, this direction-quantized pattern is divided into 11 vertical subsections × 11 horizontal subsections = 121 subsections, and a histogram is created for each subdivision so that the vertical 11 × horizontal 11 × 4 subdivisions = 484-dimensional subdivision Get the histogram. A two-dimensional Gaussian filter (5 × 5) having overlap with this direction index histogram is associated and dimensionally compressed to obtain a weighted direction index histogram of vertical 6 × horizontal 6 × 4 directions = 144 dimensions. For details, refer to "Corresponding to the collapsed characters in the weighted direction exponential histogram method" (Shingaku Giho, PRU90-128, pp21).
~ 26).

In the character recognition device using the feature vector, the feature vector of the character image (input pattern) to be recognized is extracted and the corresponding dimensions of the feature vectors of the standard pattern stored in the character recognition dictionary are compared. The degree of dissimilarity or the degree of similarity is calculated, and the character type of the standard pattern in which the sum of the dissimilarities of each dimension is the smallest or the sum of the similarity is the largest is selected as the recognition result. When trying to recognize a Japanese sentence, even if it is limited to JIS Level 1 Kanji, it is necessary to calculate the degree of difference or similarity with a character type close to 3,000, and a highly accurate recognition result is obtained. If this is done, the dimension number of the feature vector will inevitably increase, and there has been a problem that the processing time increases and becomes enormous as the number of character types used and the number of dimensions increase.

As a countermeasure against this, a preliminary rough selection is performed using a feature vector having a small number of dimensions (for example, 4 × 4 × 4 = 64 dimensions), and a feature having a large dimension only for a limited number of selected character types. Although a two-step recognition method in which final selection is performed using a vector is conceivable, there is a drawback that it requires two types of character recognition dictionaries and two types of feature vector extraction.

On the other hand, Japanese Patent Laid-Open No. 63-780 proposes a method of suppressing an increase in processing time due to an increase in the number of dimensions by using only one type of character recognition dictionary. This method is used when recognizing an unknown pattern by rearranging the arrangement order of the feature vectors stored in the character recognition dictionary and registering them in order of the dimension with the largest standard deviation or variance for all the character types registered in the dictionary. , The feature vector of the unknown pattern is extracted and rearranged in the same order, and compared with all character types using a certain number of dimensions from the higher-dimensional component to perform rough selection, and for the selected character type By comparing the components of the lower dimension only with the above, the type of characters and the number of dimensions to be operated are suppressed. That is, rough selection is compared with all character types by using only a limited number of feature vector components from the top,
Those whose degree of difference or similarity exceeds a certain limit are excluded from the candidates, and the final candidate is selected using the feature vectors of all dimensions only for the candidates remaining in the rough selection. The rough selection may be repeated in several stages, and then the final selection may be performed.

[0009]

However, in the method described in Japanese Patent Laid-Open No. 63-780, the rough selection is performed using only a limited fixed number of feature vector components. It is necessary to register the vector arrangement order by rearranging it in the order of the dimension having the larger standard deviation or variance, and rearrange the feature vector of the unknown pattern in the same order. In other words, for all of the multidimensional feature vectors, the standard deviation or variance must be calculated to determine the array order, so if the number of target character types increases, a large processing man-hour will be required and if the number of target character types is changed. It has the drawback of having to update the entire character recognition dictionary.

An object of the present invention is to perform character recognition in which only one type of character recognition dictionary can be used to reduce the number of operations for the dissimilarity and speed up the recognition of Japanese sentences without changing the arrangement of feature vectors in the character recognition dictionary. To provide a device.

[0011]

A character recognition device according to claim 1 is a character recognition device for recognizing Japanese characters using a multidimensional feature vector having a plurality of feature element components for each of the divided small areas. Yes, character group identification information indicating which of a plurality of character groups each of which consists of a plurality of character types with a small degree of difference between the feature vectors and the characteristic vector of the standard pattern of each character type, and a representative character that is the center of each character group The character recognition dictionary that stores representative character identification information that indicates whether or not the character pattern input unit that captures the character pattern of an unknown input character to be recognized, and the size and line width of the captured character pattern A character pattern normalization unit that normalizes so that it can be handled uniquely, a feature vector extraction unit that extracts a feature vector from the normalized character pattern, and an extracted feature vector And a feature vector of a representative character of the character recognition dictionary are compared, and an initial candidate selection unit that selects a representative character that is an initial candidate under a predetermined condition in order of increasing degree of difference, and a difference from the selected representative character After initializing the candidate characters in ascending order of the degree, the intermediate summation of the dissimilarity of each feature element component is performed for each character type in the order of the determined feature element components for all the character types of the character group to which the selected representative character belongs. However, it is configured to include a final candidate selection unit that replaces candidate characters at any time by adding judgments and a recognition result output unit that at least finally outputs the character positioned in the highest rank of candidate characters as a recognition result. ing.

The character recognition device according to a second aspect is the character recognition device according to the first aspect, wherein the initial candidate selection unit selects all representative characters having a difference degree smaller than a predetermined limit value as an initial candidate. It is characterized by that.

A character recognition apparatus according to a third aspect is the character recognition apparatus according to the first or second aspect, in which the initial candidate selecting section extracts the extracted feature vector and the feature vector of a representative character of the character recognition dictionary. Is compared with each other in the order of the determined characteristic element components, and for each representative character, the intermediate degree of the dissimilarity is calculated for each characteristic element component, and a judgment is added each time to select a representative character as an initial candidate.

The character recognition device according to claim 4 is the character recognition device according to claim 1
Alternatively, in the character recognition device according to claim 3, the final candidate selection unit initially sets only a fixed number of candidate characters in the order of decreasing dissimilarity among the representative characters selected by the initial candidate selection unit. I am trying.

A character recognition apparatus according to a fifth aspect is the character recognition apparatus according to each of the first to fourth aspects, in which the final candidate selection unit selects an intermediate tabulated value of the dissimilarity for each characteristic element component. Compared to the dissimilarity of the lowest rank character of the character,
The feature is that it is sequentially determined whether or not the calculation of the dissimilarity of the remaining feature element components is continued.

A character recognition apparatus according to a sixth aspect is the character recognition apparatus according to each of the first to fifth aspects, in which a plurality of feature element components forming a feature vector are skeleton patterns ( Or, it is a four-direction element indicating the direction of the line element of the outline pattern of the character image), and the order of performing the intermediate summation of the dissimilarity for each feature element component is such that the vertical and horizontal processing is performed before the diagonal processing. The feature is that it is set to be.

[0017]

Embodiments of the present invention will now be described with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention.

As shown in FIG. 1, the character recognition apparatus of the present embodiment, together with the feature vector of the standard pattern of each character type,
A character recognition dictionary 1 that stores identification information indicating which of a plurality of character groups composed of similar characters with a small dissimilarity and identification information indicating a representative character at the center of each character group and a character pattern are stored. A character pattern input unit 2 for capturing, a character pattern normalizing unit 3 for normalizing the size and line width so that the captured character pattern can be handled uniquely, and a multidimensional feature from the normalized character pattern. A feature vector extraction unit 4 for extracting a vector is compared with the representative character of the character recognition dictionary 1 for all components of the extracted feature vector, and a representative character whose dissimilarity is smaller than a predetermined limit value is set as an initial candidate. Pair the initial candidate selection unit 5 to be selected with all the character types of all the character groups to which the representative character belongs after initializing all the selected representative characters as candidate characters. As a final candidate selection unit 6 that sequentially performs intermediate calculation of dissimilarity for each characteristic element component in the order of the determined characteristic element components and replaces the candidate characters as needed while making a determination to continue processing. And a recognition result output section 7 for outputting the character positioned at the highest rank as a recognition result.

In the following, for the Japanese language, characteristic elements in four directions are assigned to the skeleton or outline line elements of a character, and vertical X × horizontal Y × 4 directions = 4XY dimensions (4 elements for each XY small area).
The description will be made assuming that the feature vector is extracted and the character recognition is performed based on the difference.

In the character recognition dictionary 1, the standard vector feature vector components of all character types to be recognized are determined for each character code in the determined small area order (for example, from the upper left to the lower right). The elements are registered in the given element order (for example, vertical "|", horizontal "-", +45 degrees "/", -45 degrees "\"). Here, the component of the feature vector registered in the character recognition dictionary 1 is represented by Gmik, the subscript m is the order of the character types registered in the dictionary, i is the order of the small areas,
It is assumed that k indicates the order of the direction of the characteristic element. When the order of the small area and the characteristic element is represented by numbers in the order illustrated above, i and k are i = 1 to M (M = XY) and k = 1, respectively.
~ 4, and each component of the feature vector for each character type is G
m11, Gm12, Gm13, Gm14, Gm21, Gm22, Gm23, Gm24, ...
It is arranged like GmM1, GmM2, GmM3, GmM4.

In addition to this, the character recognition dictionary 1 also includes character group identification information indicating to which one of a plurality of character groups composed of similar characters having a small dissimilarity for each character code, and each character code. Representative character identification information indicating whether or not the character is the representative character that is the center of the group is stored. That is, the character types registered in the character recognition dictionary 1 are classified into several character groups, each character group is composed of a plurality of similar characters with a small dissimilarity, and each character group has one representative character. It is specified. When a Japanese sentence including JIS level 1 kanji is targeted, the number of character types to be recognized is about 3300, and when the number of character groups is 200, the average number of character types per character group is 165. Although various methods are conceivable as a method of organizing a character group and the method is not particularly limited, it is necessary that the degree of difference between the representative character and the character type in the character group does not exceed a certain value. It should be noted that there may be some overlap in the range of each character group.

The character pattern input unit 2 extracts a rectangular area for one character in a document image read by an image scanner as a character pattern. The character pattern taken in here is the normalization of the size or the line. It is passed to the character pattern normalization unit 3 to normalize the width.

The normalized character pattern is decomposed into a plurality of characteristic elements in the characteristic vector extraction unit 4 to extract a characteristic vector. That is, a skeleton line drawing pattern or contour line pattern is extracted according to the property of the feature vector, each line element is aggregated for each direction (feature element) for each division unit, weighted and merged, and the feature vector of each small area The ingredients are calculated.

The extracted feature vector is passed to the initial candidate selection unit 5, the degree of difference from the standard pattern of the representative character of each character group registered in the character recognition dictionary 1 is calculated, and the detailed comparison is performed by rough selection. The number of target characters to perform is limited. That is, the representative characters of each character group are sequentially selected with reference to the representative character identification information of the character recognition dictionary 1, the same dimensions of the feature vectors are compared to calculate the sum of the dissimilarities, and the sum of the dissimilarities is calculated in advance. Character groups that exceed the defined selection criterion are excluded, and only representative characters whose sum of differences is within the selection criterion are passed to the final candidate selection unit 6 as initial candidates. The calculation of the dissimilarity in the initial candidate selection unit 5 may be an ordinary method for all dimensions because the target is only the representative character, but if the method of the final candidate selection unit 6 described below is applied, the number of calculations is further reduced. It can be reduced and increased in speed. The selection criterion is set in consideration of the number of character groups and the size of the character group, and a smaller value is advantageous to simplify subsequent processing. It is necessary that the difference is larger than the maximum value.

The final candidate selecting section 6 sorts all the representative characters first passed from the initial candidate selecting section 5 in the order of decreasing dissimilarity and initializes them as candidate characters. After the initial setting, all the character types of the character group to which the passed representative character belongs are targeted, and the intermediate summation of the feature vector dissimilarity is sequentially performed for each feature element in the order of the determined feature elements. The continuation of the process is determined by comparing the dissimilarity of the lowest rank character of. If the intermediate aggregate value of the dissimilarities exceeds the dissimilarity of the lowest rank character, the calculation for the character type is aborted and the next character is moved to. If the sum of the dissimilarities of all the feature elements is smaller than the dissimilarity of the lowest rank character of the candidate character, the candidate character is replaced with the character type. By repeating the above process, the initially set candidate character is
Characters having a smaller total sum of dissimilarities in the target character group are sequentially replaced, and the character groups are arranged in the order of smaller total sum of dissimilarities.

The recognition result output unit 7 may output and display the character positioned at the highest rank of the final candidate characters as the recognition result. If there is no difference in the difference between the highest rank and the second rank or lower, or if there is a slight difference, these characters are memorized and a mark is added to the display character of the recognition result. The second and third rank candidate characters may be configured to be presented as alternative candidates.

FIG. 2 is a flow chart showing details of the processing of the final candidate selecting section 6. Hereinafter, the operation of the final candidate selection unit 6 will be described again with reference to FIG.

The representative character passed from the initial candidate selecting section 5 is sorted as a candidate character in the ascending order of the dissimilarity, and the initial setting is performed to start the processing. Here, the number of candidate characters (the number of character groups) is J, and the total number of characters included in the corresponding character group is N.
Then, since the dissimilarity has already been calculated in the initial candidate selection unit 5 for the representative J character among the N characters, the dissimilarity calculation is sequentially performed for the remaining N-J characters, and the character type having a small dissimilarity is calculated. If there is, the process of replacing the initially set candidate characters and leaving the J character type as the final candidate character in the ascending order of difference is performed according to the procedure shown in FIG. First, N character types to be targets of difference calculation are ordered, representative characters are set to # 1 to #J, and a value of a counter n indicating the order of target character types is set to n = J + 1 in step S1.

Next, in step S2, a counter k indicating the characteristic element (direction) and a register Δk-for storing the aggregate value of the dissimilarity.
Initialize 1 and Δk. Then, in step S3, the sum of the squares of the difference between the feature vector component Fik of the input pattern and the feature vector component Gnik of the standard pattern is used as the degree of difference, and all the small regions are totalized for each feature element. M (= XY) is the number of dimensions per one characteristic element (direction component) of the characteristic vector, the dissimilarity is totaled for each characteristic element, and the calculation result is cumulatively added to the register Δk.

In step S4, the dissimilarity ΔcJ (the subscript c indicates the candidate character and J indicates the rank) of the lowest rank character of the candidate character and the current register each time the difference for each one-direction component is calculated. Compare with the difference of Δk. Here, if the dissimilarity of the register Δk is already larger than the dissimilarity ΔcJ,
The process advances to step S8 to advance the character type position by one, and if the counter n does not exceed the target total number N of characters in the determination in step S9, the process returns to step S2 to compare with the next character type. If the counter n exceeds the total number N of characters, the process ends there.

If the dissimilarity of the register Δk is smaller than the dissimilarity ΔcJ, the process proceeds to step S5, the counter k is incremented by one, and the target direction is changed. At this time, if the difference degree calculation in the four directions has already been completed in the comparison in step S6, the candidate characters are replaced in step S7, and then the difference degree calculation with other character types is performed. If the difference degree calculation of another direction component is not yet processed in the determination of step S6,
The process returns to step S3, the degree of difference of the directional component is calculated and added, and the processing from step S4 is repeated.

FIG. 3 is a flow chart of the candidate character replacement process of step S7 in FIG. The replacement process will be described in detail with reference to FIG.

First, in step S71, the number of candidate characters J is set in the counter j, and the candidate character is sequentially replaced from the lowest rank to the highest rank. First, the candidate character having the lowest rank is moved out of the candidate and a new candidate character is inserted at the lowest position. That is, in step S72, the dissimilarity ΔcJ and the dissimilarity Δ4 (Δk with k = 4) of the new candidate character are exchanged. Here, Δcj + 1 represents the degree of difference outside the candidate.

Subsequently, in step S73, the degree of dissimilarity with the candidate character having the rank higher by one is compared. If the degree of dissimilarity is higher than the candidate character with the rank higher by one, the replacement process ends. If the dissimilarity is smaller than the candidate character of the next higher rank, the process proceeds to step S74. Step S
In 74, the difference degree saving variable Δo saves the difference degree of the candidate character one rank higher and replaces the candidate rank.

When the replacement of the candidate ranks is completed, the value of the counter j is returned by one in step S75. At this time, in step S76, the counter j has the highest rank (j =
If it points to 1), the process ends there. Otherwise, if the counter j has not yet reached the highest rank, the processing from step S73 is repeated.

As described above, when the comparison with the standard patterns of all the character types of the character group corresponding to the representative character selected by the initial candidate selection unit 5 and the replacement of the candidate characters are completed, the final candidate selection unit 6 The process of (1) is ended, the character positioned at the highest rank of the candidate characters is finally output from the recognition result output unit 7, and the character recognition is ended.

As described above, in the present embodiment, instead of calculating the dissimilarity of feature vectors in order from the first component and obtaining the sum at a time, the feature elements are calculated for each feature element in a predetermined order. It is executed while performing the intermediate totaling, and each time, it is determined whether to abort or continue the processing.
As described above, in the character recognition dictionary 1, the characteristic vector components of each character type are registered in the determined characteristic element order in the determined small area order. Therefore, the intermediate summation of the dissimilarity of feature vectors for each feature element can be easily performed without rearranging the feature vector components. That is, if the characteristic element is in four directions of the line element, it is sufficient to take out every four components, calculate the difference between the respective corresponding components, and perform aggregation. When the feature vector components of each character type are registered in the character recognition dictionary 1 in the order of feature elements and in the order of small areas, the continuous components are counted for each number of small areas.

Japanese sentences include kanji, hiragana, katakana,
Alphanumeric characters and symbols are mixed, and it is considered that 30 to 40% of the used characters are Kanji. Focusing on the characteristic element of the direction of the line elements that make up a character, it can be said that most of the kanji characters are diagonal or curved straight line components, but most of the kanji are vertical or horizontal straight line components. Considering the nature of Japanese sentences and the fact that nearly 3,000 first-level character types are Kanji, the vertical and horizontal components are first and diagonal (4
It is effective to add up the components of ± 45 degrees when considered in the direction later.

In the above description of the embodiment, the initial candidate selection unit 5 is assumed to pass all the representative characters whose dissimilarity is less than the selection criterion to the final candidate selection unit 6 as initial candidates. From the viewpoint of limiting the number of characters to be processed in 6, it is possible to pass not only all the representative characters but only a certain number of representative characters from the one having the smallest difference. However, in the method of the above-described embodiment in which all the representative characters are passed, when the recognition target character such as JIS second level kanji is additionally registered in the character recognition dictionary 1, the organization of the registered character group is changed. Moreover, there is an advantage that the character groups of the additional characters can be independently set without considering the overlapping of ranges. On the other hand, in the case of the method of passing only a fixed number, some kind of countermeasure is required, such as changing the organization of the character group or changing the number of representative characters to be passed.

Further, the final candidate selecting section 6 sorts all the representative characters passed from the initial candidate selecting section 5 in the order of decreasing dissimilarity and initializes them as candidate characters. However, the number of candidate characters to be sequentially replaced in the final candidate selection unit 6 may be a limited fixed number (one may be preferable, but a plurality is preferable) instead of the total number of representative characters. In this case, the number of candidate characters and the number of target character groups do not match, and “J” in step S1 of FIG. 2 and “J” of ΔcJ in step S4 are not the same, and the latter is “J” in FIG. It is replaced with "Jo"(J> Jo).

Further, in the above-described embodiment, the total difference degree of the lowest rank of the candidate character is used as the criterion for the continuation processing of the difference degree in the final candidate selecting section 6 for each intermediate tabulation. Therefore, the judgment reference value changes as the processing progresses, but this judgment reference value may be a fixed value. In this case, the fixed value is preferably small, and at least a value smaller than the selection criterion in the initial candidate selection unit 5 is suitable.

[0043]

As described above, in the character recognition device of the present invention, the character to be recognized is divided into a plurality of character groups consisting of characters having a small dissimilarity, and the character recognition dictionary and the character group and the representative character are stored in the character recognition dictionary. The identification information is registered, rough selection is performed using representative characters, and the degree of difference in the final selection after rough selection is divided for each feature element component to calculate the difference for one feature element component. Since it is configured to terminate the calculation of unnecessary feature element components by inserting a judgment each time it is finished, it is possible to omit extra calculation and reduce the replacement process of candidate characters, and use only one type of feature vector. The recognition speed of the character recognition can be significantly improved without changing the arrangement order of the feature vector components of the character recognition dictionary.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention.

FIG. 2 is a flow chart showing a process of calculating a dissimilarity degree in a final candidate selecting unit in FIG.

FIG. 3 is a flowchart showing details of a candidate character replacement process in FIG.

[Explanation of symbols]

 1 character recognition dictionary 2 character pattern input unit 3 character pattern normalization unit 4 feature vector extraction unit 5 initial candidate selection unit 6 final candidate selection unit 7 recognition result output unit

Claims (6)

[Claims] 1. A character recognition device for recognizing Japanese characters using a multidimensional feature vector having a plurality of feature element components for each of the divided small areas, and the feature vector together with the feature vector of a standard pattern of each character type. Character group identification information indicating which of a plurality of character groups composed of a plurality of character types having a small difference degree and representative character identification information indicating whether or not the character is a representative character at the center of each character group The stored character recognition dictionary, the character pattern input part that captures the character patterns of unknown input characters to be recognized, and the character pattern normalization part that normalizes the size and line width of the captured character patterns so that they can be handled uniquely. A feature vector extraction unit for extracting a feature vector from the normalized character pattern, a feature vector extraction unit for extracting the feature vector and a feature vector of a representative character of the character recognition dictionary. After initializing the candidate characters from the selected representative character in the ascending order of the dissimilarity, the initial candidate selection unit that selects the representative character as an initial candidate under a predetermined condition by comparing All the character types of the character group to which the selected representative character belongs are targeted in the order of the characteristic element components determined for each character type. A character recognition device, comprising: a final candidate selection unit for replacing and a recognition result output unit for at least finally outputting a character positioned at the highest rank of candidate characters as a recognition result. 2. The character recognition device according to claim 1, wherein the initial candidate selection unit selects, as initial candidates, all the representative characters whose dissimilarity is smaller than a predetermined limit value. 3. The initial candidate selection unit compares the extracted feature vector and the feature vector of a representative character of the character recognition dictionary in a predetermined order of feature element components, and for each representative character, for each feature element component. The character recognition device according to claim 1 or 2, wherein the representative character that is an initial candidate is selected by adding a determination each time while performing the intermediate totalization of the dissimilarity. 4. The final candidate selecting section initializes only a fixed number of candidate characters in the descending order of the degree of difference among the representative characters selected by the initial candidate selecting section. The character recognition device according to claim 2 or claim 3. 5. The final candidate selection unit compares the intermediate tabulated value of the dissimilarity of each feature element component with the dissimilarity of the lowest rank character of the candidate character, and calculates the dissimilarity of the remaining feature element components. The character recognition device according to each of claims 1 to 4, wherein it is sequentially determined whether or not to continue. 6. A plurality of characteristic element components constituting a characteristic vector are four-direction elements indicating the directions of line elements of a skeleton pattern (or a contour line pattern of a character image) obtained by thinning a character image, and each characteristic element component 6. The character recognition device according to each of claims 1 to 5, characterized in that the order of performing the intermediate totalization of the dissimilarity is determined such that the processing in the vertical and horizontal directions comes before the processing in the diagonal direction. .