JPS61198380A - Character recognition equipment - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a character recognition device, and particularly to a handwritten character recognition device that can perform classification and identification processing at high speed.

Disclosure Types Input characters are rapidly classified and identified by hardware by extracting multiple different types of features for each input character and comparing these features to the features of each standard character. Characteristics of input characters and standard characters are examined for each type. Each type of feature contains multi-dimensional feature components,
Feature components of the same type and corresponding dimensions are sequentially compared to check the degree of discrepancy. Each time feature components of corresponding dimensions within the same type are compared, the degree of mismatch is accumulated,
Each time, the accumulated results are compared with the threshold set for that type. When the cumulative result exceeds the threshold, at that point, the comparison with the current standard character is stopped and the comparison with the first standard character is skipped. Therefore, there is no need to always perform comparison calculations on all feature components within the same type, and the one-class identification process can be speeded up.

In addition, the characteristics of standard characters are grouped by character category and stored serially in dictionary memory, and characters that do not belong to a specific category such as blank characters or are used in combination with other characters are distributed in multiple locations. be done. Therefore, by specifying the search start position and search end position in the dictionary memory, the recognition target character range can be specified.

Furthermore, recognition processing is facilitated by treating a blank character as one standard character.

[Prior Art] Japanese Patent Laid-Open No. 58-201184 discloses a handwritten character recognition device that uses three different types of features to perform character recognition in three stages: major classification I, major classification ■, and identification. O
Input characters read by a CR or the like are binarized, normalized, and thinned. Feature extraction is performed by scanning thinned character patterns. The three types of features are the amount of character lines in each direction in the character pattern.

These are the amount of depth from the four sides of the character pattern frame to the actual character line, and the amount by direction and position of the character line. In major classification 1, the input character pattern and standard character pattern are classified based on the amount of character lines in each direction. The final candidate characters are selected based on the character dose.

The technique disclosed in the above-mentioned Japanese Patent Application Publication No. 2003-120002 achieves high recognition accuracy using a relatively simple algorithm. However, in general,
When handwritten characters including Chinese characters are to be recognized, there are many types of characters, so the time required for classification and identification is much longer than for pre-processing such as 5-normalization and thinning. In the above-mentioned JP-A-Sho, the first narrowing down is performed at the stage of major classification 1, the second narrowing is performed at the stage of major classification ■, and finally the final candidates are selected at the identification stage. Narrowing down results in a considerable speed-up of processing. However, each of the above three types of features includes feature components of many dimensions, and in the above-mentioned patent application, the error from the standard pattern for each of the feature components of many dimensions included in one type, i.e. Since the total sum of the degree of inconsistency is compared with the threshold of the relevant type at the stage of calculating it, even if only the degree of inconsistency of the feature components of some dimensions exceeds the threshold, the sum is calculated, which is wasteful.

Furthermore, there is a problem in that the classification and identification processing becomes slow when processing is performed only by software. Although it is possible to implement the classification and identification mechanism in hardware, the above three types of features are each expressed by calculation formulas that include different operations, so if this were directly implemented in hardware, it would require an enormous amount of hardware. and the control becomes complicated.

In addition, in order to improve single character recognition accuracy and speed up recognition processing, the character type of the standard character pattern to be compared is specified according to the input character type, and characters other than the specified character type are compared or selected as candidates. It is common practice to ensure that this does not occur. There is a primary method of conventional character type selection.

(b) A method in which a code that identifies the character type (alphabet, number, kana, etc.) is added to a standard character pattern, and the standard character pattern with the specified code is targeted for recognition.

(b) A method of using a character type selection table to set a flag specifying the character type to be used for each field of an OCR form, and then referring to the character type selection table to identify characters to be recognized for each field.

However, method (a) requires adding an extra bit field to each standard character pattern, and requires extra hardware to decode this bit field. The disadvantage was that the processing speed was slow due to fetching and decoding. On the other hand, in the case of method (b), the capacity of the character type selection table is
For conventional alphanumeric, kana and symbol OCR, 4096 bits (=512 bytes) were sufficient for 128 character types x 32 fields, but for Kanji OCR, at least 2000 character types x 32 fields = 64000 bits (=8000 bytes) are required. This not only increases the burden on the hardware, but also requires access to a character type selection table and control of access to standard patterns using the table, resulting in a problem of reduced processing speed.

[Problems to be Solved by the Invention] Accordingly, an object of the present invention is to provide a character recognition device that can perform recognition processing at high speed.

Another object of the present invention is to provide a character recognition device capable of high-speed processing in which character types to be recognized can be specified flexibly and easily.

Another object is to provide a character recognition device that can perform classification and identification processing using a plurality of types of features at high speed with simple hardware.

Measures for Solving Problem c] Character recognition is performed by comparing features extracted for input characters with features of standard characters. Comparisons are made by type of feature. A feature of one type includes feature components of multiple dimensions, and feature components of the same type and corresponding dimensions are sequentially compared to check the degree of inconsistency. Each time feature components of corresponding dimensions are compared, the degree of mismatch is accumulated, and the accumulated result is compared each time with a threshold set for that type. When the cumulative result exceeds the threshold, at that point the comparison with the current standard character is stopped and the comparison with the next standard character is skipped.

Dictionary memory that includes features of standard characters has characters that are not specified in character categories, such as blank characters, distributed in multiple locations as one standard character pattern. You can easily specify the target character range.

[Example] The flow of processing in the recognition device of the present invention will be explained with reference to the flow chart of FIG.

First, the characters are read by a scanner. In the next preprocessing, each character is separated from the read pattern, and processing such as normalization and thinning is added as necessary. Since the normalization and line thinning processes are well known, the details will be omitted. In feature extraction, feature parameters for use in the next classification identification are extracted from the single character pattern processed in this way.

The following classification and identification mechanism is an improved part of the present invention.

In order to facilitate understanding of the present invention, first of all, the above-mentioned Japanese Patent Application Laid-Open No.
The feature extraction process shown in No. 201184 will be briefly explained, and the present invention will be explained below using the features extracted by this process as an example.

In the above-mentioned Japanese Patent Application Publication No. Sho, recognition processing is performed in three stages: major classification (2), major classification (2), and identification. In major classification I, the amount of 0 character lines in each direction in a thinned 60x60 dot character pattern is used as one type of feature parameter. , a connected element consisting of two adjacent black dots is used as the minimum unit of a character line, and the total number of connected elements is calculated for each direction of connected elements. There are four types of connected elements: horizontal, vertical, downward to the right (or upward to the left), and downward to the left (upward to the right), and the number of connected elements in each direction is totaled.

In Major Category I, the distance defined by the following formula is calculated from the four connected primes accumulated for each direction and the corresponding connected primes of each standard character, and this is the threshold set for this type of feature. Character categories that exceed the above will be dropped.

The meanings of variables and constants in formula (1) are as follows.

i = 4-way number of connected elements (1 to 4) w, = weighting coefficient considering the length of connected elements (2 for horizontal and vertical connections; 3 for downward right and left downward connections) g, = Direction-specific connection factor of the input character pattern f+ = Direction-specific connection factor of the standard character pattern relative to g+ = Weighted total connection factor of the input pattern α = Constant to reduce the influence of dose (generally ,2
(approximately 00) The characteristics of the type used in major classification ■ are, in this case, i = 1 ~
Contains 4 four-dimensional feature components.

In the major category (2), the depth from each side of the character pattern to the character line is used as the second type of feature parameter. 60
Divide each side of a character pattern area of 60 dots into, for example, 6 sections, and as each side moves toward the opposite side, it hits the first black dot of the character line or reaches a certain depth. The total number of white dots counted up to that point for each section (corresponding to the area) is used. This second
In this case, the type of feature has a feature component of 4 sides x 6 sections = 24 dimensions.

In the major classification ■ of recognition processing, the distance distance defined by the following formula is calculated for two or more character categories left in the major classification ■, and if this distance is greater than the threshold set for this type of feature. Character categories that have changed are dropped.

In equation (2), g+J represents the total value of the depth in the j-th section of the i-th side of the input character pattern, and f+
J represents the corresponding value of the standard character pattern.

In the final identification, the amount by direction and position of the character line is determined by the third
Used as a type feature. This feature is, for example, 60
O@, 22.
5°, 45m, 67.5°.

Eight look directions of 90@, 112.5', 135'' and 157°5° are set, and the character pattern is divided into 12 sections for each look direction, horizontal, vertical, The total value of connected elements in four directions, downward to the right and downward to the left, is calculated.

Therefore, the third type of feature includes feature components of 4 connection directions x 8 look directions x 12 partitions = 384 dimensions in this case.

In the identification process, a distance DI defined by the following equation is calculated for two or more character categories left in the major classification ■, and the category with the smallest value is output as a recognized character.

In equation (3), g+J, t is i of the input character pattern
represents the connected prime direction, jth look direction, and connected prime quantity in the th partition, f+J11 represents the value of the standard character pattern corresponding to g+Jk, and W+J represents the i-th connected element direction and the jth look direction. It represents the weight in the direction.

In the above-mentioned Japanese Patent Laid-Open Publication No. 2003-120023, the distances DI and D are compared with threshold values at the stage of calculating the sum of errors for all feature components of each type in the major classification I and the major classification ■. in fact,
Errors in some feature components alone may exceed the threshold.
It is a big time loss to constantly calculate all feature components. It is also possible to implement the above classification identification process of JP-A-Sho with hardware, but equations (1), (2) and (3)
), since the calculation format of each process is different, if it is directly implemented in hardware, it is necessary to prepare separate hardware for each process, making the device expensive and the control complicated.

By the way, Equations (1), (2), and (3) are common in that they calculate the sum of the absolute values of the difference values for each feature component of the input cover pattern and the standard pattern, and Equations (1) and (3) are By transforming as shown in equations (4) and (5) below, respectively,
Equations (1) to (3) can be converted into a common arithmetic format.

i=L j=1 k==1 Since Wl and W+J are weight constants uniquely determined by i, i, and j, respectively, the extracted feature components g2 and g
+Ji+ as each weight constant W 1% W 1 Jl
The values W+g+ and W+Jg+J corrected by
, and W+Jf+Jk in the dictionary memory, and the sum of the absolute values of the difference values of these corrected feature components may be calculated at the time of major classification I and identification processing. The feature components of major category (■) are not modified, and g+J, f, and J are directly used. Therefore, by using weighted values as feature components for major classification I and identification, it becomes possible to perform all operations using only addition and subtraction.
The burden on hardware can be significantly reduced, and when each type of processing is executed in series, common hardware can be used, further simplifying the hardware.

FIG. 1 shows an embodiment of a classification and identification mechanism according to the present invention, which is constructed based on the above idea. The input character pattern is read by the character reading section 2, subjected to preprocessing such as normalization and thinning in the preprocessing section 4, and processed into the above-mentioned major classification I, major classification ■, and identification by the feature extraction section 6. Three types of features are extracted. Although any technique may be used for feature extraction, the techniques disclosed in Japanese Patent Application Nos. 1970-1981 and 1970-16-1988 may be advantageously used for extracting features for major classification (1) and identification, respectively.

Further, as the normalization technique, the technique disclosed in Japanese Patent Application No. 148725/1988 can be advantageously used.

The extracted features are stored in a feature buffer memory 8 for each type.
However, as mentioned above, the major classification I and the feature components for identification are stored in the feature extraction unit 6 as W+
The weighting is corrected to the form g Is W + Jg + Jk,
This modified value is stored in the feature buffer memory 8 as a feature component of the major category (2) and identification type. The feature component g+J of the major category ■ is stored as is.

In addition, in the above-mentioned JP-A-Sho, (4 dimensions of major classification I) + (24 dimensions of major classification ■) + 384 dimensions by Ctft), a total of 4
Although 12-dimensional feature components are used, in this example, the look directions for identification are O'', 45@, and 90''.

The cylinder is abbreviated in 4 directions of 135 degrees (j = 1 to 4), so in this example, (4 dimensions of major classification ■) + (2 dimensions of major classification ■)
A total of 220 dimensional feature components (4 dimensions) + (192 dimensions for identification) are used.

On the other hand, the corresponding feature components of each type extracted for the standard character pattern are stored for each character in the dictionary memory 10 configured as a RAM. Of course, the major classification I and identification feature components are each weighted in the form W+f+% W+
It is stored as Jf+Jw, and the feature component (■) of major classification (■) is stored as is in the form of LJ.

Here, the configuration of the dictionary memory 10 will be explained. Third
The figure shows an example of the arrangement of standard character patterns in the dictionary memory 10. In the vertical direction, standard pattern dictionaries for all the characters (N) to be used are arranged in series, and in the horizontal direction, each character 256 bytes are allocated for each, and the 4-dimensional features of major classification I (w+fI) and the 24-dimensional features of major classification ■ (LJ
) - 192-dimensional features of identification (W + Jf + J) are stored in 1 byte per dimension, followed by a 2-byte character code that identifies the standard character pattern, and the remaining 34
The byte is a reserved area.

One feature of the configuration of the dictionary memory 10 is that the identification target range (S-E) of the standard character pattern to be compared with the features stored in the feature buffer memory 8 can be specified simply by specifying the starting character position and the final character position. The point is that the layout of the standard patterns has been devised to make it possible to do so. FIG. 4 shows a specific example of the arrangement of standard character patterns. This is for example JI
S
Special characters not in the IS 8-bit code system (×.

÷, eyes, 〒, etc.), the kana symbol in item number 6 is the special character for kana [11,,・ (midpoint), r,'', - (long sound)],
The half-width symbol in item number 14 is item number 0, 4, which is a special character (:y; t”to, etc.) included in the JIS 8-bit code system.
．． 8.13 is a standard pattern for white space characters.

Conventionally, the recognition process for blank characters (spaces) is skipped and special processing is performed as spaces, but since blank characters (spaces) appear frequently on forms, recognition processing and hardware are required. This was causing complications. The same preprocessing and feature extraction are performed regardless of whether the read character is a blank character, and the blank character is
By using it as a standard pattern for one character category, characters can be recognized using common processing without having to be aware of blank characters. In addition, blank characters are grouped into categories (items 1 to 3).
, item numbers 5 to 7, item numbers 9 to 12, and item number 14) as one character category, the search range can be specified only by the search start character position and the search end character position.

FIG. 5 shows an example of selecting the character categories shown in FIG. 4 depending on the type of character. The standard pattern is stored in the dictionary memory 10 in character series, so by specifying the first character position and last character position of the character category range to be selected, the character type can be selected by scanning the recognition target range once. Classification and identification can be performed.

How to classify the character categories stored in the dictionary memory 10, how to arrange each character category in the dictionary memory, and where to insert blank characters are determined by
It can be optimally selected depending on the OCR used. Also, placing the same standard pattern in multiple locations is not limited to blank characters; characters that are used regardless of character type or used in combination with other characters may also be placed in multiple locations. Possible 6 For example, if there is a request to read the yen symbol "¥" in the numeric field of a form, the standard pattern for the ¥ symbol should be placed in the category number 10 in addition to the category number 14 in Figure 4. Also include it in
Alternatively, the ¥ symbol can be inserted as one character category next to the category of item number 10.

Note that if the recognition target consists of hiragana and numbers, for example, and the standard pattern is arranged as shown in Figure 4, it can be processed by specifying items 5 to 9, but the katakana category will also be included, so there will be no waste. You will have to do some processing. In such a case, it is possible to perform the identification twice and merge the results, but if RAM is used as the dictionary memory, the standard pattern can be flexibly optimized depending on the usage status. Can be rearranged.

Now, returning to FIG. 1 and continuing the explanation, when starting the actual recognition process, it is necessary to set predetermined parameters before starting this classification and identification mechanism. it is,
These are the identification target range (S to E in FIG. 3) and the threshold value. The address S of the first character position of the range to be identified containing the desired character category is loaded into the start pointer 22, and the character position E+1 next to the last character position of the range to be identified is loaded into the end pointer 24. Also, major classification I, major classification ■
and thresholds in identification, i.e., equations (1'), (2)
, (3') are thresholds THI and TH that define the upper limit of the right-hand side value.
2 and TH3 are loaded into threshold registers 40, 42, and 44, respectively.

Then, the feature components of each type extracted from the input character pattern (w+g l+g
+J, W+Jg +J11) and then starts the classification identification mechanism, the value of the start pointer 22 is first loaded into the pointer H12. Pointer H12, in combination with pointer L14, generates an address for accessing dictionary memory 10. Pointer L14 is 0~
221 and is connected via line 16 to feature buffer memory 8 and dictionary memory 10. The pointer L14 points to the feature memory 8, as will be described later.
Each time the step of comparing the corresponding pair of feature components of the standard pattern specified by the input cover turn and the pointer H12 is completed, the pointer L is incremented by 1 by the increment signal L+1 from the timing/control circuit 28. , feature buffer 8 and dictionary memory 10
, so the pointer L14 points to the line 16 at the address 0 of the 220-dimensional feature component.
Gives ~219. Pointer L14 further points addresses 220 and 221 on line 1 to read the 2-byte character code of the standard pattern when the standard pattern is determined as a candidate character as a result of the comparison of the 220-dimensional feature components.
Occurs on 8th.

The pointer H12 is a counter having a count value sufficient to specify all characters stored in the dictionary memory 10,
If a candidate character is found as a result of the comparison of the 220-dimensional feature components, it is incremented by 1 by the overflow output from pointer L14, and the standard pattern currently being compared is incremented by 1 by the overflow output from pointer L14. If it is found that the character cannot be a candidate character, it is incremented by 1 by the pointer bi-increment signal H+1 from the timing/control circuit 28. The count value of pointer H is applied to dictionary memory 10 via line 20 to select one character.

In operation, the first feature component of the feature buffer 8 read by the pointer L14 (the feature component of the first dimension of the major classification ■) and the corresponding feature component of the standard pattern specified by the pointer H12 are processed by the difference absolute value calculator. 32. The output of the absolute difference calculator 32 is added to the output of the addition register 36 by an adder 34. The addition register 36 stores the count O14 of the pointer L14 at the start of comparing the feature components of each type (major classification I, major classification ■, identification).
, 28.

At the start of operation, since pointer L=O, addition register 3
6 has become 0, and the output of adder 34 is then loaded into addition register 36.

The output of the addition register 36 is applied to one input of a comparator 38, and the other input of the comparator 38 is applied to a multiplexer M.
The output of PX46 is given. Multiplexer 46 is responsive to the output of pointer L14 with a count of 0.4.28;
When comparing the feature components of major category I (pointer L=O to 3), the threshold THI is gated to the comparator 38, and when comparing the major category ■ (pointer L=4 to 27), the threshold TH2 is gated. , the threshold value TH3 is gated to the comparator 38 during identification comparison (pointer L=28 to 219).

Comparator 38 compares the value of addition register 36 with multiplexer 4.
Comparator 38 provides an output signal to timing/control circuit 28 when the value of addition register 36 exceeds threshold value THI. Timing/control circuit 28 is responsive to this signal;
A reset signal R is generated to reset the pointer L14, and a pointer H increment signal H+1 is generated to increment the pointer H12 by 1.

Therefore, as soon as an output from the comparator 38 indicates that the standard pattern currently being compared is significantly different from the input cover pattern, the next standard pattern is accessed;
A comparison between the input cover pattern and the next standard pattern is started.

If no output is generated from comparator 38, timing/control circuit 28 continues normal operation, generating a pointer increment signal L+1 to advance the count of pointer L by 1, and increasing feature buffer 8 and dictionary memory 10.
The output of the adder 34 is accumulated in the addition register 36 and compared in the comparator 38 for each comparison. When the threshold value THI is exceeded, as described above.

Immediately move on to comparison with the next standard pattern.

If the cumulative difference value of the addition register 36 does not exceed the threshold THI even after the comparison of the four-dimensional feature components of major classification ■ is completed, the addition register 36
is cleared to 0, and the multiplexer 46 sets the threshold TH
2 is selected, and comparison of the major classification ■ is started for the current standard pattern. In major classification (2), comparison calculation processing steps are sequentially performed for 24-dimensional feature components.

When the accumulated difference value of the addition register 36 exceeds the threshold TH2 during the period of major classification H, the comparison with the next standard pattern is immediately started, and if it does not exceed the threshold TH2, the pointer L
The addition register 36 is reset by the count 28 of
The multiplexer 4° selects the threshold value TH3, enters the identification process, and sequentially repeats the comparison calculation process steps for the 192-dimensional feature components.

If the cumulative difference value exceeds the threshold TH3, this means that the current standard pattern cannot be a candidate character, and the comparison with the next standard pattern is immediately started. However, the threshold TH
If the count does not exceed 3, the timing/control circuit 28 increments the pointer L14 and sets the count value 22 on the line 18.
0,221 is given, the 2-byte character code of the standard pattern is read from the dictionary memory 10, and the character code and the accumulated difference value of the addition register 36 are stored in the result buffer memory 50 as a distance value. Counter 4B provides an address to result buffer memory 50 and is incremented by the final count 221 of pointer L14 to point to the next storage address in result buffer memory 50.

Also, the overflow output of pointer L is pointer H1
2, it is incremented by 1 and a similar comparison step is performed for the next standard pattern starting from major category I.

The operation is repeated for each standard pattern in order to search for all candidate characters in the designated identification range, and the standard patterns that have passed the identification process and their distance values are sequentially loaded into the result buffer 50. The count value of pointer H12 is the comparator 2
6, and when the count value of pointer H12 becomes equal to the set value of end pointer 24, comparator 2
6 is the timing to end the search for the specified identification target range/
The control circuit 28 is notified and the classification identification process is terminated. At this time, the result buffer 50 contains one or more candidate characters, and the contents of the result buffer 50 and the value of the counter 48 (
(indicating the number of candidate characters) are read by a control processor (not shown), sorted in descending order of distance value, and the first candidate character is output to a printer, CRT (not shown), or the like.

By selecting a threshold value to be set in the threshold value register 40, 42, 44, rapid narrowing down can be performed.

In addition, in the above embodiment, the common hardware is classified into major categories I,
The feature components of the three types of major classification ■ and identification are processed in series, and this allows the hardware to be extremely simplified, but if desired, multiple absolute difference calculators 32 can be provided in parallel. By performing the processing, it is possible to improve the processing speed. For this we need feature buffer 8
After reading a plurality of data simultaneously from the dictionary memory 10 and calculating them with a plurality of absolute difference calculators,
A circuit is required to add the results.

Other parts may be almost the same. For example, two absolute difference calculators are provided, and the sum of these difference value outputs is added to the adder 34.
If the threshold value is exceeded, the processing speed will be doubled, and the processing speed can be further improved without increasing the complexity. The benefits of scoping can be maintained.

[Effects of the invention (1) Comparison of each feature component of multi-dimensional feature components that are continuously processed. Compare the cumulative difference value with a threshold value at each processing step, and when the threshold value is exceeded, compare with the current standard pattern. Canceled,
By immediately skipping to comparison with the next standard pattern, processing speed can be increased.

(2) Major classification ■ and ■ for the purpose of narrowing down candidate characters and identification processing for the purpose of recognition can be performed on the same hardware, and hardware simplification and processing speed can be realized.

(3) By configuring the dictionary memory so that the recognition target range can be specified using only the search start address and search end address, the search range can be narrowed down without the need for special additional circuits for character type selection. This can save you a lot of calculations.

(4) By handling blank characters without distinguishing them from normal characters, special processing and hardware for blank characters are not required, and hardware and control can be simplified.

[Brief explanation of the drawing]

Fig. 1 is a diagram showing an embodiment of the present invention, Fig. 2 is a flow chart showing the flow of character recognition processing, Fig. 3 is a diagram showing the configuration of a dictionary memory, and Fig. 4 is a diagram showing a standard pattern in the dictionary memory. □A diagram showing the arrangement of turns, and FIG. 5 are diagrams showing a selection configuration of standard patterns in the dictionary memory. 8...Feature buffer memory, 10...Dictionary memory, 12.14...Pointer, 22...Start pointer, 24...End pointer, 28...Timing/ Control circuit, 32... Absolute difference calculator,
34... Adder, 36... Addition register, 38
... Comparator, 40.42.44 ... Threshold register, 46 ... Multiplexer, 50 ... Result buffer memory. Applicant International Business Machines Corporation Agent Patent Attorney Jinro Yamamoto (1 other person) Figure 2 Figure 4

Claims (4)

[Claims] (1) A dictionary memory means storing at least one type of feature including multi-dimensional feature components for each standard character, and storing features to be compared with the above-mentioned features extracted for the input character to be recognized. feature memory means; address means for sequentially reading the feature components stored in the feature memory means in order of dimension, and sequentially reading feature components of corresponding dimensions from the dictionary memory means for each standard character to be compared; a first comparing means for sequentially comparing feature components of corresponding dimensions read from the feature memory means and the dictionary memory means and generating a comparison output indicating a degree of mismatch; threshold means for providing a threshold value set for the type of feature component being compared by the first comparison means; a second comparing means for comparing the second standard character with a threshold, and in response to the output of the second comparing means, when the output of the accumulating means exceeds the threshold, a comparison is made with the next standard character in the dictionary memory means; a control means for skipping access of the address means so as to allow the address to be accessed by the address means; (2) The dictionary memory means each stores a plurality of types of features including feature components of multiple dimensions, and the address means sequentially stores feature components of corresponding dimensions from the feature memory means and the dictionary memory means in order of type and order of dimension. reading, said threshold value means comprising register means for holding a threshold value set for each type and means for selecting a threshold value associated with the type of feature component currently being compared by said first comparison means; A character recognition device according to claim (1). (3) The dictionary memory means stores characteristics of the standard characters to be used grouped by character category in character series, and includes at least one character in a plurality of locations, and the address means specifies a search start position. Claim 1 or 2, characterized in that it has a pointer and an end pointer that specifies a search end position, and the features of standard characters are accessed within the range specified by both pointers. Character recognition device. (4) The character recognition device according to claim (3), wherein the characters arranged at the plurality of locations include blank characters.