JP2906583B2 - Pattern identification device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a pattern recognition device such as a character recognition device.

[Prior Art] In a conventional pattern identification apparatus, first, a similarity between all standard feature data stored in a standard feature data storage unit and feature data extracted from a pattern is calculated, and a category to which each pattern belongs is determined. After deciding, correction and correction were performed using knowledge of words and the like as post-processing. For example, in "pattern recognition" (edited by Kenichi Mori, Corona Publishing Co., Ltd., p. 75), it is described as post-processing of character recognition.

[Problems to be Solved by the Invention] For this reason, it is difficult to determine a correct category from all the standard feature data in the standard feature data storage means, and the calculation time of the similarity becomes long. The time required for correction using knowledge is added to the time required for pattern identification, which is a factor that significantly increases the overall processing time. The emergence of a device that can perform pattern identification and correction processing using word knowledge in a short time Was desired.

According to the present invention, in order to solve such a problem, a frequently used category is registered in advance by using the fact that characters used are limited in identifying a specific format pattern such as a telephone number and a postal code in an input pattern. First, the matching process is performed only for the standard patterns of the most frequently occurring category, then the correction process is performed using the knowledge of the format of each specific pattern, and only the input patterns for which the correction could not be applied are compared with all the standard patterns. It is an object of the present invention to significantly reduce the number of times of collation with a standard pattern by performing processing, and to provide a pattern identification device having a high identification rate while executing the pattern identification device at high speed.

[Means for Solving the Problems] A pattern identification device of the present invention is a pattern identification device that separates an input pattern sequence into individual pattern signals, extracts feature data from the pattern signals, and identifies the feature data as a predetermined category. Standard feature data to be compared with the feature data, category string storage means for storing frequent category strings constituting words or the like extracted in advance, and storing a part of the standard feature data in the category string. Collation category selecting means for selecting based on the category column stored in the means; collating part of the standard feature data selected by the collation category selecting means with the feature data; First category identification means for determining a category candidate; and an identification result identified by the first category identification means. A category sequence determining unit that performs a comparison with the frequent category sequence and determines the feature data present in the frequent category sequence; and a part of the feature data other than the category data determined by the category sequence determining unit. A second category identification unit that performs matching including the standard feature data other than the standard feature data and determines a category candidate to which the feature data belongs.

Means for designating a specific category related to the frequent category column; determining whether the feature data belongs to the specific category; if the feature data belongs to the specific category, permitting this as a recognition result Specific category permitting means, wherein the second category determining means is the characteristic data other than the characteristic data determined by the category sequence determining means and the characteristic data permitted by the specific category permitting means. In this case, the matching is performed by including the standard feature data other than a part of the standard feature data, and a category candidate to which the feature data belongs is determined.

[Operation] In the present invention, an input pattern sequence to be identified is input to pattern input means, converted into an electric signal sequence, and stored in a memory. After the stored electric signal sequence is separated into individual signals by the pattern separating means, the characteristic data of the pattern is extracted by the characteristic extracting means. Matching of individual feature data extracted by the feature extracting means with standard feature data selected by the matching category selecting means from the standard feature data stored in the standard feature data storage means is performed by the category identifying means. Here, the category selected by the collation category selecting means may be the category used in the category string stored in the category string storage means plus the category used in the pattern of a specific format. Next, the category judging unit checks whether there is a category sequence that matches the category sequence stored in the category sequence storage unit, and flags a pattern of the matching category. Further, if the category column indicates a specific format, a pattern that matches the format is also flagged. Next, only the pattern for which the flag is not set is compared with the standard feature data of all the categories stored in the standard feature data storage means by the category identification means, and the category to which the pattern belongs is determined. Thus, all the categories to which the individual patterns in the pattern sequence belong are determined.

Embodiment FIG. 1 is a block diagram showing a configuration of an embodiment in which a pattern identification device according to the present invention is applied to character recognition.

The character string pattern input from the scanner, which is the pattern input means 1, is sent as a binary data signal to the character cutout device, which is the pattern separation means 2, and is separated into characters by extracting the circumscribed frame of each character. Next, feature data of each character is extracted by the feature extracting means 3. Based on the characteristic data, a character category which is most likely to belong to each character is determined by the category identification means 4,
It is stored in the pattern identification result storage means 7. However, at this time, among the character categories of the standard feature dictionary which is the standard feature data storage means 10, only the character category selected by the matching category selection means 8 is to be compared. Next, the identification result of the character string is determined by the category string determining means 5 to determine whether or not a word registered in the word dictionary as the category string storing means 9 exists in the character string. Is set up. Further, if the word is a word indicating a specific format, such as a telephone number or a postal code, the characters constituting the number portion in the specific format are also flagged. Finally, if the character string does not have a flag set, the matching process is performed on all character categories stored in the standard feature dictionary by the category identifying means 6 only for the character pattern. The character category to which the pattern belongs is determined. With the above configuration, it is possible to convert an input character pattern string into a character category string to which each character pattern belongs without checking against all character categories in the standard dictionary.

FIG. 2 shows an example of a configuration as hardware in which a pattern identification device for implementing the present invention is applied to character recognition. The contents of 20 are a CPU as a central control device, and a RAM and a ROM as memories. A scanner for optically scanning an object to be recognized, an auxiliary CRT such as a display CRT, a floppy disk device, a keyboard, and a bus line connecting these components are shown.

Next, "Head office TEL0263 (36) 1212" shown in FIG.
A detailed description will be given by taking recognition of the character pattern “representative” as an example.

About 100 words are registered in the word dictionary as the category string storage means 9.
It is assumed that only three words “L”, “Headquarters”, and “telephone” are registered [FIG. 4 (f)]. Each word has an attribute flag. For words such as "telephone" and "TEL", which are always followed by a digit string (number), the attribute flag is set to "1".

Step F1 in the flowchart of FIG.
The character string pattern input from the scanner (pattern input means 1) is separated one by one by the character cutout means (pattern separation means 2) in step F2. Steps
F3 moves the pointer to the "book" which is the first character in FIG. 4 (a).
Set to. In step F4, the features of the character pattern pointed to by the pointer are extracted (feature extraction means 3). In step F5, one character is read from the collation target character list shown in FIG. 4 (f) (collation category selecting means 8), and in step F6, standard feature data corresponding to the character is dictionary (standard feature data storage means 10). Read from In step F7, the similarity between the standard feature data and the feature data extracted from the character pointed to by the pointer is calculated, and the similarity is stored in the storage device. In step F8, it is checked whether or not the similarity has been calculated for all the character categories of the collation target character list. If any character category remains, the process returns to step F5.

In this way, the processing from steps F5 to F7 is performed for all the characters in the collation target character list. Characters in the collation target character list include characters used in two words registered in the word dictionary,
Parentheses are added. The result of calculating the similarity between these characters is stored in the storage device, and in step F9, the character having the highest similarity is selected (category identification means 4). In the case of “book”, “book” is naturally selected as the most likely character. In step F10, it is determined whether the character pointed to by the pointer is the last character. If not, the pointer is set to the next character "company" in step F11, and the process jumps to step F4. The result of successively recognizing characters in this manner becomes one of the collation target character lists as shown in FIG. 4 (b). In other words, the two characters “representative” are not in the collation target character list, and thus become “shaden”.

Next, the category string determining means 5 checks whether a character string that matches any of the three words “TEL”, “Headquarters”, and “telephone” exists in the category string of the recognition result. First, in step F12, a pointer is set to the first character. In step F13, it is checked whether a word starting from the character indicated by the pointer exists in the word dictionary. In this case, since "headquarters" matches, the character corresponding to "book" and "sha" is flagged in step F14, and the pointer is set to "sha" corresponding to the last character of the word in step F15. In step F16, it is checked whether the character pointed by the pointer is the last character of the document. If not, the pointer is set to the next character "T" in step F17, and the process jumps to step F13. When the pointer reaches the end of the last "den" at the last "den", the "connector B" in FIG. 3A is connected to the "connector B" in FIG. Go to F18. FIG. 4C shows the state of the character string at the stage where the processing up to this point has been completed. An underline indicates a character string determined to be a word, and a dot above a character indicates that the flag is set.

After setting the pointer to the first character in step F18, a word search is performed in step F19, and "head office" is found. In step F20, it is checked whether an attribute flag is set for the found word. In the case of "head office", since it has not been completed, the process proceeds to step F25, after performing an end check, returns to the search for the word in step F19, and finds "TEL" as the next word. Since the attribute flag of "TEL" is set, a numeric string to be the next number should follow. Therefore, in step F21, the pointer is set to the next character of the word, and the process proceeds to check the number portion. In step F22, if it is a character such as an arithmetic numeral, a Chinese numeral, a parenthesis, or the like, it is determined that the character is a number portion, and in step F23, the character is flagged.In step F24, the pointer is set to the next character. Return to If it is not an arithmetic numeral, a Chinese numeral, or a parenthesis in step F22, the process proceeds to step F25, in which it is determined whether the character pointed to by the pointer is the last character. If not, the process returns to the word search in step F19. If it is at the end, the category column determination processing ends. FIG. 4 (d) is a diagram showing the end state of the category string determination process, and a flag is also set in the character of the telephone number in comparison with the state of FIG. 4 (c).

Finally, the “representative” part is recognized by the category identifying means 6. In step F26, the pointer is set to the first character. In step F27, check whether a flag is set for the character pointed to by the pointer.
After checking if it is the last character in F31, if it is not the last character, the pointer is set to the next character in step F32 and the process returns to step F27. In the case of FIG. 4 (d), since the flag is set from the first "book" to the last "2" of the telephone number portion, the pointer advances to "shaden" of "shaden".
If the flag is not set, the feature of the character pointed to by the pointer is extracted in step F28, and after calculating the similarity with all the standard feature data registered in the standard feature data storage unit 10 in step F29, In step F30, the standard feature data having the highest similarity is selected. In step 31, it is checked whether "sha" is the last character, and since it is not the last character, the pointer is moved to the next character "den" in step 32.
And returns to step F27. Since no flag is set for "den", the recognition processing in steps F29 and F30 is performed, and as a result, "shaden" is correctly recognized as "representative" [FIG. 4 (e)]. In the above example, for simplicity, the word dictionary assumed that there were three words, “TEL”, “Headquarters”, and “telephone”. However, when actually applied to business card recognition, the number of words was 100, The number of characters is about 200 characters.

[Effects of the Invention] According to the present invention, the number of categories used in the category strings stored in the category string storage unit is about 100 to 200, which is one tenth of the number of categories 3000 in the standard dictionary. ,
Category string part (word part) existing in input pattern string
By significantly reducing the number of matching calculations in the recognition of words, it is possible to execute recognition at high speed, and by reducing the number of matching categories in the recognition of word parts, it is possible to improve the recognition rate of word parts. became.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a pattern identification device according to the present invention. FIG. 2 is a hardware configuration diagram of the pattern identification device according to the present invention. FIG. 3 is a flowchart of a character recognition process according to an embodiment of the present invention. FIG. 4 is an explanatory diagram showing a recognition process in a character recognition process according to an embodiment of the present invention. 1: pattern input means 2: pattern separation means 3: feature extraction means 4: category identification means 5: category sequence determination means 6: category identification means 7: pattern identification result storage means 8: collation category selection device 9: category sequence storage device 10: Standard feature data storage

Claims (2)

(57) [Claims] 1. A pattern identification apparatus for separating an input pattern sequence into individual pattern signals, extracting feature data from the pattern signals and identifying the feature data as a predetermined category, wherein a standard feature to be compared with the feature data Data, a category string storage means for storing frequently-used category strings constituting words or the like extracted in advance, and a part of the standard feature data based on the category strings stored in the category string storage means. Matching category selecting means to be selected; a part of the standard feature data selected by the matching category selecting means; and first category identifying means for comparing the feature data and determining a category candidate to which the feature data belongs. Comparing the identification result identified by the first category identification means with the frequent category column, A category sequence determining means for determining the feature data present in the category sequence; and for the feature data other than the category data determined by the category sequence determining means, the standard feature data other than a part of the standard feature data is also included. And a second category identification means for performing a comparison to determine a category candidate to which the feature data belongs. 2. The pattern recognition apparatus according to claim 1, wherein: means for designating a specific category related to the frequent category sequence; and determining whether the feature data belongs to the specific category. And a specific category permitting unit that permits the recognition as a recognition result when belonging to the specific category. The second category determining unit includes the feature data and the feature data determined by the category sequence determining unit. The feature data other than the feature data permitted by the specific category permitting unit is collated with the standard feature data other than a part of the standard feature data, and a category candidate to which the feature data belongs is determined. A pattern identification device characterized in that the pattern identification device is a means for performing pattern identification.