JPS5932080A - Recognizing system of character pattern - Google Patents

Abstract

PURPOSE:To improve the recognizing efficiency of character patterns, by comparing the distances between features of input character pattern with a standard pattern, studying the context between preceding and succeeding characters when plural candidate characters are produced and deciding the above-mentioned distance with the threshold value when it is impossible to decide a candidate character. CONSTITUTION:An unknown character pattern train 1 is supplied and binary coded 3 after a photoelectric conversion 2 to be stored in a storage circuit 4. A sorting circuit 5 extracts the feature quantity out of the data of the circuit 4 and compares it with a standard pattern to calculate a distance to an input character. Then plural characters are sorted 1 or selected and stored 6. A collating circuit 7 collates the data of the storage circuit 6 with the language knowledge (word, grammar, meaning, etc.) for decision 9. When it is impossible to decide the type of the character, a detailed discrimination 10 is carried out through a comparison of distance values and based on the data of the circuit 6. Then a recognized result 11 is delivered. Thus the decision and discrimination are carried out based on the threshold value after sorting and language processing. This reduced the processing quantity and improves the recognizing accuracy.

Description

DETAILED DESCRIPTION OF THE INVENTION (1) Description of the field to which the invention pertains The present invention utilizes linguistic knowledge regarding the shape characteristics of character patterns and character strings for character patterns in character pattern strings constituting words and sentences. It concerns the recognition method.

(2) Description of conventional technology Conventional character recognition methods for text input involve performing recognition processing on each input character pattern, including the identification of similar characters using shape characteristics, and then , when it is impossible to determine the character type, the character type of the input character pattern is determined by comparing the character strings before and after the target character with a dictionary of words, grammar, or meanings prepared in advance. It has been known.

However, this recognition method has the following drawbacks. That is, since character pattern recognition processing and post-processing using linguistic information were performed separately, recognition was impossible if a unique candidate character type for the character could not be determined based on linguistic knowledge. In addition, when identifying similar characters based on differences in shape between patterns, for example, it is possible to identify combinations of many character types that have very similar shapes, even though they are not valid as words or sentences. It is necessary to perform identification processing for.

The advantage of being able to narrow down candidates based on language knowledge is not taken advantage of,
Discrimination ability was not fully demonstrated.

(3) Purpose of the Invention In order to eliminate these drawbacks, the present invention performs post-processing using linguistic knowledge after recognition processing based on the rough shape characteristics of character patterns. If the pattern is not determined, recognition based on threshold determination and recognition processing based on fine shape differences between patterns are performed.The drawings will be described in detail below.

The figure shows one embodiment of the gR method for character patterns according to the present invention. In the figure, 1 is an unknown input character pattern string to be recognized, which is scanned and a photoelectric conversion circuit 2 obtains an electrical signal proportional to the shading of the character pattern.

This electric signal is compared with a reference value in the binarization circuit 3, converted into a black and white binary character pattern, and stored in the storage circuit 4.

5 is a classification circuit that realizes a known method (for example, Japanese Patent Laid-Open No. 54-34638), which first extracts the feature amount of the input character pattern, and then compares and matches it with a standard pattern to classify each of the input character patterns. Calculate the distance value for the character type. Furthermore, based on the distance value obtained here, a character or a plurality of character types that satisfy the predetermined criteria are selected as classification results and stored in the memory circuit 6.
Stack on. The following commonly used criteria are known.

d(CI)<do d(Cz)-d(Cx)≧d1(
1) Here, d(Ci) is the distance value for the character type Ci, and d(CI) r d(C riff d(Criku...). Also, dotct is a constant. (1) is satisfied. When a character type exists, it is determined as the recognition result. However, if there is no character sC1 that satisfies 1), d(Ci) - d(C1) < d2 (di is a constant) (II) is satisfied. Multiple characters 1m (Ci)
are stacked in the storage circuit 6 as the classification results.

The matching circuit 7 matches characters that can be linguistically connected to the previous and subsequent recognition results from among the character types of the classification results (Ci) outputted by the classification circuit 5 with the linguistic knowledge accumulated in the memory circuit 8. Selected based on the following. As linguistic knowledge, word knowledge, grammatical knowledge, semantic knowledge, etc. are used. The following processing is generally performed for this collation/processing. However, Ai (i=±1.±2....mountain) is the recognition result. Each Ct (1=1t 2....
..., N), consider N character strings. First, these character strings are compared with a word dictionary stored in the memory circuit 8. As a result, character strings for which no corresponding word exists are rejected.

Further, for character strings in which all characters correspond to words, a check is performed against grammatical knowledge and semantic knowledge. At this time, the character string is a word string...W-2W-1wow1w21・(wi:word)o
V). First, it is checked whether the word string Ov) matches the grammar stored in the memory circuit 8. Next, the semantic connection of each word is sometimes performed by checking it with a semantic dictionary.

Through the above matching process, a character string that becomes a correct sentence or word is extracted from the N character strings shown in Oi+). The character type Cj corresponding to this extracted character string is newly defined as an i-complementary character type group (Di).

(If Dip includes two or more characters, the determination process is performed again in the determination circuit 9. In other words, the memory circuit 9 determines whether the condition (1) is satisfied for [)i).
The determination is made using the distance value stored in .

If (1) is not satisfied in the determination circuit 9,
Detailed identification processing is performed in the identification circuit 10. The identification circuit 10 uses a known method (for example, Japanese Patent Application No. 55-00
This is an identification circuit based on the difference between patterns that realizes 6688), and performs detailed identification on a group of candidate character types (Di) and recognizes them based on character patterns.

An example of the above operation is shown below.

Example 1: Assume that "one side" is input as an unknown input character pattern string, and the classification circuit 5 outputs °C (word, tsume, edict 2, ki) as the classification result for "word". At this time, as a result of checking with the word dictionary in the checking circuit 7, “
Since the character strings ``word'' and ``single'' become words, the candidate character group is narrowed down to two (word, kiki).In the classification circuit 5, the distance difference between ``word'' and ``single'' is small. To (1)
is not satisfied, but if the distance difference between "word" and "ki" is large, it is recognized by the determination circuit 9.

Example 2: Assume that the unknown input character pattern string is "analysis". The classification results for "analysis" are (analysis, folding, material).

body, prayer). 6 Analysis ``The word ``disassembly'' is collated, and the (analyzed 2 bodies) are outputted as a candidate character type group from the collation circuit 7. If the criterion is not satisfied in the judgment circuit 9, the discrimination circuit 10 uses the difference in patterns to If all classification results are to be classified, "analysis" is performed.

It is necessary to carry out the process including discrimination between character types in which the difference in characters is relatively small, such as ``1 fold'' and ゛material''.

However, since the identification circuit 10 only discriminates between the six analyzes and bodies, which have relatively large differences, highly accurate recognition is not possible.

(B) Description of effects As explained above, in the character pattern recognition method according to the present invention, after classification and language processing, judgment processing and identification processing based on threshold values are performed, so if ambiguity remains due to language processing. However, it has the advantage of being able to perform recognition with high accuracy. In addition, since identification processing is performed only between candidate character types that have been sufficiently limited through classification and language processing, the amount of processing required for identification processing is reduced compared to the case where identification processing is performed on all candidate character types selected through classification processing. Another advantage is that it can be significantly reduced and effective processing can be performed.

[Brief explanation of drawings]

The figure is a diagram showing one embodiment of a character pattern recognition method according to the present invention. In the figure, 1 is an unknown input character pattern string, 2
3 is a photoelectric conversion device, 3 is a binarization circuit, 4 is a storage circuit, 5 is a classification circuit, 6 is a storage circuit, 7 is a collation circuit, 8 is a storage circuit, 9 is a judgment circuit, 10 is an identification circuit, 11 is a recognition circuit Show the results. Patent applicant Hiroshi Mori, patent attorney representing Nippon Telegraph and Telephone Public Corporation

Claims (1)

[Claims] In a character pattern recognition processing method that recognizes a character pattern string that constitutes a sentence character by character, the distance value between the feature extracted from the input character pattern and that of the standard pattern is calculated, and based on this, the a first means for selecting a candidate character type; and, when a candidate character cannot be uniquely determined by the first means, whether or not there is a linguistic connection between the candidate character type and the candidate character type for the preceding and succeeding character patterns;
A second means of verifying by comparing and comparing with linguistic knowledge such as words, grammar, and semantic dictionaries, and narrowing down the candidate character types to only those that are linguistically valid from the group of candidate character types; When a character type is selected, a means for performing recognition by determining a threshold value between the candidate character types based on the distance value obtained by the first means, and a third means for performing identification based on a difference in pattern between the character types. [Character pattern recognition processing characterized by comprising]A.