JPS6344287A - Character reader - Google Patents

Abstract

PURPOSE:To increase processing speed by deciding the type of character that will be read next time from the code of a read character and eliminating the need that all patterns should be collated. CONSTITUTION:A pattern collating means 10 outputs a distance value obtained by collating the pattern of an input character pattern and the character code of a corresponding standard pattern. A standard feature vector storage means 11 stores a standard feature vector generated from the standard pattern of a character to be read. A read-out result output means 12 evaluates the distance values of plural standard patterns obtained by the pattern collating means 10 and outputs final read-out results of the input character pattern. A collation candidate selecting means 13 selects the type of character to be collated for a character to be read next time and by judging from the connecting relation between mutual characters comprising a word or sentence to be read and the code of the read character, inputs the corresponding standard feature vector to the pattern collating means 10, and selectively collates patterns.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to a character reading device that performs pattern matching processing, and particularly to a character reading device that efficiently performs pattern matching processing when reading continuous characters such as words and sentences. The present invention relates to a character reading device that improves speed and reading accuracy.

[Conventional technology]

2. Description of the Related Art Character reading devices are used as means for high-speed input of large amounts of handwritten or printed documents and forms into computers and the like. This type of character reading device generally performs preprocessing on a photoelectrically converted character pattern, which consists of noise removal, normalization of position and size, etc., and then performs a preprocessing process based on the preprocessed character pattern based on a predetermined algorithm (procedure). A feature extraction process that generates an input feature vector using t1m, and a distance calculation between the standard feature vector generated from a standard pattern based on the character type targeted for t1m and the input feature vector obtained by the feature extraction process. Post-processing is performed to determine the character code that will be the reading result based on the pattern matching process and the distance value for each e feature vector obtained by the pattern matching process.

These processes exclude pattern matching processing (pre-processing, feature extraction processing, and post-processing can be processed with the amount of calculation proportional to the number of characters read, but pattern matching processing uses input feature vectors and all input features for each input character). Since the distance is calculated between standard feature vectors, a processing amount proportional to (number of characters read) x (number of types of standard pakun to be read) x (number of dimensions of feature vectors) is required.

In particular, character reading devices that target handwritten kanji, etc. use hundreds of dimensional feature vectors to read thousands of types of complex character patterns unique to kanji with high precision by absorbing the deformation caused by handwriting. It is necessary to perform matching with the PJ quasi-feature vector, and the amount of calculation required for pattern matching processing to find the distance between vectors is enormous.

On the other hand, conventional character reading devices perform brute force pattern matching between the input feature vector and the standard feature vectors of all categories to be recognized. It has been difficult to discriminate similar characters such as "-" and the long kana "-" with high accuracy. Therefore, in order to distinguish such similar characters, when filling out the form in advance, for example, "0" for Alphabea)
Something like adding a line above the .

As a result, Li! Although it is possible to improve the m rate, expanding the range of characters to be recognized to include Kanji will dramatically increase the number of similar characters, making it difficult in practice to distinguish between similar characters when filling out forms. .

[Problem that the invention seeks to solve]

In conventional character reading devices, each input character is recognized as 4! The amount of calculation is proportional to the product of the number of types of quasi-patterns and the number of dimensions of the feature vector, and when dealing with handwritten kanji, etc., the number of dimensions is large (and a large number of standard feature vectors and input features are required). Since distance calculations are performed between vectors, the amount of calculation required for pattern matching becomes enormous, and there is a problem in that the pattern matching process limits the reading speed of the entire device.

On the other hand, since a round-robin pattern matching process is performed between the input feature vector and a standard feature vector generated in advance from a standard pattern for each character to be recognized, it is possible to use There was a problem in which it was difficult to accurately distinguish similar characters such as .

[Means for solving problems]

In the character reading device of the present invention, a table is prepared in advance that stores the connection relationships between characters constituting a Japanese sentence or word to be read, and in the process of reading consecutive characters, the reading is performed first. The present invention is characterized in that it includes a matching candidate selection means that searches a table showing connection relationships prepared in advance from the character code of the character read, and determines the type of character to be matched for the next character to be read. Distance calculation is performed by pattern matching between the standard feature vector corresponding to the character to be matched obtained by the matching candidate selection means and the input feature vector, and the obtained distance value is evaluated to determine whether it is sufficient for the input character pattern. Defeat! ! When a quasi-pattern is detected, the character code corresponding to the standard pattern is output as a reading result, and the subsequent matching process is terminated.

[For production]

A character reading device with the above characteristics reads a word or sentence consisting of multiple consecutive characters, and estimates in advance the type of character to be matched against the next character to be read from the character code of the next read character. Therefore, the character reading device of the present invention performs selective pattern matching processing between a standard feature vector and an input feature vector. There is no need to perform matching on a hit basis, and the amount of pattern matching processing, which was a factor that limited processing speed in the past, can be reduced. Furthermore, if multiple similar characters exist, the matching process is proceeded from the character type with the highest degree of connection, and the matching process is stopped when a sufficiently matching standard pattern is detected by evaluating the degree of matching with the input character. Even if it is, the connection M obtained from the connection relationship between characters! It is possible to reduce reading errors for similar characters.

〔Example〕

An embodiment of the present invention will be described below with reference to the drawings. Although not shown in Figure 1, which shows the basic configuration of a character reading device according to an embodiment of the present invention, the input character pattern is subjected to preprocessing such as noise removal and normalization of position and size. , an input feature vector is generated based on a predetermined algorithm. This input feature vector is
The pattern matching means 10 calculates the inter-vector distance between the input feature vector and the standard feature vector read from the standard feature vector storage means 11. Therefore, with respect to the input character pattern, the pattern matching means 10 outputs the character code of the standard pattern that corresponds to the distance value obtained by pattern matching.

The standard standby vector storage means 11 stores standard feature vectors generated from standard patterns of characters to be read,
By giving the character code number, the corresponding character t! It is possible to read quasi-feature vectors every time.

The reading result output means 12 evaluates the distance values for the plurality of standard patterns obtained by the pattern matching means 10, and outputs the final reading result for the input character pattern. In other words, if there is one type of standard pattern with a sufficiently small distance from the input character pattern, the character code corresponding to that standard pattern is output, and if there is no standard pattern with a small distance from the input character pattern, The final reading result is unreadable. In addition, if the distance between the input character pattern and multiple standard patterns is almost the same, and it is impossible to determine which standard pattern is the closest, the reading result may not be limited to one type, but multiple standard patterns. The reading result indicating 0 or more is outputted as the output of the character reading device, and at the same time, the present invention uses the following matching candidate selection means. It is assumed that there are 130 inputs.

The matching candidate selection means 13 is a means for controlling the order of pattern matching in the pattern matching means section 10, and the process of successive character reading will be explained in detail below using the drawings.

Figure 2 shows the text to be read as ``Connections between characters...''
'' is input and the second character from the beginning is read. At this stage, the first character "sentence" has already been read. The matching candidate selection means 13 determines the character type to be matched for the next input character pattern from the obtained character code of the "sentence".

Fig. 3 shows a part of an example of a matching candidate table for estimating matching candidate character types, a matching FA complement table, and the next connected character for each character code, that is, the next input character in the character reading process. Contains codes for characters that may appear as character patterns.

In the example in Figure 3, for the characters ``Sun'', ``Tuesday'', ``Book'', which are read first, ``Book, month, of...'', ``Yo, power, star...'', ``person'', etc. , word, body...'' indicates that it is connected as the next character. therefore,
By using this matching candidate table, when reading consecutive characters, it is possible to obtain the character code of a character that may appear as the next character to be read from the character code of the previously read character.

In the example of FIG. 0, a standard feature vector prepared in advance is read out from the standard feature vector storage means 11 based on the character code of the matching candidate character type determined by the matching candidate selection means 13, and is input to the pattern matching means 10. ``Science'', ``Chapter'', ``Sentence'' that may be connected to
The standard feature vectors corresponding to the characters "..." are sequentially input to the pattern matching means 10, and distances between the vectors are calculated.

On the other hand, after performing processing such as noise removal and position/size normalization on the character pattern, a preprocessing and feature extraction means 14 that generates an input feature vector based on a predetermined algorithm is used to perform the next reading. The input feature vector for the character pattern to be performed is input to the pattern matching means 10.

In the pattern matching means 10, the input feature vector and the
The distance between the vectors is calculated by pattern matching with the standard feature vectors corresponding to "chapter", "character", etc., and the distance to each standard pattern is determined.

To calculate the distance between vectors, City Pro adds the absolute values of the differences between each element of two vectors and takes the average.

The Euclidean distance, which averages the sum of squares of the differences between the elements of two vectors, is known. These methods of finding distances between vectors require computations between the elements of the two vectors, so finding the distance to one standard feature vector requires repeated computations several times in the dimensions of the vectors. . Therefore, when one input character is compared with all the standard feature vectors, (number of standard feature vectors) x (number of dimensions of feature vectors) iterative calculations are required. In the example in Figure 3,
Assume that (letter, distance M) is obtained as the distance to the standard Pakun. The reading result output means 12 outputs the character code of the "character" with the smallest distance obtained as the reading result.

In the above process, the reading result for the input character can be output simply by performing pattern matching with the character types to be matched limited by the matching candidate selection means 13.

FIG. 4 shows a second embodiment of the character reading device of the present invention. In this embodiment, the matching candidate selection means 13 shown in the first embodiment is replaced with the matching result evaluation means 15 and the matching order specifying means. 16 and a matching candidate table 17.

The matching result evaluation means 15 evaluates the distance between the feature vectors obtained by the pattern matching means 10, and when a standard pattern that sufficiently matches the input character pattern is detected,
A signal is output to the matching order specifying means 16 to terminate the subsequent pattern matching process. At this time, the reading result output means 1
In step 2, the character code for the I quasi-pattern for which the abort instruction has been issued is output as the final reading result.

The matching order specifying means 16 uses a matching candidate table 17 shown below to sequentially input corresponding standard feature vectors from the first candidate to the pattern matching means 10. When receiving an abort instruction from the matching result evaluation means 15, the pattern matching means 10 is forcibly initialized to enable processing for the next read character.

The matching candidate table 17 contains characters that have a high probability of appearing as the next character to be read based on the code of the next character to be read, that is, characters that have a high probability of being connected next in the connection relationship of consecutive characters. By configuring the collation candidate table 17 in this way, in which character codes are written as the first candidate, second candidate, etc., characters with a high probability of being output as a reading result by the collation order specifying means 16 can be selected. You can proceed with the matching process from

FIG. 5 shows a third embodiment of the character reading device of the present invention. This embodiment is characterized in that a matching candidate updating means 18 is added to the matching candidate selecting means 13 shown in the second embodiment.

The verification candidate updating means 1 extracts connection relationships between characters in input sentences or words in the character reading device of the present invention during continuous reading processes, and updates the verification candidate table 17.
It is a means of updating. The matching candidate table 17 can be created in advance by analyzing a large amount of Japanese text, etc., but the range of character types to be matched can be limited by updating the matching candidate table 17 in the process of reading successive characters. This reduces the amount of calculation required for pattern matching.

As described above using the embodiments, the character reading device of the present invention detects the character that appears as the next character to be read from the code of the previously read character in the pattern matching process of the character reading device shown in the prior art. It is characterized by providing a matching candidate table for limiting. Therefore, components other than the matching candidate table 17 can be easily configured using conventional techniques. The matching candidate table 17 can also be easily created by analyzing a large amount of Japanese sentences and the like in advance. Furthermore, the collation candidate table 17 can be updated in the process of continuous character reading using the character reading gl of the present invention.

〔Effect of the invention〕

In the character reading device of the present invention, a table (matching candidate table 17) is prepared in advance that extracts the connection relationships between characters constituting a sentence or word to be read, and in the process of reading consecutive characters, A matching candidate table is searched based on the character code of the read character, the character type of the next character to be read is determined, and an inter-vector distance is calculated between the corresponding standard feature vector and the input feature vector. If a character that sufficiently matches the input character is detected from the obtained distance value, the subsequent matching process is terminated.

Therefore, in conventional character reading devices, in order to read one input character, distance calculation in pattern matching is required, which requires a calculation amount proportional to the product of (the number of character types to be read) and (the number of dimensions of the feature vector). However, this can be reduced to the point where characters that sufficiently match the input characters are detected and the matching process is terminated. Therefore, it is possible to reduce the amount of pattern matching processing, which involves distance calculations for an extremely large number of standard patterns with high-dimensional feature vectors, which is a problem when dealing with handwritten characters and the like. As a result, the problem that pattern matching processing in a character reading device limits the overall reading speed can be solved.

Furthermore, there is a problem when performing brute force pattern matching with standard feature vectors to be recognized, which has been done in the past.
Distinguishing similar characters such as the kanji numeral "-" and the long kana "-" uses a matching candidate table to proceed with the matching process starting from the most frequently connected characters, and then evaluates the degree of match to the input character and performs the matching process. By terminating the characters, it is possible to reduce reading errors for similar characters that are connected less frequently.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing a character reading device according to a first embodiment of the present invention, FIG. 2 is a diagram showing a character reading procedure,
FIG. 3 is a diagram showing an example of the configuration of a matching candidate table, FIG. 4 is a block diagram showing a character reading device according to a second embodiment of the invention, and FIG. 5 is a diagram showing a third embodiment of the invention. It is a block diagram showing a character reading device.

Claims (3)

[Claims] (1) A pattern matching means that calculates the distance between an input feature vector generated from an input character pattern and a standard feature vector generated in advance from a standard pattern of a character type, and an input feature obtained by the pattern matching means. A reading result output means that evaluates the distance value between the vector and the standard feature vector and outputs a character code corresponding to the input character; Select the character type to be matched for the next character to be read from the character code of the character read, input the corresponding standard feature vector to the pattern matching means, and selectively perform pattern matching between the standard feature vector and the input feature vector. A character reading device comprising a verification candidate selection means for performing the following. (2) The matching candidate selection means determines the matching target character type of the next read character based on the connection relationship between the characters constituting the word or sentence to be read and the character code of the previously read character. a collation order specifying means for sequentially selecting characters with a high connection probability and inputting corresponding standard feature vectors to the pattern matching means; and a distance value between the input feature vector obtained from the pattern matching means and the standard feature vector. 2. A character reading device according to claim 1, further comprising a verification result evaluation means for evaluating and determining whether to terminate or continue verification processing in the pattern verification means. (3) The matching candidate selection means extracts connection relationships between characters from character code strings obtained as reading results corresponding to sequentially input character patterns, updates connection relationship information prepared in advance, and matches the pattern. Claim 1 or 2 further comprising a matching candidate updating means for changing the selection procedure of character types to be matched in the means.
Character reading device as described in section.