JPH03150692A - Word collation system - Google Patents

Abstract

PURPOSE:To prevent false retrival due to the difference of the sort of a character by calculating the product of ratio between the distance value of the candidate character of first order among the candidate characters and the distance value of each candidate character and a standard value set beforehand, and sort-processign the points-obtained of a word obtained by a summing the points-obtained of the character as considering a normalized calculated result the points-obtained of the character. CONSTITUTION:Distance ratios to higher order such as th distance ratio between the first and the second of the candidate characters and the distance ratio between the first and the third and so on are calculated successively by inter-character candidate distance ratio calculation 102a. Further, the product of a constant value determined experientially and the distance ratio obtained in the processing of a preceding stage is calculated by normalized points-obtained calculation 102b. Next, in order to calculated the points-obtained of the word by using the points-obtained of the character candidate obtained by this processing, character points-obtained cumulative summation and word points-obtained sorting are executed, and a word candidate is fixed. Thus, at the time of the retrieval of the word in which plural sorts of the characters are mixed, the false retrieval of a word dictionary due to the difference of the sort of the character can be prevented.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to a word matching method that searches a word dictionary to find the word with the highest luminance based on character recognition results, and is particularly applicable to handwritten character strings. We will develop an S-word matching method that can improve recognition accuracy.

[Conventional technology]

Traditionally, handwritten character recognition has been used to create a database by scanning a document with kanji or kana handwritten on a piece of paper using a computer, or to input documents such as slips and manuscripts. However, because the structure of kanji is complex, it is difficult to obtain sufficient accuracy through character recognition alone.
It is common practice to correct the results of character recognition using a terminology dictionary.

For example, “Matsuo et al.: [Fuzzy word retrieval method using associative integration type matching”, Information Society of Japan, 34th National Conference.

4E-7 (1986)'', as shown in Figure 2, the distance value of the character obtained by character recognition is subtracted from a constant value201°202), and the difference is calculated as the character distance value. As the score of the character indicating the confidence level of the character, the score of the character is added to the word that includes the character candidate obtained as a result of character recognition (203). Perform sorting process (20
4), the words with the highest luminosity were output in descending order of the score.

Note that the distance value is the difference between the standard pattern vector of the one-character dictionary and the vector indicated by the candidate character.

Further, the constant value is a value indicating a range within which the difference is allowed, and it is determined that the larger the difference between the constant value and the distance value, the higher the luminous intensity. In addition, the category kanji is a serial number assigned to a kanji character, in addition to a code such as a JIS code, in order to facilitate management.

In addition, as a conventional character recognition method for handwritten kanji, for example, “Hagita et al.: “Identification of handwritten kanji using outer direction contribution characteristics” Journal of the Institute of Electronics and Communication Engineers, Vol.
J66-D Na1O (1983)'', the contour direction contribution method is mentioned.

With this method, the distribution of black pixels that make up a character is measured vertically and
Examination is performed from multiple directions horizontally and diagonally, and the amount of consecutive black pixels in each direction is used as a feature amount (hereinafter referred to as a feature vector). In character recognition, distances are calculated between a character dictionary obtained from this feature vector as the average value of many character samples and feature vectors obtained from the character image data to be recognized, and characters are selected in order from the one with the smallest distance. Judged as a candidate.

In this way, in a character recognition method based on the distribution of black pixels, the distance value obtained as a result of distance calculation is difficult to distinguish between a complex-shaped kanji and a simple-shaped alphanumeric character, etc. and alphanumeric characters tend to be smaller. Note that a complex shape means that, when viewed at the image level, there are many black pixels or that there is a large amount of variation in the shape depending on the sample.

[Problem to be Solved by the Invention J] The above prior art does not take into account the fact that the distance value obtained from the result of five-character recognition is affected by the difference in one character type. In the case of a word containing a mixture of , there was a problem that the correct word could not be retrieved due to the increased weight of non-kanji character strings.

For example, when recognizing the handwritten character 'Kineko Masutori' as shown in FIG. 3, the distance value of each candidate character obtained through character recognition is larger for the Kanji candidate.

As a result, the result of the difference calculation is the kana characters “Ne”, “Nu”,
and the kanji ``子'' have a high score, and from the cumulative addition result of character scores, ``Kinuko Masahata'' is output with a higher rank than ``Kimeko Masutori''.

SUMMARY OF THE INVENTION An object of the present invention is to provide a word matching method that can improve such problems and prevent erroneous searches due to differences in character types when searching for words that include a mixture of multiple character types.

[Means to solve the problem]

In order to achieve the above object, the word matching method of the present invention uses a character dictionary in which the average value of the feature values of characters of categories l - N is stored, and the feature values of handwritten characters and the characters stored in the character dictionary are Calculates the distance from the average value of the feature values, determines character candidates for the input character in order from the category with the smallest distance value, searches the word dictionary based on the information on the obtained character candidate string, and selects the most In a word search method that searches for words with high luminosity, the distance value of the first candidate character and the distance value of each candidate character are calculated by calculating the distance ratio between character candidates, and the ratio is calculated by calculating the normalized score. Calculate the product with a predetermined standard value (a constant value determined empirically), use the calculation result as the normalized character score, and add the character scores by accumulating the character scores to obtain the word score. It is characterized by sorting processing.

Another feature is that a plurality of the above reference values are provided, the complexity of the character is determined based on the feature amount of the candidate character obtained through character recognition, and the reference value is selected according to the complexity.

[Operation] In the present invention, when calculating the score of candidate characters obtained by character recognition, the ratio of the distance value of each candidate character in the character candidate string to the distance value of the first candidate character is calculated. The product of this ratio and the standard value (a constant value determined empirically) is the normalized character score.

As a result, when searching for a word that includes a mixture of multiple character types, it is possible to avoid erroneous searches in the word dictionary due to differences in character types.

Furthermore, a plurality of empirically determined constant values are prepared for each category, and are appropriately selected depending on the feature vector of the candidate character when calculating the normalized score. This is because the feature vector is calculated based on the amount of continuous black pixels, so it approximately represents the complexity of the character.

This makes it possible not only to distinguish between Kanji and non-Kanji characters, but also to take into account the complexity of the characters, thereby further improving the search accuracy.

〔Example〕

An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a flowchart showing the word matching method in the first embodiment of the present invention, and FIG. 4 is an explanatory diagram showing the word matching process in the first embodiment of the present invention.

In this embodiment, as shown in FIG. 1, a character dictionary is used to perform character recognition on a handwritten character string to be recognized (101).
.

Next, normalization processing is performed based on the category number and distance value obtained at this time (l O2).

That is, by calculating the distance ratio between character candidates (102a),
Distance ratios with higher rankings are sequentially calculated, such as the distance ratio between the first and second candidate characters, the distance ratio between the first and third candidate characters, and so on. moreover,
The normalized score calculation (102b) calculates the product of the empirically determined constant value and the distance ratio obtained in the previous process.

Next, in order to calculate a word score using the character candidate scores obtained through this process, character score cumulative addition and word score sorting are performed (103 and 104) to determine word candidates.

The normalized score for the first-place candidate is determined by calculating the distance from the character dictionary for samples for which the correct category is known in advance, and calculating the distribution of the calculated distance values for each kanji and non-kanji character. .

A specific example of this word matching process is the 45th! Shown in J.

In this embodiment, as in FIG. 3, the handwritten character string "Kineko Masutori" is character recognized and the distance value of each candidate character is obtained.

Furthermore, when normalization processing is performed, the distance ratio of the first candidate character is l, so the constant value = 2000 becomes the character score as it is. That is, the score of the first candidate character becomes a constant value. Therefore, the score of the first candidate character is "2000" without being affected by the difference between kanji and non-kanji characters.

By cumulatively adding up the character scores obtained in this way for the corresponding characters in the word dictionary, "Kimeko Masutori" registered in the word dictionary is output at a higher rank than "Kinuko Masuhata".

FIG. 5 is a flowchart showing the word matching method in the second embodiment of the present invention.

In this example, a plurality of empirically determined constant values are set using feature vectors that approximately represent the complexity of characters.
Performs normalization processing according to the complexity of character shapes.

As shown in FIG. 5, character recognition is performed in the same manner as in the first embodiment (501).

Next, in addition to the category number and distance value obtained at this time, normalization processing is performed using a feature vector that can be easily obtained as progress data (502).

That is, by calculating the distance ratio between character candidates (502a),
The ratio to the first candidate character is calculated from the distance value obtained by character recognition. Furthermore, by constant selection (502b), the sum of each dimension of the feature vector obtained by character recognition is calculated,
The optimum constant value is selected from among a plurality of constant values prepared as the score of the ``1st candidate character'' according to this total value.

Using the constant value thus selected, the product with the distance ratio obtained in the previous process is calculated (502C).

Thereafter, the same processing as in the first embodiment is performed (503, 50
4) Output word candidates.

〔Effect of the invention〕

According to the present invention, the distance value written as a result of character recognition can be normalized in a simple way, so when searching for a word that contains a mixture of multiple character types, the characteristics of the distance value will differ depending on the character type. It is possible to avoid erroneous word dictionary searches that occur from

[Brief explanation of the drawing]

FIG. 1 is a flowchart showing the word matching method in the first embodiment of the present invention, FIG. 2 is a flowchart showing the conventional word matching method, and 31! I is an explanatory diagram showing the conventional word matching process, FIG. 4 is an explanatory diagram showing the word matching process in the first embodiment of the present invention, and FIG. 5 is an explanatory diagram showing the word matching method in the second embodiment of the present invention. FIG. Diagram word candidates! s'fji

Claims (2)

[Claims] (1) Using a character dictionary that stores the average value of the feature values of characters in categories 1 to N, calculate the distance between the feature values of handwritten characters and the average value of the character feature values stored in the character dictionary. In the word search method, character candidates for the input character are determined sequentially from the category with the smallest distance value, and a word dictionary is searched based on the information of the obtained character candidate string to search for the word with the highest luminosity. Among the above character candidates, the product of the ratio of the distance value of the first candidate character to the distance value of each candidate character and a preset reference value is calculated, and the calculation result is normalized as the character score. A word matching method characterized by sorting the scores of words obtained by adding character scores. (2) A plurality of the reference values are provided, the complexity of the character is determined based on the feature amount of the candidate character obtained through character recognition, and the reference value is selected according to the complexity. Word matching method.