JPH0438586A - Device and method for recognizing character - Google Patents

Abstract

PURPOSE:To prevent errorneous reading owing to the ambiguity of a discriminating feature by weighting the feature element of an input character pattern and comprehensively judging with the whole feature element. CONSTITUTION:A feature element extracting means 4 extracts a feature element from the image data of an input character pattern read by a character pattern reading device and next, a probability value converting means 5 converts feature existing into a maximum probability value and feature not existing into a minimum probability value based on the discriminating feature found by the feature element extracting means 4. A means 8 to calculate difference of every feature element calculates the absolute value of the difference of the probability value of every feature element of the input character pattern and the probability value of the corresponding feature element in a pattern dictionary (dictionary B). Then, a valid distance calculating means 10 calculates valid distance by taking a sum and so on at every respective registered character pattern for the absolute value of the difference of the feature element found on the feature element of the respective registered character patterns.

Description

[Detailed Description of the Invention] [Summary] A character recognition device and a character recognition method that perform character recognition by comparing input character patterns such as handwritten characters with character patterns registered in a dictionary, which reduce misreading due to ambiguity of determination features. The purpose of the invention is to obtain a character recognition device and a character recognition method, including a device for reading an input character pattern, a feature extraction means for extracting feature elements from the read input character pattern, and a feature element extracting means for extracting feature elements from the read input character pattern, and a method for detecting the appearance of all feature elements for each registered character pattern. A pattern dictionary in which probability values based on frequencies are registered, a probability value conversion means for converting extracted feature elements into the maximum value of the above-mentioned appearance probability and non-extracted feature elements into the minimum value of the same appearance probability, and an input character pattern. and a feature element difference finding means that takes the absolute value of the difference in probability values for each corresponding feature element for all the feature elements of the registered character pattern, and a prize for each registered character pattern. and a recognition candidate output means that outputs recognition candidates by assigning a higher rank to the one with the smaller calculated effective distance.

[Industrial application field]

The present invention relates to a character recognition device and a character recognition method that perform character recognition by comparing input character patterns such as handwritten characters with registered character patterns in a dictionary.

Conventionally, human character pattern recognition in a character recognition device involves extracting its characteristic elements, comparing the extracted characteristic elements with the characteristic elements of registered character patterns in a dictionary, and selecting registered character patterns with registered character patterns whose characteristic elements all match. It outputs the recognition results using characters as answers.

Therefore, in conventional character recognition devices, if there is ambiguous data that allows the characterizing elements of an input character pattern to be determined to be either characteristic or non-featured, this may cause misreading.

The present invention provides a character recognition device with high reliability by eliminating the influence of ambiguous feature element numbers on feature determination by making comprehensive judgments on all feature elements.

[Conventional technology]

A conventional character recognition method will be explained with reference to FIG.

Figure (a) shows a conventional determination method.

In the figure (al), 5I shows an example of the determination feature when 10 feature elements (denoted as features in the figure) are extracted from the input character pattern, with 1 indicating the presence of a feature and O indicating no feature. Pattern dictionary (
An example of dictionary A) is shown.

In pattern dictionary A, the determination of whether a feature exists or not is determined by, for example, collecting a large amount of data for the letter AJ, and if the appearance frequency of each feature is 50% or more, it is set to 0 if it is less than 1.50%. .

Conventionally, the determination features of the input character pattern, that is, the features l,
2.5.7.8 is set to 1, and the others are set to 0.Reference is made to the pattern dictionary (dictionary A) 52 for registered character patterns that have the same judgment features as the judgment features, and those that completely match are selected in the illustrated example. outputs the recognition result using the registered character pattern (a) as the answer.

Figure (b) shows the flow of a determination method in conventional character recognition.

The flow will be explained according to the illustrated numbers.

■ Input the input character pattern as image data.

■ Extract the feature elements (N) and find the determination features.

In figure (a), N=10.

(2) Compare the N determination features with the determination features of each registered character pattern in dictionary A.

(2) If there is a complete match in which all the corresponding feature elements match, then in (2), the recognition result is output using the characters of the registered character pattern that matched with the dictionary A as the answer.

If there is not a complete match, the determination feature for the next registered character pattern in dictionary A is extracted.

[Problem to be solved by the invention]

As mentioned above, in conventional character recognition methods, when the judgment features of the input character pattern and the judgment features of the registered character pattern match completely for each feature element, the matched character is output as a recognition result. Ta.

Therefore, the recognition accuracy of conventional character recognition devices depends on the authenticity of the determination feature obtained when extracting features from the input character pattern, and if there is ambiguity in the determination feature, it may be misread.

An object of the present invention is to obtain a character recognition device that does not cause misreading due to ambiguity of determination features.

[Means to solve the problem]

The present invention prevents misreading caused by ambiguous features by weighting the feature elements of an input character pattern and making a comprehensive judgment using all the feature elements.

In addition, as a result of comprehensive judgment of all feature elements, if the difference between the data for judgment as a recognition result is small, among the plurality of recognition candidates obtained first, among the plurality of elements, the one with the highest degree of effectiveness as a judgment feature The reliability of recognition is improved by selecting only things and making comprehensive judgments based on multiple factors.

The basic configuration of the present invention will be explained with reference to FIG.

In the figure, 1 is an input character pattern, 2 is a character pattern reading device such as an image scanner, 3 is image data of the input character pattern, and 4 is a characteristic element extracted. is set to 0, and 5 is a probability value converter that converts the elements with features obtained by the feature extractor 4 into maximum probability values, and the elements without features into minimum probability values, For example, the maximum probability value is 100 and the minimum probability value is O, 6 is the probability value for each feature element obtained by the probability value conversion means 5, and 7 is the appearance frequency of each feature for the feature element in each registered character pattern. For example, in a pattern dictionary obtained as a probability value based on
%, Feature 2 was 5%, etc., so its appearance frequency was registered as a probability value. 8 is the probability value of the feature element of the input character pattern and the probability value of the corresponding feature element of the registered character pattern in the pattern dictionary. Difference calculation means for each feature element that calculates the absolute value of the difference for each feature element by comparing it with each registered character pattern, 9 is data of the difference for each feature element calculated for each feature element for each registered character pattern , 10 is an effective distance calculation means for calculating an effective distance of an input character pattern for each registered character pattern from difference data for each feature element, for example, from difference data for each feature element, for each registered character pattern. 11, which calculates the sum of each characteristic element for each
is a recognition candidate output means that outputs recognition candidates to the effective path 8 in ascending order of the peak value, starting from the highest value.

[Effect]

The operation of the basic configuration of the present invention will be explained with reference to FIG. Refer to FIG. 1 if necessary.

Figure (a) shows an example of probability value conversion of feature elements.

The figure shows that 10 types of feature elements are extracted from the input character pattern, and feature elements 1.2.5.7.8 are considered to have features, and feature elements 3.4.6.9 and 10 are considered to have no features. Indicates the case of determination.

Then, the probability value conversion means (5) converts each feature element into a maximum probability value of 100 for a characteristic element and a minimum probability value of 0 for a characteristic element, indicating that weighting has been performed.

In the figure (b), 20 indicates an example of the pattern dictionary (dictionary B) of the present invention.

The pattern dictionary (dictionary B) 20 is a pattern dictionary in which probability values are obtained for feature elements in each registered character pattern based on the frequency of appearance of each feature, for example,
In the figure, as a result of collecting data on patterns in which the answer is "a", the frequency of appearance of feature 1 is 90%, and the frequency of feature f12 is 5.
%, etc., so the frequency of appearance is registered as a probability value.

From the image data of the input character pattern read by the character pattern reading device, the feature element extraction means (4
) extracts feature elements, and creates determination features for the 10 feature elements, setting 1 if the feature exists and 0 if the feature does not exist for the 10 feature elements, for example, as shown in Figure (a).

Next, the probability value conversion means (5) converts the characteristic element extraction means (4) into a probability value conversion means (5).
), the presence of features is converted to the maximum probability value, and the absence of features is converted to the minimum probability value.

Figure (The converted value in the barrel is the case where the appearance frequency of the feature element is expressed in %, and the maximum probability value is 100 and the minimum probability value is weighted as 0.

The difference calculating means (8) for each feature element calculates the absolute value of the difference between the probability value for each feature element of the input character pattern and the probability value of the corresponding feature element in the pattern dictionary (dictionary B). Then, the effective distance calculating means (10) calculates the effective distance by, for example, summing the absolute values of the differences between the characteristic elements determined for the characteristic elements of each registered character pattern, for each registered character pattern.

For example, when comparing with the registered character pattern "a",
Effective distance is l 100-90 l+l 100-51+
l0-801+l0-101+l 100-75
+ 0-60 + 100-100 +
100-0 + 0-20 + 0-85
Find =485.

This effective distance is calculated for all registered characters "i", "tsu", etc., and is calculated for each registered character pattern.

Then, the recognition candidate output means (11) outputs recognition candidates in ascending order of the effective distance value.

According to the present invention, since the characteristic elements of the input character pattern are weighted and comprehensive judgment is made using all the characteristic elements, misreading due to ambiguous characteristic elements does not occur.

〔Example〕

The present invention will be explained in detail with reference to FIGS. 3 and 4.

Figure 3 shows the device of the present invention! An example of the configuration is shown.

In the figure, 31 is a microprocessor, 32 is a feature extraction unit that extracts feature elements and weights them, and 33 is a dictionary matching recognition unit that matches an input cover pattern with a pattern dictionary using feature elements, calculates an effective distance, and performs character recognition. Processing section, 34
ffi! is an effective feature when the recognition result cannot be determined because there is little difference in the effective distance between the recognition candidates obtained from the effective distance calculation results. A feature selection section for extracting feature elements that extracts only the elements and calculates the effective distance again using the extracted feature elements; 35 is an effective distance meter unit for calculating the effective distance; 36 is an image of reading the input character pattern 37 is an image data storage memory for storing image data read by the image scanner; 38 is a dictionary memory for storing registered character patterns; and 39 is a shared memory used for program processing and the like.

FIG. 4 shows a flowchart of an embodiment of the apparatus configuration shown in FIG.

The flow will be explained in the order of the numbers shown.

■ Read the input cover pattern with an image scanner and input the image data.

■ Extract feature elements (number of features to be extracted is N)
.

(2) Calculate the effective distance from each registered character pattern by referring to Dictionary B in which probability values of the registered character patterns are registered for the N characteristic elements.

■ Determine whether there is any ambiguity in the recognition candidates obtained based on the effective distance.

■ In processing when the effective distance between recognition candidates is small and there is ambiguity, from the N features obtained in ■, feature elements and probability values with small differences in probability values of feature elements between recognition candidates are M feature elements effective for determination are selected, excluding ambiguous feature elements around 50 features.

(2) Calculate the effective distance to each registered character pattern in dictionary B using the M selected valid feature elements.

■ Outputs the recognition result using the registered character pattern with the closest effective distance as the answer.

〔Effect of the invention〕

According to the present invention, the entire extracted feature elements are compared with registered character patterns and character recognition is performed based on a comprehensive judgment, which reduces misreading caused by ambiguous features and significantly improves recognition accuracy. .

[Brief explanation of drawings]

FIG. 1 is a diagram showing the basic configuration of the present invention. FIG. 2 is an explanatory diagram of the operation of the present invention. FIG. 3 is a diagram showing an embodiment of the device configuration of the present invention. FIG. 4 is a diagram showing the flow of an embodiment of the device configuration of the present invention. FIG. 5 is a diagram showing a conventional character recognition method. In FIG. 1: input character pattern, 2: character pattern reading device, 3: image data, 4: feature element extraction means, 5: fi rate value conversion means, 6: probability value for each feature element, 7: pattern dictionary, 8: feature Difference calculation means for each element; 9: Difference data for each feature element; 10: Effective path II award output means; 11: Recognition candidate output means.

Claims (3)

[Claims] (1) A reading device (2) for the input character pattern (1), a feature element extraction means (4) for extracting feature elements from the read input character pattern, and a feature element extracting means (4) for extracting feature elements from the read input character pattern; a pattern dictionary (7) in which probability values based on the probability values are registered; and a probability value conversion means (5) that converts the extracted feature elements to the maximum value of the above-mentioned appearance probability and the feature elements that are not extracted to the minimum value of the same appearance probability. and feature element difference calculation means (8) that calculates the absolute value of the difference in probability values for each corresponding feature element for all the feature elements of the input character pattern and the registered character pattern, and the calculated value of the difference calculation means (8). effective distance calculation means (10) for calculating the sum of the sum for each registered character pattern, and recognition candidate output means (11) for outputting recognition candidates with the smaller calculated effective distance as a higher rank. Character recognition device. (2) In a method of character recognition by comparing an input character pattern and a registered character pattern with respect to characteristic elements, a pattern dictionary is provided in which probability values are registered based on the frequency of appearance for each characteristic element for the registered character pattern, and the input character Extract feature elements from the pattern, convert the extracted feature elements into the maximum value of the above-mentioned appearance frequency probability, convert the unextracted feature elements into the minimum value of the same appearance frequency probability, and calculate the probability of the feature elements of the input character pattern above. The difference between the value and the probability value of the feature element of the registered character pattern in the dictionary pattern is taken for each feature element, the sum of the absolute values of the differences is found for each character pattern, and the one with the smaller sum value is selected as a recognition candidate with a higher rank. A character recognition method characterized by outputting . (3) In the character recognition method according to claim 2, if the difference in the calculated value of the sum among the plurality of recognition candidates is small, the difference in the probability value of the feature element between each recognition candidate is small. Probabilities are calculated for the feature elements of the registered character pattern that correspond to the feature elements of the input character pattern, excluding feature elements and ambiguous feature elements of features whose probability value is around 50 when the maximum probability is 100 and the minimum probability is 0. , calculate the sum of the absolute values of the differences for each registered character pattern, and output multiple recognition candidates with the smaller sum as the higher rank, or the one with the smallest sum as a recognition candidate. Characteristic character recognition method.