JPH02193280A - Character recognizing device - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] <Industrial Field of Application> The present invention relates to a character recognition device, and more specifically, the present invention relates to a character recognition device, and more specifically, to recognize characters, symbols, etc. (hereinafter representatively referred to as “characters”) through an image input device or a communication medium such as a facsimile. ”), extract its feature amount, perform calculations based on the feature amount, select a candidate character, and select the candidate character that is closest to the character to be read from among the candidate characters. The present invention relates to a character recognition device that can identify characters in detail and output signals representing the characters.

<Conventional technology> Character recognition using conventional character recognition devices extracts the characteristics of input characters from input character signals, compares the degree of difference with the characteristics stored in the recognition dictionary owned by the user, This is done by selecting the character with the smallest degree of difference.

For example, as shown in FIG. 4, an image containing characters is input using an image input means (11) such as a scanner, and each unit character is cut out using a character cutting means (12). and,
A feature extraction means (13) extracts feature amounts based on the cut out character image. The detailed identification means (15) uses this feature amount and the recognition dictionary (1B) (recognition dictionary (16)
In addition to the features, for example, the average value of the features, the state of the distribution,
The ranking of each feature amount influencing recognition is stored.

) is compared point by point with the information on the feature quantities for the number of characters stored in , the degree of dissimilarity or degree of similarity (hereinafter representatively referred to as "degree of dissimilarity") is calculated, and the results are output. The ranking determining means (17) rearranges the characters in descending order of dissimilarity, and outputs, for example, the character having the smallest dissimilarity as a recognition result.

FIG. 5 is a block diagram showing another conventional example. After the feature extracting means (13) extracts the feature amount based on the cut out character signal, the major classification discriminating means (14) Narrow down multiple candidate characters using a method. The detailed identification means (15) obtains information regarding the feature amount of the candidate character stored in the recognition dictionary (16) for each candidate character sent from the major classification identification means (14). ), calculate the degree of difference for each, and output the results. Ranking determining means (17
) rearranges the characters in order of decreasing degree of difference using the same method as above, and outputs, for example, the character with the smallest degree of difference as the recognition result.

In particular, a specific example of the character recognition device shown in FIG. 5 is the character recognition device described in Japanese Patent Publication No. 83-28915. In this character recognition device, the major classification identification means selects dictionaries of categories with similar biases, and the detailed identification means masks the biased parts of the characters in the dictionary of the selected candidate category. , only the labor part is identified in detail. This also improves the recognition rate for groups of characters with similar biases.

<Problem to be solved by the invention> However, in the configuration shown in FIG.
) compares the features of all the characters stored in the recognition dictionary (16), which increases the amount of calculations and increases the recognition time for literature recognition devices that recognize many types of characters. . For example, in Kanji OCR, which recognizes 3,000 character types at the first level of JIs, the degree of dissimilarity is a to recognize - characters.
, ooo times. In addition, the accuracy of calculating the degree of dissimilarity (for example, the number of terms in the series required to calculate the degree of dissimilarity) can be adjusted so that characters that are most difficult to recognize with the character recognition device can be concatenated (correctly recognized). In addition, the same recognition process is performed for characters.

Therefore, even input characters that are easy to recognize are calculated with the above accuracy, which results in wasted time.

In the character recognition device shown in FIG. 5, the above-mentioned drawbacks have been solved to some extent. In the configuration shown in FIG. 5, the major classification identification means (14
) to select multiple candidate characters similar to the character you want to recognize. However, in this case too, in a character recognition device that recognizes many character types, it is necessary to perform calculations for selecting candidates for all character types, but since the time required for one calculation is short, it is necessary to select candidate characters. The overall time required will be reduced. Then, the candidate characters are sent to the detailed identification means (15), and the degree of dissimilarity is calculated for the candidate characters in the same manner as the character recognition device shown in FIG.

Here, n is character recognition device b (number of recognizable character types, nl
The number of candidate characters output by the major classification identification means (14), tl
Assuming that t2 is the time per character required for major classification character identification and t2 is the time per character required for detailed classification character identification, the character recognition device shown in Fig. 4 takes t2 In the character recognition device shown in FIG. 5, it takes tl Xn+t2 Xnl time. If tl is much smaller than t2 and nl is much smaller than n, the calculation time can be reduced.

However, the character recognition device shown in Fig. 5 is the same as the character recognition device shown in Fig. 4 in that the calculation accuracy is adjusted to the character for which detailed classification characters are most difficult to identify. This does not mean that the time t2 per character is shortened. For example, in the character recognition device described in Japanese Patent Publication No. 63-26915 mentioned above, when calculating the degree of feature difference in the labor part, calculations up to the same number of dimensions (in the same publication (1)
) is calculated by taking Σ of the formula from −1 to n.
Actual calculation takes time.

Therefore, it can be said that in conventional character recognition devices, there is a certain limit to the reduction in calculation time.

An object of the present invention is to provide a character recognition device that can shorten the overall character recognition time by employing a simple method.

Means for Solving the Problems> The character recognition device of the present invention for achieving the above objects has the following features:
As shown in FIG. 1, an image signal acquisition means (1) that acquires an image signal representing an object to be read including characters, and a feature that extracts the feature amount of the character to be recognized in the image based on the image signal. a quantity extraction means (2), a major classification identification means (3) that performs calculations based on the feature quantities and selects one or more candidate characters, and a recognition device that stores information regarding the feature quantities necessary for recognizing each character. detailed identification means (4) that compares the feature amounts of the character to be recognized with the feature amounts and other information of the candidate characters stored in the recognition dictionary (6) and performs a feature amount comparison calculation; , a recognized character output means (5) which selects a character according to a certain standard from among the characters identified by the detailed identification means (4) and outputs a signal representing the character; A calculation accuracy correspondence table (7) that stores constants specifying calculation accuracy when the identification means (4) performs feature value comparison calculations, and one or more candidates from the above candidate characters based on a predetermined standard. Select a character, search the calculation accuracy corresponding to the selected candidate character from the calculation accuracy correspondence table,
The calculation accuracy determining means (8) determines one calculation accuracy based on the retrieved calculation accuracy and sends it to the detailed identification means. The detailed identification means (4) has a function of performing feature value comparison calculation with the calculation accuracy specified by the calculation accuracy determination means (8).

Effects> According to the character recognition device configured as described above, the feature amount extracting means (2) extracts the feature amount (several elements) are extracted. The major classification identification means (3) selects one or more candidate characters by a simple method using this feature amount. The above candidate characters are composed of a group of characters similar to the character to be recognized, and it is thought that regular continuation characters exist among these candidate characters.

By the way, if the character you are trying to recognize is complex, the candidate characters will also consist of similar complex characters, and if the character you are trying to recognize is simple, the candidate characters will also easily consist of similar simple characters. It can be inferred that Here, "complex" and "easy" refer to whether there is a high or low possibility of misreading other characters during recognition. Even if a character has a large number of strokes, it can be said to be an "easy" character if it has unique characteristics and there are almost no similar characters.

Therefore, in order to recognize "complex" characters, the detailed identification means (4) must compare them in detail with the feature values and other information stored in the recognition dictionary (6). rate cannot be achieved.

However, when attempting to recognize "simple" characters, it is not necessary to perform the same comparison as for the "complex" characters described above.

In the present invention, the detailed identification means (4) is provided with a function to compare and calculate feature quantities with specified precision, and also perform feature quantity comparisons according to the "simpleness" and "complexity" of each candidate character. Constants that specify calculation accuracy when performing calculations are stored in calculation accuracy correspondence table (7).

The candidate characters selected by the major classification identification means (3) are notified to the calculation accuracy determination means (8), and the calculation accuracy determination means (8) selects one or more candidate characters from among the notified candidate characters based on predetermined criteria. Select candidate characters. Then, with reference to the calculation accuracy correspondence table (7), one calculation accuracy is determined based on the calculation accuracy corresponding to the selected candidate character, and the detailed identification means (4) is determined.
). The detailed identification means (4) performs comparative calculations for each candidate character with the specified calculation accuracy while referring to the recognition dictionary (6). The recognized character output means (5) selects a specific character from among the characters identified by the detailed identification means (4) and outputs a signal representing the character.

In this way, the calculation accuracy is determined according to the "simpleness" or "complexity" of the candidate character, and detailed calculations are performed using this calculation accuracy.

<Example> Embodiments will be described in detail below with reference to the accompanying drawings showing embodiments.

FIG. 2 is a block diagram showing the configuration of the character recognition device of the present invention, in which (la) shows an image camera such as a bidicon that captures the entire document, and the output signal thereof is binarized to obtain a formatted signal. It represents an image input section consisting of a binarization circuit. The image signal is subdivided character by character by character cutting means (1b). The feature extraction means (2) extracts the feature amount of each character. For example, it is a histogram of directional vectors of character outlines or a distribution of blank area amounts. This feature quantity is input to the major classification discriminating means (3), which uses a simple discriminating function that takes a short calculation time, such as a city block function or a linear-dimensional number, for all character types. Calculate the degree of difference between Then, a plurality of candidate characters are selected.

The data of each candidate character is sent to the calculation accuracy determination means (8), and the calculation accuracy determination means (B) uses the calculation accuracy correspondence table (7) to determine the highest character among the candidate characters (the character with the lowest degree of difference). Search the calculation accuracy from the calculation accuracy correspondence table (7) for the character with the highest number) or the character in the top number, determine one calculation accuracy based on these calculation accuracies, and use the detailed identification means (4)
send to. The contents of the calculation accuracy correspondence table ω may be determined in advance when creating the calculation accuracy correspondence table (7), but it is preferable to update the contents according to the results of actual use. For example, a learning function may be added to the character recognition device, and the character recognition device may be updated through learning. Alternatively, the information may be updated artificially without using the learning method.

The detailed identification means (4) calculates the degree of dissimilarity for each candidate character based on the calculation accuracy determined above.

For example, the distance between the feature amount of the candidate character included in the recognition dictionary (6) and the feature amount obtained by the feature extraction means (2) is determined. These feature quantities are generally vectors (dimension is n), but when the above calculation accuracy is low, there is no need to calculate up to n dimensions, and it is sufficient to calculate distances in a lower dimension commensurate with the calculation accuracy. It is. On the other hand, when the calculation accuracy is high, it is necessary to find the distance in a correspondingly high dimension.

In this way, since the degree of dissimilarity is determined for each candidate character, the ranking determining means (51) rearranges the candidate characters in descending order of degree of dissimilarity and outputs the character code of the first rank.

Next, we will change the operation of the above character recognition device to "It's sunny today".
An example of recognizing the first character "hon" in a manuscript written as "hon" will be explained.

When the image input unit (1a) reads a manuscript that says "Today is sunny", the character extraction means (1b) reads "Book".
Image signals are converted into single characters such as ``Sun'', ``Ha'', ``Sunny'', ``Ten'', ``Na'', and ``Su''. And feature extraction means (2) c. Feature amount of "book" X (Xi, X2.-, X
Extract n). The major classification identification means (3) uses the method described above to select multiple candidate characters such as ``book,''``tree,''``water,'' and so on.
・Select. The calculation accuracy determining means (8) searches for the calculation accuracy of, for example, the highest character "hon" among the candidate characters, and obtains a result of accuracy 3 from the calculation accuracy correspondence table (7). The detailed identification means (4) performs distance calculation based on the accuracy 3 described above. For example, candidate characters in the recognition dictionary (6)
The feature amount of "book" is A (Al, A2, -An), the feature amount of "tree" is B (B1.B2...Bn),...
・Then, the distance X is calculated by the city block function below.
-A I -[(Xi-Al)2 + (X2-
A2) 2+ (X3-A3) 2 ] 112X
-B l -[(Xi-BL)2+(X2-82)2
10 (X3-B3)2 ] ” 2 etc. will be found.

Note that this city-block function is sometimes used in the case of the above-mentioned major classification identification, but in the case of major classification identification, it is difficult to perform very rough calculations, such as calculating only the first term. This is different from the calculation method for detailed classification.

A ranking determining means (51) rearranges the characters in descending order of distance and outputs the character code of the first ranking.

As described above, FIG. 3 shows the time required to recognize all the characters "It's sunny today."

FIG. 3(a) shows the time required for recognition by a conventional character recognition device. Since the same calculation time t is required for each character, it takes X7 time for all characters. On the other hand, FIG. 3(b) shows the time required for recognition using the character recognition device of this embodiment, and the calculation time t1. t 2゜...
X7 is different. As a result, it can be seen that the overall recognition time is 1-Σtn, which is shorter than in the case of FIG. 3(a).

In addition, in the above embodiment, an example was shown in which the calculation accuracy determining means (8) determines the calculation accuracy based on the highest character "hon" among the candidate characters, but the invention is not limited to this. The calculation precision determining means (8) may determine the precision of each of the top several characters, and determine one calculation precision based on the plurality of precisions found. For example, the accuracy is 3, 4, and 5 for the top three characters ``book,''``tree,'' and ``water,'' respectively.
If so, take the average value of these accuracies, or
One calculation accuracy may be determined by taking the maximum value or the minimum value. This is because each candidate character included in the candidate character group realized as a result of the major classification is considered to have approximately the same calculation accuracy in the calculation accuracy correspondence table (7), so only the highest character has a calculation accuracy of Even if you decide
This is because even if the calculation accuracy is determined for the top few characters, the results will not differ much.

Further, in the above embodiment, the detailed identification means (4) simply calculates the distance, but it is of course possible to use a more complicated calculation formula in order to increase the recognition rate.

Even in this case, the number of terms to be calculated can be limited depending on the "simpleness" of the candidate character, so the present invention is very effective in shortening calculation time.

Further, in the above embodiment, the image camera was used to input the document image, but the invention is not limited to this.
The upper part of the image may be input through a communication line such as a facsimile.

Various design changes can be made without changing the gist of the invention.

<Effects of the Invention> As described above, according to the character recognition device of the present invention, the accuracy of comparison calculation for each character is determined in advance according to the degree of "complexity" and "simpleness" of the character, and the calculation accuracy correspondence table is created. The calculation accuracy is then stored and the calculation accuracy determined by the calculation accuracy determination means to be applied to the candidate character group classified into major categories.

Therefore, whereas conventional calculations were performed with maximum accuracy for each character, resulting in wasted calculation time, the present invention enables comparative calculations for each character with the minimum accuracy based on the determined calculation accuracy. Therefore, when recognizing an object to be read containing a mixture of "complex" and "simple" characters, the average recognition time per character can be shortened. In other words, if the same recognition time is allowed, the accuracy of character recognition can be increased. Therefore, it is possible to improve the recognition performance of the character recognition device.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of the character recognition device of the present invention, FIG. 2 is a block diagram showing an embodiment of the character recognition device, FIG. 3 is a comparison table of character recognition time with the conventional example, and FIG. 5 are block diagrams showing the configuration of a conventional character recognition device. (1)...Image signal acquisition means, (2)...Feature amount extraction means, (3)...Major classification identification means, (4)...Detailed identification means, (5)...Recognition Character output means, (6)・
...Recognition dictionary, (7) ...Calculation accuracy correspondence table, (8)
... Calculation accuracy determining means Applicant: Sumitomo Electric Industries, Ltd.

Claims (1)

[Claims] 1. Image signal representing the object to be read including characters an image signal acquisition means for acquiring the image signal; Trying to recognize images based on image signals Feature extraction to extract the features of characters Perform calculations based on the means and the above feature amounts 1 Broad classification identification that selects two or more candidate characters methods and features necessary for recognizing each character. A recognition dictionary that stores information about Recognizes the features of the characters to be recognized Candidate character information stored in the dictionary Detailed identification means for calculating feature value comparison and among the characters identified by the detailed identification means. Select characters based on certain criteria and write the relevant sentence. Recognized character output that outputs a signal representing a character In a character recognition device having means, Calculation accuracy when performing feature comparison calculations Memorizes the specified constant corresponding to each character. Calculation accuracy correspondence table and the above candidate characters one or more based on predetermined criteria from among Select a candidate character and use this selected candidate character. Calculate the calculation precision corresponding to the complementary character vs. calculation precision Search from the table and based on the searched calculation accuracy. Determine the calculation accuracy and perform detailed identification Equipped with means for determining the calculation accuracy to be sent to the stage and the above detailed identification means has calculation accuracy. Depending on the calculation accuracy specified by the determining means The feature is that feature comparison calculation is performed using character recognition device.