JPH02242391A - Character recognizing system - Google Patents

Abstract

PURPOSE:To improve the recognition precision of characters by extracting the density distribution feature quantity and the outline feature quantity of the character image of a character to be recognized and comparing them with feature quantities preliminarily stored in a dictionary to output a character code corresponding to recognition results. CONSTITUTION:Density distribution feature quantities and out-line feature quantities of character images are preliminarily stored in a dictionary 25, and image data of a document is outputted by an image scanner 21 and a character is segmented from this data by a character segmenting part 22 and is sent to a density distribution feature quantity extracting part 23 and an outline feature quantity extracting part 24. Feature quantities extracted by both extracting parts 23 and 24 are compared with feature quantities in the dictionary 25 by a retrieving part 26, and feature quantities of the smallest difference are taken out and are outputted to a code converting part 27, and the converting part 27 outputs the character code corresponding to specified feature quantities through an output device 28.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to a character recognition method, particularly a character recognition method that extracts features from read character images and compares them with dictionary features stored in advance in a dictionary. Regarding the method of identification.

In recent years, with advances in computers, word processors, databases, etc., documents are increasingly used in encoded form. Conventionally, documents have generally been encoded manually using a word processor. however,
This method is complicated and takes a lot of time. Therefore, character recognition has recently been widely studied in order to efficiently encode documents.

Character recognition involves processing the image data obtained by scanning a document with a scanner, extracting character image features, storing them in a dictionary in advance, and comparing them with the dictionary features to find the dictionary with the smallest difference. The corresponding code is output and stored in a magnetic disk device or the like.

[Conventional technology]

FIG. 6 is a block diagram showing an example of a conventional character recognition device. In the figure, image data of a document read by an image scanner 1 is provided to a feature extraction section 2. The feature history extraction unit 2 performs preprocessing for special@quantity extraction.
Normally, characters are extracted, normalized, etc. From the character image of each character obtained through this preprocessing, the feature extracting unit 2 extracts the feature ff1ffi of the character image (that is, the character feature). Various methods have been proposed for extracting this feature amount. Among these, methods that focus on the contours of characters yield good results for characters with many contours, such as kanji. Various methods are known for extracting character contour features. For example, the outline of a character image is traced from the starting point and a direction code is assigned. Therefore, the outline feature amount of a certain character is obtained as a direction code string.

The feature amount obtained in this way is given to the search unit 3. The search unit 3 stores this characteristic mu in a dictionary in advance and compares it with a dictionary feature amount. For example, if a contour feature is used as a feature, the dictionary feature is also expressed as a contour feature. As a result of the comparison, the search section 3 sends the one with the closest difference (distance) between the two to the code conversion section 5 as a recognition result. The code converter 5 selects a character code corresponding to the recognition result,
The data is sent to an output device M6 such as a magnetic disk device, display, or printer.

[Problem to be solved by the invention]

However, such a conventional character q recognition method has the following problems. Although character recognition is suitable for characters with many contours using the above-mentioned contour feature extraction, it is suitable for characters with few contours, such as “, 2, .;:”, and characters with similar shapes, such as “YuYu It is difficult to distinguish the contours of Pa and Yu, resulting in a low recognition rate.

Therefore, it is an object of the present invention to solve the problems of the prior art and to provide a character recognition method that can accurately recognize characters with few characteristics or characters with similar shapes.

[Means to solve the problem]

FIG. 1 is a block diagram of the principle of the present invention.

The first special yi mouth extraction means 11 extracts the density distribution characteristic mouth of the character image of the character to be recognized.

The second feature extraction means 12 extracts the contour feature amount of the character to be recognized.

The dictionary 13 stores density distribution features and contour features of character images as dictionary features of characters.

2! The identification means 14 extracts the density distribution features of the character images extracted by the first and second special extraction means 11.12 and the contour time t! Verify the character by comparing the amount with the corresponding feature in the dictionary.

[Effect]

In the present invention, the feature values are extracted not only from the outline but also from the density distribution, so it is possible to clearly distinguish between periods, '', commas, etc. Also, since the density clearly indicates the size of the character, small ``'' It is also possible to distinguish between ``Y'' and large ``Y''.As a result, the recognition rate is particularly improved for characters and symbols with little contour information, and for hiragana and katakana, which have differences in size.

〔Example〕

Hereinafter, one embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 2 is a block diagram of one embodiment of the present invention. In the figure, an image scanner 21 optically scans a document and outputs image data of the document in the form of an electrical signal. A character cutting unit 22 cuts out characters from image data of a document. Various methods are known for cutting out characters, but any method can be used in this embodiment. For example, characters are cut out by projecting image data of a document in the horizontal direction (main scanning direction) and vertical direction (sub-scanning direction). The image data of the extracted characters is sent to the density distribution characteristic am extraction section 23 and the contour feature amount extraction section 24.

The density distribution feature extracting unit 23 extracts a feature based on the m-component distribution of the character image, that is, a density distribution feature. For this reason, the density distribution feature extraction unit 23 equally divides the extracted character image (rectangular pattern) vertically and horizontally into small regions.

This situation is shown in FIG. The illustrated example is a case where a character image is divided into eight equal parts vertically and horizontally.The 6'a degree distribution feature amount extracting unit 23 obtains the density distribution for each of the 64 small areas divided into eight equal parts vertically and horizontally. The density distribution feature ω is calculated by dividing the number of black dots in a small area by the horizontal length of the small area. Focusing on the small area at the bottom left of Figure 3, the number of focal points in this area is 45, and the horizontal length is 8 (the number of dots), so
The density distribution feature amount is approximately 5. Here, the reason why the number of black points is divided by the horizontal length of the small area is to achieve a balance with the contour time mu described below.

The contour feature extracting unit 24 extracts contour time points for each small region. The contour feature amount is based on the connection state of the contour points on the character contour, and the contour points are set to the left and right (0), downward to the right (1), upward (2), and upward to the right (3).
(direction code), and the number of direction codes in the small area is n1l (An, Bn, Cn, Dn) at the time of contour.

This situation is shown in FIG. Focusing on the small area at the bottom left,
The direction code is shown as an arrow. Therefore, the contour feature amount (An, an.

Cn, Dn)=(11, 3, 2, 4).

In the manner described above, the density distribution feature value and contour time mu are obtained for each character and sent to the search unit 26 which functions as a recognition unit. The search unit 26 compares the density distribution feature and the contour feature J in a dictionary stored in advance in the dictionary @25.

Figure 5 (A> shows the feature values obtained for each direction code in the examples of Figures 3 and 4. Figure 5 (B) also shows the density distribution obtained in the example of Figure 4. Indicates the feature amount.The dictionary 25 has registered in advance the density distribution feature amount and contour feature amount extracted using the same method as the special y1 maximum extraction method described above using a reference sample for the dictionary.Then, the search Part 26 is 8X8X, which is a combination of the contour feature R and the density distribution feature.
The difference (distance) between the features is reduced using the city block distance of the 4+3X8-320 dimensional vector, and the dictionary with the smallest difference is selected Y! im is extracted and output to the code converter 27.

The code conversion unit 27 has a character code corresponding to each dictionary feature, and outputs the character code corresponding to the dictionary feature specified by the search unit 26.

The output device 28 is composed of a magnetic disk device or the like, and stores or outputs the character code output from the code converter 27.

When a document of approximately 450 characters was recognized using the character recognition device configured as described above, the recognition rate for punctuation marks was 13 characters/21 characters to 19 characters/21 characters compared to when only contour features were used. It was confirmed that this had improved.

One embodiment of the present invention has been described above. In the above embodiment, each character image is divided into 8 equal parts vertically and horizontally to utilize the 11 degree distribution and the directional component on the outline of the character, but this is not necessary.

Further, the direction code is not limited to four directions, and a direction code indicating eight directions may be used.

Further, in the above embodiment, the density distribution feature and the contour feature are used in the same ratio, but they may be weighted. For example, by setting the weighting of the contour feature large (e.g., 1 to 1.5 times), the proportion of the contour feature given to the recognition result is increased. This is because when each of the textiles is used alone, the contour feature provides a higher recognition rate.

Further, in order to improve the recognition rate, known image correction processing such as normalization or feature mouth correction processing may be used.

Furthermore, the distance calculation method is not limited to city blocks, and Maharano-Heath distance or the like may also be used.

Furthermore, there may be a plurality of dictionary features for each character (for example, for Mincho typeface and Gothic typeface).

〔Effect of the invention〕

As described above, according to the present invention, since both the density distribution and the contour are used as feature quantities, the recognition rate for commas, periods, and similar characters is improved.

[Brief explanation of drawings]

Fig. 1 is a diagram for explaining the principle of the present invention, Fig. 2 is a block diagram of an embodiment of the present invention, Fig. 3 is a diagram for explaining extraction of density distribution feature times, and Fig. 4 is a contour feature amount. 5(A) and 5(8) are diagrams showing an example of contour time mu and density distribution features, respectively, and FIG. 6 is a diagram illustrating an example of a conventional character recognition device. FIG. 2 is a block diagram of a configuration example. In the figure, 1 is the first feature extraction means, 2 is the second feature extraction means, 3 is the dictionary, 4 is the recognition means, 1 is the image scanner, 22 is the character segmentation unit, and 23 is the density distribution feature extraction 24 is a contour feature amount extraction unit, 25 is a dictionary, 26 is a code conversion unit, and 27 is an output device. Patent applicant: Fujitsu Ltd. Diagram for explaining the extraction of density distribution features Figure 3 Diagram for explaining the extraction of contour features Figure 4 An example of the configuration of a conventional character recognition device Block diagram Figure 6

Claims (1)

[Claims] A dictionary (13) that stores density distribution features and outline features of character images as dictionary features of characters; and a first dictionary that extracts density distribution features of character images of characters to be recognized. a feature extracting means (11), a second feature extracting means (12) for extracting a contour feature of a character image of a character to be recognized, and first and second feature extracting means (
a dictionary search means (14) for comparing the density distribution feature and contour feature of the character image extracted in steps 11 and 12) with the corresponding feature in the dictionary and outputting the result as a character code corresponding to the recognition result; A character recognition method characterized by: