JPH0632070B2 - Character recognition device - Google Patents

Description

TECHNICAL FIELD OF THE INVENTION The present invention relates to a character recognition device that enables adaptive character recognition processing according to the type of character.

[Technical background of the invention and its problems]

In performing character recognition, it is very important to know the character type in advance in order to improve the recognition processing efficiency and the recognition rate. For example, in a character recognition device for sorting mail,
It is known that the object to be recognized is a postal code indicated by numbers, but in a character recognition device for general office processing, the character description format differs depending on the type of input form. Therefore, conventionally, in this type of office processing apparatus, a character description format (format type) determined for each type of the input form is stored in advance, and when recognizing the character, the input form is first input. The form number is read to identify the character description form, and subsequently, the characters described in the input form are recognized.

However, recently, the character types to be recognized have been expanded from alphanumeric characters and katakana to kanji, and the description types have been diversified. In particular, it has been attempted to directly read the written characters such as the destination address and the address printed or handwritten on the letter mail to recognize the characters. However, the above-mentioned destination address in the letter mail is often written in Chinese characters, but is also often written in Katakana. Furthermore, in the so-called Chinese cultural area overseas,
In many cases, the above-mentioned destination address etc. are written on every other line in Kanji and English characters. For such various character description styles, it is extremely unrealizable to deal with the character recognition by preregistering the information of the styles as seen in the above-mentioned office processing apparatus.

However, compared with the character recognition processing for alphanumeric characters and katakana, the kanji recognition processing is extremely complicated because it requires the use of context information peculiar to kanji. Therefore, in order to improve the processing efficiency, even in such a case, it is important to know by some means whether the character to be used in the recognition process is a kanji character or a non-kanji character. Will have. Further, in order to perform the recognition accurately, for example, it is necessary to correctly recognize the characters of the same isomorphic type such as (force-power), (engineering-d), (mouth-lo), and (ni-ni). Is necessary, and even in this case, it is important to know the character type in advance.

[Object of the Invention]

The present invention has been made in view of such circumstances, and an object of the present invention is to determine the type of characters written on a document for each character line or field, and to determine the type according to the determined character type. It is an object of the present invention to provide a character recognition device capable of adaptively and efficiently performing character recognition.

[Outline of Invention]

The present invention scans a document surface, extracts a character writing area in the document surface from a quantized signal obtained by photoelectric conversion / quantization, analyzes a character pattern in the character writing area, and analyzes each character line or field. The character type is determined, and the recognition processing is performed for each character forming the character string or field according to the determined character type.

In particular, in the present invention, even when the margins between characters are irregular, such as handwritten characters, and the sizes of the characters are various and uneven, the height of each character line (character (Width) V is measured, and the character width is multiplied by a predetermined count α to set the margin length between the characters, whereby the margin length between characters and the size of each character are normalized to a constant value. The character pattern of the generated character line is analyzed to determine the character type, and each character is recognized according to the character type determined here.

〔The invention's effect〕

Thus, according to the present invention, since the character type of each character pattern is determined for a character line in which the inter-character space length and the size of each character are normalized to be constant, the character type of the character string is accurately determined. Can be judged,
The recognition processing of each character according to the determined character type can be performed with high accuracy. For example, the character recognition processing is performed after determining the character type of each address line or field of a postal item, for example, by determining whether the character type of the character line is Kanji or non-Kanji. Can be done adaptively. Therefore, for character lines consisting of Kanji, efficient and highly accurate recognition processing is performed by utilizing context information peculiar to Kanji, while conventional recognition processing technology is used for character lines consisting of non-Kanji. The recognition can be easily performed.

Moreover, since the character type of each character line is detected from the structural complexity of the character, the accuracy of the character type determination is high, which greatly contributes to the improvement of the processing efficiency in character recognition. In addition, it is possible to effectively perform character recognition for a letter mailpiece in which Chinese characters and English characters are written every other line, which is very effective in practice.

Example of Invention

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

FIG. 1 is a schematic configuration diagram of the embodiment apparatus. Documents such as letter mails that have characters written by printing or handwriting are
The document surface is scanned by the scanning photoelectric conversion quantizer 1,
After being photoelectrically converted, it is quantized and read and input. At this time, as shown in FIG. 2, the character description area B in the document surface A is predicted according to the reference position data (u, v) and the size information (x, y), and only the image information of the area B is predicted. Read in. The predicted character writing area B is defined as an area for writing an address, an address, etc., with a certain allowance for the document surface A of the letter mail. Then, the image information (pattern data) of the area B thus read is temporarily stored in the scanning pattern memory 2 and then used for the subsequent processing. The scanner 3 scans the image information stored in the memory 2 with, for example, a coarse scanning line as shown in FIG. 3 to obtain the projection pattern. In this projection pattern, the density level is high in the part where the character lines are present, and the density level is low between the lines. Therefore, the positions of the character lines C and D existing in the area B are coordinate data respectively. (Y ₁ ,
it is possible to detect a y ₂ ~y _4). If no clear change in density level is recognized in this projection pattern, it can be considered that scanning is not performed in the character line direction. In this case, the scanning direction should be changed. Good. By changing the scanning direction in this manner, it is possible to detect a vertically written character line and a horizontally written character line, respectively.

Regarding the start position and the end position of the character lines C and D, coordinate data (x ₁ , x _{2 to} x ₄ ) of the signal detection position on each scanning line can be obtained.

For such position detection of the character lines C and D, various techniques proposed in conventional character recognition processing can be appropriately used.

Then, the image information of the character lines C and D whose positions have been detected as described above is selectively extracted and stored in the compressed character line memory 5. At this time, under the control of the scanner 6, the normalization processing is performed on each character forming the character lines C and D. That is, the characters forming the character lines C and D are, for example, the fourth characters.
As shown in FIGS. 5 and 5, generally, the margins between the characters are irregular, and the sizes of the characters are various and uneven. Therefore, the controller 6 is provided with information when the scanner 3 scans the image forming the character lines C and D stored in the memory 2, and the character lines C and D are supplied.
The overall height (character width) V is measured. This character width V is multiplied by a predetermined coefficient α to set the length αV of the margin to be normalized between the characters. Based on the margin length αV, the scanner 5 controls writing of images of the character lines C and D to the compressed character line memory 4 to obtain a compressed character line with a constant inter-character margin length. At this time, the size of the character is also normalized to a certain size suitable for the processing described later. As a result, the redundant blank area, which is information unnecessary for character recognition, is removed and the size of the character is made constant.

This normalization operation by compression of the character line is performed mainly for the Chinese character as shown in FIG. 4 or for the katakana character as shown in FIG. Is done in the same way.

Thereafter, the complexity calculation circuit 7 performs horizontal scanning and vertical scanning on the character line image stored in the compressed character line memory 4 with a thin scanning line, and the number of crossing between the scanning line and the character line portion is n. _x and n _y are calculated respectively. At the same time, the length L of the character line is calculated. According to these information n _x , n _y , L, the complexity calculation circuit 7 determines the complexity of the character line structure for the entire character line in the horizontal direction c _x = n _x / L and the vertical direction c _y = n _y / V. Is calculated as The number of intersections n _x and n _y indicates the complexity of the character line structure. Generally, the kanji is more complicated and the kanji has a larger value than non-kanji such as katakana. However, since the number of intersections n _x and n _y is obtained for the entire character line, the larger the number of characters forming the line, the larger the value. Therefore, if the number of times of intersection is normalized by the length of the character string normalized by the compression, the complexity of the average character line structure of each character forming the character string is c _x , it will be obtained as c _y.

Then, the information c _x and c _y thus obtained are summarized, and the value (c _x + c _y ) is used as the evaluation value indicating the complexity of the character structure. The evaluation value indicates the degree of complexity for the entire character such that the character structure is complicated, that is, the character structure is simple in the horizontal direction, but the character structure is complicated in the vertical direction. Therefore, this evaluation value (c _x + c _y )
Is discriminated according to a predetermined threshold value a, as shown in the relationship in FIG. 6, it is possible to discriminate very easily whether the characters forming the character line are kanji or non-kanji. . This determination is performed by the determination circuit 8, and it is determined that the character is a Chinese character under the condition of c _x + c _y > a and the character is a non-Kanji character under the condition of c _x + c _y a.

The character recognition device body 9 reads out the image pattern of the character line from the scanning pattern memory 2 by using the kanji recognition unit 10a or the non-Kanji recognition unit 10b according to the information of the kanji and non-kanji determined as described above. The recognition process.

As described above, according to the present apparatus, a character string is cut out from the document surface, a redundant blank portion of the character string is removed, and the size (height) V of the character is made constant, and then the horizontal direction and Find the number of intersections n _x and n _y with the character line by scanning in the vertical direction,
Moreover, since this is normalized by the length L of the character line, it is possible to obtain the evaluation values c _x and c _y that well reflect the complexity of the average character structure of each character of the character line. The evaluation values c _x , c _y take a large value when the character type is Kanji, and take a small value when the character type is a non-Kanji character, so that the character line is a Kanji character or a non-Kanji character. Can be easily determined. Then, according to this information, it is possible to adaptively perform the character recognition for the character line, so that the character recognition can be performed extremely efficiently.

The present invention is not limited to the above embodiment. For example, the characteristic peculiar to the character type of the character structure may be obtained from information other than the number of times of intersection described above. In other words, the number of strokes of the character line and information unique to each character such as splashes may be detected. By detecting such a character-specific characteristic pattern, for example, it is possible to distinguish Greek letters from English letters. In the example, the evaluation values are c _x and c _y
Although the sum of is calculated, it is also possible to use the weighted sum, product, etc. as the evaluation value. In short, the present invention can be variously modified and implemented without departing from the scope of the invention.

[Brief description of drawings]

FIG. 1 is a schematic configuration diagram of an embodiment apparatus of the present invention, and FIGS. 2 to 6 are views for explaining the operation of the embodiment apparatus. 1 ... Scan photoelectric conversion quantizer, 2 ... Scan pattern memory, 3, 5 ... Scanner, 4 ... Compressed character string memory, 6 ...
... controller, 7 ... complexity calculation circuit, 8 ... determination circuit, 9 ... character recognition device body, 10a ... kanji recognition unit,
10b ... Non-kanji recognition unit.

Claims (2)

[Claims] 1. A means for scanning and photoelectrically converting and quantizing a document surface on which characters are printed or handwritten, a means for extracting a character writing area on the document surface from the quantized signal, and a character. By measuring the character width for each line and setting the margin length between each character, the means for normalizing the inter-character margin length and the size of each character to a certain level, and these normalized character lines or fields It is characterized by comprising means for analyzing the character pattern for each character to determine the character type, and means for recognizing each character forming the character line or field according to the determined character type. Character recognizer. 2. The means for analyzing a character pattern to determine a character type is a method for detecting, for each character line or field, a characteristic peculiar to a character type forming the character line. The character recognition device according to item 1.