JPS62295193A - Type recognizing method - Google Patents

Abstract

PURPOSE:To recognize the European language character of a type with high accuracy by checking a size after an unknown pattern is recognized by utilizing a code to designate the size of the horizontal direction classified by a character and the starting position of a pattern to be connected. CONSTITUTION:On the dictionary (template) of a character pattern, a code designating a size in the horizontal direction classified by a character and the starting position of the pattern to be connected is provided, and after an unknown pattern is recognized by utilizing this, the size is checked. At the time of segmenting it with a segmenting candidate and the character size is not coincident, rejection is obtained. Consequently, the pointer for segmenting is shifted to the next candidate segmenting position and characteristic extraction and matching are executed again. In such case, as well as the preceding case, not only the matching between characteristic vectors, but also the matching of the character size can be executed.

Description

Detailed Description of the Invention 3. Detailed Description of the Invention ■Technical Field The present invention relates to a character recognition device (OCR), particularly character recognition in typesetting of printed materials such as magazines, and proportional pitch recognition in type printing. It is something.

■Prior art A prerequisite for OCR is to cut out characters individually.

This cutting may fail when characters touch each other.

To prevent this, in the past, multiple cropping candidates were prepared from the minimum value of the histogram projected in the vertical direction, recognition was attempted at each cropping position, and the one with the smallest similarity distance was selected.

For example, in the case of the pattern shown in Figure 1, each candidate point P, , P, , P4 is prepared with P□ as the end point, and each interval OL is
Among 1g and 1g, candidate 110'', which exhibits the maximum mantength in P1 and P2, was output as a result.

However, as shown in FIG. 2(1), when there is a defect such as blurring in the pattern, a new cutting candidate P is generated. In this case, the pattern cut out at P5 will be recognized as 11 C#1, which will have an adverse effect on subsequent recognition.

Furthermore, when adjacent characters overlap as shown in FIG. 2 (2), it is extremely difficult to cut out the characters.

In the above cases, it is judged as unrecognizable (rejected),
Although an application has already been filed for starting the cutting process again, the explanation will be omitted here.

For normal printed matter and typesetting output, the font size and left edge position of the next character are determined for each character, as shown in Figure 2 (1).
) is a candidate for cutting out, the font size of “CTt is usually different from the P-P5 interval, so the 0 recognition result is “C7”.
It will never be 1.

■ Purpose The present invention is characterized by recognizing printed European characters with high precision. In particular, characters printed by typesetting often touch each other, and it is required to recognize these with high precision. ing.

■Configuration Figure 3 shows the overall processing flow of the present invention, and Figure 2 (1)
), the character sizes do not match, resulting in a rejection.

Therefore, the pointer for clipping is shifted to the next candidate clipping position P2, and feature extraction/matching is performed again.

Needless to say, at this time, like last time, not only matching between feature vectors but also character size matching is performed.

FIG. 4 shows an example of a dictionary of feature vectors for character recognition.

In the figure, 1 is the registration code, 2 is the output code, 3 is the size,
4 is a start shift, and 5 is a feature vector.

Namely.

Registration code (ASCI I for English text, JIS code for Japanese text is common) 1° Output code (This corresponds to the combination typeface used in one typesetting, such as "fi" in Figure 2 (2) is 1. (Although the output result is separated into "tsf" and "i" and output) 2. Size (a numerical value indicating the horizontal size of the character, the unit is arbitrary) 3. Start shift ( As shown in FIG. 2 (2), when the pattern on the right is included inside the pattern on the left, it consists of a value 4 indicating its size and a normal feature vector 5.

In FIG. 5, A is size 3 and B is start shift 4.

In order to deal with character pattern overlap (FIG. 2 (2)), not only the above-mentioned two-character dictionary is prepared, but also a separate recognition method is used to prevent segmentation errors.

When matching each chord W and D to the pattern of width A in Figure 5, for chords whose start shift value is greater than 0, the right end of the pattern is deleted by this value before matching. Execute.

FIG. 6 is a detailed explanation of the "pattern extraction" and "feature extraction &matching" portions of FIG. 3.

In the above explanation, the character size needs to be given in advance. However, this value differs depending on the type of font.

For this reason, it has a separate memory in which data as shown in the table is held.

Therefore, at the beginning of recognition, data for frequently used fonts is registered in the dictionary, and if a large number of rejections occur, this data can be changed to find the most suitable one. can correspond to

By the way, when recognizing an overlapping pattern, matching is performed on the entire overlapping pattern, and a specific registered code is used as a candidate.

Next, delete part of the pattern based on Ss (right).

(left) and performs matching for each part with the feature vector of the character corresponding to the registered code.

When the input cover turn is fi, fi becomes a candidate, then f
Recognize with and i.

This increases recognition accuracy.

(2) Effects As explained above, the present invention has the effect of being able to recognize printed characters with high recognition accuracy.

[Brief explanation of the drawing]

Fig. 1 is a diagram showing an example of a conventional pattern, Fig. 2 is a diagram for explaining problems in character recognition in conventional characters, Fig. 3 is a diagram showing the overall processing flow of the present invention, and Fig. 4 is a diagram showing an example of a conventional pattern. The figure shows an example of a dictionary of feature vectors for character recognition, Figure 5 shows an example of character pattern overlap, and Figure 6 shows the "pattern extraction" and "feature extraction &matching" methods in Figure 3. ” This is a detailed explanation of the part. 1...Registration code, 2...Output code, 3
...Size, 4 ...Start shift, 5
...feature vector. Patent applicant Ricoh Co., Ltd. Figure 1 Figure 2 (+) (2) Bow P5P2P3 Figure 3 Figure 4 Figure 5 Figure 6

Claims (4)

[Claims] (1) A device that optically recognizes printed characters has a code on a dictionary (template) of character patterns that specifies the horizontal size of each character and then the start position of the connected pattern; A type recognition method that uses this to recognize an unknown pattern and then performs a size check. (2) For character patterns that partially overlap when the patterns are projected in the vertical direction, part of the pattern is deleted and then the feature vectors are matched. The printed character recognition method described in section 1). (3) The printed character recognition method according to claim (1), characterized in that the character size and numerical values specifying the starting position of the next character are stored in separate tables. (4) The printed character recognition method according to claim (2), characterized in that when performing matching, only specific characters constituting an overlapping pattern are recognized.