JPS6362084A - Character segmentation system - Google Patents

Abstract

PURPOSE:To prevent wrong recognition and to improve a recognition rate by removing a noises which exceeds the minimum width of a character when the character is segmented. CONSTITUTION:For example, a number '4' which becomes blurred and is separated into two is segmented. Then, a vertical projection pattern 2 is scanned from left to right to find an interval WLEN to a next pattern. The WLEN is compared with predetermined character pitch and when the WLEN is larger than the character pitch, black length BLEN is found. This BLEN is compared with the predetermined minimum width CHMIN of a character and when BLEN > CHMIN, the vertical projection pattern 2 is further scanned continuously to find the front white length FWLEN. This FWLEN is compared with predetermined white length GAP and when FWLEN >= GAP, it is judged that the FWLEN is caused by the blurring and further front black length FBLEN is found. The sum HW of those BLEN, FWLEN, and FBLEN is calculated and compared with the maximum character width and when the sum is smaller, it is judged as the segmentation width of one character pattern.

Description

DETAILED DESCRIPTION OF THE INVENTION [Technical Field] The present invention relates to a character extraction method in OCR (optical character reader) and the like.

[Prior art]

Generally, in OCR and the like, vertical projection and horizontal projection are used for character extraction. However, there is a problem in that when a single character pattern is separated into a plurality of parts due to partially faded characters on the document, or when noise is present around the pattern, it may be cut out incorrectly.

〔the purpose〕

An object of the present invention is to prevent character patterns from being erroneously cut out due to separation of character patterns caused by noise existing around the character pattern or blurred characters when cutting out a character pattern in OCR or the like.

〔composition〕

According to the present invention, when a narrow pattern is detected during character extraction, forward projection scanning is performed to match the standard size of one character, and it is determined whether it can be integrated as one character pattern. In addition, when a pattern is detected by scanning within the determined character width range from above or below, scanning is resumed from the position skipped by a predetermined height, and when a pattern is detected again, the previous detected position Characters are extracted by comparing and removing noise contained above or below the character pattern.

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

FIG. 1 is a diagram for explaining the character cutting method of the present invention. FIG. 1(a) is a diagram when determining the character width of a character pattern, and FIG. This figure shows a case where noise present below is removed. In FIG. 1, it is assumed that one character pattern is composed of seven segments. In Figure 1(a), 1
is one read character pattern line, and 2 is its vertical projection pattern. As is well known, vertical projection pattern 2
can be obtained by checking character pattern row 2 sequentially from left to right, for example, to see if there is a black dot in the vertical direction, and setting it to "1" if there is a black dot, and "O" if there is no black dot. Also,
In FIG. 1(b), 3 is one character pattern, and 4 is noise existing above the character pattern.

First, the process of determining the horizontal character width of a character pattern to be cut out will be explained with reference to FIG. 1(a). Second
The figure shows the processing flow in this case.As an example, in 2N, the number "4" in FIG. 1(a) (this is
(separated into two due to blurring etc.) shall be cut out.

After cutting out the number "5" pattern, scan the vertical projection pattern 2 from left to right to find the distance WL from the next pattern.
EN is determined (step 101).

This WLEN is compared with a predetermined character pitch PITCH (step 102), and WLEN<PITCH.
If it is CH, proceed directly to step 104 and use WLE.
If N≧PITCH, space processing (step 103
), the process proceeds to step 104. In step processing, if WLEN≧PITCH, the interval region is counted as a space. The spanene number is WLEN/PITCH
Find it with If it is better not to output spaces in post-processing, specify PITCH=O, check PITCH=O before calculating the number of spaces,
If H=O, the process of calculating the number of spaces is skipped and space suppression is performed.

In step 104, the next black length BLEN is determined. This BLEN is compared with the predetermined minimum width CHMIN of characters (step 105), and if BLEN≦CHMIN, this black area is regarded as noise and is added to >WLEN (step 106), and BLEN>CHMIN In this case, the vertical projection pattern 2 is further scanned to obtain the front white length FWLEN (step 107).
Compare WLEN with a predetermined white length GAP (step 108), if FWLEN≧GAP.

It is determined that FWLEN is caused by blurred characters, etc., and the front black length FBLEN is further determined (step 1).
09). And BLEN+FWLEN+FBLEN=
Calculate the HW (step 110), compare this HW with a predetermined maximum character width CHMAX (step 111),
If HW≦CHMAX, this HW is set as the cutting width for one character pattern (step 112).On the other hand, if HW>CHMAX, the BLEN and FBLEN parts are set as separate character patterns (step 11).
3), this also applies when it is determined in step 108 that FWLEN>GAP.

Next, referring to FIG. 1(b), a process for removing noise existing above or below a character pattern to be cut out will be explained. FIG. 3 shows the processing flow in this case.

Once the character width (HW') of character pattern 3 has been determined by the process shown in Figure 2, the range of the character width HW' is scanned, for example, from the top to the bottom, and black pixels are detected first. Stop the scanning at the position and move to that position (xt-y□
) (Step 201) 0 Next, add the minimum height CHMIN of the character pattern to the Y coordinate Y1 of the position (Step 202), restart scanning from that position, and then restart scanning when a black pixel is detected. Stop and move to the corresponding position (X, Y,
) is determined (step 2o3). Now, compare the X coordinate X1 of the first stopped position and the X coordinate x2 of the next stopped position (step 204). If X1=X, then the first stopped position (XY, Yl) is the vertex of the pattern (step 205). On the other hand, if X1 to X2, 2
The position at which it stopped the second time is regarded as the position at which it stopped for the first time, that is, x2→X□, Y2→Y□ (step 206)
, X□=x2 is found, the processing from step 202 onward is repeated. Thereby, noise existing above the pattern 3 as shown in FIG. 1(b) can be removed. Similarly, when cutting out a single character pattern whose bottom point is found by scanning from the bottom to the top within the range of the found character width, its vertical width is within the range of the top and bottom points. And it is sufficient.

FIG. 4 is a schematic block diagram of an example of a hardware configuration for realizing the method of the present invention. In FIG. 4, the image memory 14 stores character pattern data and its vertical projection pattern. Now, when vertical projection scanning is specified from CFULL, the scan control section 12 loads the initial value of the vertical projection pattern storage address in the image memory 14 into the address generation section 13, and also loads the address generation section 13 with the initial value of the vertical projection pattern storage address.
3 to count enable state. As a result.

a vertical projection pattern in image memory 14 is scanned;
The images are sequentially taken in by the scan control section 12. The scan control unit 12 detects black runs and white runs of the captured vertical projection pattern, and sends the start point/end point addresses to the CPU II. The CPU II executes the process shown in FIG. 2 based on the start/end addresses of the black run and white run sent from the scan control section 12 to determine the character width (horizontal direction) of the pattern to be cut out.

Next, in the case of processing to remove noise existing above or below the character pattern to be cut out, the CPU 11 specifies the character pattern scan and its scanning range to the scan control unit 12. In response to this, the scan control section 12 loads the initial value of the character pattern storage address in the digital memory 14 into the address generation section 13, and puts the address generation section 13 into a count enable state. As a result, the corresponding character pattern area in the image memory 14 is scanned from the top to the bottom or from the bottom to the top, and the results are sequentially fetched into the scan control section 12. When the scan control unit 12 detects a black pixel, the scan control unit 12 temporarily suspends the count up of the address generation unit 13, that is, suspends the scan, sends the address to the CPU II, and instructs the next scan start position from the CPU 1. and resume scanning. When a black pixel is detected, scanning is stopped again and the address is transferred to CP.
Send it to U1l and wait for the next instruction from C, FULL.

The CPU 1 executes the process shown in FIG. 3 based on the black pixel address sent from the scan control section 12, and instructs the scan control section 12 to resume scanning.

Find the top and bottom points of the pattern.

〔effect〕

As is clear from the above description, according to the present invention, when cutting out characters, it is possible to efficiently remove noise that exceeds the minimum width of characters, and even when character patterns are separated due to blurring, etc., it is possible to easily integrate them. This results in an improvement in the recognition rate in OCR and the like.

[Brief explanation of the drawing]

FIG. 1 is a diagram explaining the principle of the character extraction method of the present invention, FIGS. 2 and 3 are flowcharts of an embodiment of the character extraction method of the present invention, and FIG. 4 is an example of a hardware configuration for realizing the method of the present invention. FIG. 11... CPU, 12... Scanning control section, 13.
...Address generation diagram, 14...Image memory. FWLEN Figure 2

Claims (2)

[Claims] (1) When a narrow pattern is detected in a method of character segmentation using vertical or horizontal projection, forward projection scanning is performed to match the standard size of the character, and it is determined whether or not it can be integrated as a single character pattern. A character cutting method featuring (2) When the character pattern is scanned and the first black pixel is detected, the scanning is restarted by shifting a predetermined position, and the noise part is extracted from the positional relationship between the position of the first detected black pixel and the next detected black pixel. The character cutting method according to claim 1, characterized in that the character is removed.