JPS61182182A - Character recognizing device - Google Patents

Abstract

PURPOSE:To execute the correct character recognizing by providing a black base extracting means besides the stroke extracting means. CONSTITUTION:When a tracking point exists at a picture element 31 shown by labels d and l, the bit [the bit is '1' since it includes (d) in the label.] of a position A1 in the labels (d) and l is noticed, and then, the bit '1' does not exist at the position A2, the bit '1' exists at a position A3 and therefore, the tracking point at an adjoining picture element 32 displayed by the label including the bit '1'. In the sequence of B2-B4, C2-C4 and D2-D4 respectively, the bit '1' of the label is detected and tracked. It is decided whether the information, which is not extracted as a stroke edge point, exists in the information to show respective linked black base parts. When the information, which is not extracted as the stroke edge point, exists, the respective linked black parts are not extracted by the stroke, and therefore, they are extracted as a black base lump.

Description

[Detailed Description of the Invention] <Technical Field of the Invention> This invention reads a pattern of unknown characters, numbers, symbols, etc. (in this specification, these are collectively referred to as "characters"), and The present invention relates to a character recognition device that recognizes an unknown character by converting it into black and white binarization to obtain an input image, extracting its geometric features by a feature extraction unit, and comparing the extraction results with a standard pattern.

<Summary of the Invention> This invention solves the problem that when strokes are extracted by tracing the boundary between a black background and a white background of an input image, a series of black backgrounds remain without stroke extraction, which causes incorrect character recognition. It prevents

<Background of the Invention> As shown in FIG. 9, a conventional character recognition device includes a reading section 2 that optically reads an unknown character 1 and converts it into an image, and a smoothing process for the input image from the reading section 2. a preprocessing unit 3 that performs preprocessing such as the following; a feature extraction unit 4 that extracts geometric features of the preprocessed image; and a feature extraction unit 4 that compares the extracted features with standard patterns stored in a dictionary 5 in advance. and a dictionary matching section 6 that recognizes unknown characters.

In general, the feature extraction unit 4 focuses on black pixels in the input image, thins the input image G to obtain a thin line pattern g, and extracts the features of the unknown character from this thin line pattern g. An extraction method is used.

However, this method requires image thinning processing, and in recent years, methods have been proposed in which features of unknown characters are directly extracted from the input image. In this method, for example, when extracting the features of the character pattern shown in FIG.
(indicated by thick lines in the figure), and by tracing this boundary in the four directions A-D shown in FIG. 12, the paired sub-strokes (A1.A2) (B1.B2),
Extract (C□, C2). In this case, pairness of substrokes is recognized only if the interval between them is less than or equal to a predetermined value. The boundary tracking is performed using the input image G as shown in FIG.
is scanned in a predetermined direction (indicated by an arrow in the figure), and in this scanning process, the preset A-
When a pattern extending in any direction of D is encountered, tracking of the boundary in that direction is started, and tracking is continued until the pattern in that direction disappears. Once paired substrokes are obtained in this way, an approximate pattern f as a set of strokes as shown in FIG. 14 is determined based on them.

However, according to this method, "tsu", "tsu", "shi"
, "n", "n", and other characters that are likely to have very short and wide strokes, errors in feature extraction may occur.

FIG. 15 illustrates a character pattern for "tsu", in which the portion indicated by H is an extremely short and wide stroke. In this stroke, sub-strokes C3, C4 and Dl, B2 are too far apart from each other by a predetermined value or more, so pairability does not hold between these sub-strokes. Therefore, the black background area indicated by H (
Hereinafter, this will be referred to as a black base mass. ) is not extracted as a stroke, and the obtained approximate pattern is incorrect as shown in FIG. Note that in the approximate pattern in FIG. 16, the strokes indicated by dotted lines correspond to black background blocks that were not extracted.

As described above, conventional character recognition devices have the disadvantage that accurate character recognition cannot always be performed, especially for characters that tend to have extremely short and wide strokes.

<Object of the Invention> The present invention is intended to overcome the above-mentioned drawbacks, and aims to provide a character recognition device that can perform accurate character recognition.

<Structure and Effects of the Invention> In order to achieve the above object, in this invention, a black background extraction means is provided in addition to the stroke extraction means, the stroke extraction means first extracts a stroke from an input image, and then the stroke extraction A connected black background portion in the input image that has not been converted into a stroke is extracted as a black background block by the means.

According to this invention, even if a black background block occurs in character patterns such as "tsu", "tsu", "shi", "n", "n", etc., the black background block is used as stroke alternative information. The information represented is extracted as a character feature, and this can be used to perform accurate character recognition.

<Description of Embodiments> FIG. 1 is a block diagram showing a main part of a feature extraction section of a character recognition device which is an embodiment of the present invention.

In FIG. 1, an image memory 7 is a memory for storing an input image obtained by reading the geometric pattern of unknown characters and converting it into black and white. The boundary labeling circuit 8 is for labeling black pixels located at the boundary between a white background and a black background in the image stored in the image memory 7, as will be described later. Further, the substroke extraction circuit 9 is a circuit for extracting substrokes from the data of the character pattern after labeling, and the stroke extraction circuit 10 extracts paired substrokes from the data regarding the substrokes. This is a circuit that extracts strokes and provides stroke data. The character feature point extraction circuit 11 is a circuit that examines the positional relationship between strokes based on stroke data and extracts geometric features such as intersections and branch points, which are basic features of characters. Furthermore, RAM12 is
The program of the CPU 13, the boundary labeling circuit 8, the sub-stroke extraction circuit 9, and the stroke extraction circuit 10 are executed by the CPU 13 in order to perform the process of extracting the black ground blocks for which no strokes have been extracted from the data after stroke extraction.
a program for controlling the data, labeling information obtained by the labeling circuit, substroke data and stroke data obtained by the substroke extraction circuit 9 and the stroke extraction circuit 10, and black background block data that is the object of the present invention. It is a memory that stores.

First, a method of labeling processing by the boundary labeling circuit 8 will be explained with reference to FIG. 2.

FIG. 2 shows an enlarged view of the part of the character pattern in FIG. 15 that corresponds to part H. The pairing between the strokes D□ and D2 is not established. Therefore, no strokes are extracted from this portion H, and the portion H remains as a black background block.

In Figures 2 and 15, X and y are pixel numbers (coordinates) for identifying each pixel.Also, the pixels corresponding to the character pattern are on a black background, and the other parts are on a white background. I think it is.

First, the boundary labeling circuit 8 labels each pixel (01
, (rl, (/1. As shown,
The pixels adjacent to the top of that pixel are white (Wl), and the pixels adjacent to the right, left, and bottom are black (bl).
This is a label indicating a pixel that looks like this. Actual labeling requires a total of 4 bits for each of the top, bottom, left, and right sides.
It is expressed using binary values of 1" and 0. In the example in the same figure (a), the pixels adjacent to the top and left are white (-)
and by r=right,
It shows which of the pixels adjacent to each pixel has a white background, and the symbol (0) means that there is no white background among the adjacent pixels. In this way, the pixels in FIG.
It is stored in an area called NC:OT in the AM 12 as shown in FIG.

When storing this, first, the unknown pattern in Figure 15 is scanned in the direction shown by the arrow in Figure 13, and from the first pixel of the (u, r) label reached, the character pattern is moved to the left. Track around. Here, tracking in the counterclockwise direction means, as shown in FIG. 5, paying attention to the bits A1 to D1 indicating @1" of the four bits of the label of the pixel where the tracking point is located at that time, and When the position of the bit indicating "1" is A1, the bit indicating "1" is placed at the positions A2 to A in that label or the label adjacent to that label.
Detects whether or not the bit exists in the order of this person 2 to A4, and if it exists, moves the tracking point to the coordinate position of the label that includes the detected bit @1'', and moves the tracking point to the coordinate position of the label that includes the detected bit @ 1" to perform the next tracking.

For example, pixel 31 indicated by label (d, Z) in FIG.
If there is a tracking point at , the position A in this label (d, /)
If we look at bit 1 (°1" because the label includes d), there is no bit @1" in position A2, and bit 11" exists in position A3. Therefore, the tracking point is moved to the adjacent pixel 32 represented by the label containing "1". 5) to (D) are similar, and bit 11'' of the label is detected and tracked in the respective orders of B2 to B, C2 to C4, and D2 to D4.

Through this tracking, when pixels located at the outer boundary of the character pattern are sequentially detected, the position coordinates and label type are stored in the memory area NC0T shown in FIG.
It will be stored above. On the other hand, when a tracking point goes around one connected part of a character pattern, it returns to the original tracking starting point, forming a closed loop. 4. Provide an index labeled "stopper" in Figure 4. Thereafter, the next connected portion is tracked and the same process is repeated. Therefore, the number of "stoppers" in Figure 4 matches the number of connections in the character pattern, and data stored between one stopper and another stopper means data regarding the same connected part. It turns out.

Then, to identify each piece of data in this series,
In front of each series of data is a concatenation number cor1゜C0T2.
Add...

FIG. 6 shows ST for storing data of pixels forming the end points of each stroke in the storage area in the RAM 12.
A part of M area is shown. For example, ASTM refers to an area for storing end point information for incoming strokes, and pixel data corresponding to the four end points of an incoming stroke, for example, stroke SAl, is stored as a pack in substroke units. Data is stored in the same manner in the area C3TM that stores the end point information of the stroke in the C direction.

Next, the processing in this embodiment will be sequentially explained. first,
The reading unit 2 shown in FIG. 9 reads the geometric pattern of the unknown character. This pattern is stored in the image memory 7 (first
The boundary labeling circuit 8 performs the above-described labeling, and the serial data is stored in the NCOT area of the RAM 12 in the manner shown in FIG. Thereafter, the substroke extraction circuit 9 traces the continuity of the labeled black background pixels in four directions to obtain substroke data, and then the stroke extraction circuit 9 refers to this substroke data. , the pairing between substrokes is determined, and the substroke data is classified for each stroke in each direction.

Next, this stroke extraction circuit 9 determines the stroke end points of each stroke from the extracted stroke data, and stores the end point information in the S memory in the RAM 12 in the manner shown in FIG.
Store in the TM area.

Next, the CPU 16 stores N in the RAM 12 as described above.
Using the continuous data and stroke end point data stored in the COT area and the STM area as shown in FIGS. 4 and 6, respectively, extraction processing of black background blocks that have not been converted into strokes is executed. The processing procedure of this black background block extraction process will be explained below with reference to the flowchart of FIG.

In step 21, it is determined whether all stroke end point data in FIG. 6 has been checked. At first °N
o'', the process advances to step 22 and the stroke end point data is loaded.

In the example of FIG. 6, the end point coordinate data of (4, 11) in the ASTM area is loaded first. Next, in step 23, the NCOT area is scanned to determine whether the end point coordinates exist within the NCOT area. In the example of FIG. 4, although not shown in the figure, a match is made in the C0T2 region.

Continuing (In step 24, it is determined whether a match has been made, and if there is a match, the process proceeds to step 25, and the corresponding COT
Label the number, C0T2. This labeling is performed by flagging a specific bit associated with the COT number ON. Then, the process returns to step 21 to check whether all stroke end points have been checked. Note that when the determination in step 24 is "No", the process returns from step 24 to step 21.

When the above processing is executed for all end point data in the 57M area, the determination in step 21 becomes "YES" and the process proceeds to step 26.

In step 26, the COT number areas within the NCOT area are sequentially scanned. Subsequently, in step 27, the COT number scanned in step 26 is replaced with an unlabeled COT number.
Determine whether it is an OT number. When the COT number is not converted to label/)j, it suggests the existence of a black background part of the connection for which no complete stroke information exists, so the process advances to step 28 and the N of RAM 12 as shown in FIG.
The number of black background blocks and their COT numbers are stored in the BLACK area. When the processing in step 28 is completed, the process proceeds to the next step 29, but the determination in step 27 is not completed.
If "NO", step 28 is skipped and the process moves to step 29. In step 29, it is determined whether the entire COT number area of the NCOT area has been scanned,
If @YES', the process ends; if "No", the process returns to step 26 and continues the process of step 3 and steps 26 to 29 described above.

By executing the above processing, it is determined whether or not there is a stroke end point that is not extracted as a stroke end point in the information representing each continuous black background portion. If there is information that has not been extracted as a stroke end point, the continuous black background portion has not been extracted as a stroke, and will therefore be extracted as a black background block. This extracted black background block information is stored in the RAM 12 as described above.
This is stored in the NBLACK area of , and by referring to it in later dictionary matching processing, it becomes possible to perform reliable character recognition.

In this case, the above-mentioned black background block information is stored in advance in the standard pattern in the dictionary.
It can be said that there is a high need for dictionaries for patterns that tend to have black background blocks, such as ``ノ'', １し'', １しょ'', １ん'', etc.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the main parts of the feature extraction unit in an embodiment of the present invention, FIG. 2 is a diagram showing a part of an input character pattern, FIG. 3 is an explanatory diagram of labeling, and FIG. Figures 6 and 8 are explanatory diagrams of storage areas, Figure 5 is an explanatory diagram of counterclockwise tracking, Figure 7 is a flowchart showing the procedure of black background block extraction processing, and Figure 9 is a diagram of a conventional character recognition device. The block diagram shown in FIG. 10 is a diagram showing thinning processing of an input image, and FIG.
12 is an explanatory diagram of the tracking direction; FIG. 13 is an explanatory diagram showing an example of scanning the input image; FIGS. 14 and 16 are diagrams showing approximate patterns; FIG. 15 is a diagram showing an example of an input character pattern. 8...Boundary labeling circuit 9...Substroke extraction circuit 10...Stroke extraction circuit 12...RAM 13...CPU Patent applicant Tateishi Electric Co., Ltd. φ1 吋・Su block 6v 云3 fig. ``2 fig. Manpower and enjoyment to the turn of the turn.
Moon L4'76Ei21 Ancient) Seireihiki to Seven-Moon Surface ÷ Moj EnJゎRiu'' 5 straddles, Su L K Moe Side''A q 12] 't Medium'A, IKj1.4
a+I*6*=, <ro-rqrA67 diagram

Claims (2)

[Claims] (1) Read the character pattern of unknown characters, convert the pattern to black and white to obtain an input image, extract the geometric features of the input image using the feature extraction section, and then output the output from the feature extraction section to the dictionary section. A character recognition device that recognizes the unknown character by comparing it with a standard pattern stored in the input image, wherein the feature extraction unit extracts a directional stroke by adding a boundary between a black background and a white background of the input image. and a black background block extraction unit that extracts, as a black background block, a continuous black background part in the input image that is not extracted as a stroke by the stroke extraction unit. Device. (2) The character recognition device according to claim 1, wherein the standard pattern is formed including information regarding the black background block.