JPH01292487A - Device for recognizing character - Google Patents

Abstract

PURPOSE:To furthermore improve a character recognizing speed by executing the smoothing processing of a character pattern extracted from image character information. CONSTITUTION:Smoothing processing to invert the data of an objective picture element PEOBJ when a comparing pattern consisting of the objective picture element PEOBJ and its adjacent picture elements PENF1-PENF8 coincides with a smoothing reference pattern PTREF is executed. When the objective picture element PEOBJ is a black character part projecting from a black character part, when white picture elements are mixed with the black character part, when an independent white picture element is mixed in the black character part, or when an independent black picture element is mixed in a white background part, respective parts concerned are removed or embedded in accordance with necessity to furthermore simplify the peripheral rugged shape of a character pattern. Since the identifying accuracy of a character identifying part can be improved, the whole character identifying speed can be furthermore improved.

Description

[Detailed description of the invention] The present invention will be explained in the following order.

A. Industrial field of application B. Overview of the invention C. Conventional technology Problems to be solved by the invention (Figure 2) E. Means for solving the problems (Figures 1, 3, 6 and 6) 8th
Figure) 2 Effects (Figures 1 and 8) G Example (G1) Example of Character Recognition Device (Figures 1 to 10) (
G2) Other Embodiments H Effects of the Invention A Field of Industrial Application The present invention relates to a character recognition device, and is particularly applicable and suitable for recognizing printed characters.

B. Summary of the Invention According to the present invention, character recognition speed can be further improved by smoothing a character pattern extracted from image character information in a character recognition device.

C. Conventional technology In the past, it has been proposed to digitize and file a large amount of printed documents, or create a database to build an information network that can be used for a variety of purposes. Instead of character information input devices that require input operations, it has been considered to use character recognition devices that do not require manual input operations.

The reason is that printed character recognition devices generally optically read the printed characters on a printed document, digitize them as two-dimensional image information, extract the printed characters from the image information, and output the corresponding character code. is being done.

By encoding image information into character codes in this way, you can search for words using a computer,
It is possible to realize a document processing system that automatically executes decoding processing such as understanding meaning.

In this way, for example, it is possible to compress the data extracted from the image information and use it for decoding processing as necessary, thereby making it possible to improve the document processing speed as necessary.

Problems that the invention aims to solveHowever, the currently used printed kanji recognition devices actually have a high recognition ability of about 97 to 99% in terms of recognition rate, but the recognition speed is about 10 to 30 characters/second. However, it is still insufficient as an input means for digitizing a large amount of printed documents.

Conventionally, as a method to improve identification speed, there has been a large classification method in which candidate characters are selected from a large classification dictionary by detecting special @ quantities based on the characteristics of input information extracted from image information (for example, peripheral features on four sides). A character recognition method has been proposed in which the most similar characters are determined by executing processing and subclassifying the candidate characters based on pattern matching (Japanese Patent Application Laid-Open No. 62-18
6390).

In this way, character recognition methods that hierarchically perform major classification processing and subclassification processing further improve identification ability, especially when identifying characters that do not have special characteristics such as shape or position (i.e., normal characters). It has been confirmed that there are advantages that can be achieved.

However, in practice, image scanners are generally used as a means of obtaining two-dimensional image information from printed documents, but when converting optical light information into electrical binary information, fine details are generated in the peripheral areas of printed characters. There is an unavoidable problem that unevenness occurs.

For example, as shown in Figure 2, in the character pattern PTIIOJI formed from image information obtained by reading the character ``娃'' with an image scanner, the upper side of the character part or the boundary of the upper side slopes slightly upward to the right. It is observed that a concavo-convex pattern portion DBKO is formed in which a few dots of "black" dots are arranged with "white" dots of a few dots in between.

This uneven pattern portion DEKO is generated from image information obtained from standard printed characters to be identified, and exhibits an uneven shape determined by the arrangement of image sensors provided as photoelectric conversion elements in the image scanner. The character pattern PT of the character "娃" when viewed microscopically. It can be said that this represents the characteristics of Jl.

However, the character pattern PT. When trying to recognize characters using a peripheral feature extraction method that grasps the uneven shape seen from the periphery of 1 as a feature, the uneven pattern portion DBKO that occurs at the periphery of the character part is regarded as an element of the peripheral feature. Extracting the amount of characters may make character features unnecessarily complicated.0 If the character features become complex, the general classification process becomes complicated, and as a result, the probability of misrecognition of characters increases. This may cause deterioration of recognition speed.

The present invention has been made in consideration of the above points, and it is an object of the present invention to simplify the shape of an uneven pattern portion included in image character information generated by an image scanner without impairing the main characteristics of the character pattern. This paper attempts to propose a character recognition device that can prevent causes of deterioration in character recognition speed.

E Means for Solving the Problem In order to solve the problem, in the present invention, the pixel of interest P is
The comparison pattern PTco consisting of E, , □ and its adjacent pixels PENFI to PENr* is compared with the smoothing standard pattern PTmty, and when they match, the pixel of interest PE0IJ
By inverting the data, a smoothing processing pattern F'Tssz is obtained, and character pattern output data D A T consisting of the smoothing processing pattern PTs and 2 is obtained.
Character identification processing is executed based on Aout.

F action pixel of interest PE0IJ and its adjacent pixels PENFI to PE
When the comparison pattern PTcox consisting of NFI matches the smoothing standard pattern PTmtr, the pixel of interest PE0I
By performing smoothing processing such as inverting the data of J, it is possible to find that the pixel of interest PE0IJ is a black part that protrudes from the black character part, white background pixels are mixed in the black character part, or white background pixels are isolated in the black character part. In cases where pixels are mixed or isolated black pixels are mixed on a white background, this can be removed or embedded as necessary, further simplifying the uneven shape around the character pattern. obtain.

By thus being able to improve the identification accuracy in the character identification section, the character identification speed of this branch can be further improved.

G example An embodiment of the present invention will be described in detail below with reference to the drawings.

(Gl) Embodiment of character recognition device In FIG. 3, 1 indicates the character recognition device as a whole;
The image read signal S1 obtained in the original reading section 2 is applied to the noise removing means 4-.

The noise removal means 4 operates to avoid the possibility of erroneously recognizing the so-called isolated points as part of a character by removing noise about the so-called isolated points included in the image read signal St. do.

The noise removed output S2 of the noise removing means 4 is given to the rotation correcting means 6 of the character cutting section 5, and after correcting the rotation error of the document, the corrected output S3 is supplied to the character string extracting means 7.

The character string extraction means 7 separates the character area from other areas (for example, areas of photographs, drawings, etc.) in the printed document, extracts only the image character data included in the character area, and extracts only the image character data included in the character area. After confirming that the character string to be written is horizontal writing, extract the character string.

This character string extraction is performed in the character area A, as shown in Figure 4.
The position of each dot in R in the column direction (horizontal direction)
The character string ARI constituting the character area AR is expressed by xy coordinates consisting of the axis and the y axis taken in the row direction (vertical direction).
, AR2...... by projecting the sum of the logical "1" level dots (representing the dots in the black character part) onto the y-axis (this is called y-projection). Obtain a signal S.

Here, since there is no black character part between the lines between the character strings ARI, AR2, etc., the signal level of the y projection signal Sy becomes "0" level, whereas the signal level of the character string A
At positions on the y-axis corresponding to RI, AR2, . . . , each character included in the character string has a signal level corresponding to the total number of dots on the in parallel to the X-axis. Therefore, the y projection signal Sy is compared with a predetermined threshold level, and the logic is "1" during the interval above the threshold level.
Character string cutout data CL that rises to the level can be obtained.

The character string extraction means 7 uses this character string extraction data CL to extract the signal portion of the correction output S3 given from the rotation correction means 6 during the period in which the character string extraction data CL is logic "1". , character strings ARI, AR2 in each line...
. is supplied to the character cutting means 8 as a character string signal S4.

In this way, the period in which the character string extraction data CL rises to the logic "1" level is the character string ARI, AR2, . . .
Maximum height HL of the rows (=HL1, HL2...
), and the position of the characters included in each line in the height direction (i.e., y-axis direction) is the maximum height H of the line.
This will occur in the range of L (-HLI, HL2...).

The character cutting means 8 extracts the character string API, AR2, etc. of each line.
..., the position and range of each character and its constituent character parts (in the case of separate characters) in the X direction and y direction.
By detecting the direction, as shown in FIG. 5, processing is executed to cut out a rectangular area CHR surrounded by a circumscribing frame WAKU of height h and width W that circumscribes each character and its constituent character parts. do. In addition,
P represents the average character pitch of full-width characters, and d represents the rectangular spacing.

Thus, the rectangular area CH surrounded by the circumscribing frame WAKU
Inside R, as described above with reference to FIG. 2, there is a time period in which the dots in the black character part of the character pattern MOJI are represented by the logic "1" level, and the dots in the white background part are represented by the logic "0" level. Character pattern input data DATA, , having a serial structure of two typefaces is formed.

The character cutting means 8 receives character pattern input data DAT A.
By subjecting IN to smoothing processing by the smoothing circuit 15 shown in FIG. 6, character pattern output data D AT A ouy as shown in FIG. 7 is obtained, and this is sent to the character identification section 9 as input character information S5. .

The smoothing circuit 15 uses character pattern input data DA.
For each dot pixel that constitutes TAIN (Fig. 2), as shown in Fig. 8, the pixel is designated as the pixel of interest PE0IJ.
and surrounding eight adjacent pixels PE,4
Comparative cover turn PTC composed of □ ~ PENFII
O) Compare I with the smoothing standard pattern PT□ shown in FIG.
By performing an exclusive OR operation on the data rPEomJJ of 0IJ and the logic rlJ, a smoothing process is executed to invert the data of the pixel of interest in the smoothing process pattern PTs, 4z.

Smoothing standard pattern P T in case of Fig. 1 (A)
++wv (= P T * Er++ ~ PTIEF14
) is character pattern input data D AT A IN (
This is applied when a comparative cover pattern PTcos has arrived, in which the target pixel PE0IJ protrudes outward from the black text area by one pixel in the concavo-convex pattern portion DEKO in Figure 2).In this case, smoothing is applied. circuit 15
is a smoothing processing pattern P Tsn by performing smoothing processing to remove the protruding pixels.
z (−P Tstz+”) is obtained.

Also, the smoothing standard pattern PT*t in FIG. 1(B)
v (-PTatrz+~PTxtrz*) is
Comparison cover turn PTC in which the white background part cuts into the black character part by one pixel in the part of the target pixel PE0IJ. In this case, the smoothing circuit 15 performs a smoothing process to embed the white pixel in question with a black pixel, and then creates a smoothing process pattern PT*Mz (=PTsw).

).

Also, the smoothing standard pattern PTtt in FIG. 1(C)
tv (=PTmtrs+ ~PT□.4) is similarly applied when a comparison cover pattern PTco14 arrives, in which a white background pixel digs into a black character part by one pixel, and at this time, the smoothing circuit 15 A smoothing process pattern PTsxz (=PT3□,) is obtained after performing a smoothing process similar to that of embedding the white pixel part with black pixels.

Also, the smoothing standard pattern PT*t in FIG. 1(D)
y (=PT□F4) is applied when a comparison cover pattern PTcox in which only one white pixel is isolated in a black text area has arrived.
At this time, the smoothing circuit 15 performs smoothing processing to embed a black pixel into the white pixel, and then obtains a smoothing processing pattern PT□z ("PTs.4)."

Smoothing standard pattern PT*wv in Figure 1 (E)
("PTmz□) is a comparison pattern PTCO in which one black pixel is isolated on a white background.
This is applied when M arrives, and at this time, the smoothing circuit 15 performs a smoothing process to erase the isolated black pixel, and then creates a smoothing process pattern P.
Tsnz (-PTsxzs) is obtained.

Smoothing standard pattern PT*tr in Figure 1 (F)
(-PT□F&) is applied when a comparison input cover pattern PT Cog in which only one white pixel is mixed in the center pixel of the cross-shaped line part, such as in the case of intersecting line parts, is applied. , At this time, the smoothing circuit 15 performs a smoothing process to embed the white pixel with a black pixel, and then creates a smoothing process pattern PT. (”
P T s□, ) is obtained.

The smoothing circuit 15 (FIG. 6) sequentially transfers character pattern input data DATA ray consisting of time series data to shift circuits DLII and DL12.1 for a 1-bit delay.
Line delay shift circuit DL21.1 Bit delay shift circuit DL22, DL23, l Line delay shift circuit DL31.1 Bit delay shift circuit DL32, DL;
33, one bit at a time, and thus the 1-bit delay shift circuit DLI
L DLI2, DL22, DL23, DL32, DL3
From the input end and output end of 3, sequential comparison type cover turn P
Pixels P Eur*, P EMF configuring TCON
? , PENF & -, P ENrs s P Eo
mJSP ENF4, P ENF3, P ENFrz
, P E□, is extracted from the ROM (read
The signal is supplied to a smoothing processing circuit SMZ having a (only memory) conversion circuit configuration.

The smoothing processing circuit SMZ is shown in Figures 1 (A) to (F).
) described above for the smoothing standard pattern PTmt
It has a conversion table that converts r into a smoothing processing pattern PT□2, so that each time the character pattern input data DATA IN is shifted through the shift circuit one bit at a time, the comparison input cover pattern PT,
. , is compared with the smoothing standard pattern PT□7, and when a matching pattern arrives, it is converted into a smoothing processing pattern PTsxz and sent as character pattern output data DATA out.

In this embodiment, the character identification unit 9 identifies candidate characters through major classification processing based on peripheral features, and then performs hierarchical identification processing such as performing subclassification processing based on pattern matching on the candidate characters. It is made to be.

Here, as a method for character recognition based on peripheral characteristics, the method disclosed in Japanese Patent Application Laid-open No. 186390/1983 can be applied.

Incidentally, the character identification section 9, as shown in FIGS. 9 and 10,
Input character MOJI surrounded by a quadrilateral circumscribing frame WAKU
, the left and right sides (these are called sides A and B) are the first and second sides seen from WAKUA and WAKUl sides.
Based on the peripheral features of and the third and fourth peripheral features seen from the upper side and the upper side (these are called 0 side and 0 side) WAKUc and WAKUゎ, the character created by the black character part on a white background. A feature amount DcNII is formed by encoding the structure, and similar candidate characters are selected from the major classification dictionary based on this feature amount DC□.

The peripheral feature on side A is human power character MOJI
Among them, when the black character part of the third dot line along the left side WAKU a is viewed from the direction indicated by the arrow ARA, the lengths of the black character part on the left side of the dot line LA are dsI and ds2 , ds, is encoded to form a feature amount D cMt.

In the case of Figure 1O, there is no dot in the upper left corner because the length C3 of the scanning start portion (this is called the start corner) at the upper left corner of the left side character portion MOJIa is quite long. can be expressed.

On the other hand, since the length Ce of the lower left corner at the end of scanning (this is called the end corner) is short, it can be indicated that there is a dot at the end corner.

Thus, when the lengths Cs and Ce of the start corner and the end corner are shorter than the predetermined threshold length, the data of logic "1" representing that there is a black character part at the start corner and the end corner is set to the 0th and 1st of the feature quantity DCMI. The data of the th bit B and B.

On the other hand, the length of the black character section between the start corner and the end corner is ds,,da!・・・・・・ds
When n is shorter than a predetermined threshold length, data of logic "0" is set (this indicates that the black character part is a dot), whereas when n is larger, data of logic rlJ is set (this This means that the black character part is a long dot).

The data corresponding to this length da, Sds, ...... is the data of the second, third ...... bit Bz
sB, ......, the way the lengths of the black character part are arranged (therefore the way the dots and dashes are arranged), that is, the characteristics of the character MOJI, can be expressed as the feature amount Dcm of the variable length code. It can be expressed as

In this way, the character identification unit 9 obtains feature value numerical data consisting of feature values DC□ for the A side, B side, 0 side, and D side in advance for standard characters, and stores them as standard character codes in the major classification dictionary. Then, when the input character information S5 arrives, a standard character code having the same feature amount Dell as the feature amount D CHI of the input character information S5 is delivered to the subclassification processing circuit as candidate character information.

The subclassification processing circuit reads the character pattern for the candidate character information from the subclassification dictionary, performs pattern matching with the character pattern of the input character information S5, identifies the most likely character, and uses the identification result as a recognized character. It is sent as information S6.

According to the above configuration, the smoothing process is executed in the smoothing circuit 15 of the character cutting means 8, so that the character pattern PT is cut out from the character string signal S4. Concave and convex pattern part DE on Jl (Fig. 2)
Even if a KO occurs, it can be smoothed, resulting in character pattern output data D A that does not have an uneven pattern around the black character part, as shown in FIG.
After converting it into T A out, it can be delivered to the character recognition section as input character information S5.

As is clear from comparing Figures 2 and 7,
Character pattern output data D after smoothing processing
Character pattern PT14゜J represented by A T A oat
IK is character pattern PT processing of character pattern input data DATAIM before smoothing process. , I is significantly simplified compared to the uneven shape in the case of I.

Moreover, the character pattern PT simplified in this way
There is no risk that the uneven shape around MOJIX will significantly change the main shape features of the input character MOJI, so when extracting the peripheral feature amount D eNR in the character recognition unit 9, the feature amount is , □ can effectively avoid the possibility of unnecessarily complicating the data, and can improve the processing efficiency of this fractional classification.
The character Ill speed can be improved as a whole.

Incidentally, if the data of the feature quantity Dell regarding the peripheral feature becomes complex, even minute characteristic changes in the input character MOJr will be incorporated into the feature quantity Dell, so the character recognition unit 9 will extract the peripheral feature extraction result. However, according to the above-described embodiment, such fluctuation factors can be effectively eliminated.

(G2) Other Embodiments (1) In the above-described embodiments, a case was described in which the character identification unit 9 performs identification processing based on peripheral characteristics, but when performing other identification methods as a character identification method, The present invention can also be widely applied.

(2) In the above-mentioned embodiment, a case has been described in which the smoothing process is performed in the character extraction means 8, so that the smoothing process is performed after the processes in the rotation correction means 6 and the character string extraction means 7 are performed. However, in addition to or in place of this, the same effect as in the above case can be obtained even if the noise removing means 4 executes smoothing processing.

(3) In the above embodiment, eight adjacent pixels PE, □ to PEu surrounding the pixel of interest PBOIJ
Although the smoothing process is performed using all of rs, some of the eight adjacent pixels may be omitted.

H Effects of the Invention As described above, according to the present invention, by smoothing the uneven pattern portion that occurs on the periphery of the character pattern of image character information, the outer shape of the character pattern can be simplified, thereby achieving - A character recognition device that can efficiently perform character identification processing can be easily realized.

[Brief explanation of the drawing]

FIG. 1 is a route map showing a smoothing pattern used in the smoothing process executed in a character recognition device according to the present invention, FIG. 2 is a pattern diagram showing a character pattern of character pattern input data, and FIG. 3 is a character diagram according to the present invention. FIG. 4 is a block diagram showing the overall configuration of the recognition device. FIG. 4 is a route diagram for explaining the character string extraction operation. FIG. 5 is a route diagram for explaining the rectangular area extraction operation. FIG. 6 is a route diagram for explaining the rectangular area extraction operation. A block diagram showing the configuration, FIG. 7 is a pattern diagram showing character patterns of character pattern appearance data subjected to smoothing processing, FIG. 8 is a route map showing comparative cover turns to be subjected to smoothing processing, and FIGS. 9 and 1O. FIG. 3 is a route map showing the principle of extraction of the ferrule feature. 1...Character recognition device, 2...Document reading unit, 3...Character recognition processing unit, 4...Noise removal means, 5...・Character cutting part, 6...
・Rotation correction means, 7...Character string extraction means, 8.
...Character cutting means, 9...Character identification section,
15...Smoothing circuit, DLII to DL3
3...Shift circuit, SMZ...Smoothing processing circuit.

Claims (1)

[Claims] A comparison pattern consisting of a pixel of interest and its adjacent pixels among character pattern input data consisting of binary data cut out from a character string signal is compared with a smoothing standard pattern, and if they match, the data of the pixel of interest is A character recognition device characterized in that a smoothing processing pattern is obtained by reversing the pattern, and character recognition processing is performed based on character pattern output data formed by the smoothing processing pattern.