KR930001471B1 - Character recognition method and apparatus for handwrite and printer - Google Patents

Abstract

No content.

Description

Method and apparatus for recognizing handwritten and printed characters

1 is a block diagram showing a processing sequence of a conventional pattern or character recognition system.

2 is a view showing an embodiment of an image processing apparatus to which the present invention can be applied.

3 shows a sequence for determining external and internal contours and contour tracking starting points in accordance with the present invention.

4 shows a directional code at each contour pixel when the present invention is applied.

5 is a diagram showing a sequence of direction code extraction in contour tracking according to the conventional art.

6 is a view showing a sequence of extracting the director code from the direction code applied to the present invention.

7 is a view showing a sequence of extracting the director code of the first smearing range from the director code according to the present invention.

8 is a view showing a sequence of extracting a director code string into a sequential order to the R-summing range according to the present invention.

9 is a view showing a procedure of processing replacement from a shape code string extracted from a director code string of a predetermined smoothing range according to the present invention.

10 shows a sequence for recognizing a pattern of text in accordance with the present invention.

FIG. 11 is a diagram illustrating a procedure for counting the number of successful pattern recognition in each smoothing range during the process of FIG. 10.

12 shows an image pattern of the English character "E".

12A-12F show the shape of the result of processing the image pattern of FIG. 12 in accordance with the present invention.

FIG. 13 shows an image pattern of an English character "E" in a different style.

13A-13C show the shape of the result of processing the image pattern of FIG. 13 in accordance with the present invention.

14 is a view showing an image pattern of the Chinese character character "可".

14A-14J illustrate shapes of processing the image pattern of FIG. 14 in accordance with the present invention.

15 is a view showing the image pattern of the Chinese character character "可" in a different style.

15A-15C show the shapes of the image pattern of FIG. 15 processed in accordance with the present invention.

FIG. 16 shows the image pattern of the Korean character "oyster", in which the white pixel "0" is omitted.

16A and 16B show shapes in which the image pattern of FIG. 16 has been processed in accordance with the present invention.

FIG. 17 shows an image pattern of the Japanese character "S", in which the white pixel "0" is omitted.

17A-17D illustrate shapes in which the image pattern of FIG. 17 is processed in accordance with the present invention.

FIG. 18 is a view showing an image pattern of Japanese character "S" of a different style, in which the white pixel "0" is omitted.

18A-18D show the shapes of the image pattern of FIG. 18 processed according to the present invention.

The present invention relates to a method and apparatus capable of recognizing a pattern in a handwritten character, a print character, and the like, in particular, a pattern recognition method that can be used as a human-machine interface of an electronic device such as a computer, a typewriter, a speech synthesizer, and the like. And to an apparatus. Various systems have been proposed for pattern recognition. The pattern recognition system typically inputs data indicative of a pattern to be recognized, and performs a predetermined operation for comparing with a known pattern in order to recognize the input pattern. 1 is a block diagram showing a processing procedure of a conventional pattern recognition system. The input pattern 10 is a pattern to be recognized. Digitizer 11 Digitizer 11, used in such devices as, for example, facsimile machines, electronic copiers and optical character recognition systems, has a binary data " 0 " corresponding to the bag to store input patterns in system memory. And digital image data of binary data " 1 " corresponding to black. It is treated as 1 when the gray level is greater than the threshold and 0 when the gray level is less. The input image data is segmented into each character (step 12), and specific extraction of each segmented character is performed (step 13). Data representing the extracted feature is compared to match one of the known pattern data (step 14). The matched pattern-data is converted into output data such as ASCII or the like provided to an output device such as a printer, a display device and a speech synthesizer (step 15). Specific extraction methods proposed by known techniques are thinning and contour tracking. The specific extraction by the thinning method consumes a lot of time in the process of extracting the skeleton from the characters written in thick strokes, and leads to the difficulty of accurate extraction of the skeleton due to stroke irregularity or noise sensitivity. The specific extraction by the contour tracking method tracks the black or binary " 1 " at the boundary of the stroke using the directional code, thereby enabling faster processing than the thinning method.

However, in the contour tracking method, processing of a specific extraction that can remove the nonuniformity or noise of the stroke is required. One prior art of the contour tracking method is November 1972, IEEE Trans, Comput, Vol. C-1, No. 11, pp 1206-1216 and 1988, named "Method of optical character recognition". No. 4,773,098, issued September 20, to WCScott.

In this patent, the input pattern digitized from the image is transferred to the memory. Only data representing character boundaries is read from the data in memory. During contour tracking, character variables such as height, width, perimter and area are determined, which are compared with the reference pattern that matches the character. Pattern recognition is performed by the output of the reference pattern in the presence of a matching reference pattern. However, this method is a specific extraction method irrespective of the size of the character, but no further processing is performed to remove noise due to the error reading of the scanner, the quality of the paper or the smearing of the ink. In this method, it is also difficult to recognize characters that have changed in various forms depending on the personal writing style, the characteristics of different strokes according to the typeface, and the quality of paper and ink.

In general, a character can produce various formations according to a writing style, a printing style, and the like. Complex shapes that develop at the edges of strokes, partially bleed contours, changes that progress from a straight line to other shapes (for example, "木" in Chinese characters), and character shapes consisting of one or more phonemes (for example, " Standing, cow, place ", etc.). Therefore, one character can be written or printed with various modified characters. Therefore, when a specific extraction method using a simple contour tracking method is used, various reference patterns for one same character must be stored in a read only memory (ROM) or a specific storage means. In this case, since the amount to be stored in the ROM or storage means is not limited, the overall ROM capacity is greatly increased. In particular, serious problems arise in the case of Chinese characters, Korean characters, and Japanese characters having various strokes.

Because some characters, especially characters such as Chinese characters, have external and internal contours, it is required to extract internal and external contours in certain extraction of such characters. It is therefore an object of the present invention to provide a method and apparatus for recognizing a character that eliminates noise or non-uniformity on a line in character-specific extraction using the contour tracking method.

It is still another object of the present invention to provide a method and apparatus capable of recognizing a pattern regardless of the style or size of a written or printed character in a character specific extraction using the contour tracking method.

It is another object of the present invention to provide a pattern recognition method and apparatus for minimizing the storage capacity of a reference pattern regardless of the styles or sizes of handwritten or printed characters.

It is another object of the present invention to provide a method and apparatus for tracking the outer and inner contours of an image pattern.

It is still another object of the present invention to provide a pattern recognition method and apparatus capable of performing a specific extraction of an image pattern accurately by a contour tracking method regardless of a thin line or a thick line.

In order to achieve the object of the present invention as described above, the present invention extracts a direction code string that traces the contour of the pattern from the stored image data, and changes the direction of each contour pixel point from the direction code string. And extracting a director coded string of a predetermined smoothing range from the director coded string to remove noise of the pattern. The method may further include scanning a plurality of lines in one direction with respect to the image pattern, detecting a first contour point that meets the contour of the image pattern during the scanning, and contouring from the contour point when the first contour point is detected. The second contour according to the process of tracking, the process of detecting the second uncontoured contour point on the image pattern during the same line scanning, and the number of the contour points detected from the second contour point during the same line scanning. Characterized in that the process consists of detecting whether the point is an internal or external contour. The method may further include storing a replacement code in the memory device, extracting a director code string of a predetermined smoothing range from the director code string, and selecting a first code string selected from the extracted director code strings with the replacement code. It is characterized in that the process of comparing and replacing can be performed.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. The term " image " used in the present invention should be understood to indicate a character, a sign, a figure, or the like. 2, an embodiment of an apparatus for realizing an image processing method according to the present invention is shown. Reference numeral 20 is an image scanner, which converts an image into digital data. The scanner 20 used in the present invention is MS-2, 300A manufactured by MICROTEE Co. The scanner 20 may process black and white pixels having 300 dots per inch (DPI). However, it should be understood that the present invention is not limited to such scanner 20. Reference numeral 22 is a microprocessor that controls image processing in accordance with the present invention. The memory 24 is composed of a read only memory (hereinafter referred to as " ROM ") and a random access memory (RAM). The ROM stores dictionaries, i.e., reference patterns and data for substitution, which will be described later, for comparison with the data processed for the image processing program and the character according to the present invention.

The image data digitized from the scanner 20 is stored in the RAM after being processed as a matrix of binary digits in a known preprocessor (not shown) or by the processing of the microprocessor 22. If one page of text is scanned by the scanner 20, the data stored in the RAM is separated into each character by a known image segmentation process. Image data of each segmented character is processed to perform operations such as internal and external contour extraction, specific extraction, and pattern recognition of text under the control of the microprocessor 22 according to the characteristics of the present invention described below. After such specific work of the present invention is performed, pattern recognized data is converted into ASCII code and output to an output device 26 such as a printer, display, speech synthesizer, or the like.

According to another preferred embodiment of the present invention, a pattern recognition system using the scanner 20 and a general purpose computer may be used. In such a case, the image data digitized from the scanner 20 is processed into a matrix of binary digits by a preprocessor such as a digital signal processing board embedded in a general-purpose computer, and stored in a buffer memory. The operation for performing the pattern recognition operation according to the present invention is performed by the operation procedure of a storage device such as a hard disk or a floppy disk which stores a program input to a keyboard. The general purpose computer used in this case is "SPARC station 1" manufactured by "SUN Microsystem".

[External and External Contour Extraction]

Specific extraction for recognizing a pattern basically uses the contour tracking method in the present invention. Before contour tracking is performed, the contour tracking starting point must be found, and some types of patterns have internal contours. Patterns that have an inner contour and whose inner contour plays an important role in determining the characteristics of the pattern must not only find the starting point of the inner and outer contours, but must also distinguish between the inner and outer contours. In addition, as described above, a pattern having an inner stroke or a shape separately in which the outer stroke or the shape is stacked. When one character is combined with unconsonant consonants and vowels, like Hangeul, it is necessary to find the starting point of contour tracking of each phoneme in order to extract each phoneme. Referring to FIG. 3, there is shown a flow chart of a method for finding the inner and outer contours and their contour tracking starting points in accordance with the present invention. For ease of explanation, the Chinese characters shown in FIG. 14 will be described with the flowchart of FIG. 3, for example. The character of FIG. 14 is an image pattern stored in RAM or buffer memory by segment processing as described above, where binary "0" is an image value of a white pixel of an input pattern, and binary "1" represents an image value of a black pixel. , Symbol R represents a row and symbol C represents a column. The illustrated character is therefore an image pattern stored in the form of a 38x38 NxM matrix in rows and columns. A column or vertical scan is initiated starting from the first column C = 1 under the control of the microprocessor 22 to land the outer and inner contours.

The contour can be found by reading and comparing the image values IPD (R, C) and IPD (R + 1, C) of two adjacent pixels during the scanning (process 302-304). For example, while scanning the first column (continuously performing steps 302-306), the sequence from step 305 to step 307-309 proceeds sequentially when R = 38 (= N) because there is no binary change. . Therefore, the scanning of the first column ends and the scanning of the next column (C = 2) is started. Scanning the second column, if the fifth type is selected (R = 5), then the procedure proceeds from step 304 to step 310 since IPD (5, 2) ≠ IPD (6, 2). From the detection of the binary change, the count number CNT becomes two (step 310). Since the pixel at the pixel position SP1 (6, 2) has never been contour tracked before, the procedure proceeds to step 312 through step 311. Since CNT (= 2) is even in step 312, in step 313, position SP1 (6, 2) is stored as the external contour tracking start point and contour tracking proceeds from step 314 from this starting point SP1 (6, 2). . While tracking the pixels in the contour, the directional code described below is determined and marks the traced pixels by tagging the traced pixels. The row is then incremented by one in step 315 and proceeds to step 302. When R becomes 8 due to the increase of the row (R = 8), the next binary change occurs, so that the count number CNT becomes 3 in step 310.

However, process 315 is performed after process 311 because the pixels at positions 8 and 2 are already traced points. Therefore, since there is no binary change from R = 9 to R = 37, steps 302-306 are repeatedly performed. If R = 38, the count number CNT is initialized to 1 by executing steps 305 and 307-309, and the next column (C = 3) is scanned. If a binary change is detected during the scanning from the third column to the seventh column, the corresponding pixel is already contour tracked, so step 315 is performed after step 311, and no new contour tracking start point is found. If the tenth row is selected while scanning the ninth column, the count value CNT becomes 4 in step 310 because it is IPD (10, 8)? IPD (11, 8). However, since the pixel at pixel position SP2 (11, 8) has never been tracked before, in step 313 the position SP2 (11, 8) is determined and stored as another new outer contour tracking starting point, and in step 314 Contour tracking is performed for directional code extraction.

If the fifteenth row (C = 12, R = 15) of the twelfth column is selected while scanning in the above-described manner, the process value 310 is not equal to the image value IPD (16, 12). The count number CNT becomes 5 at. However, in step 312, the location SP3 (15, 12) is determined and stored as an internal contour starting point, because the pixel at location SP3 (15, 12) has never been tracked before, and in step 314 the internal contour direction codes are contoured. Extracted by the tracking method. When the pixels at the positions 37 and 38 are selected while scanning in the above-described manner, the extraction of the outer and the contour and the trace starting point is terminated through the processes 305 and 307 and the process is moved to the next task.

As a preferred embodiment of the present invention, the present invention has described a method for finding internal and external contour trace starting points by column scanning, but a row scanning method may be used. In addition, the character may have a thin thickness, and may not return to the starting point by encountering a contour traced point during contour tracking. At this time, the procedure of extracting the inner and outer contours is performed by setting the contour point outside the contour point already tracked while returning to the starting point after the tracked point.

[Specific extraction]

In accordance with the present invention, the contour tracking method is basically used for specific extraction of the input image pattern. As described above, the contour tracking starting point is determined by the method of FIG. Then proceeding in a clockwise direction from the contour tracking starting point, by step 314 of FIG. 3, the direction codes between two adjacent contour pixels are sequentially determined until they return to the starting point. As a preferred embodiment of the present invention, the directional codes used in the present invention are the codes shown in FIG. In other words, Dongeun 1, Northeast 2, North 3, Northwest 4, West 5, Southwest 6, Remnant 7, Southeast 8 and 8-degree codes.

(1) direction code

After the above-described determination of the contour tracking start point, the direction code sequence extraction method is shown in the flowchart of FIG. Such a method is disclosed in the prior art, for example, in U.S. Patent No. 4,773,098, as described above, and no further details will be given in connection with FIG. See also Digital Image Processing, Rafael C. Gonzalez / Paul Wintz, Addison-Wesley Publishing Co., 1987, especially section 8.1.1, and in November 1972, IEEE Trans, Comput. Vol. See C-1, No. 11, pp 1206-1216.

(2) director code

After extracting the aromatic code sequence, a procedure for extracting the director code sequence is performed. Referring to FIG. 6, the order of finding the director code sequence Δθ (1) from the direction code sequence θ (1) is shown. In the figure, 1 represents a position and S represents a final position. Method for extracting fragrance cords is disclosed in Digical Image Processing, Rafael C. Gonzalez / Paul Wintz, Addison-Wesley Publishing Co., 1987, in particular Section 8.1.1 and Section 8.2.2. However, an embodiment in which the method of extracting a director code string in the present invention will be described with reference to the image data (data stored in RAM) of the English character “E” shown in FIG.

After the contour tracking start point SP is determined by the method of FIG. 3, the direction code string θ (I) is obtained as shown in the attached Table 1 while contour tracking in the clockwise direction along the contour point. In step 601, the position I of the directional code string θ (I) is initialized (I = 1). In step 602, the next position (I = 2) is designated, and in step 603 after step 603, the direction code values [theta] (1) and [theta] (2) of the first and second positions are read from the RAM or the storage of the computer. Then, in step 605, the difference θ (2) -θ (1) of the two values is calculated and the resulting value (0) is given as the directional mark value Δθ (2) at the second position. Then in step 606, the position value is increased by one in step 602 (I = 3), and Δθ (3) = Δθ (3) -θ (2) is calculated in the above-described manner. In other words, a change in direction is obtained at the position of each contour pixel. If the director code is found to be greater than 4, then the value obtained is subtracted from 8, and if it is less than -4, 8 is taken. 5 is equal to -3 and -5 is equal to third. If the position I of the directional code value θ (I) is the last position 3, the next position is the same chain code as the first position, and the directional code string θ (I) must be treated. Therefore, if I = S + 1 in step 602, step 607 is processed after step 603. In the above-described manner, the direction difference code Δθ (I) of each pixel is calculated from the direction code θ (I) as shown in Table 1 attached thereto. In addition, the following formula must be satisfied about director code | column (DELTA) (theta) (I).

As described above, the determination of the director code value is made from one selected contour point and two contour points adjacent to it. Therefore, the direction coder A of the first contour point A (Δθ (I) = A) is such that the direction vector from the first contour point to the I + 1th contour point is the first to the I-1th pixel. The change of direction of A x 45 degrees from the heading direction vector is caused by the first contour point. If the sign of A is positive, the direction change is counterclockwise, and if negative, it is clockwise. In terms of the shape of the pattern, the positive director code value is concave, and the negative director code value means a convex shape.

(3) Directional code of smoothing range

After the above-described determination of the director code string Δθ (I), the director code strings up to a predetermined smoothing range may be obtained in order according to the characteristics of the present invention.

In general, characters or figures written by hand on the paper have unevenness or noise at the straight part or the end of the outline due to smearing of ink depending on the quality of the paper. Such noise can be caused not only by the quality of the paper but also by the error sensing of the scanner. In addition, there are significant differences in the serif depending on the style of the writer and other printed materials. In order to solve such problems, a method capable of eliminating noise and arithmetic and font style differences without damaging the characteristics of the character can be realized by extracting the director code of a predetermined smearing perception according to the present invention.

In the present specification, the term "N-th smoothing range" is defined as the minimum number N of contour points at which no change in direction occurs between two contour points or pixels at which direction change occurs in the director code string. Here, it should be understood that a plurality of contour points having no change in direction form a straight line. Referring to FIG. 12, it can be easily seen that the image data of the English character " 8 " has a lot of noise, especially at the ends. Hereinafter, a method for removing noise of the character "E" will be described as an example. The portions underlined by the dotted lines along the contour points of FIG. 12 correspond to the director code values of the underlined portions of the code string of Δθ (I) in Table 1 attached from the contour tracking starting point SP, respectively. Each of these parts are parts with consecutive nonzero director values. Most of these parts are noises to be removed. In order to deal with such noise, the sum of the nonzero director values is continuously calculated and the sum of the results is placed at the center position of the positions corresponding to the director values while all remaining positions set to zero. For example, successive non-zero director values -1, -2, -1 and 1, -1, 1, 1 at the second and third from the starting point SP are 0, -4, 0 and Treated as 0, 2, 0

Also, since the code string is always treated as a chain code string, the last consecutive non-zero director values 1, -2, and -2 (starting point values) are code values 0, -3, and 0 of the first smoothing range. (Starting point value of the result). The resultant code sequence, i.e., direction directional code sequence Δθ (1, I) of the first smoothing range, is shown in Table 1. Where I represents the position. The shape code string is an array in which values indicating direction changes (non-direction code values) are extracted from the director code string. Since each direction difference code value in the shape code string has a specific shape due to the angle change as described above, the shape code string does not indicate the size of the pattern or the length of the line, but determines the shape of the overall contour of the pattern. do. According to an embodiment of the present invention, since the present invention uses an eight-way code, the direction difference code value -2 is a convex portion of 90 degrees, 2 is a concave portion of 90 degrees, -4 is a convex portion of a round shape, and 4 is a round shape. The concave portion of the image and the like are easily understood. Therefore, the overall contour shape is determined by drawing the contour shape until returning to the starting point while sequentially selecting the direction difference values of the shape code string clockwise from the starting point.

FIG. 12A shows the pattern shown based on the shape code string extracted from the direction difference code string Δθ (1, I) of the first smoothing range shown in the attached Table 1. As shown in FIG. When compared with the pattern of FIG. 12, it is easy to remove a lot of noises and the structure of the font stroke end.

Referring to FIG. 7, a detailed detailed flowchart for extracting the direction difference code sequence Δθ (1, I) of the first smoothing range from the direction difference code sequence Δθ (I) as described above is shown. The meanings of the variables used in FIG. 7 are as follows.

I = location

R = range

Δθ (I) = direction difference value at position I of contour point

A = position with zero direction difference value from the beginning

Δθ (R, I) = direction difference value of position I in the R range

S = last position

SB, SC, SD = Sum of consecutive nonzero direction difference values

B = first position where the direction difference value is non-zero in consecutive non-direction values

C = increment from position B

[] = Gaussian Sign

7 will not be further described as the flowchart can be readily understood by one of ordinary skill in the art in connection with the foregoing description.

Looking at the pattern of FIG. 12A, the corner portions 1210-1215 require further processing. Noise in the straight portion has been eliminated, but the inadequate shape of the edge portion still remains. If these shapes are not removed, a problem arises in that the shape code values of the reference pattern to be stored in the ROM for pattern matching are greatly increased. Therefore, the direction difference code string must be extracted in which the noise is gradually eliminated while increasing the smoothing range until an appropriate shape is obtained. In order to solve such a problem, the direction difference code sequence of the direction difference code sequence (DELTA) (theta) (1, I) of a 1st smizzing range from the predetermined range R is extracted sequentially.

8A to 8E, a flowchart for extracting the direction difference code sequence Δθ (R, I) up to a predetermined range R is shown. Before explaining the flowchart in detail, a method of extracting the direction difference code sequence Δθ (R + 1, I) of the next range from the direction difference code sequence Δθ (R, I) of the R-th range will be briefly described. Consecutive direction difference values in which zero Rs are selected between two non-zero direction difference values among the direction difference code sequences Δθ (R, I) of the R-th range are selected. The sum of the nonzero direction difference values is then calculated and the sum is placed at the central position of the consecutive direction difference values and all other positions are set to zero. Δθ (R + 1, I) is extracted by repetition of such calculation while treating the code Δθ (R, I) as a chaincode.

Referring to FIG. 8, starting from the starting point (I = 1) through steps 811 to 814, a non-zero direction difference value is searched for in the direction difference code sequence Δθ (R, I) of the R range. If found, the position value is stored as J (step 815). Next, it is judged whether the stored position value J has reached the position S-R-1 of the tail portion of the code sequence Δθ (R, I) (step 816). If not reached, a check is made for the direction difference code sequence of the R-th range in steps 817 and 818. If the code string is not a column of the R range, the values Δθ (R, I) from J to J + R are moved to Δθ (R + I, I) corresponding to the position through steps 819 to 824. do. Position I is then incremented by one (process 825) and position I is compared with the last position S (process 826). If I is greater than S, I becomes I-S (= 1) (step 827). If not, a determination is made as to whether Δθ (R, I) is zero (step 828). If Δθ (R, I) = 0, the value of Δθ (R + 1, 1) continues to be set to 0 until a position I satisfying Δθ (R, I) ≠ 0 is found (steps 825-829). If Δθ (R, I) ≠ 0, the procedure shifts from step 828 to step 815.

If the code row is a row of the R range, the procedure moves from step 818 to step 830. Thereafter, the processes 830 to 834 reset all Δθ (R + 1, A) corresponding to the J + R position from J to zero. The sum SUM calculated at step 818 is then placed at the center of the code sequence (step 835). In the next process, it is determined whether position A is the last position S (step 836). If A? S, then position I becomes position J + R + 1 (= A) (step 837) and step 825 is performed. If A = S, process 836 goes to process 850.

If the position J of the non-zero direction difference value is larger than S-R-1, the process 816 proceeds to the process 838. The subsequent process is to process the tail portion of the direction difference cord row and the head portion connected thereto. Processes 838 and 839 are similar to processes 817 and 818 described above. If the direction difference code of the R range is not satisfied, the values Δθ (R, J) of the J position are set to the Δθ (R + 1, J) value of the same position through the processes 840 to 849, and the remaining positions J +. The values of Δθ (R + 1, I) in 1 to S and 1 to J + R + 1-S are all set to zero. The judgment of process 850 is then made. If R is not the designated swimming ring, in step 851 the range R is increased by 1 and step 811 is performed.

If the direction difference code command selected in step 839 is an R range, the procedures of FIG. 8d are performed. Step 852 is the central position of the selected code string. Is in the head or tail. If it is at the tail, the sum SUR is placed at the center position in step 859, and the value of Δθ (R + 1, I) from position J to the last position S except the center position is reset to zero in step 857. The value of Δθ (R + 1, I) at the positions 1-J + R + 1-S of the head portion is reset to zero through steps 860-863. Process 850 is then performed through process 861.

On the other hand, if the central position is at the head in step 852, then step 852 goes to step 864. Through steps 864 to 868, the value of Δθ (R + 1, I) from the position J to the last position S is set to zero. In steps 869 to 874, the head is a flowchart of the processing. In step 874, the sum SUR is placed at the center position, and the value of Δθ (R + 1, I) at positions other than the center position is reset to 0 in step 871. Process 850 is then performed.

Referring to Table 1, the direction difference code strings up to the fifth range and the shape code strings corresponding thereto are shown as a result of using the above-described algorithm for the image pattern of FIG. Also, patterns corresponding to the shape code strings of R = 1 to R = 5 are shown in FIGS. 12A to 12E. As can be seen from the figure, it is easy to see that the noise is gradually removed with the increase of the smsing range. However, referring to Fig. 12E, the ends 1270 to 1272 still remain. However, these portions 1270 to 1272 can be matched with the reference pattern by removing them by the substitution process described later.

The attached Table 2 shows the final direction difference and shape code strings processed up to the smearing range 6 for the image pattern of the English character "E" shown in FIG. Referring to FIGS. 13A to 13C with reference to the attached Table 2, a shape corresponding to smoothing range R = 2 or 3 is 13a, a shape corresponding to R = 4 is 13b, and R = 5 Or a shape corresponding to 6 is shown in FIG. 13C. As can be seen in the figure, it can be seen that the shape code string of the range 5 or 6 can be used as the reference pattern of the English character "E".

Attached Table 3 shows the code strings that process the image pattern of the Chinese character character shown in FIG. 14A to 14E are shapes corresponding to the shape code strings from the range R = 1 to R = 9 of the attached table 3 obtained by contour tracking from the starting point SP1 of the pattern of FIG. 14, respectively. 14J are likewise corresponding shapes respectively to the ranges of the attached table 3 of the starting point SP2.

The attached table 4 shows the code strings of the phonemes corresponding to the contour trace starting points SP1 and SP2 for the character of FIG. 15 representing the character written by the brush as the same character as the Chinese character character of FIG. FIG. 15A is a correspondence shape of the range 7 corresponding to the start point SP1 in the attached table 4, FIG. 15B is a correspondence shape after the substitution process described later, and FIG. 15C is an R = 5, 6 or 7 correspondence shape.

The attached table 5 shows the code strings resulting from the processing of the Hangul image pattern of FIG. 16, FIG. 16A shows a shape corresponding to the range 7, and FIG. 16B shows a shape after replacement.

The attached table 6 shows a code sequence resulting from processing the Japanese image pattern written with the brush of FIG. 17, and FIGS. 17A to 17D show shapes corresponding to the respective ranges of the attached table 6. FIG.

The attached Table 6 shows a code string resulting from processing a Japanese image pattern that is the same as the character of FIG. 17 but differs only in style, and FIGS. 18A to 18D show shapes corresponding to each range of the attached Table 6.

As can be seen from the above drawings, a noise-free shape can be obtained by extracting the direction difference code sequence from the image pattern to a predetermined range, and a predetermined reference shape can be obtained regardless of writing or printing style. have.

[Replace]

Pattern recognition uses a pattern matching method, such as comparing the reference patterns (pre) stored in a ROM or specific storage means with processed data. Therefore, reducing the preliminary capacity not only reduces the cost of the hardware but also the computational time of the system. In order to solve this problem, it is important to reduce the data amount of the reference pattern corresponding to each test image pattern. The present invention provides a solution by using an alternative method.

Referring to FIG. 12E, there are many changes in the ends 1270 to 1272 according to the style of the handwriting and the shape of the printing font. If each of these ends can be replaced to have one shape code value, the data amount of the reference pattern corresponding to this pattern can be reduced. Such replacement may be accomplished by storing the replacement table shown in Table 8 below in a ROM or storage means and using the microprocessor 22 programmed in the flowchart shown in FIG.

Table 8

[Replacement table]

In order to replace, the above-described shape code sequence ΔK (R, I) must be extracted from the direction code sequence of the selected range R. I is the position of the code value. When the direction difference code sequence Δθ (R, I) of each range is sequentially calculated from the direction code sequence Δθ (I), a non-zero direction difference value may move to the head portion. In this case, the subsequent selection of the shape code string is executed until it returns to the starting position while sequentially selecting a non-zero direction difference value following the moved direction difference value.

Referring to FIG. 9, in step 910, an initial value of position I is given, and three values ΔK (R, I), ΔK (R, T + l) and ΔK (R, which are continuous from position I in steps 912 to 814. , I + 2) are stored as A, B, and C, respectively, and a determination is made at steps 915 and 916 as to whether (A, B) and (A, B, C) match the replacement code in the replacement table. . If it doesn't match I Run steps 912-916 until T-2. Where T represents the total number of ΔK (R, I) values. If there is a replacement code, a determination is made in step 917 whether MF x R > Where MF is the magnification given in the table, R is the selected range and L is the number of zeros (hereinafter referred to as distance) between the outer director values of the matching replacement code. If the condition of step 917 is satisfied, it is replaced with the corresponding replacement code (step 918) and position I is increased by one (step 919). Process 911 is then performed. If not satisfied, go back to step 911. For example, the end portion 1270 of FIG. 12E has a direction difference value string of -2, -4, and 2. This column is in the substitution table and the corresponding magnification MF is three. The number of zeros between the outer direction difference of -2 and 2 is 7 + 5 = 12 as can be seen in FIG. The distance L is 12. Since R = 5, MF x R > L is satisfied and the direction difference value is replaced by -4 instead of -2, -4 or 2.

In that way, after replacement, the shape of FIG. 12E is replaced with the shape of FIG. 12F. Therefore, if the direction sequence (-4, 2, 2, -4, 2, 2, -4, -2, -2) of the English character "E" is stored in the ROM as data of the reference pattern, the direction difference of a predetermined range Accurate pattern recognition can be achieved by using a replacement code of the shape code selected from the code.

If in course 911 I If T-2 is not satisfied, the processes of Fig. 9b are performed. The processes in FIG. 9b are processes for treating the tail and the head of ΔK (R, I) because ΔK (R, I) is treated as a circulating code. Since the maximum number of replacement codes is 3, steps 912 to 919 are performed when position I is T-1, and steps 920 to 924, 918 and 919 are performed when position I is T. These processes are basically the same as the processes 912 to 919 described above.

On the other hand, the reason why MF × R > L is given for the replacement is to avoid the change of the field by restricting the replacement only when the important direction difference values, which are the characteristics of the pattern, are adjacent to each other.

It can be seen that the shape code of R = 4 shown in FIG. 13B is converted in the same manner as the shape code of R = 5 or 6 by replacing the end portion 1310. In addition, the shape of R = 9 of FIG. 14e becomes the shape of FIG. 14f by substitution. Similarly, the shape of FIG. 15a is converted into the shape of FIG. 15b and the shape of FIG. 16a is converted into the shape of FIG. 16b.

[Pattern recognition of text]

It is not only inefficient but also time-consuming for characters on one page or one original to be pattern-recognized in their respective smoothing ranges. However, because the characters printed on one page of the printed matter are printed in a certain size, and the characters written by hand on a single manuscript have the same personal writing style, the pattern recognition of one page is substantially constant. Can be run with According to the present invention, the inventors have found that the optimum smoothing range is 5 for newspapers, 5 or 6 for fairy tales or novels, and 4 for dictionaries with small characters. Therefore, it has been found that, in general, larger and thinner characters have a smaller value, and thicker characters have a larger value.

Hereinafter, a method of determining an optimal smoothing range from a single page of text and recognizing a pattern will be described in detail with reference to FIG. 10.

In step 110, the first line on the first line is selected so that an optimal smoothing range can be determined from the characters on the first line of text. In step 111, the microprocessor 22 inputs the image data of the first character on the first line into the RAM in the microprocessor 22.

The microprocessor 22 then finds the presence of the inner and outer contours and the contour trace starting point described in connection with FIG. 3, and extracts the direction coder from the found contour trace starting point. After extraction of the directional code string, the director code strings are extracted and stored in the memory 24 from the extracted directional code string to the designated smoothing range in step 113. It should be noted that the method of extracting the director codes has already been described in detail with reference to FIGS. 5 to 8. The designated value of the smoothing range is preferably a value sufficient to recognize various types of patterns, for example, 7. However, the designated value is not limited to this, and when the computer is used, the designated value can be appropriately adjusted by an input means such as a keyboard according to the type of text. After extraction of the direction code strings, steps 114 and 115 are performed. Detailed flowcharts of fixed 114 and 115 are shown in FIG.

Referring to FIG. 14, the smoothing range is set to 1 in step 141. Thereafter, in step 142, the direction difference code of the first range stored in the memory is accessed, and the shape code sequence is obtained by the method of substitution procedure already described with reference to FIG. Then, in step 143, the shape code string is compared for pattern matching with data of a reference pattern stored in a dictionary, that is, a ROM. If it is in the dictionary, the pattern is recognized in step 144, SN (1) = 1 counts in step 145, and in step 146 the range R is incremented by one. If not, the success number SN (1) maintains the initial state value of 0, and the range R is increased to 2 in step 146. If the range R is less than or equal to the designated smoothing range value in step 147, steps 142 to 146 are executed again for R = 2. Therefore, the success number SN (R) for the first character up to the specified range is determined. If the range R is larger than the designated smearing range value, the procedure goes to the next procedure, process 116 of FIG. A determination is made as to whether the selected character in step 116 is the last character of the first line. If not, then the next character on the first line is selected in step 117 and steps 112 to 116 are executed again for the next character. In that way, after steps 112 to 117 are repeatedly executed for each character on the first line, the success number The maximum value of SN (R) can be obtained, and the optimum smoothing range OSR can be determined by reading the range corresponding to the maximum value (step 118).

After the determination of the optimum smoothing range OSR, pattern recognition from the next line to the last line of text is executed based on the flowcharts of FIGS. 10B and 10C. After the first character of the second line is selected in steps 119 and 120, for this character in step 121 the presence of external and internal contours, extraction of the contour trace starting point and extraction of the direction code string are performed. The direction difference code strings are extracted and stored from the direction code extracted in step 122 to the optimum smoothing range OSR. Then, in step 123, the direction difference code sequence of the optimum smoothing range is selected and the shape code sequence is extracted therefrom. The extracted shape code string is replaced by a comparison example with a replacement table in step 124, and then compared with the aforementioned dictionary in step 125. If it is in the dictionary, a comparison with the aforementioned dictionary is made in step 126. If it is in the dictionary, a comparison with the dictionary described above is made at fixed 126. If so, the character is recognized in step 126, and a determination is made whether it is the last character of the line selected in step 127. If it is not the last character, the next character is selected in step 128 and the processes below fixed 121 are performed. If it is the last character of the selected line, then in step 129 a determination is made whether it is the last line of the page. If it is not the last line, it is selected in step 119 of the next line and a process below step 120 is executed. If the last line, text pattern recognition is terminated.

In step 125, if not in advance, the procedure of FIG. 10C is executed. In this case, since the pattern recognition is lost due to the optimum smearing range, the smearing range needs to be changed. In such a case, pattern recognition has been found to be substantially successful in the smoothing range close to the optimum smoothing range. Referring to FIG. 10C, steps 130 to 136 are steps for pattern recognition while decreasing the smearing range AR by one in the optimum smearing range OSR. If the pattern recognition is not successful when the smoothing range AR is 1, the procedure proceeds from step 135 to step 137, and pattern recognition is performed by increasing the smoothing range BR by one in steps 138 to 142. At this time, in step 139, the direction difference code sequence of the smoothing range BR may be calculated from the direction difference code sequence of the optimum smoothing range already calculated in step 122. However, since the navigation difference code sequence of the smoothing range AR in step 132 is selected from the direction difference code sequences calculated in step 122, no calculation is required.

Hangul consists of 24 phonemes, 10 vowels and 14 consonants. Each phoneme has its own independent form. However, one Hangul character is composed of two or more phonemes, so it has a mixed voice of initial and neutral or mixed of first and neutral and final. In addition, vowels are used only as neutrals, and consonants are used as primary and final stars only. And one Hangul character always has such a positional relationship with Choseong above the Jongsung. Therefore, if the Hangul character to be recognized is composed of three unconnected letters, if the feature value sequence of the reference pattern stored in the ROM is in the order of initial, neutral, and final, then the extracted row values of the contour trace start points It is possible to distinguish between the initial and the final, so that the pattern recognition is possible by processing according to the present invention in the order of initial, neutral, and final.

Chinese characters can also be formed by a combination of unconnected phonemes, but Chinese characters have a structural order formed of sides, rooms, heads, moths, feet, feet, and bodies. Pattern recognition can be achieved by using such a structure. For example, in the case of the Chinese character shown in FIG. 14, if the arrangement of the feature value strings of the reference patterns of the phonemes is made according to the extraction order of the contour tracking starting points, the pattern recognition is possible by processing according to the present invention in the starting point extraction order. do.

As described above, the present invention has the advantage of being able to recognize patterns regardless of the style of the character and also to recognize the pattern regardless of the direction in which the paper is scanned or the position of the paper because the processing is performed with a director code. Can be.

[Table 1]

[Table 2]

[Table 3]

[Table 4]

[Table 5]

[Table 6]

[Table 7]

Claims (10)

CLAIMS 1. A method of processing an image pattern from a memory storing image data of a pattern to be recognized, comprising: extracting a direction code string from which the contour of the pattern is traced from the stored image data; Extracting a direction difference code sequence providing a change in direction of each contour pixel point from the direction code sequence; And extracting a direction difference code sequence of a predetermined smoothing range from the direction difference code sequence in order to remove a change and noise caused by a characteristic of a pattern and a font difference of the pattern. A method of processing an image pattern for determining external and internal contours from an input image pattern, the method comprising: scanning a plurality of lines in one direction with respect to the image pattern; Detecting a first contour point that meets the contour of the image pattern during the scanning; Tracking from the contour point when the first contour point is detected; Detecting a second contour point not contour tracked on the image pattern during the same line scanning; And detecting whether the second contour point is an inner or an outer contour according to the number of contour points detected before the second contour point is detected during the same line scanning. A method of processing an image pattern from a direction code sequence that traces the contour of the pattern from this data of the input pattern and a direction difference code sequence that provides a change in direction of each contour point extracted from the direction code sequence. Storing the code in the memory device; Extracting a direction difference code sequence for a predetermined smoothing range from the direction difference code sequence; And comparing and replacing the selected first code sequence among the extracted direction difference code sequences with the replacement code. A text pattern recognition method for recognizing a pattern by comparing data of image data of an input pattern with data of a reference pattern, the method comprising: selecting image data of a predetermined number of patterns among text patterns; Extracting the processed first group of data from the data, comparing the first data group with the data of the reference pattern, obtaining a pattern recognition success number, and data of the data group having the highest success number And recognizing the remaining patterns that are not selected according to the processing method. A method of recognizing a pattern by comparing the data processed by the image data from a memory device storing image data of a text input pattern with data of a reference pattern stored in a memory, the method comprising: image data of a predetermined number of patterns among the patterns Selecting a process; Extracting a direction code sequence from which the contour of each pattern is traced from each of the selected image data, and a direction difference code sequence providing a change in direction of each contour point extracted from the direction code sequence; The pattern recognition success rate is extracted by extracting the direction difference code sequence from the extracted direction difference code sequence to a predetermined range of each pattern and comparing the data from each extracted direction difference code sequence with the data of the reference pattern. The process of selecting a high range; And extracting the direction difference code string of the selected range from the remaining patterns and comparing the data with the reference pattern data. An apparatus for processing an image pattern from a memo that stores image data of a pattern to be recognized, the apparatus comprising: means for extracting a direction code string from which the contour of the pattern is traced from the stored image data, and the direction code string Means for extracting a direction code sequence that provides a change in the direction of each contour pixel point from and a predetermined smoothing from the direction code sequence to remove noise and variations caused by the characteristics and font differences of the aquatic pattern of the pattern And means for extracting the direction difference code sequence of the range. An apparatus for processing an image pattern for determining external and internal contours from an input image pattern, the apparatus comprising: means for scanning a plurality of lines in one direction with respect to the image pattern, and a method that meets the contour of the image pattern during scanning; Means for detecting a first contour point, means for contour tracking from the contour point when the first contour point is detected, means for detecting a second contour point that is not contour traced on the image pattern during the same line scanning; And means for detecting whether the second contour point is an internal or external contour according to the number of contour points to be detected before the second contour point is detected during the same line scanning. An apparatus for processing an image pattern from a direction code sequence that traces the outline of a pattern from image data of an input pattern and a direction difference code sequence that provides a change in direction of each contour point extracted from the direction code sequence. A means for storing a code in a memory device, a means for extracting a direction difference code sequence for a predetermined smoothing range from the direction difference code sequence, and a first code sequence selected from the extracted direction difference code sequence with the replacement code And an image pattern processing apparatus comprising a means for replacing. A text pattern recognition apparatus for recognizing a pattern by comparing data of image data of an input pattern with data of a reference pattern, comprising: means for selecting a predetermined number of image data of a text pattern, and each of the selected images Means for extracting data of the first group processed from the data, means for comparing the first data group with data of the reference pattern, and obtaining pattern recognition success numbers, and data of the data group with the highest success number. And a means for recognizing the remaining patterns that are not selected in accordance with the processing. A device for recognizing a pattern by comparing the data processed by the image data from a memory device storing image data of a text input pattern with data of a reference pattern stored in a memory, wherein the image data of the pattern of the small and medium number pattern is determined. Means for selecting, means for extracting a direction code sequence that traces the contour of each pattern from the selected respective image data and a direction difference code sequence that provides a change in direction of each contour point extracted from the direction code sequence; The pattern recognition success rate is extracted by extracting the direction difference code sequence from the extracted direction difference code sequence up to a predetermined range of each pattern and comparing the data from each extracted direction difference code sequence with the data of the reference pattern. Means for selecting a high range and the direction difference code of the selected range. Extracting the turn and text pattern recognition apparatus, characterized by a means adapted to compare the data of the reference pattern.