JPH0516632B2 - - Google Patents

Description

[Detailed description of the invention]

[Industrial Field of Application] The present invention relates to a character recognition system, and more specifically, to a character recognition system that classifies and normalizes handwritten characters including consonants and consonants into a plurality of types. [Prior Art] Recently, with the wave of OA, various Japanese information processing systems have been developed, but the major bottleneck in Japanese information processing systems is data input. Until now, various Japanese input methods have been proposed, such as the tablet method, key touch method, and kana-to-kanji conversion method using a keyboard, and have achieved some degree of success. It is not suitable for inputting large amounts of data such as Therefore,
Character recognition using OCR, especially handwritten character recognition, is highly expected to play a role in Japanese data input methods. In particular, in the future, OCR devices will be combined with OA equipment,
Applications that use OCR to read handwritten characters, display them, edit them, process them, and print them are likely to become popular, but it is essential for such applications to have the ability to clearly distinguish between capital letters and accents, and to recognize and process them. It is thought that. However, conventional OCR for alphanumeric characters and kana usually does not distinguish between uppercase and lowercase letters (press sounds) of kana, and processes uppercase and lowercase letters of the same shape as the same character. Furthermore, since the handwritten kanji OCR to date has mainly been used for applications such as reading addresses and names, little consideration has been given to the recognition of accentuated sounds. One possible method for distinguishing between uppercase and lowercase letters is to determine the size and position of the input character pattern and use this information as feature data for classification and identification. Since the size and position of the characters must be used as additional feature data, there are problems in that the recognition process becomes complicated, difficult to implement in hardware, and the recognition time tends to be long. In general, handwritten character recognition using OCR is
It is performed through the steps of character reading, preprocessing, feature extraction, and classification identification.The preprocessing normalizes the size of the input character pattern, but in conventional OCR, which does not consider uppercase or lowercase letters, all input character patterns are It is common to normalize all at once. When uniform normalization is used, lowercase letters or symbols such as periods and commas that originally have small dimensions are unnecessarily enlarged, which not only involves unnecessary processing but also eliminates unnecessary characteristics such as minute irregularities in lines. , and unnecessarily increases the amount of character pattern data to be recognized.
There is a problem that the amount of data processing for classification identification increases. JP-A No. 55-10624 uses uppercase letters, numbers,
A character identification processing method for identifying special symbols is shown. In this processing method, first, in the first step, short special symbols such as periods and commas are separated and identified as they are, and the remaining symbols are normalized in height and then roughly classified and identified. Next, the width of the remaining items is normalized for further classification and identification. This patent application uses the normalization of character height and width separately, but this is normalization to a constant size in stages based on the previous identification result, and unlike the present invention, the size of the input character pattern,
It does not show normalization to different sizes and positions based on position, nor does it show the identification of capital letters and accents. [Problems to be Solved by the Invention] Therefore, an object of the present invention is to easily and quickly distinguish between uppercase and lowercase letters (especially accents) and to recognize them without complicating the recognition process. That's true. [Means for Solving the Problems] The present invention classifies and normalizes input characters into a plurality of types based on the size and position of the input characters at the normalization stage of preprocessing. . That is,
Based on the size and position of the read binary input character pattern, the input characters are classified into multiple types such as regular characters and accent sounds, and normalized by making at least one of the size and position different for each type. The image area is normalized, and features are extracted from the entire normalized image area. Therefore, according to the present invention, the size and position parameters of the input character pattern are inherent in the normalized character pattern itself, so in the subsequent feature extraction, classification, and identification processes,
There is no need to consider parameters such as the size and position of input characters. Therefore, a unified recognition algorithm can be used for all character categories, simplifying feature extraction and classification/identification processing, and increasing recognition speed. Furthermore, since the same recognition process can be used for all character categories, it is easy to implement the character recognition system in hardware, and the recognition speed can be further increased. [Example] Next, a preferred example of the present invention will be described with reference to the drawings. FIG. 1 is a functional block diagram of a character recognition system according to the present invention. First, in the character reading section, characters handwritten on a form are read by an OCR scanner. In this example, the form used had a 10mm x 10mm character frame. The read binary character pattern data is temporarily stored in a memory as a dot pattern, and a character frame portion is cut out by a character cutting section. Increase OCR scanner resolution to 8 dots/mm
Then, the size of the cut out character frame will be 80 x 80 dots. The outer circumference detection section detects the minimum and maximum values (Xmin, Xmax,
Ymin, Ymax) are detected. The coordinates of the upper left corner of the cut out character frame are set to (0, 0). Since the above-mentioned character reading, character cutting, and outer circumference detection are well-known general processes, a detailed explanation thereof will be omitted. FIG. 1 shows the normalization classification section and normalization section improved by the present invention. The normalization classification section calculates the width in the X and Y directions and the center of gravity in the X and Y directions as follows based on the Xmax, Xmin, Ymax, and Ymin of the circumscribed rectangle obtained by the perimeter detection section. X-direction width ΔX = Xmax-Xmin+1 Y-direction width ΔY = Ymax-Ymin+1 X-direction center of gravity Gx = 1/2 (Xmax + Xmin) Y-direction center of gravity Gy = 1/2 (Ymax + Ymin) Classifying input characters into multiple types based on pattern size and position. Table 1 below shows an example of classification when the resolution of the reading scanner is 8 dots/mm and the character frame is 80 x 80 dots.

[Table] In Table 1, "left part" in the column for center of gravity Gx represents the left part when the character frame is divided into three equal parts in the X direction, and "upper", "middle", and "lower" in the column for center of gravity Gy represents the upper, middle, and lower parts when the character frame is divided into three equal parts in the Y direction. Also, a blank column means "don't care." If the four conditions ΔX, ΔY, Gx, and Gy (excluding blank columns) are satisfied, the corresponding type number is selected. Here, we will explain the meanings of the above 10 types. In this example, the basic recognition targets are kanji, kana, kanji, uppercase letters (excluding lowercase letters), numbers, and some special characters included in the JIS C6235 Japanese input dial layout. It will be appreciated that various types can be created based on location and size using combinations of letters or symbols. Type 1 is for special characters such as single quotation marks "'" and "'" written in small letters at the top of the text frame, and Type 2 is for special characters such as single quotation marks "'" and "'" written in small letters at the center of the frame.
Type 3 is for special characters such as a period "." written in small size at the bottom of the frame.
Type 4 is for special characters such as double quotation marks "", "'', which are written slightly larger at the top of the frame, and type 5 is for special characters such as double quotation marks "" and "'', which are written in the center of the frame.
Type 6 is for accents, punctuation marks, etc. Type 7 is the kanji "", exclamation mark "!",
Type 8 is for vertically long characters such as the colon ":" and the number "1", and type 8 is for the kanji ichi "ichi",
It is for negative sign "-" etc., type 9
is an underline ” is for. Type 10 is for other kanji, uppercase English letters, uppercase kana letters, numbers, and larger special characters such as "%" and "\".
Type 5 has smaller dimensions than Types 4 and 6, but this is because the characters written in the center of the character frame are likely to be regular characters even if they are a little smaller, and these regular characters are smaller than Type 5. This is to prevent it from being classified as a repeating symbol "〃". For example, type 7 and type like the number "1"
A case may occur where both of the above 10 apply, but in this case, the priority encoder gives priority to the one with the smaller type number. Furthermore, for special characters and accentuated sounds, the size and position within the character frame are important factors. Therefore, when filling in forms, it is desirable to give the filler a guide on how to fill in the forms, but in the case of this type classification, it is recommended to It is sufficient to provide guides such as "Special characters should be written in appropriate sizes" and "Special characters should be written separately at the top, middle, and bottom of the character frame depending on their type." The normalization unit in FIG. 1 converts input characters into
At least one of size and position is made different for each type and normalized to a predetermined normalized image area. In this example, the binarized character pattern within the extracted 80 x 80 bit character frame is normalized to a 64 x 64 bit normalized image area. FIG. 3 illustrates how each type of character is normalized into a 64 x 64 bit normalized image area. The hatched area indicates the area where the normalized image is written, and the white area around it indicates the white image area. Types 1 to 6 are for lowercase sized characters;
6 uses a normalized size of 30, types 7 to 10
In this case, a normalized size of 60 is used. However, for types 1, 2, 3, 7, 8, and 9 of elongated character patterns, these character patterns are 30x30 or 60x60.
When normalized to , the shape characteristic of elongation is lost, and especially when thinning the line after normalization, minute irregularities on the line edge in the length direction are emphasized by enlarging, distorting the original characteristics. Therefore, types 1, 2,
In 3 and 7, the value of ΔX is used as is, and in type 8,
9, the value of ΔY is used as is for normalization. After being normalized as described above, the normalized character pattern is subjected to feature extraction in the feature extraction section from the entire 64×64 bit normalized image area including the surrounding white background. Next, in the classification identification section, pattern matching is performed between the extracted features and features prepared in advance for standard character patterns that have been similarly normalized, and the input character pattern is identified. As a feature extraction method, differences in the size and position of normalized character patterns within a certain normalized image area are reflected as feature differences when features are extracted from the entire normalized image area. If it is something,
That is, any feature that is size and position dependent can be used. Conventionally known methods include methods for determining the number or distribution of black dots by position or by section when viewed from the X direction, Y direction, or diagonal direction, and methods for determining the number or distribution of black dots by position or by section when viewed from the , how to find the number of connected black dots in vertical and diagonal directions, how to find the number of white dots from each side (top, bottom, left and right) to the character pattern (depth), or the total number of white dots (area) up to a certain depth. This is known, but one example is Japanese Patent Application Laid-Open No. 58-201184, which calculates the number of connected black dots for each direction and section and the area of a white region along each side. These feature extraction and classification/identification processes themselves are common in pattern matching-type recognition methods, so detailed explanations will be omitted. If necessary, please refer to the above-mentioned JP-A-Sho. Next, a good normalization mechanism according to the present invention will be explained with reference to FIG. Type selector 1
0 corresponds to the normalization classification section in FIG. 1, and the other sections correspond to the normalization section in FIG. The type selector 10 is the coordinate value of the input character pattern circumscribing rectangle within the 10 mm x 10 mm (80 x 80 dots) character frame,
Based on Ymin, Ymax, Xmin, and Xmax, the type of input character (type 1 to type 10 in Table 1) is determined, and a control signal necessary for normalization is generated according to each type. These control signals will become clear later. The normalization mechanism includes a normalization ROM 18, 28, 128 x 128 bit image buffer 22, and a 64 x 64 bit normalization image area which is the normalization image area.
It has a buffer 52. image battle 22
contains the input character pattern of the cut out character frame.
The input character pattern is stored with the upper left corner of the character frame aligned with the coordinates (0, 0) when the coordinates of the upper left corner of the image buffer 22 are (0, 0). The image buffer 22 is set to 128 x 128 bits to facilitate addressing, but it may be of any size as long as it includes the character frame. The purpose of this normalization mechanism is to normalize the input character pattern to a predetermined size and position shown in FIG. 3 in accordance with the determined type of input character and store it in the normalized image buffer 52. The normalization ROMs 18 and 28 each have two normalization matrices A and B, where matrix A is for normalized size 30 and matrix B is for normalized size 60. Matrices A and B are selected by the type selector 10 according to the determined type. The functions of the normalization ROMs 18 and 28 are such that when a character pattern in the image buffer 22 is normalized to a normalized size of 30 or 60 by reduction/enlargement and stored in the buffer 52, depending on the size of the character pattern,
Normalize buffer 52 which dots in the character pattern
It is to generate an address that indicates what to write to. In the case of reduction, the dots are thinned out and read, and in the case of enlargement, the same selected dots are read repeatedly,
Alternatively, in some cases, the image may be outputted from the image buffer 22 as it is without being enlarged/reduced. For this purpose, Ymin and Ymax (7 bits each) are given to the subtracter 12, and its output (ΔY−
1) is applied to the upper address (H) that selects the row position (Y position) of the ROM 18 via the A input of the multiplexer MPX14. The lower address (L) for selecting the column position (X position) of the ROM 18 is given from the upper bits ²⁶ to ²¹¹ of the address counter 30. The B input "59" of the MPX 14 is selected by the type selector when the normalized size is 60, and is used when the character pattern of the image buffer 22 is output as is with the Y direction enlargement ratio=1. Since the output of the normalization ROM 18 indicates the Y address in the character pattern circumscribing rectangle, the output of the ROM 18 is added to Ymin in the adder 20 and converted into a Y address for actually addressing the image buffer 22. Similarly, Xmin and Xmax (7 bits each) are given to the subtracter 24, whose output (ΔX-1) is passed through the A input of the multiplexer MPX26 as
It is given to the upper address (H) that selects the row position (Y position) of the ROM 28. ROM28 column position (X
The lower address (L) for selecting the location) is given from lower bits ²⁰ to ²⁵ of address counter 30.
The B input “59” of MPX26 converts the character pattern of image buffer 22 to X when the normalized size is 60.
This is selected by the type selector 10 when outputting as is with the direction enlargement ratio=1. MPX2
The C input "29" of No. 6 is selected by the type selector 10 when outputting the character pattern of the image buffer 22 with the X direction enlargement ratio=1 in the case of the normalized size of 30. The output of the ROM 28 also indicates the X address in the circumscribed rectangle, so the adder 32
It is added to Xmin and converted into an X address for actually addressing the image buffer 22. Here, the configuration of the normalized ROM matrix will be explained with reference to FIG. In this example, to simplify the explanation, the character frame dimensions are 10 x 10 bits (Figure 5A), and the normalized size is 4 x 4 bits (Figure 5A).
Normalized ROM matrix 18 for Figure B)
28 (FIG. 5C). Element Ekl of the kth row and lth column of the normalization matrix
The value of is determined by Ekl=INT(k×l/L+0.5). Here, L is the maximum value that the address index can take (3 in FIG. 5C), and INT(x) is the integer part of x. Row positions 0 to 9 are size indices corresponding to ΔX-1 or ΔY-1, and correspond to the upper addresses of the normalization matrix. Column positions 0 to 3 correspond to lower addresses of the normalization matrix. Generally, the number of rows in a matrix for normalization in the X direction is the number of bits in the X direction of the character frame (10 in the case of Figure 5), and the number of columns is the number of bits in the X direction of the normalized size (in the case of Figure 5). 4), the number of rows of the matrix for normalization in the Y direction is the number of bits in the Y direction of the character frame (10 in the case of Figure 5), and the number of columns is the number of bits in the Y direction of the normalized size. (4 in case of Figure 5)
Therefore, in the practical example shown in FIG. 4, the number of rows of the normalization matrices 18A, 28A is 80, the number of columns is 30, and the number of rows of the normalization matrices 18B, 28B is
80, and the number of columns is set to 60. In the case of Figure 5, the character frame and normalized size X,
Since the number of bits in the Y direction is the same, the normalization matrix shown in FIG. 5C can be used in common in the X and Y directions. In operation, for example, if the dimension in the X direction of the input character pattern Xmax-Xmin=ΔX-1 (size index) is 2, the value of the X normalization matrix is 0,
1, 1, 2 are read, so X address =
1 is generated twice and expanded. When the size index = 3, the dimension in the X direction of the circumscribed rectangle = the normalized size in the X direction, which corresponds to an enlargement ratio of 1. When the size index=9, X addresses=0, 3, 6, 9 are generated and thinning reading is performed. The Y-direction normalization matrix operates similarly. However, since the normalization matrix generates an address that converts the character pattern to the normalized size by setting the address of the upper left corner of the circumscribed rectangle as (0, 0), the actual X, Y address for accessing the image buffer 22 is Ax and Ay are given by the following equations. Ax = Xmin + matrix (ΔX-1, ix) Ay = Ymin + matrix (ΔY-1, iy) Here, matrix (ΔX-1, ix) and matrix (ΔY-1, iy) are ΔX-1 and ΔY-, respectively.
1 is the size index, ix, iy are the address
This is a matrix value obtained as an index. Adders 20 and 32 in FIG. 4 perform this addition. Therefore, the normalization buffer 5 of 64×64 bits
In the case of Fig. 4 using 2, normalization ROM18, 28
generates an address signal 4096 times (=64×64) and reads the image buffer 22. However, the third
In order to normalize the data and write it into the normalization buffer 52 as shown in the figure, some ingenuity is required. Comparator CMP4
2, 44 and adders 46, 48 are for that purpose. First, the principle will be explained with reference to FIG. Let the X and Y addresses Anx and Any of the normalization buffer 52 be given by the following equations. Anx=ixαx Any=iyαy Here, indicates addition of binary numbers (2 bits in FIG. 6, 6 bits in FIG. 4), ignoring overflow. Therefore, for example αx=0,
The data written as shown in Figure 6A when αy=0 is
When αx = 1 and αy = 2, data can be written by wrapping around starting points (1, 2) as shown in Figure 6B, and data to be written when ix > 1 and iy > 1 can be written. If masked, it will be written as shown in FIG. 6C. That is, the normalized buffer 5 with αx and αy
Specify the writing start point of 2, ix > limx, iy > limy
By masking the write data, the read output of the image buffer 22 can be masked and written to any position in the normalization buffer 52 with any size. The limit values limx and limy at this time are taken as limit count values. Adders 46, 48 and comparators 42, 44 control writing to normalization buffer 52 based on the above principle. Adders 46 and 48 receive addresses corresponding to the above iy and ix as one input from the address counter 30, and receive addresses corresponding to αy and αx as the other inputs from the multiplexer.
Receive from MPX38,40. The type selector 10 controls the MPXs 38 and 40 and selects the corresponding starting point address αx(2,
17, 23 or 27), αy (2, 17, 25, 32 or 47)
to adders 48 and 46. Adder 46 provides an upper address (H) for selecting the Y address of normalization buffer 52, and adder 48 provides a lower address (L) for selecting the X address. Therefore, the dot data read from the image buffer 22 is written from the starting point (αx, αy). Comparators 42, 44 receive as one input the addresses corresponding to iy, ix from address counter 30, and the limit count value from multiplexer MPX 34, 36 as the other input.
Type selector 10 gates the limit count value to MPX 34, 36 depending on the determined type. Comparators CMP42 and CMP44 operate AND gate 50 when iy and ix are less than the limit count value.
energized and exceeds the limit count value
AND gate 50 is prohibited. Note that the outputs of the type selector 10 are “2”, “17”, “23”, “25”,
“27”, “29”, “32”, “47”, “59”, “ΔX-1
”,
“ΔY−1” is MPX14, 26, 34, 36, 38, when these outputs are generated according to the type determination.
40 to select input of the corresponding number. Table 2 below is used for normalization of each type 1 to 10.
Indicates the ROM matrix (for normalized size 30...A, for normalized size 60...B), size index (X, Y), limit count value (X, Y), and starting point (X, Y). There is.

[Table] As shown in Figure 3, types 1, 2, 3, 7
Now, the character pattern of image buffer 22 is X.
The enlargement ratio in the direction is set to 1 and the data is output as is.For types 8 and 9, the enlargement ratio in the Y direction is set to 1 and the data is output as is. Therefore, types 1, 2, 3, 7
The X limit count value of type 8 and type 9 is set to ΔX-1, and the limit count value of types 8 and 9 is set to ΔY-1. Also, when the expansion ratio is 1, the size index is the maximum value of the address index of the normalized matrix (29 for normalized size 30,
60 is equal to 59), so the X size of types 1, 2, and 3 using normalized size 30.
The index is 29, the X size index of type 7 with normalized size 60 is 59, and the Y size index of types 8 and 9 with normalized size 60 is 59. Therefore, for types 1, 2, and 3, the type selector 10 selects the size index 29 in the multiplexer 26 and accesses the normalization matrix 28A with an X-direction enlargement ratio of 1; The index 59 is selected and the normalization matrix 28B is accessed with the X-direction enlargement ratio of 1, and in the case of types 8 and 9, the size index 59 is selected in the multiplexer 14 and the normalization matrix 28B is accessed with the Y-direction enlargement ratio of 1.
Access 8B. According to the above normalization mechanism, while the address counter counts 4096 (=2 ¹² ), the image
One input character pattern in the buffer 22 is normalized and stored in the normalization buffer 52 with a size and position depending on its type. Also, in the case of a character frame of 80 x 80 bits, the normalization ROM
A 7-bit byte is sufficient for each of the X and Y addresses 18 and 28, but adding 1 bit to make an 8-bit byte and using this 1 bit to control the output AND gate 50 will result in unintentional expansion. can be prevented. That is, for example ΔX=26, ΔY
When a small, horizontally elongated character pattern such as =5, Gy center of gravity = middle part is input, this character pattern is determined to be type 5 by the type selector 10 according to Table 1. Therefore, this character pattern is normalized to 30 x 30 bits, resulting in the elongated pattern being transformed into a square and losing the geometric features of the original elongated pattern. At this time, the magnification ratio in the Y direction is 6 (=30/5),
When creating a normalization matrix, make sure that no more than a predetermined number of the same addresses are lined up in the same row of the normalization matrix, and add the above addition 1 to the addresses in the remaining positions.
By specifying inhibition of the output AND gate using a bit, careless expansion can be prevented. FIG. 7 shows a simple example of a normalization matrix with a maximum magnification of 2. A in Figure 7 shows the case where the magnification rate is not suppressed; a line with a width of 1 dot (size index = 0) is expanded six times, and a line with a width of 2 dots (size index = 0) is expanded six times.
In the case of 1), the image is expanded three times. B in Figure 7 shows the case where the magnification rate is suppressed to a maximum of 2, and an address with the output AND gate prohibition bit H set is placed at both ends, and when this address is read from the normalization matrix, the output AND gate is prohibited. It is something to do. Therefore, size index=0,
1, the expansion can be suppressed to 2 times. In FIG. 4, lines 18H, 2 to output AND gate 50
8H is this output gate inhibition bit. By arranging addresses having output gate inhibit bits at both ends of the normalization matrix, the character pattern can be normalized by aligning it with the center of the normalization area (shaded area) in FIG. According to the present invention, which specifies the character type at the normalization stage, the number of misrecognitions due to type errors is reduced to 1.
However, according to the results of an experiment in which a rough guide was given to the person filling out the form in advance and the person filled it out with normal attentiveness, it was found that it was sufficiently practical. If you want to further increase the recognition rate, for example
In the format program for reading OCR forms, the character type is specified for each field, and selection of types 1 to 6 is prohibited in fields that do not allow lowercase letters, and uppercase letters written smaller in the character frame are normalized to lowercase size and erroneously. To prevent character patterns from being recognized or to easily cause type classification errors, prepare multiple templates that include the characteristics when normalized for each type that may be classified. I can do it too. After normalization, the character patterns in the normalization buffer 52 are read out, features are extracted, and classification and identification processing is performed. [Effects of the Invention] The present invention divides the input character pattern into a plurality of types based on the size and position of the input character pattern, and normalizes the input character pattern into a normalized image area with a different size and position for each type. Therefore, the character pattern itself in the normalized image area includes size and position information, and therefore, the size and position of the input character pattern are not taken into account at all in the subsequent feature extraction and classification/identification stages. Since there is no need to do so, recognition processing becomes simple and fast, and uppercase letters, accents, and lowercase letters can be clearly distinguished.

[Brief explanation of the drawing]

Fig. 1 is a functional block diagram of a character recognition method using the normalization technology according to the present invention, Fig. 2 is a diagram showing a circumscribed rectangle of a character pattern in a character frame, and Fig. 3 is a diagram showing a circumscribing rectangle of a character pattern in a character frame.
4 is a diagram showing a normalization device according to the present invention; FIG. 5 is a diagram showing an image buffer, a normalized image buffer, and a normalized ROM matrix; FIG. Figure 6 shows the storage pattern of the normalized image buffer.
FIG. 7 is a diagram showing a storage pattern of a normalized ROM matrix.

Claims (1)

[Claims] 1. Binarization of a recognition target character pattern including special characters, small characters including consonants and consonant characters, and large characters occupying a large area in a character frame, and binarization of cut out characters. In a character recognition method that recognizes characters by normalizing character patterns, extracting features from the normalized character patterns, and classifying and identifying characters based on the extracted features, the above-mentioned binary character patterns are divided into sizes and character frames. means for classifying into a plurality of types based on the position of at least a predetermined special character type, a predetermined lower case type including consonant and consonant characters, and a predetermined upper case type; and means for normalizing the normalized character pattern into a normalized image area of a constant size by varying at least one of the size and position of the character pattern normalized area for each type, the entire normalized image area. A character recognition method characterized by extracting features from the target.