JPS5814710B2 - pattern classification device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a pattern classification device, and more particularly to a pattern classification device that classifies character patterns using information surrounding the characters.

It is essential to develop a kanji reading device in Japan, where the daily language is kanbun and kana mixed sentences.

Conventionally, various devices have been known for reading characters, but
Recognition was limited to relatively simple patterns and few types of characters, such as numbers and katakana.

The difficulty in reading kanji is that there are many types of characters, many characters are complex, and the amount of information per character is several times that of alphanumeric characters.

Therefore, if conventional character reading methods were used as they were, the storage capacity of dictionary patterns would increase and the recognition speed would decrease, making it difficult to realize a practical recognition device.

In order to solve these problems, characters to be recognized are classified in advance into a plurality of character groups based on the characteristics of their patterns, and it is determined which character group the input character pattern belongs to.

Thereafter, devices have been developed that determine and recognize the degree of similarity between each candidate character within the character group and the input character pattern.

One method for classifying characters into multiple character groups is to quantize the amount of character lines that exist on the four sides (top, bottom, left, and right) into three levels of 0, 1, and 2, and then quantize the amount of character lines that exist on the four sides of the character into four digits, each digit consisting of three values. Code (all 8
Type 1).

Hereinafter, this will be referred to as a four-sided code. For example, in the case of the character shown in FIG. 1, the character code is coded clockwise from the left side according to the amount of character lines in frames 1 to 4, and becomes "2010".

However, the code of the input character pattern changes due to noise (for example, positional shift), so even in such cases, in order to classify correctly, only the number of expected code changes for each character is used as a standard code. The drawback was that it had to be prepared in a dictionary.

For this reason, the storage capacity of dictionaries has become enormous.

An object of the present invention is to provide a pattern classification device that can perform accurate classification with a simple device.

FIG. 2 is a block diagram showing the overall configuration of a character reading device as an embodiment of the present invention.

Reference numeral 10 denotes a document on which characters to be read are printed, and each character is converted into an electrical signal via a photoelectric conversion unit 11.

This electrical signal is supplied to the preprocessing section 12 and subjected to a quantity equalization process.

This quantized signal is supplied to the major classification section 13, which detects the four-sided code of the human-written character and determines the character group to which it belongs.

The recognition units 1 and 4 determine the degree of similarity between each candidate character in the character group to which the input character belongs and the input character, and output the code of the candidate character that is most similar to the input character as a recognition result.

FIG. 3 is a configuration diagram of the major classification section 13.

The quantized signal obtained by the preprocessing section 12 is stored in the pattern memory 20.

Hereinafter, the quantized signal in the memory 20 will be referred to as an input character.

This input character is supplied to the manual character encoding circuit 21,
The four-sided code is found.

The input character encoding circuit 21 generates a four-digit code depending on the amount of character lines in each area on the four sides of the character, as described in connection with FIG.

As mentioned above, since each digit consists of three values, the whole is represented by an 8-bit code.

The input character code obtained by this circuit 21 serves as one input to a mismatch detection circuit 22.

The other input of the mismatch detection circuit 22 is a dictionary code,
This is stored in the dictionary code memory 23.

FIG. 4 shows the format of the information stored in the dictionary code memory 23.

Looking at the contents of address A in the memory 23, for convenience, the C field is written as "upper", which represents the character code corresponding to the character "upper".

Further, each of the 8-bit fields F1, F2, F3, and F4 represents a four-sided code obtained when the character "upper" changes due to positional displacement or the like.

Further, 1-bit fields S1, S2, S3, and S4 are provided corresponding to each field Fl, F2, F3, and F4.

Hereinafter, this will be referred to as a control bit.

This control bit indicates that the four-sided code of the corresponding field is in a mismatch tolerance mode that allows mismatches within a Hamming distance of 1.

In FIG. 3, the contents of memory 23 are set in registers 25 and 26 by the operation of controller 24.

The contents of field C are stored in the register 26, and the other fields are stored in the register 25.

Under the control of the controller 24, the F1 field is first supplied to the mismatch detection circuit 22.

The mismatch detection circuit 22 compares the four-sided code of the human-written character with the contents of the field F1, and outputs a match signal on line 27 if they do not match, and outputs a mismatch signal on line 28 if they do not match.

At this time, if control bit S1 is set, gate 2
If 9 is selected and not set, gate 30 is selected.

If neither gate 29 or 30 currently generates an output signal, controller 24 supplies the contents of field F2 from register 25 to mismatch detection circuit 22.

At this time, gate 29 or 30 is selected depending on the content of control bit 32.

If one of gates 29 and 30 now generates an output signal, the contents of register 26 are stored in character code buffer 33 via gates 31 and 32.

At the same time, the output of the gate 31 causes the controller 24 to transfer the contents of the next address in the memory 23 to the register 25 . It is accommodated in 26.

In this way, if any of the contents of the F1 to F4 fields of each address matches the input character code, the matched character code is sequentially stored in the character code buffer 33.

When all dictionary codes and input character codes are compared, the character code buffer 33 stores codes corresponding to decimal characters to which human characters may belong.

The recognition unit shown at 14 in FIG. 1 determines which character the input character belongs to by matching the contents of the pattern memory 20 with each character indicated by the character code of the character code buffer 33. be done.

Next, the effects of this invention will be explained using a specific example.

FIG. 5a shows a normal pattern for the character "upper".

The four-sided code in this case is represented by ``0011''.

When this character shifts upward as shown in FIG. 5B, the four-side code becomes ``0010'', and when it shifts downward as shown in FIG. 5C, it becomes ``0012''.

However, when the character "upper" is shifted left or right, the four-side code does not change.

Therefore, for this character, the change in code due to positional shift is limited to the fourth digit.

Therefore, the F1 field at address A shown in FIG. 4 contains "0011" as the representative code, and the control bit S
Set it to 1.

Further, in the F2 field, a modified code specific to noise related to line width is checked and its representative code is similarly stored.

In this way, the dictionary code memory 28 stores the representative codes of all the characters to be read by checking in advance the unique deformation codes of the characters for various types of noise.

Assume that the input character encoding circuit 21 in FIG. 3 encodes the input character "upper" and outputs "0012".

This value is compared with the dictionary code output from the dictionary code memory 23.

At this time, a control bit attached to the dictionary code indicates whether the mode is a complete match mode or a mismatch tolerance mode.

Although the detailed configuration of the mismatch circuit 22 in FIG. 3 is omitted, when the two input signals completely match, the line 2
The comparison circuit provides an output signal to line 7 and provides an output signal to line 28 if the Hamming distance of the two input signals differs within 1.

In other words, the value of each digit of the input character code is Pi (
i=1 to 4), and the value of each digit of the dictionary code to Ai (i=1 to 4)
4), then K=lA-Pl-4. −
I A2-P21 +l A3-P31 + IA4-
When P41≦1, an output signal is generated on line 28.

Assume that the input character encoding circuit 21 encodes the input character "upper" and outputs "0012".

This value is sequentially compared with the dictionary code in the dictionary code memory 23.

If the control bit of this dictionary code is in the reset state,
The character code of the character having the modified code “0012” is stored in the character code buffer 33.

(exact match mode). In addition, if the control bit of the dictionary code is in the set 1 state, even if it is not "0 0 1 2", if the Hamming distance is within 1, the corresponding character code is sent to the character code buffer 33 in the mismatch permission mode.
to be accommodated.

For example, assume that the F1 field of field A- is supplied to the mismatch circuit 22.

As mentioned above, the contents of this field are: Although it is "0011" and does not match the four-sided code of the input character, since K=1, an output signal is obtained on line 28.

On the other hand, since the control bit S1 is set, it is the mismatch permission mode, so the character code of "top" is buffer 2.
It is accommodated in 3.

As described above, according to the present invention, the capacity of the dictionary code storage section can be reduced, and input characters can be roughly classified while taking into account noise such as misalignment.

Furthermore, since the number of searches for dictionary codes can be reduced, it is possible to classify patterns at high speed.

[Brief explanation of drawings]

FIG. 1 is a diagram showing an example of encoding input characters, FIG. 2 is an overall layout diagram of an embodiment of this invention, and FIG. 3 is a configuration diagram of the main part of an embodiment of this invention. FIG. 4 is a diagram showing an example of the format of a dictionary code used in one embodiment of the present invention, and FIGS. 5 a, b, and c are diagrams for explaining the effects of the present invention. 21... Input character encoding circuit, 22...
. . . Mismatch detection circuit, 23 . . . Dictionary code memory.

Claims (1)

[Claims] 1. A dictionary code storage unit that stores a large number of dictionary codes in which a large number of character patterns are encoded in advance according to their characteristics, and a control bit that is provided for each dictionary code and specifies the matching mode of the corresponding dictionary code; means for encoding based on the characteristics thereof; and first matching means for detecting a complete match between the input code corresponding to the input pattern obtained by this means and the dictionary code sequentially read from the dictionary code storage unit; a second matching means for detecting that the Hamming distance between the input code and the dictionary code is within 1; and a second matching means for detecting that the Hamming distance between the input code and the dictionary code is within 1; A pattern classification device comprising: means for selecting the output of the first matching means; and means for selecting the output of the first matching means when the control bit indicates a mismatch tolerance mode.