JPS5949630B2 - pattern recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (1) Field of Application of the Invention The present invention relates to a pattern recognition device suitable for cases where there are many similar patterns such as Chinese characters.

(2) The technical problems in conventional printing kanji recognition are as follows:
These problems include a reduction in recognition speed due to the large number of character types (2,000 to 4,000), and a reduction in recognition accuracy due to the presence of many similar character sets such as "question" and "ma".

Conventionally, the former problem, that is, increasing recognition speed, has been solved using the hierarchical pattern matching method (IEICE Transactions No. 56).
- Vol. D, p. 365).

However, regarding the latter problem, the recognition accuracy of similar characters,
Pair judgment method (Transactions of Institute of Electronics and Communication Engineers Vol. 56-D, p. 5
45) have been proposed, but no satisfactory solution has been reached. For example, when recognizing voiced and hand-voiced characters, ``pa'' and ``ba'', the difference is often less than 10% of the total sunspot area, and conventional unified methods cannot sufficiently focus on the difference. It is often misread or even unrecognizable. (3) Purpose of the Invention Therefore, the purpose of the present invention is to correctly separate (recognize) character pairs (sets) that are difficult to separate using conventional pattern matching methods.
The goal is to provide the means to do so.

(4) General explanation of the invention Generally, observed character patterns contain so-called noise, such as differences in thickness due to printing pressure and ribbon condition, and quantization errors during observation. and the standard pattern will never be an exact match.

Therefore, during recognition, the degree of similarity expressing the closeness between the observed unknown character pattern and the standard pattern is calculated, and the character in the standard pattern with the highest degree of similarity is classified into the category (character type) of the unknown pattern. It is determined that However, in the case of similar characters, as in Figure 1 a and b or d and e, the area of the different part is different from that of the other part (in the case of a, "gate",
In the case of d, it is smaller than the area of ``ha'').

Therefore, even if we assume an ideal case where there is no noise, the difference in similarity is not large between similar characters.
For example, when the unknown pattern "ba" is input, the degree of similarity with respect to the standard pattern "pa" is 0.900, and with respect to the standard pattern "ba" it is 0.897. Moreover, since noise actually exists, confusion may occur between similar characters. However, the standard patterns of similar characters must be created so that the parts that should sound the same (in the case of Figure 1, the parts of ``mon'' and ``ha'') are exactly the same. For example, it is possible to approach the ideal state described above.

However, the disadvantage of this strategy is that it is extremely difficult to automatically or semi-automatically create a standard pattern that satisfies the above conditions because there are many combinations of similar characters as in the example above. Or, at the stage of printing, in the case of the above example, the patterns for ``mon'' and ``wa'' may be slightly different, so making them exactly the same pattern may not necessarily produce a good result. Furthermore, even if a standard pattern satisfying the above conditions can be created, a sufficiently large difference in similarity cannot be obtained. Therefore, in the present invention, when a character with a similar character set is selected as the first candidate character, a mask pattern (C, C in FIG. 1,
f or i) is selected, and the degree of similarity between the unknown pattern and the standard pattern is calculated again only in the area masked by the same mask pattern, and a final decision is made.

Here, when multiple candidates are listed in the first stage of judgment, only in areas where the similarity between those characters is low,
The method of performing pattern matching again and making a final judgment is known (
Publicly known example: Japanese Patent Application No. 41-42363 [Special Publication No. 46-97
No. 27]), but the features of the present invention are as follows.

That is, in the above-mentioned known example, if there are multiple character categories whose similarity as a result of the first stage pattern matching exceeds a predetermined threshold, pattern matching is performed again only in areas where the similarity is low. , in the present invention, the first or second character listed in the hierarchical pattern matching determination process is a character type that has similar characters (for example, fat, ball, question, ha, ba, etc.). , and only when the difference in their similarity is within a certain threshold, re-pattern matching is performed using a mask pattern indicating a predetermined portion (region) with low similarity.

Then, the character types that have no similar characters are ranked first or second.
If the difference in similarity is less than or equal to a certain threshold among the top candidates, it is determined that determination is not possible, and pattern matching is not performed again. The above features of the present invention are important for the following two reasons.

(1) If pattern matching is always performed again when multiple candidates are listed, is the reason why multiple candidates were listed due to poor printing quality or because the characters are extremely similar? There is no such distinction.

Therefore, if characters that are originally not similar are selected as top candidates due to poor print quality or noise, the characters that should have been impossible to determine can be determined by re-pattern matching. There is a high risk of misrecognition. (2) The number of categories is 2000 as in the case of kanji recognition.
4000, the number of combinations of candidate characters that are ranked first and second is enormous, and it is necessary to specify in advance areas with low similarity between characters in any combination of characters, so the similarity It is practically impossible to perform pattern matching again whenever the values are close to each other. Therefore,
In order to realize the above characteristics, the device of the present invention includes a means for storing whether or not a pre-specified similar category set exists, and a method for selecting which mask pattern among a plurality of mask patterns to re-pattern match. It has a means of remembering what to use.

These can be stored in a storage device in the form of a table, and it is only necessary to have a memory capacity equal to the number of categories. The contents of the table are
If a very similar character set exists for each character type, enter the number of the mask pattern to be used; if none exists, enter the value zero, and memorize it in one table. . To summarize the advantages of this method, there is no need to make any special modifications to each standard pattern, a relatively small number of mask patterns can be created, and some mask patterns can contain multiple characters. (For example, mask pattern C in FIG.
It can be used not only for "threshold, fight, seki, close, open, listen, kaku, etc.", but also for "tree, book, circle, nine"), and a mask. - Patterns are written on paper and photoelectrically converted in the same way as character patterns and stored in the mask pattern storage device as electrical signals! Create a mask pattern,
It is easy to make additional corrections, and regarding recognition speed, it is not necessary to always perform this processing, and it is only necessary to perform this processing on unknown patterns that have similar character sets, so there is little reduction in recognition speed. Since it is only necessary to calculate the degree of similarity between a pattern obtained by multiplying and 1 mask pattern, and a pattern obtained by multiplying the standard pattern and the same mask pattern, the required hardware is required in addition to the conventional recognition device. The wear is small.

1 (5) Example Hereinafter, the present invention will be explained in detail with reference to Examples.

FIG. 2 is a block diagram of an apparatus according to one embodiment of the present invention.

In the figure, a signal 10 is a form on which printed characters are printed. The form is input to the character observation section 1. The character observation unit 1 optically scans the characters on the form, performs photoelectric conversion, and
Cut out each character and create a character pattern f(I,j)(I,j
= 1, 2, ..., 40) is output to the classification device 2 as an electrical signal 11. The character pattern is 40 points x 4
It is a binary pattern of 0 points, with black points versus white points on the paper.

The classification device 2 receives the binary character pattern f(I, j), and performs blurring processing to generate a blurred pattern g(K,l)(K
, l = 1, 2, 3, ..., 8), sequentially read the classification standard pattern gΔ(K, l) 31 from the classification standard pattern storage device 21, and calculate the similarity ρ1 (G, G
ω) is calculated by the similarity calculating circuit 5 according to the following equation (3).

however, Further, ω is a category number, and M is the number of character types.

Note that the blurring process referred to here means that the binary pattern f(I, j) on a 40x40 mesh is increased by increasing the number of 1 bits in a 7x7 small area to create a "blur, 8" It refers to obtaining a multivalued pattern g(K,ll) of ×8 meshes. Expressed mathematically, it becomes. 2 is further calculated by the degree of similarity ρ1 (
Select the top N characters from among ω = 1, 2, ..., M), and select the number N of candidates and the same top N character numbers 12 (ω1, ω2
, ω3, ..., ωN) are transferred to the recognition unit 3.

3, when N=0, there is no candidate, so a message indicating unrecognizability is outputted from the terminal 14.

When N=1, the same character number is output from the terminal 14 as the recognition result. When N≧2, recognition is performed using an unknown binary pattern f(I, j) among N candidate characters. 22 is a standard pattern storage device for recognition, and the recognition unit 3 stores N standard patterns 32, f0 (I, j) (ω
:ω1 ωι2, ゜゜゜゜゜゜, ω÷N),
The similarity calculation circuit 6 calculates the similarity ρl according to the following equation (6).

However, here again, ωμ (μ=1, 2, . . . , N) is a candidate character number.

The recognition unit 3 further searches for the maximum similarity ρ and the next-largest similarity ρM from among the calculated similarities ρ, and makes the following determination based on the absolute threshold δ1 and the relative threshold ε1.

(1) If ρ = δ1 and ρ blue - ρiε1, there is sufficient similarity and the difference between the 1st and 2nd place is wide enough, so the character number ωμ that gives ρ blue is recognized as the terminal 14 Output from

(2) ρ.

In the case of .delta.1, the similarity is not large enough to satisfy the requirements, so a result indicating that recognition is not possible is outputted from the terminal 14. (3) ρ
If 1≧δ1 and ρBlue1ρi<ε1, N' character numbers, ω(ν=1,2,ゝ ν・・・, Output N●13
As such, it is transferred to the final determination section 4.

However, the character numbers to be transferred are arranged in descending order of similarity.

In step 4, it is checked by the mask pattern number table whether or not there is a similar character to the character ω1 that gives ρ-lock or the character ω2 that gives ρ-blue.

The mask pattern number table is stored in the storage device 23, and for each character number, if there is no similar character, the value is 0,
If similar characters exist, a mask pattern number indicating a different part is registered. Therefore, 4 is ω1
Check the mask pattern number J (.1) of , and if the value is O, then check the mask pattern number J (ω2) of ω2, and if this value is also O, there is no similar character, so it cannot be determined. The result is output from the terminal 14 as a recognition result.

J(Ci)1), when O, read the J(ω1)th mask pattern hμ(1,j)34 (μ=J(ω1)) from the mask pattern storage device 24, and perform the following processing. A final judgment will be made.

(5) Mask the unknown pattern.

f(1,j)=f(1,j)・hμ(1,j)(8)(
6) Mask the N' standard patterns.

1 go (1, j) = f: v (1, j) · hμ (1, j)
ν=1, 2,..., N' (9) -X (7) f and F using the similarity calculation circuit 6.

) is calculated. However, the calculation formula is (6)
Same as expression. (8) Search for the maximum similarity ρ↓ and the next largest similarity among the N' calculated similarities.

(9) The following determination is made using the absolute threshold value δ2 and relative value ε2 for ρ and ρ1.

UO) When ρ1≧δ2 and ρichi-ηzuku≧ε2, the character number giving ρ1 is output from the terminal 14 as the recognition result.

O1) If the above conditions are not met, an unrecognized result is output from the terminal 14.

The above is the overall flow of the operation of the apparatus of the present invention, and the contents of the processing of the recognition unit 3 and final determination unit 4, which are the center of the present invention, are shown in flowcharts in FIGS. 3a and 3b.

3 and 4 are actually constituted by a microprocessor and perform the above-mentioned processing according to a program.

The same portion 40 will be explained in more detail using FIG. 4. In FIG. 4, 41 is R for storing the microprogram.
OM (ReadOnjyMemO-Rv) and a sequence control unit, and 42 is an ALU (arithmetic logic unit). 42 receives data through a data bus 44 and performs recognition and final judgment.

The output 12 of the classifier 2 puts the classification results on the Zetabus through an interface 121.

A standard pattern storage device 22 used during recognition is connected via an interface 221. The mask pattern number table 23 and the master pattern 24 are stored at different addresses in the storage device 200. An unknown pattern 25 is also stored in 200.

In the final judgment, the pattern multiplier 43 multiplies the product of the standard pattern obtained from 22 and one mask pattern from 24 by 1? Furthermore, the product f of the unknown pattern 25 and the same mask pattern is obtained by 43, and 1
? The similarity calculation circuit 6 calculates the similarity between and f. The determination result 14 is output through the interface 141.

(6) Summary As explained above, according to the present invention, similar characters, which conventional character recognition devices could not accurately understand, can be recognized with high accuracy by emphasizing each difference using a mask pattern. can be recognized.

Additionally, the present invention is efficient in doing so, as mask patterns can be created, added, and modified relatively easily, and less hardware is required than in the past. In the above embodiment, a measure of similarity is used to increase the similarity between patterns, but a measure of dissimilarity may be used conversely.

At this time, the overall configuration does not change much and the configuration is self-evident, so the explanation will be omitted. It is also clear that when calculating the degree of similarity ρ or ρ in the recognition unit and the final determination unit, the accuracy can be further improved by mutually correcting the positions.

(Regarding position correction, see Information Processing Society of Japan, Vol. 16, N.
O. l2, plO64) Also, instead of the process 4 of the above embodiment, the pattern f(I,
j)・FU(I,j), a mask pattern hμ(I
, j) to obtain the difference amount f(I
, j)·f: (I, j)·hμ(I, j)AO).

In addition, as a simplification of the above embodiment, there is a method of omitting the classification device 2 (in this case, the character numbers given to the recognition unit 3 are all character numbers), or a method in which the classification device 2 uses macros such as the number of sunspots and moments. It is also possible to use methods that use other features.

Furthermore, the recognition unit 3 can be omitted and the character number having the maximum similarity in the classification device 2 can be used instead of the output of the recognition unit 3. Furthermore, in the above embodiment, the recognition unit 3 and the final determination unit 4
In the above, we assumed that binary pattern matching was performed, but it is clear that multivalue pattern matching may also be used, and the mask pattern is not a binary pattern of 1 and 0, but a pattern whose values change gently at the boundary surface. It is also clear that a multi-valued pattern like this can be used. Furthermore, the present invention is applicable not only to kanji but also to general patterns, and in that case, the part ``character'' in the above explanation may be replaced with ``pattern/power lever 1 month''.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing examples of similar character patterns and examples of mask patterns for them, for explaining the principle of the present invention.

Claims (1)

[Scope of Claims] 1. A pattern recognition device for determining a standard pattern corresponding to an inputted unknown pattern, comprising means for outputting a category name of a standard pattern for the inputted unknown pattern; means for determining whether or not a similar category set exists; means for storing a pattern difference area for each similar category set; A pattern recognition device comprising means for extracting a corresponding pattern difference region and making a final decision in the region. 2. The pattern recognition device according to claim 1, wherein the storage means stores a mask pattern for masking a similar area for each similar category group.