JPS59154579A - Pattern recognizing device - Google Patents

Abstract

PURPOSE:To facilitate the generation of dictionary pattern and similarity calculation and to recognize a pattern securely even when there is similar category for an input pattern by preparing respective dictionary patterns by the method of regular canonization and adaptive canonization. CONSTITUTION:A pattern input circuit 1 converts an input pattern to be recognized into an electric signal, which is processed preliminarily by a preprocessing circuit 2, whose output is segmented, syllable by syllable, or, character by character, and outputted by a detecting and segmenting circuit 3. The output of the circuit 3 is canonized by a regular canonizing circuit 4 and supplied to a recognizing circuit 5. This circuit 5 is provided with a buffer 10 and a similarity calculating circuit 11. The input pattern from the regular canonizing circuit 4 is stored in the buffer 10 temporarily, and the circuit 11 calculates the composite similarity to the input pattern in the buffer 10 and either of dictionary patterns M1 and M2 from the 1st-stage and the 2nd-stage dictionary memories 6 and 7. The recognition result K of this circuit 5 is applied to a deciding circuit 9 to recognize the pattern securely.

Description

DETAILED DESCRIPTION OF THE INVENTION [Technical Field of the Invention] The present invention relates to a pattern recognition device used in character reading devices and the like.

[Technical background of the invention and its problems]

There is a superimposition method as one of the recognition methods of pattern recognition means. This method calculates the degree of similarity between a representative pattern for each category (hereinafter referred to as a jigon pattern) stored in a memory in advance and an incoming cover pattern to be recognized.

The category whose similarity, which is the result of this calculation, is at the maximum and above a certain threshold value is recognized as corresponding to the input pattern.

In such a method, a simple similarity method, a composite similarity method, a mixed similarity method, etc. have been conventionally used based on a similarity calculation power formula, a dictionary pattern creation method, and the like. In the simple similarity method, the dictionary pattern creation method and similarity calculation are simple, but the recognition process becomes unstable due to pattern deformations that tend to occur in handwritten characters and speech. Further, the composite similarity method has a function of absorbing pattern deformation so as to solve the above-mentioned drawbacks of the simple similarity method. However, on the other hand, each feature of a similar category (for example, in the case of characters, "5" and rsJ, "0" and rOJ, etc.) will be absorbed as part of the pattern, and the same category and pIg texture will be absorbed. There is an inconvenience in doing so.

There is also a mixed similarity method which is a complete improvement of this composite similarity method. This method absorbs normal pattern deformations, but emphasizes and identifies deformations (features) that increase the discriminability between similar categories without absorbing them. However, the mixed similarity method has drawbacks such as the complexity of dictionary pattern creation and similarity calculation.

[Purpose of the invention]

This invention was made in view of the above circumstances, and its purpose is to:
To provide a pattern recognition device which can easily create a dictionary pattern and calculate the degree of similarity even when a category similar to an input pattern exists, and can also perform pattern recognition reliably.

[Summary of the invention]

This invention provides a pattern recognition device that performs a composite similarity meter A second dictionary memory is provided. 2nd IJ memo! j u, stores a dictionary pattern created using an adaptive canonical method for a set of similar categories that are difficult to identify in the dictionary pattern of the first deweighting memory; Then, the recognition means calculates the degree of similarity between the input cover pattern and the dictionary pattern of the first dictionary memory,
If it is difficult to identify similar categories, use the input pattern and the second
A composite similarity calculation is performed with the dictionary pattern in the dictionary memory. Thereby, even if a similar category exists for the input pattern, pattern recognition can be performed reliably and easily.

[Embodiments of the invention]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing the configuration of a pattern recognition device according to the present invention. In the figure, reference numeral 1 denotes a pattern input circuit which converts input patterns to be recognized such as characters and voices into electrical signals and outputs the electrical signals. The preprocessing circuit 2 performs preprocessing such as filtering and sampling on the input pattern P output from the pattern input circuit 1.

The detection cutout circuit 3 cuts out a pattern corresponding to one syllable or one character from the input pattern P output from the preprocessing circuit 2 and outputs it. The extracted pattern is subjected to normalization such as alignment and level correction in a normalization circuit 4, and then provided to a recognition circuit 5.

As shown in FIG. 2, the identification circuit 5 includes a buffer 10 and a similarity calculation circuit 1, and temporarily stores the input pattern P from the normalization circuit 4 in a buffer 1o. Further, in the similarity R1n circuit 1, the input cover pattern P in the buffer 10 is combined with either one of the binding patterns M1 and M2 outputted from the first, second stage, or second unloading memory 6.7. Perform similarity calculation. In this case, the Jirei pattern Ml,
M2 is selected by a selection gate circuit 12 controlled by a split signal G output from a control circuit 8, for example. Recognition circuit 51-1. , outputs the similarity calculation result K to the determination circuit 9. The determination circuit 9 determines whether the calculation result (ie recognition result) K output from the recognition circuit 5 is valid or not, and if it is valid, outputs it as an answer. The control sub-circuit 8 is a judgment circuit 9
Based on the determination result, the control signal G is outputted to control the selection gate circuit (12 in FIG. 2).

By the way, a method of creating each dictionary pattern M1 and M2 in the first and second stage dictionary memories 6.degree. 7 will be briefly explained. First, the first stage dictionary memory 6 stores dictionary patterns of all the categories to be Mm, which are created by the foot canonicalization method in the normal composite similarity method. Regarding the definite canonicalization method in this case,
This will be briefly explained with reference to FIG. In Figure 3, r
UJ and rVl are pattern vectors of different categories, rEJ is a constant vector whose elements are all equal in size and is 1, and T is a hyperplane orthogonal to rEJ.

[RJ, [(>J are definite canonicalizations of rUJ + rVJ, respectively, and are represented by the following equations (11 and (2)).

rJ −1”UJ −(1trJ, rwJ)・r
EJ・・・・・・(1) Evil J −rVJ −(1”VJ
, lfJ )・rEJ (2) and,
The similarity between “8” and “♀” is ([target “J”) − (rUJ, IVJ) − (1t
JJ, l”EJ) (Seki, Dai) - (ftJJ,
rVJ )-1fITJl-1rVJ 1eosβeo
sγ・・・・・・(3) Now, if β and r are in the range of 90° or less, according to the above formula (3), <rO", ~J) < (rUJ, "VJ)
...between songs (4). That is, from the similarity between the similar original pattern rUJ and "v", the canonical pattern rtu+
, r9'' becomes smaller, and it can be expected that the discrimination ability becomes higher. In this case, the discrimination ability is "U".

It depends on the positional relationship between rVJ and rEJ, and is no longer valid especially for similar categories.

Next, the first-stage dictionary memory 6 generates a shoulder circumference pattern created by an adaptive canonicalization method for a set of similar categories that are likely to be misrecognized and rejected in the classification based on the dictionary pattern in the first-stage dictionary memory 6. It is stored. This adaptive canonicalization method will be briefly explained with reference to FIG.

In Fig. 4, rUJ and rVJ are average pattern vectors of similar categories, and rWj is (rUJ + "V
J) 12, ``WeJ is a vector with the same direction as rWJ and a magnitude of 1.S is a hyperplane orthogonal to rWJ.rUJ, rVJ are ``U''.

It is an adaptive canonicalization of "■", and the following is the formula (5) of F,
Displayed at f61.

rVJ - "VJ - (rvJ, "WeJ) -r
wa'',, -・------(to 6+, and the similarity between rUJ and rVJ is < ny,, +, θ'') - (Mentenma Kano (rUJ, Kuma J) (Figure 2 Prefectural phrase )-(River: J, I'VJ)-1rU
J lth - VJ l cosβcosβ・・・・・・
...(7) becomes. According to the above equation (7), if rUJ and rVJ are similar categories, β is 45° or less and the above equation (7) is satisfied.
The value of is definitely smaller than the value of equation (3) above, and the discriminative power of similar categories is large.

The operation of this configuration will be explained. The input cover pattern P, which is the object to be recognized, is converted into electrical 4tJ by a pattern input circuit J. Then, as described above, the input cover turn P is provided to the recognition circuit 5 via the preprocessing circuit 2, the detection extraction circuit 3, and the normalization circuit 4. In the recognition circuit 5, the input cover pattern P is temporarily stored in a buffer 10, and then provided to a similarity calculation circuit 11.

The similarity calculation circuit 11 calculates the dictionary pattern Mlf in the first stage dictionary memory 6 by the control signal G from the control circuit 8.
Then, the similarity between the dictionary pattern M and the input cover pattern P is calculated. The determination circuit 9 determines whether the similarity measurement result (recognition result) K of the recognition circuit 5 is valid. If this judgment result is valid, the recognition result K,j
! 11 Output the category of dictionary pattern M as the answer.

Further, when the determination circuit 9 determines that the recognition result K based on the dictionary handling pattern M1 of the first stage dictionary memory 6 is negative, it instructs the control circuit 8 of the determination result. That is, this is a case where there is a possibility that the input pattern P is erroneously recognized due to a conflict between similar categories, or a case where it is determined to be an abridged pattern. In this case, the dictionary pattern M2 outputted from the second stage dictionary memory 7 is given to the similarity calculation circuit 1 of the recognition path 5 by the control signal G of the control circuit 8.

As a result, the recognition circuit 5 outputs the input cover pattern P from the repeating buffer 10 to the similarity calculation circuit 1, and the similarity between the input cover pattern P and the yakuzae pattern M2 is 7'+! - Do. By calculating this degree of similarity dt, a similar category to the input cover turn P has been identified. and,
The determination circuit 9 selects the portrait pattern M2 of the first stage dictionary memory 6.
The recognition result K is output as the answer.

To specifically explain such an operation using an example of recognition processing of handwritten alphanumeric characters, the first stage dictionary memory 6 is preliminarily stored with (0 to 9. dictionary pattern 7M2 is stored. Furthermore, the second stage dictionary memory 7 contains similar categories that are often misread in the normal composite similarity method (IO, O, DJ, Is, 5J, "8.BJ").
.

1 t
is given, this is mostly recognized by the dictionary pattern M1, and the result is output as the answer. Next, when rBJ is given as the input cover turn, this becomes “B”
This results in a rough high similarity result for both "8" and "8". Therefore, 5 recognition circuits, dictionary pattern M
Similarity calculation by 2 (M2c) Full re-entry cover turn rB
J, and it is determined whether it is "8" or rBJ. This identification result will be output as an answer.

〔Effect of the invention〕

As detailed above, according to the present invention, in a pattern recognition device that performs composite similarity calculation, all dictionary patterns are prepared using fixed canonicalization and adaptive canonicalization techniques. Even if similar categories exist, it is possible to reliably identify similar categories by creating a simple opening pattern and calculating the degree of similarity. Therefore, reliable pattern recognition can be performed with a simple configuration and recognition processing.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of a pattern recognition device according to an embodiment of the present invention, and FIG. 2 does not show the specific configuration of the recognition circuit in FIG. 1. Soi
z It is a diagram for explaining a method of creating each dictionary pattern in the first and second stage dictionary memories in FIG. 1. 1... Pattern input circuit, 5... Recognition circuit, 6...
・1st stage dictionary memory, 7...2nd stage dictionary (q memory, 1
1...Similarity calculation circuit. Applicant's agent Patent attorney Suzu 'Lushi Takehiko Figure 1 456 Figure 3

Claims (1)

[Claims] An input means for inputting an input pattern to be recognized and converting it into an electric signal, a first dictionary memory for storing a dictionary pattern created in advance using a constant pressure method, and a dictionary for this first dictionary memory. a second dictionary memory that stores a dictionary pattern created by an adaptive canonical method for a set of similar categories that are difficult to identify by the pattern; and an input cover pattern output from the input means and a dictionary pattern of the first dictionary memory. pattern recognition means for performing a composite similarity calculation between the input cover pattern and the dictionary pattern of the second dictionary memory when it is difficult to identify a similar category. recognition device.