JPH0632085B2 - How to create a pattern recognition dictionary - Google Patents

Description

The present invention relates to a method for creating a pattern recognition dictionary.

[Background of the Invention]

In pattern recognition, one answer category is determined or one or more candidate categories are selected by comparing an input pattern with a standard pattern representing a plurality of categories.

Specifically, for example, a candidate category is selected and an answer category is determined as follows. See FIG.

Input pattern u and a plurality of categories C ₁ ... C
_A standard pattern f that is a representative of _k (here k1)
(C ₁ ), ..., f (C _k ) and the distance d (u, f (C
₁ )), ..., d (u, f (C _k )).

A category satisfying d (u, f (C _i )) θ [f (C _i )] is collected with respect to a threshold value θ [f (C _i )] predetermined for each standard pattern f (C _i ). .

If there is only one category, it will be output as the answer category. If the collected categories are two or more, they are used as the candidate categories for the answer. In the procedure from to, the pattern information for each category C _i is the standard pattern f (C _i )
And a threshold value θ [f (C _i )]. A collection of these sets for all categories is called a dictionary.

By the way, conventionally, in order to sequentially create a dictionary as the number of learning samples increases, the processes of the following (1) to (2) are performed for each category C.

(1) First step (initial creation) For the learning samples u ₁ , ..., U _m (here m1),
The standard pattern f _m and the threshold value θ [f _m ] are calculated according to the following two equations.

(2) a second step (create additional) learning samples u _{m + 1,} ..., when u n _(n> m) is added, the standard pattern f _n and the threshold θ additional before and after the addition [f _n] all the learning samples u _{1, ...,} with reference to the u _n to create according to the following formula.

This dictionary creation method has the following advantages and disadvantages.

(a) Advantages As shown in FIG. 2, the maximum distance between the standard pattern f _n and all the learning samples u ₁ , ..., U _n is set as the threshold value. Therefore, when recognizing the learning sample according to the procedure of selecting the answer category and determining the answer category,
The correct category can be selected as one of the candidate categories for all learning samples.

(b) Disadvantages In order to calculate the second step threshold θ [f _n ], not only the additional learning samples u _{m + 1} , ..., U _n, but also the learning samples u ₁ , ..., U _m before the addition are stored. It is necessary to do so, and the memory capacity becomes enormous. For example, in character recognition, 30 to 300 bytes in pack format are required to store one learning sample, so the number of learning samples increases and n = 2 × 10 ⁵ (= 20
(00 categories x 100 samples / category) requires a memory capacity of 6 to 60 MBytes, which is not practical.

In addition, since the number of times of distance calculation n in the equation (2-3) is 2 × 10 ⁵ , the distance calculation time is 0.1 to
Even if it is 1 sec / time (using the microprocessor 8086), it takes 2 × 10 ⁴ to 2 × 10 ⁵ sec in total, which is not practical.

In particular, when the dictionary is updated sequentially on-site, the above-mentioned drawbacks in memory capacity and threshold calculation time are unacceptable.

The conventional dictionary creation method is discussed in the following two documents.

IEICE Technical Report, PRL80-16, Seiichiro Naito, Comprehensive Analysis of Handwritten Kanji Data Technical Report of IEICE, PRL80-25, Shigeru Akamatsu, Classification Experiment of Handwritten Kanji Characters by Structure Accumulation [Objective of the Invention] It is an object of the invention to provide a method for creating a pattern recognition dictionary which solves the above-mentioned drawbacks of the conventional invention.

[Outline of Invention]

In the present invention, in order to achieve the above object, in the additional creation of the dictionary, only the dictionary before the addition and the additional learning sample are referred to, and it is equivalent to the case where all the learning samples before and after the addition are referred to. It allows you to create a dictionary of the performance of.

The principle of the present invention will be described below. For each category C
The dictionary is sequentially created by performing steps (1) and (2).

In the present invention, it is assumed that the dictionary is composed of a canonical pattern, a threshold value, and the number of learning samples.

(1) First step (initial creation) For the learning samples u ₁ , ..., U _m (here m1),
The standard pattern f _m and the threshold value θ [f _m ] are calculated according to the following two equations.

The initially created dictionary has a standard pattern f _m and a threshold θ.
_{[F m],} a training sample number m.

(2) Second step (additional creation) When the learning samples u _{m + 1} , ..., U _n (n> m) are added, the standard pattern f _n and the threshold value θ [f _n ] are added to the dictionary f before addition. _{m, θ [f m]} and additional learning sample u _{m +} 1,
, U _n are created according to the following equation.

The additionally created dictionary has a standard pattern f _n and a threshold θ.
[F _n ] is the number of learning samples n.

In the present invention, in the additional creation of the second step, it is possible to save the pre-added learned samples u ₁ , ..., U _{m in} the memory and to omit the reference in the threshold calculation, so that the drawbacks of the conventional invention are solved. You can

Further, even when the dictionary created by the present invention is used, just as in the case of the conventional invention, by setting a sufficiently small threshold value θ [f _n ], the correct answer category is set as one of the candidate categories for all learning samples. Can be selected as The proof is shown below. The point is to use the triangle inequality for distance (see Fig. 3).

[Proof that correct answer category can be raised to candidate category for all learning samples] About pre-addition learning samples u _i , i = 1, ..., M In equation (5-1), the triangle inequality is applied to the distance.

For post-addition learning sample u _i , i = m + 1, ..., N For all learning samples u _i , i = 1, ..., N (5-
2), from equation (5-3), d (f _n , u _i ) In other words, equation (5-4) is used for all learning samples u before and after addition.
_{It is shown that i} (where i = 1, ..., N) is within the distance θ [f _n ] from the standard pattern f _n . (End of proof) Further, as standard patterns to be stored in the dictionary, cumulative standard patterns u _m and u _n , where u _m = m · f _m u _n = n · f _n are used instead of the average standard patterns f _m and f _n. Also in the case, a dictionary can be similarly created.

Example of Invention

Next, a first embodiment of the present invention will be described with reference to the above principle using the configuration example of FIG. The configuration example includes a pattern observation unit 11, a feature extraction unit 12, a learning sample memory unit 13, a dictionary calculation unit 14, a dictionary memory unit 15, and a control unit 16. The pattern observing unit 11 is a unit that converts an input pattern 10a such as a character pattern or a voice pattern into an electric signal 11a, and includes a scanner, a microphone, and the like. The feature extraction unit 12 performs preprocessing, normalization, on the input electric signal 11a,
A feature pattern 12a is output by performing feature extraction. The learning sample memory unit 13 stores the characteristic pattern output by 12 as a learning sample. Learning samples u ₁ , ..., U _m are stored in the case of initial generation, and samples u _{m + 1} , ..., U _n are stored in the case of additional generation. The dictionary calculation unit 14 is a unit that calculates a dictionary. In the initial creation, calculations are performed using equations (3-1) and (3-2). In this case, the learning sample u _{1 of} 13a
, U _m are input, and the dictionaries f _m , θ [f _m ], m of 14a are output. For additional creation, calculations are performed using equations (4-1) and (4-2). In this case, the additional learning sample u _{m + 1} of 13a to 14
..., dictionary before additional learning corresponding group to u _n and 15a f _m,
θ _{[f m],} m is input. The dictionaries f _n , θ [f _n ], n corresponding to 14a are output. The dictionary memory unit 15
This is a part for storing the dictionary 14a created in 14. Control unit 16
Controls signals such as 10a, 11a, .... In FIG. 4, control signal lines are omitted to avoid complication of the drawing.

Examples of the feature pattern 12a used in character recognition include marginal distribution feature, stroke density function feature, peripheral feature, directional pattern feature, and others.
An example of is shown in FIG. 5 (a), and an example of is shown in FIG. 5 (b).
An example of is shown in (c) and (d) of FIG. 5, and an example of is shown in FIG.
Shown in (b), (c), (d), and (e).

Next, a second embodiment of the present invention will be described using the configuration example of FIG. The configuration of FIG. 7 is obtained by adding a matching unit 17, a determination unit 18, and a correction unit 19 to the configuration of FIG. Since these are added, the dictionary can be created even when the category of the input pattern is not known in advance. (In the first embodiment, it was necessary to know the category of the input pattern in advance.) Means that the pattern recognition process and the correction process by 17, 18, and 19 parts are correct categories of each input pattern. Because you can get

This will be described in detail below. When an example of the input pattern 10a such as a character pattern or a voice pattern is input to the pattern observation unit 11, an electric signal 11a is output. The feature extraction unit 12 performs preprocessing, normalization, feature extraction, and the like on the input electric signal 11a, and outputs a sequence u _{m + 1} , ..., U _n of the feature pattern 12a. The matching section 17 evaluates the distance between each feature pattern u _i (where i = _{m + 1} , ..., N) with all standard patterns belonging to the dictionary 15a before the input pattern sequence is input. . Then, the determination unit 18 based on the distance data
Calculates answer categories and candidate categories and displays them on the display. In the correction unit 19, the user of this device corrects the answer category while referring to the input pattern, the answer category, the candidate category, and the like. By the above processing, the learning sample u _{m + 1} obtained by adding the answer categories modified, ..., u _n are obtained. These are stored in the learning sample memory unit 13. Thereafter, the dictionary calculation unit 14 uses the additional learning samples u _{m + 1} , ..., U _{n of the} input 13a and the dictionary 15a before the additional learning to create an additional dictionary. Dictionary 14a after additional creation as output
Is obtained.

In the present embodiment, even when the category of the input pattern is not known as described above, the user can successively create the dictionary while performing pattern recognition (character recognition or voice recognition) with this apparatus.

In this embodiment, all the characteristic patterns obtained from the input pattern sequence are used as learning samples. A dictionary can be created in the same manner by using the following ~ as learning samples.

A learning sample is obtained by excluding the special jump pattern determined by the user not to be learned from all the above characteristic patterns.

A feature pattern that is an error or rejected by recognizing the input pattern is used as a learning sample.

A learning sample is obtained by excluding the characteristic pattern determined by the user not to be learned from the characteristic pattern that has become an error or rejected.

The learning sample selection and determination processing described above can be performed by the correction unit 19 in FIG.

The above embodiments can be realized by using a microprocessor, a memory, a scanner, a microphone, a display and the like.

The above embodiment is for the case where the dictionary consists of a set of standard patterns and thresholds per category. The recognition performance can be further improved by using a plurality of sets of standard patterns and thresholds for each category. The present invention can be applied to this case as well.

〔The invention's effect〕

The effects obtained by the present invention can be summarized as follows: In the additional creation of the dictionary, it is not necessary to refer to the learned sample used before the additional learning. for that reason,
Since the memory for storing the i-learned samples is not required, the memory capacity can be reduced. In addition, since it is not necessary to calculate the distance between the learned sample and the new standard pattern after the additional learning in the ii threshold calculation, the calculation time can be significantly reduced.

When recognizing learning samples, the correct category can be raised to one of the candidate categories for all learning samples. Moreover, the number of complementary categories can be made sufficiently smaller than the number of all categories to be recognized.

[Brief description of drawings]

FIG. 1 is a diagram for explaining a procedure for selecting a candidate category and determining an answer category, FIG. 2 is a diagram for explaining a procedure for creating a dictionary, and FIG. 3 is a diagram for explaining the principle of the present invention. Figure,
FIG. 4 is a diagram showing the configuration of the first embodiment of the present invention, and FIG.
6 and 6 are diagrams showing examples of characteristic patterns used in character recognition, and FIG. 7 is a diagram showing the configuration of the second embodiment of the present invention. 11 ... Pattern observation section 12 ... Feature extraction section 13 ... Learning sample memory section 14 ... Dictionary calculation section 15 ... Dictionary memory section 16 ... Control section 17 ... Matching section 18 ... Judgment section 19 ... Correction Part 10a …… Input pattern 11a …… Electrical signal 12a …… Feature pattern 13a …… Learning sample 14a …… Updated dictionary output from the dictionary calculator 15a …… Input to the dictionary calculator before update dictionary

Claims (3)

[Claims] 1. A standard pattern, which is an average value for the same category of feature patterns called a learning sample, a threshold value defined by a maximum distance between all the learning samples forming the average value and the average value, and each average. In the pattern recognition dictionary creation method that creates a pattern recognition dictionary consisting of the number of learning samples that is the number of samples used to obtain a value, when learning samples are added sequentially, the standard that becomes the dictionary after addition The pattern (f _n ), the threshold value (θ [f _n ]), the learning sample number n, the additional learning sample (u _{m + 1} ... u _n ), the standard pattern f _m that is the dictionary before the addition, and the threshold value θ [f _m ] and the number of learning samples m, A method for creating a pattern recognition dictionary, which is characterized by being calculated and obtained by. 2. A feature pattern obtained by subjecting an input pattern to pattern recognition of the input pattern by a dictionary before input of the input pattern and correction of an answer category after the pattern recognition is used as an additional learning sample. A method for creating a pattern recognition dictionary according to claim 1. 3. A feature pattern obtained by subjecting an input pattern to pattern recognition of the input pattern by a dictionary before input of the input pattern and correction of an answer category after the pattern recognition is characterized by an additional learning sample. A method for creating a pattern recognition dictionary according to claim 1.