JPH0221032B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a character recognition device, and particularly to a method for efficiently creating a plurality of dictionaries from a large number of samples.

Normally, in a character recognition device, there is one dictionary for each category for matching input signals.

In addition, although attempts have been made to use multiple dictionaries per category to perform more accurate character recognition, there are cases where there are overlaps between samples within each dictionary, significant differences in the number of constituent samples, etc. Not only was it difficult to create a dictionary, but it also required complicated procedures.

The present invention provides a novel method for creating multiple dictionaries, which will be described in detail below.

Figures 1 to 4 are examples of the process of creating multiple dictionaries.
The classification is shown using 10 samples a to j. In the figure, the two-dimensional distance between each sample is calculated by calculating a mutual distance matrix after extracting the respective features, and then schematically representing it. That is, the larger the distance between each sample, the more different the characteristics.

FIG. 1 shows the classification when one dictionary is conventionally created for one category. Regarding the creation of a single dictionary, the average A of all samples is found and a dictionary is created from there. This also applies to the creation of multiple dictionaries according to the present invention; after selecting constituent samples of each class, a dictionary is created from the average of the constituent samples.

In the present invention, when creating K plural dictionaries from N samples, from the first to {K-N
+(n ₁ *K)}th dictionary, n ₁ = |N/K
For the | {K-N+(n ₁ *K)+1}-th to K-th dictionaries, (n ₁ +1) constituent samples are selected. In addition, the selection of constituent samples involves determining samples S ₁ and S ₂ , which are the core of each dictionary.
This is done for each sample close to S ₁ and S ₂ so as not to overlap each other. That is, the two most distant initially
After selecting one sample as a core and selecting constituent samples of each dictionary using this sample as the core, mask the samples that have already been selected.
From the next step onwards, only the remaining samples will be classified.

In the above equation, N...Number of samples in one category K...Number of dictionaries created for one category _n1 ...Number of samples in one dictionary, respectively.

In addition, when creating K dictionary patterns from the above N samples, if N is an integer multiple of K, multiple dictionaries should be created using N/K (integer) samples for each dictionary pattern. It's fine, so there's no problem.

On the other hand, since there may be cases where N is not always an integral multiple of K, it is desirable to average the number of samples used to create each dictionary pattern even when such a situation occurs. be.

Therefore, when N is not an integral multiple of K, [N/K]
By creating two types of dictionary patterns: one created from N/K samples and one created from [N/K]+1 samples, the number of samples used to create each dictionary pattern is approximately averaged.

That is, the above two equations represent the number of dictionary patterns to be created from [N/K] samples, and [N/K]
This is to find the number of dictionary patterns to be created from K]+1 samples.

In Figure 2, K=2, that is, 2 out of 10 samples.
This example shows how to create dictionaries. here,
Since N=10 and K=2, n ₁ =|10/2|=5, and K-N+(n ₁ *K)=2-10+5×2=
2. Select five constituent samples for each of the first and second dictionaries (all dictionaries).

First, the distance between samples x and y is set as follows by converting an input character into a video signal and extracting various features from the input video.

d(x,y)= _n 〓 ^l=1 ｜x ^l −y ^l ｜ (where l is the number of dimensions consisting of m) In this way, calculate the inter-sample distance for each of the 10 samples and form a distance matrix. This is what I did.

In FIG. 2, since the distance between a and c has the maximum distance between samples, a and c are selected here as the core of the dictionary. Next, from either a or c, five samples are each selected that are closest to the sample, that is, those with a small inter-sample distance.

For example, when first creating a dictionary with a as the core, h, b, which has a small distance between samples with respect to a,
Select f and g, respectively. In addition, since the number of dictionaries to be created here is two, after selecting the constituent samples of the dictionary with the above-mentioned a as the core, all remaining samples may be used as constituent samples of another dictionary.

After selecting constituent samples of each dictionary, each dictionary is created by taking the average of each using conventional dictionary creation means. Since such a step is merely an application of a conventional technique, it will be omitted in the following description.

FIG. 3 shows the case where three multiple dictionaries are created from ten samples. A class for creating the first dictionary in the same process as in Figure 2 above.
k ₁ with sample a as the core, second class k ₂
After selecting constituent samples using sample c as a core, a third class _k3 is created from the remaining samples.

Here, n ₁ = | 10/3 | = 3, K-N + (n ₁
*K) = 2, for k ₁ and k ₂ , 3 configuration samples each, for k ₃ , (n ₁ + 1)
(=4) configuration samples are selected.

FIG. 4 shows a case where four plural dictionaries are created from ten samples.

Nuclei a and c are similar to the explanation in FIGS. 1 and 2 above.
, and create classes k ₁ and k ₂ respectively. Next, a, which is a sample of the core classes k ₁ and k ₂ ,
Mask h, j, and c.

Here, since n ₁ = 2, K - N + (n ₁ * K) = 2, the number of constituent samples for the first class k ₁ and the second class k ₂ is 2 each, and the number of samples for the third class is 2.
The number of constituent samples of k ₃ and the fourth class k ₄ are each (n ₁ +1=) 3.

After the constituent samples of the classes k ₁ and k ₂ are masked, they remain in the same process as above. b, e,
Classification is performed for samples d, f, g, and i. Here, since g and i have the largest inter-sample distance, these are used as the core of classification.

As mentioned above, the number of constituent samples for creating the third and fourth dictionaries is three, so for example, i is selected as the third kernel and samples b, f,
Form a class k ₃ consisting of i. So class k ₄
is composed of samples d, e, and g.

It is also possible to select sample g as the third nucleus here. If sample g is used as the third nucleus, its class will consist of b, e, and g, and the fourth class will consist of the remaining d, f, and i. Such a difference in the constituent samples occurs depending on which of the two samples with the largest inter-sample distance is used as the core to determine the constituent samples first, but in applying the present invention, each sample It is preferable to configure each class so that the sum of the distances between them is small, that is, each class is composed of adjacent samples.

As described above, the creation of multiple dictionaries according to the present invention consists of a feature extraction section that extracts the characteristics of input characters based on a signal from an observation section that converts input characters into a video signal, and an interval between each sample based on the characteristics. a distance matrix generation unit that calculates the distance of each dictionary by the above-mentioned means; a detection unit that detects the core of each dictionary; and a constituent sample detection unit that detects the samples that constitute each dictionary from the core sample. a masked distance matrix generation unit that masks the distance matrix regarding the core sample;
This is performed by a multiple dictionary creation device comprising a dictionary creation unit that creates a plurality of dictionaries from the average of the features of constituent samples classified into each class.

The basic operating principle of each part will be explained below with reference to FIG.

In the figure, 1 is the distance matrix, 2 is the selector section,
3 represents a first detection unit, 4 represents a second detection unit, 5 represents a first register, 6 represents a configuration sample detection unit, 7 represents a masked distance matrix generation unit, 8 represents a masked distance matrix, and 9 represents a second register. There is. Further, _CL1 to _CL3 each indicate the input of a clock signal.

As described above, the features of the input characters converted into video signals by the observation unit are extracted by the feature extraction unit to form a distance matrix 1. This step has been carried out in conventional singular dictionary creation devices, and the method used is not particularly limited.

Next, the selector unit 2 selects the first dictionary to be created.
If it is the second one, the signal from the distance matrix 1 is selected, and if it is the second one or more, the signal from the masked distance matrix 8 is selected. Here, as a means for making the selection, a counter k for checking the number of the dictionary currently being created may be built into or connected to the selector section 2.

In the first detection unit 3, if the dictionary to be created is an odd numbered dictionary (the value of the counter k is an odd number), that is, 2
It operates at clock 2 (the same applies for C _L 2 and below) only once among the input times, and calculates the maximum inter-sample distance in distance matrix 1 or masked distance matrix 8, and calculates the numbers of core samples S ₁ and S ₂ . Detect. For example, in the example shown in FIG. 6, the distance between c and f is maximum at 56, so samples c and f are detected as the core for dictionary creation.

When the detection at the first detection section is completed, C _L 3 is turned ON, the signal from the first detection section is taken in by the second detection section, and the distance between the core sample and other samples is stored in the first register 5. Forward. Here sample c
The distance between the sample and the other sample is transferred to the first register 5.

The configuration sample detection unit 6 includes the first register 5
The first to {K-N+(n ₁ *
K)}-th dictionary creation, n ₁ (=|N/K
|) samples (k=1~{K-N+(n ₁ *
In the case of {K-N+(n ₁ *K)+1}-th to K-th dictionary creation, (n ₁ +1) samples are detected from the one with the smallest distance.

Here, N=7, K=3, and k=
Since this is the case of 1, n ₁ =|7/3|=2, and two samples from the smallest one are detected. The detected samples are b and c.

This result is transferred to the second register 9 and held there.

The masked distance matrix generation unit 7 applies a mask of O to the distance matrix regarding the previously detected sample. That is, this is to prevent the same sample from being used as a constituent sample of a dictionary that is created repeatedly.

The data of the masked distance matrix 8 is transferred to the selector section 2.

Thereafter, the data in the second register 9 is sent to a dictionary generation section (not shown) to create a dictionary according to its contents.

As has become clear from the above embodiments, according to the present invention, it is possible to create appropriate multiple dictionaries simply and in a short time.

Further, in the above embodiment, from the first to the {K
−N+(n ₁ *K)}th dictionary, use n ₁ configuration samples, and for {K−N+(n ₁ *K)}th to Kth dictionaries, use (n ₁ +1) samples. Although a configuration sample was selected, the application of the present invention is not limited to this.

That is, for the first to {N-(n ₁ *K)}-th dictionaries, there are (n ₁ +1) {N-(n ₁ *K)}-th dictionaries.
n ₁ for the Kth dictionary from
For example, select {N−
It is possible to select (n ₁ +1) constituent samples in (n ₁ *K)} dictionaries.

[Brief explanation of drawings]

FIG. 1 shows the classification in the conventional creation of a single dictionary, and FIGS. 2 to 4 show the classification in the creation of a plurality of dictionaries according to the present invention. or,
FIG. 5 shows the main structure of the dictionary creation method according to the present invention, and FIG. 6 shows its details. In the figure, 1 is the distance matrix, 2 is the selector section,
3 is a first detection section, 4 is a second detection section, 5 is a first register, 6 is a configuration sample detection section, 7 is a masked distance matrix generation section, 8 is a masked distance matrix, and 9 is a second register.

Claims (1)

[Claims] 1. In the step of creating multiple dictionaries of a character recognition device,
A feature extraction unit that extracts the characteristics of each of the N dictionary creation character samples from the input video; a distance matrix generation unit that calculates the distance between each sample from the features; and a distance matrix generation unit that calculates the distance between each sample based on the number of the dictionary to be created. a selector section that selects a signal from the sensor section, a detection section that selects a sample S separated by the maximum distance from the signal from the sensor section and a distance between the sample S and other samples, and a detection section that selects a signal from the sensor section. a constituent sample detection unit that detects a specific number of samples starting from those having a small distance between the samples; a dictionary generation unit that generates a dictionary by calculating the average of the features of the constituent samples from the output of the constituent sample detection unit; 1. A plurality of dictionary creation device comprising: a masked distance matrix generation unit that masks signals related to sample S among signals from a selector unit to generate a masked distance matrix and transfers the masked distance matrix to the selector unit. 2. When creating K multiple dictionaries, the configuration sample detection unit selects the first to {K-N+
(n ₁ *K)}th dictionary is n ₁ (=|N/K
|) configuration samples as {K−N+(n ₁ *K)
+1}th to Kth dictionaries are (n ₁ +
1) A plurality of dictionary creation device according to claim 1, wherein each of the configuration samples is selected.