JPH0414359B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an improvement in a device for recognizing an input pattern expressed as a sequence of vectors, and more particularly to an improvement in a device that performs recognition using a vector standard pattern prepared in advance.

Generally, patterns such as sounds, figures, characters, etc. can be expressed as a series of feature vectors, but when a pattern is expressed as a series of vectors, standard patterns are prepared in advance, and these patterns and input patterns can be expressed as a series of vectors. Recognition can be performed by comparing the degree of similarity (a measure of similarity) with At this time, as seen in the Proceedings of the Acoustical Society of Japan (October 1982), p.391-p.392,
The amount of processing can be reduced by preparing a standard pattern (vector standard pattern) in which vectors are units, and expressing the category standard pattern as a dictionary using a combination of the vector standard patterns.
However, even in this case, if it is necessary to prepare many dictionaries for one category, there is a problem that the amount of processing increases because many dictionaries must be referred to during recognition.

An object of the present invention is to realize a pattern recognition device that requires a small amount of processing even when a large number of dictionaries are required for one category as described above.

In general, the pattern S is a sequence of n vectors |S ₁ , ..., _o |, and the vector standard pattern is R ₁ ,
R ₂ , ..., each vector S _i of the pattern S
By replacing S with a vector standard pattern that has the highest degree of similarity to S _i , the pattern S can be expressed as a symbol sequence [L ₍₁₎ , . . . , L _(o) ] consisting of n symbols. In other words, if the similarity between vectors S _i and R _j is X (i, j), then: X (i, L _i ) = max j (referred to as vector sequence standard patterns), the number of categories to be recognized is M, the number of vector sequence standard patterns of category m is k(m), and the vector sequence standard patterns are A ^(m) ₂ , A ^(m) ₂ , ..., A ^(m) _k(n) , each of A ^(m) _j can be expressed as a symbol sequence as follows.

A ^(m) _j = [L ^(m) _J (1), L ^(m) _J (2), ..., L ^(m) _J (n
)], j = 1, 2, ..., k (m) m = 1, 2, ..., M (2) These are called vector sequence standard pattern dictionaries.

Using these, recognition can be performed as follows. The vector sequence of the input pattern is |S ₁ ,...
..., S _o |. First, each vector S _i of the input pattern
and the vector standard patterns R ₁ , R ₂ and the similarity X(i, ₁ ), X(i, ₂ ), ...(3) is calculated. Next, the degree of similarity with each vector sequence standard pattern is determined using a dictionary as follows.

y ^(m) _j = n 〓 ⁱ⁼¹ x (i, L ^(m) _J (i)), j = 1, 2, ..., k (m), m = 1, 2, ..., M ( 4) Next, find m that maximizes y ^(m) = max j = 1, K (m) y _j ^(m) (5), and this will be the recognition result.

In the above recognition process, it is necessary to execute equations (4) and (5), but when there are many dictionaries for one category, the processing of equations (4) and (5) increases. The present invention solves this problem as follows.

Instead of expressing one vector series standard pattern as a dictionary, for a symbol j representing a certain vector standard pattern, the probability that it is at position i on the symbol series of a certain category m is expressed as P(m)(i,
j), (that is, the probability that it is the i-th symbol of category m) is defined and given to all combinations of symbol j, category m, and position i. This will be called a vector standard pattern dictionary. At this time, the input pattern is replaced with a symbol sequence by the operation of equation (1), and the result is [L ₍₁₎ ,...,
L _(o) ], the similarity for category m can be expressed as y ^(m) = n 〓 ⁱ⁼¹ P ^(m) (i, L(i)) (6).

An example of the execution of equation (6) will be shown below for the case where the pattern is represented by a series of four vectors, the number of categories M=3, and the number of standard patterns is six.

Suppose that the input pattern |S ₁ , S ₂ , S ₃ , S ₄ | is converted into a symbol sequence [5, 5, 2, 6]. Also,
It is assumed that the vector standard pattern dictionary is described as shown in FIG. In this case, the similarity y ⁽¹⁾ , y ⁽²⁾ , y ⁽³⁾
is obtained as the sum of the ○ marks in the table as follows.

y ⁽¹⁾ =0.9+0.8+0.6+0.7=3.0 y ⁽²⁾ =0.4+0.9+0.2+0.3=1.8 y ⁽³⁾ =0.3+0.4+0.3+0.1=1.1 Vector standard pattern dictionary P ^(m) (i, j) can be described by empirical knowledge, or can be obtained by a statistical method using a large number of vector sequence standard patterns. For example, as an example of the former, in the case of word speech recognition, if the vector standard pattern corresponds to the phonological standard pattern,
Since P ^(m) (i, j) represents "the probability that a certain phoneme is in a certain position of a certain word", it can be described using phonetic knowledge.

The apparatus according to the present invention includes means for storing a vector standard pattern, means for converting an input pattern consisting of a sequence of vectors into a sequence of n symbols by referring to the standard vector pattern, and representing a category as a sequence of n symbols. Then, the probability that symbol j is at position i on the symbol series of category m (that is, the probability that symbol j is the i-th symbol of category m) is calculated for all combinations of symbol j, category m, and position i. means for storing the information given as a vector standard pattern dictionary; means for obtaining the degree of similarity for each category from the symbol series by referring to the vector standard pattern dictionary; and means for determining the category with the maximum degree of similarity as a recognition result. and output means.

According to this configuration, recognition processing is executed using the formula
Equation (6) is used instead of (4) and (5), thus realizing a pattern recognition device with a small amount of processing.

The principle of the present invention will be explained below with reference to drawings of embodiments.

FIG. 2 is a block diagram showing an example of an apparatus according to the present invention.

The input pattern signal S is input to the processing section 2. The vector standard pattern data is stored in the storage unit 1, and the processing unit 2 stores the input vector sequence S.
= |S ₁ , ..., S _o | into the symbol sequence [L(1), ..., L(n)] with reference to the vector standard patterns R ₁ , R ₂ ... stored in the storage unit 1 Convert and output as a signal.

Note that in order to convert each input vector into a symbol, for each input vector, find the degree of similarity with all vector standard patterns and assign the symbol corresponding to the vector standard pattern with the maximum degree of similarity. . The storage unit 3 stores a vector standard pattern dictionary. The processing unit 4 uses the vector standard pattern dictionary stored in the storage unit 3 to calculate the similarity of the symbol sequence signal sent from the processing unit 2 to each category y ⁽¹⁾ , ..., y ^(M) Convert and output. Here, the similarity is obtained by executing equation (6). The processing unit 5 compares these similarity values and outputs the maximum m as the category number of the recognition result.

Although the present invention has been explained above, the measure of certainty used in the vector standard pattern dictionary may be any arbitrary measure, and instead of expressing the degree of similarity for a category as a sum of likelihoods as in equation (6), it can be expressed as a product or It may be expressed by maximum value, minimum value, etc.

Furthermore, the concept of "category" used in the explanation does not have to be exclusive. Therefore, the principle of the present invention also holds true, for example, when a word is made to correspond to a plurality of categories in word speech recognition.

In the explanation, both the number of vectors in the vector sequence of the input pattern and the number of symbols in the converted symbol sequence are n.
However, the number of vectors may be arbitrary as long as a certain number of symbols are obtained when converted to a symbol sequence.

[Brief explanation of drawings]

FIG. 1 is a diagram showing an example of the contents of a vector standard pattern dictionary, and FIG. 2 is a block diagram showing an embodiment of the present invention. In the figure, 1 is a storage unit that stores vector standard patterns, 2 is a processing unit that converts a vector sequence into a symbol sequence, 3 is a storage unit that stores a vector standard pattern dictionary, and 4 is a processing unit that obtains the similarity for each category. , 5 is a processing unit that compares the degree of similarity and obtains a recognition result.

Claims (1)

[Claims] 1. means for storing a vector standard pattern; means for converting an input pattern consisting of a sequence of vectors into an input symbol sequence consisting of n symbols (symbols corresponding to the vector standard pattern) by referring to the vector standard pattern; When a category is expressed as a symbol series consisting of n symbols, the probability that symbol j corresponding to a certain vector standard pattern is at position i on the symbol series of category m is expressed as all symbols j,
means for storing information given for the combination of category m and position i as a vector standard pattern dictionary; means for obtaining the degree of similarity for each category from the input symbol sequence by referring to the vector standard pattern dictionary; 1. A pattern recognition device comprising: means for outputting a category having a maximum degree as a recognition result.