JPH0522154A - Method for setting reference vector initial value of lvq - Google Patents

Abstract

PURPOSE:To facilitate learning by automatically setting an initial value corresponding to learning vector distribution among categories in the case of setting the initial value of a reference vector at an LVQ used for encoding signals in audio engineering or image engineering. CONSTITUTION:First of all, a j-th reference vector Mij in an i-th category is arranged at the centrode of each category ( and (square) and next, the most adjacent reference vector is calculated in respect to each learning vector Lij and defined as a member vector. Afterwards, a new reference vector Mi'j' is generated at the centrode as the member vector of the different category, and a ratio to occupy the same category in the member is defined as an inside potential. For the reference vector having this value smaller than a prescribed reference value, a process following to a figure (b) is repeated.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a learning vector quantization technique (learning vector quantizer) used for signal coding in voice engineering and image engineering.
section; hereinafter referred to as LVQ), and more particularly to an initial value setting method for setting an initial value according to a learning vector distribution between categories.

[0002]

2. Description of the Related Art LVQ has been conventionally emphasized in signal coding in voice engineering and image engineering, but with the rapid development of the AI method, learning has improved qualitatively and has become more important. Is coming. Vector quantization was developed by increasing the memory capacity and cost, and is a coding method that uses a group of consecutive sample values as one category. A group of N sample values is It can be considered as a point in an N-dimensional space having N components or an N-dimensional vector from the origin to that point. For example, by some method, M numbered reference vectors Yi as shown in FIG. 2 are arranged in this N-dimensional space, and a desired vector X (= x1, x2, ... xN) is input. Then, the spatial distance between this and each reference vector Yi is defined, and the input vector X is defined by the reference vector Ym having the smallest distance.
Of the reference vector Ym
Is determined (encoded). LVQ is the reference vector Y
This is a pattern recognition method in which m is obtained by learning.

[0003]

In the LVQ, the LVQ depends on how the initial value of the reference vector is set.
Performance is greatly affected. However, the conventional method of selecting the initial value is either to appropriately extract from the learning vector or to give the initial value such that the distortion rate in each category becomes small. In this case, it cannot be said that the learning vector distribution state between categories is faithfully reflected.
Even if reference vectors are learned by VQ, the number of reference vectors irrelevant to the recognition rate only increases. The present invention was devised in view of such problems, and it is possible to set an initial value according to a learning vector distribution state between categories.
It is an object of the present invention to provide a method for setting an initial value of a reference vector that facilitates learning of.

[0004]

Means for solving the above-mentioned problems in the present invention are as follows in the method of setting an initial value of a reference vector in learning vector quantization for obtaining an optimum reference vector representing each category by learning. A reference vector is arranged in a centroid of a category, a reference vector closest to each learning vector is obtained, the learning vector is set as a member vector of the reference vector, and a cent that overlaps with a member vector of a different category of the member vector. It is assumed that a new reference vector is generated by Lloyd and the ratio of member vectors belonging to the same category is used as the internal potential of the new reference vector to set the reference vector initial value.

[0005]

The present invention is an initial value setting method in which a reference vector is self-generated according to the distribution state of the learning vector of each category. Therefore, in the present invention, the reference vector is arranged in the centroid of each category, the reference vector closest to each learning vector is obtained, the learning vector is set as the member vector of the reference vector, and the different category among the member vectors is calculated. , A new reference vector is generated in the centroid which is also a member vector, and the ratio of the member vectors belonging to the same category is evaluated as the internal potential of the reference vector.

[0006]

Embodiments of the present invention will now be described in detail with reference to the drawings. FIG. 1 is an explanatory view of a process of one embodiment of the present invention. In this embodiment, each symbol is defined as follows.

M; reference vector L; learning vector i; category number (1 ≦ i ≦ n) j; reference vector number (1 ≦ j ≦ k) Therefore, Mij is referred to as the j-th reference vector in the i-th category. Similarly, the j-th learning vector in the i-th category is Lij. In this embodiment, first, as shown in FIG. 1A, the reference vectors M11, M21, M31 are arranged in the centroid of each category (Δ, □,). Next, the reference vector closest to each learning vector Lij (not shown) is obtained, and each learning vector Lij is the reference vector M11, M2 shown in FIG.
It will be referred to as a member vector of either 1 or M31. Here, since the member vector of each reference vector is determined by the spatial distance, remote parts include those of different categories, but the centroids of the member vectors of different categories include the new reference vector Mi′j ′.
Is generated. For example, in FIG. 1C, category 1 (△)
There are also category 2 (□) and category 3 () member vectors in the member vector of the reference vector M11. Therefore, in this embodiment, the former centroid is caused to generate the reference vector M22 and the latter centroid is caused to generate the reference vector M32. Similarly, the reference vector M12 and the reference vector M33 are generated also in the member vector of the reference vector M21, and the reference vector M13 and the reference vector M23 are also generated in the member vector of the reference vector M31. The internal potential Ei'j 'of these novel reference vectors Mi'j' is calculated by the following equation.

Ei'j '= number of category i among member vectors / number of member vectors For the reference vector Mi'j' whose internal potential Ei'j 'is less than a predetermined reference value, 1
(B) Repeat the subsequent procedure.

The following effects are apparent in this embodiment.

(1) The contribution of the reference vector in each category can be recognized by defining the internal potential in the reference vector.

(2) By self-generating the reference vector, the initial value can be set according to the distribution state of the learning vector between categories.

[0012]

As described above, according to the present invention, there is provided a reference vector initial value setting method capable of setting an initial value according to the distribution of learning vectors between categories and facilitating LVQ learning. be able to.

[Brief description of drawings]

FIG. 1 is an explanatory diagram of an embodiment of the present invention.

FIG. 2 is an explanatory diagram of vector quantization.

[Explanation of symbols]

M ... Reference vector, i ... Category number, j ... Reference vector number.

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁵ Identification code Office reference number FI technical display location H04N 7/133 Z 4228-5C

Claims (1)

Claim: What is claimed is: 1. In a method of setting an initial value of a reference vector in learning vector quantization, which obtains an optimum reference vector representing each category by learning, the reference vector is arranged in a centroid of each category, The nearest reference vector is obtained for each learning vector, the learning vector is used as a member vector of the reference vector, and a new reference vector is generated in the centroid that overlaps with the member vector of a different category of the member vector, and the member vector is generated. A method for setting an initial value of a reference vector, characterized in that a ratio of vectors belonging to the same category is used as an internal potential of a new reference vector.