JPH08316842A - Representative vector generating system and method for vector quantization - Google Patents

Abstract

PURPOSE: To prepare a representative vector minimizing an expected value of a distortion error by providing a specific representative vector initializing means, a specific contention learning means and a specific new representative vector generating means to the system. CONSTITUTION: A representative vector initializing means generates initially a representative vector having a value equal to an input vector selected at first at random to start processing. In the processing, a contention learning means learns contention of a representative vector having been already generated based on the input vector read sequentially. Every time the contention learning is repeated for a prescribed number of times, a new representative vector generating means calculates the total sum of distortion errors for all input vectors and their representative vectors in a dominant area of each representative vector, selects a representative vector with the highest sum and generates one new representative vector or over with the same vector value as the representative vector. Then the new representative vector is added as an object of the contention learning and the contention learning is repeated for the object.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a representative vector generation method and method for vector quantization applied to information compression of images and voices, image region division, image gradation change, pattern recognition and the like. .

[0002]

2. Description of the Related Art For various waveform data such as images and voices, data compression is performed by vector quantization to improve the efficiency of data storage and transmission. Vector quantization is used in various fields, such as performing voice recognition by converting into a pattern.

FIG. 6 conceptually shows vector quantization by using an example of a two-dimensional vector. The whole vector space is
It is divided into a dominant region 33 of a limited number of representative vectors 32 prepared in advance. When the vector 31 to be quantized is input, the input vector 31 is converted into the representative vector 32 of the dominant region 33 to which the vector 31 belongs and is output. Thus, all input vectors 31 are
It will be quantized into a predetermined finite number of representative vectors 32.

In performing this vector quantization,
It is necessary to prepare the representative vector in advance. In that case, it is desirable to prepare an appropriate representative vector according to the distribution of the input data so that vector quantization can be performed so that the expected value of the distortion error is minimized. That is,
When the sum of distortion errors (for example, Euclidean square distance) between all input vectors belonging to each representative vector and its representative vector (hereinafter referred to as partial distortion) is calculated, all representative vectors are calculated. It is desirable that the partial strain of is minimized.

A method known as competitive learning has been conventionally known as a method for preparing a representative vector. As one method of this competitive learning, for example, “selective competitive learning method for vector quantization design: equal distortion principle and its realization algorithm” (IEICE Transactions, '94 / 1
1, Vol. J77_D_II, pp. 2265-227
The technology disclosed in page 8, 1994) is known.

According to this disclosure, one learning of competitive learning selects one from a large number of input vectors and distorts the error between the input vector and each of a predetermined number of representative vectors. Is calculated, one representative vector having the smallest distortion error from the input vector is selected as the winner, and the representative vector of the winner is updated so as to be closer to the input vector. Then, such learning is repeated many times.

Further, in the same document, a theorem is shown in the principle of equal distortion. The theorem states that "when the total number of representative vectors is sufficiently large, regardless of the input vector distribution, the governing region (Voronoi region) of each representative vector is The partial distortions in) must be equal to each other. "It is necessary to realize this equal distortion in order to realize vector quantization that minimizes the expected value of the distortion error. It is shown. Then, a method for realizing equal strain is shown.

Regarding competitive learning, Japanese Patent Laid-Open No. 4-12
The technique disclosed in No. 4782 "Feature extraction method and its implementation device" is also known. This is a method in which a representative vector having the highest input vector density in the governing region is selected from a predetermined representative vector group, and a representative vector having the same vector value as the selected representative vector is newly generated. is there.

[0009]

As shown in the former known document, it is necessary to realize equal distortion for vector quantization that minimizes the expected value of distortion error. However, realization of equal distortion is a necessary condition to minimize the expected value of distortion error.
Not enough conditions. For example, when the partial distortion of any of the representative vectors is large, even if the condition of equal distortion is satisfied, vector quantization is not appropriate for the purpose of minimizing the expected value of distortion error. In this case, although the equal distortion is realized, the possibility that the solution converges to the locally optimal solution having a large error increases.

Further, as in the latter document, a representative vector having the highest input vector density in the dominant region is selected,
There is also a method of generating a representative vector having the same vector value as the representative vector, but this method alone does not realize equal distortion.

The present invention has been made in view of such a background, and an object thereof is to provide a representative vector generation method for vector quantization that minimizes an expected value of a distortion error. is there.

[0012]

A representative vector generation method for vector quantization according to the present invention comprises the following means A) to C).

A) Representative vector initialization means for initially generating a smaller number of representative vectors than the number of representative vectors required for vector quantization.

B) One input vector is selected from a large number of input vectors, and one representative vector having the smallest distortion error with the selected input vector is selected as a winner from among the representative vectors already generated. Competitive learning means for repeatedly performing competitive learning for updating the representative vector of the winner so as to approach the selected input vector.

C) Every time the competitive learning is repeated a predetermined number of times, a representative vector having a relatively large partial distortion is selected from the already generated representative vectors according to a predetermined criterion, and based on the selected representative vector. A new representative vector generating means for generating one or more new representative vectors.

[0016]

According to the method of the present invention, first, the number of representative vectors (for example, one) smaller than the required number of representative vectors is initially set, and then the process is started. First, one is selected from a large number of input vectors, the distortion error (for example, Euclidean square distance) between the input vector and each representative vector is calculated, and one representative vector having the smallest distortion error with the input vector is calculated. Competitive learning is performed by selecting as a winner and updating the selected representative vector so as to be closer to the input vector. This competitive learning is
Repeatedly selecting input vectors sequentially.

By repeating this competitive learning, each representative vector is modified so that the partial distortion in each dominant region becomes small.

Then, each time the number of repetitions of competitive learning reaches a predetermined number, a new representative vector generation process is performed. That is, a representative vector having a relatively large partial distortion is selected from the currently existing representative vectors by a predetermined criterion, and a new representative vector is generated based on this representative vector.

Since a new representative vector is generated based on the representative vector having a relatively large partial distortion in this way, the dominant region of the representative vector having a large partial distortion is divided by the new representative vector. , Its large partial strain will be small.

When a new representative vector is generated in this way, it is also added to the object of competitive learning, and the above-mentioned competitive learning is repeated again.

Here, as the method of determining the value of the new representative vector, for example, the method of determining the same value as the selected representative vector can be adopted as the simplest method, but it is not limited to this method. is not.

There are several possible variations in the method of selecting a representative vector having a large partial distortion. For example, the most basic method is to select a representative vector having the largest partial distortion among the already generated representative vectors. Also, if you want to simplify the process by omitting the calculation of partial distortion, for example, in each round of competitive learning, the variable that the representative vector selected as the winner has is the distortion error between that representative vector and the selected input data. Can be added, and a representative vector having the largest variable can be selected from the representative vectors already generated. This is because the above variable is almost proportional to the partial strain.

Further, when a representative vector having a large partial distortion is selected by the above method, only when the number of selections of the representative vector has reached a predetermined threshold number, the representative vector is based on the representative vector. A method of generating a new representative vector can also be adopted. According to this method, when a state with large partial distortion is a minor state that can be resolved by repeated competitive learning, generation of a new representative vector in the dominant region of the representative vector is avoided. Therefore, it is possible to prevent the local distortion from becoming excessively small locally due to the generation of the new representative vector, and it is possible to better achieve the condition of equal distortion.

[0024]

Embodiments of the present invention will now be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing a schematic configuration of an embodiment of a vector quantization device to which the representative vector generation method of the present invention is applied.

As shown in FIG. 1, the vector quantizer of this embodiment has an input device 1, an arithmetic processing device 2, an output device 3, and an input vector file (stored in an external memory) 6. , And a representative vector file (stored in an external memory) 7. The input vector file 6 stores a large number of input vectors. The representative vector file 7 stores a plurality of prepared representative vectors in advance.

The arithmetic processing unit 2 includes a central arithmetic processing unit (CP
U) 4 and an internal memory 5. The internal memory 5 stores control programs, processing programs and the like.

FIG. 2 is a flow chart showing the process of vector quantization in this embodiment.

When performing vector quantization, as shown in FIG. 2, input vectors are input one by one (S1), and representative vectors are sequentially taken out from the representative vector file 7 for the input vectors, Euclidean square distance between the input vector X and the representative vector W, that is,

[Equation 1] (S2), one representative vector having the smallest distance from the input vector X in the representative vector group is selected as the winner, and the vector value of the representative vector that is the winner is quantized for the input vector. Output as (S
3).

FIG. 3 is a flow chart showing a process of preparing a representative vector by the improved competitive learning method of the present invention so that the vector quantization as described above can be performed.

First, the vector quantization number n indicating the number of representative vectors to be prepared, the end number L indicating the total number of learning to be repeated until the end of the process, and the number of inspections until one representative vector is generated ( The learning number) C and are input (S11).

Next, the parameters for controlling the process are initialized (S12). Specifically, the number of representative vectors q indicating the number of generated representative vectors is defined, the initial value 1 is substituted, and the learning number t indicating the number of times of already repeated learning is defined, and the initial value 0 is set. Substituting and defining a partial number r that counts the number of learnings until reaching the inspection number C, substituting an initial value 0, and further randomly selecting one input vector from the input vector file 6 The value is defined as the value of the first representative vector (initialization of the representative vector).

Next, it is judged whether or not the partial number r has reached the inspection number C and the representative vector number q is smaller than the vector quantization number (vector quantization level number) n (S).
13). Here, if YES, the process proceeds to step S20, and if NO, the process proceeds to step S14.

In step S14, one input vector is randomly selected from the input vector file 6. Then, the Euclidean square distance between the selected input vector X and each representative vector W already generated is calculated, and the representative vector W having the smallest distance from the input vector X among the representative vectors is selected as a winner (S1).
5).

Next, the representative vector W of the winner is updated using the learning coefficient g (a small positive number) prepared in advance as W ← W + g (X−W) (S16). That is, the value of the representative vector W of the winner is brought closer to the value of the input vector X by a rate according to the learning coefficient g.

Next, after adding 1 to the partial number r and the learning number t (S17), it is determined whether or not the learning number t reaches the end number L (S18). Here, if YES, it proceeds to step S19, and if NO, step S1.
Return to 3.

The learning process from steps S13 to S17 described above is repeated until the partial number r reaches the inspection number C. When the learning for the number of inspections C times is completed, YES is obtained in step S13, and therefore, the process proceeds to step S20 and subsequent steps to generate a new representative vector.

That is, after the partial count r is cleared to 0 (S
20), partial strain calculation is performed (S21). Here, it is defined that every position in the vector space belongs to the dominant region of the representative vector closest to the Euclidean square distance. First, for each of the generated representative vectors, the Euclidean square distance between all the input vectors existing in the control area of the representative vector and the representative vector is calculated and summed. That is, the partial strain of each representative vector is calculated. Next, a representative vector having the largest partial strain in the dominant region is selected from the representative vectors.

Next, a new representative vector is generated based on the selected representative vector (S22). That is, one new representative vector (or a predetermined number of two or more) having the same vector value as the representative vector selected in step S21 is generated. Then, 1 is added to the number of representative vectors q (S23), and the process returns to step S13.

Returning to step S13, the learning process from steps S13 to S17 is repeated for the representative part vector group to which the new representative vector is added, the number of times equal to the number of inspections C. Then, after the repetition of this learning, a new representative vector is added again by the processing of steps S20 to S23.

As described above, the representative vector which was initially one is added one by one (or a predetermined number of two or more) every time the learning is repeated C times, and finally the vector quantum is added. Up to a number equal to the number n is generated.
Further, the learning (update) based on the input vector of the representative vector is repeated, and finally the learning is repeated by the number of times equal to the end number L.

When the learning number t reaches the end number, the process proceeds to step S19, the q representative vectors already generated are stored in the representative vector file 7, and the process ends.

According to the above processing, the processing is started from one representative vector having a value equal to a certain input vector, and the number of representative vectors is gradually increased. It reduces the possibility of generating a vector, that is, arranging the representative vector in a region where the input vector is not distributed,
The representative vector can be arranged according to the distribution of the input vector.

Further, by iterative learning processing in which the already-generated representative vector is updated with a large number of input vectors, each representative vector is modified so that the partial distortion in each governing region becomes small, and the partial distortion Since a new representative vector is generated in the largest dominant region, the dominant region is divided and the partial distortion is reduced. As a result, the representative vector is prepared so as to minimize the partial strain while satisfying the equal strain condition.

FIG. 4 is a flow chart showing a modification of the representative vector generation processing of FIG.

The difference from the process of FIG. 3 of this modification is the process of updating the representative vector (S16-1) and the process of calculating the partial misalignment (S21-1).

First, the representative vector update process (S16-
To explain 1), the distance between the representative vector of the winner and the input vector determined in step S15 is added to the variable of the representative vector of the winner (initially 0).
Has been initialized to). Then, the representative vector W of the winner is updated like W ← W + g (X−W) as in the update process (S16) of FIG.

Next, a partial strain calculation process (S11-1)
As will be understood from the description of step S16-1, since the variable of the representative vector is almost proportional to the partial distortion of each representative vector, paying attention to this point, the representative vector already generated , The representative vector with the largest value of the variable is selected. And
Clear the variables of all representative vectors to 0.

Next, one new representative vector (or a predetermined number of two or more) having the same vector value as the selected representative vector is generated (S22).

According to this modification, since the cumbersome process of calculating the partial distortion of each representative vector is replaced with a simple process using variables, the processing amount is reduced and the processing time is shortened.

FIG. 5 shows a further modification of the process of FIG.

The difference from the process of FIG. 4 of this modification is that the initial input process (S11-1), the partial darkness calculation process (S21-2), and the newly added winning number determination process (S).
24).

In the first input processing (S11-1), in addition to the vector quantization number n, the end count L and the check count C,
Input the number of thresholds K for determining the representative vector generation.

In the partial strain calculation process (S11-2),
First, similar to the process of FIG. 4, the representative vector having the largest variable value is selected, then 1 is added to the number of wins of the selected representative vector, and the variables of all the representative vectors are cleared to 0.

Next, the process proceeds to a winning number determination process (S24), and it is determined whether or not the winning number of the selected representative vector has reached the threshold number K. As a result, if YES, the process proceeds to S22, where one or two of the same values as the selected representative vector
Generate more than one new representative vector, and clear the winning number of the selected representative vector to zero. On the other hand, if the result of step S24 is NO, the process returns to step S13 without generating a new representative vector.

According to this modification, with respect to the representative vector having the largest partial strain selected in the partial strain calculation process, when the large partial strain is such that it is eliminated by the repetition of the learning process, The generation of a new representative vector in the dominant region is avoided, and when the large partial distortion is difficult to eliminate by repeating the learning process, a new representative vector is generated there. As a result, the local strain is prevented from becoming too small locally by the generation of the new representative vector, and the equal strain condition can be more favorably achieved.

According to the embodiment of the present invention described above, in the method of obtaining representative vectors by competitive learning, competitive learning is started with the number of representative vectors being smaller than the number of vector quantization levels. It is possible to reflect various requirements for optimal vector quantization as a criterion for determining the representative vector value, and at the same time reduce the possibility of selecting a representative vector value that causes a large distortion error at the start of competitive learning. Is something that can be done.

Further, when the learning is repeated a certain number of times, each representative vector is changed so as to reduce the partial distortion, and a new representative vector is generated where the partial distortion in the dominant region is large. , The dominant region of the representative vector in the vicinity is narrowed and the partial distortion is reduced. Therefore, while realizing the equal distortion condition of equalizing the partial distortion, which is one of the necessary conditions for optimal vector quantization, The strain can be reduced.

[0059]

According to the present invention, a representative vector suitable for vector quantization that minimizes the expected value of the distortion error can be prepared. As a result, when compressing and storing or transmitting data by, for example, vector quantization, improve the data compression rate with the same quality, or improve the storage and transmission efficiency with the same compression rate. You can

[Brief description of drawings]

FIG. 1 is a block diagram showing a schematic configuration of a vector quantization device to which the present invention has been applied.

FIG. 2 is a flowchart showing a vector quantization process of this embodiment.

FIG. 3 is a flowchart showing a representative vector generation process of this embodiment.

4 is a flowchart showing a modified example of the processing of FIG.

FIG. 5 is a flowchart showing a further modified example of the processing of FIG.

FIG. 6 is a conceptual diagram of vector quantization using a two-dimensional vector as an example.

[Explanation of symbols]

 1 input device 2 arithmetic processing device 3 output device 6 input vector file 7 representative vector file

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification code Internal reference number FI Technical indication H04N 7/24 H04N 7/13 Z

Claims (6)

[Claims] 1. A method for generating a representative vector for vector quantization, including a representative vector initializing means for initially generating a smaller number of representative vectors than the number of representative vectors required for the vector quantization, and a large number of inputs. Select one input vector from the vectors,
Competitive learning in which one representative vector having the smallest distortion error with the selected input vector is selected as a winner from the already generated representative vectors, and the representative vector of this winner is updated so as to approach the selected input vector , A competitive learning means repeatedly performed, and each time the competitive learning is repeated a predetermined number of times,
From the already generated representative vector, select a representative vector with a relatively large partial distortion according to a predetermined criterion,
And a new representative vector generating means for generating one or more new representative vectors based on the selected representative vector. 2. The vector quantization method according to claim 1, wherein the new representative vector is generated with the same value as the selected representative vector. 3. The vector quantization method according to claim 1, wherein the predetermined criterion is to select a representative vector with the largest partial distortion among the already generated representative vectors. Vector quantization method. 4. The vector quantization method according to claim 1, wherein the competitive learning means has a representative vector of the winner and the selected input as variables of the representative vector selected as the winner in each competitive learning. A vector quantization method, wherein a distortion error from data is added, and the predetermined criterion is to select a representative vector having the largest variable among the already generated representative vectors. 5. The vector quantization method according to claim 1, wherein when the new representative vector generation means selects a representative vector according to the predetermined criterion, the selected number of times of the selected representative vector is a predetermined number. A vector quantization method characterized in that a new representative vector is generated based on the selected representative vector only when the threshold number is reached. 6. A method of generating a representative vector for vector quantization, including a representative vector initialization process for initially generating a number of representative vectors smaller than the number of representative vectors required for the vector quantization, and a large number of inputs. Select one input vector from the vectors,
From the already generated representative vectors, one representative vector having the smallest distortion error with the selected input vector is selected as the winner, and the representative vector of this winner is updated so as to be closer to the selected input vector. , Repeatedly performing a competitive learning process, and each time the competitive learning is repeated a predetermined number of times,
From the already generated representative vector, select a representative vector with a relatively large partial distortion according to a predetermined criterion,
And a new representative vector generation process for generating one or more new representative vectors based on the selected representative vector.