JPH06152430A - Method for adaptively quantizing vector and device therefor - Google Patents

Abstract

PURPOSE:To enable satisfactory vector quantization by adding update processing based on the increase and extinguishment of representative vectors to the successive update processing of representative vectors. CONSTITUTION:Input vectors are sent to a quantizing part 1 and a learning data preparing part 4, sampled for every fixed cycle by the preparing part 4 and sent to a distortion calculating part 5 as learning data. The calculating part 5 calculates the distortion of the representative vectors in a code book 3 and the inputted learning data, and the representative vector with minimum distortion is defined as amost adjacent vector. The ID of this vector is sent to the part 1 and a correcting part 6. Thus, the part 1 outputs quantized data. On the other hand, the part 6 counts the number of accumulated input vectors and when this number is coincident with a certain value, frequency for each representative vector is normalized by referring to a frequency table 7. At the time of non-coincidence, '1' is added to the frequency corresponding to the ID, and the representative vector corresponding to the most adjacent vector is corrected/updated and stored in the book 3.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an adaptive vector quantization method and apparatus for efficiently compressing and coding image / sound data.

[0002]

2. Description of the Related Art Vector quantization (vector quantization) is one of the methods for compressing and encoding image data and audio data.
There is quantization; VQ). This vector quantization is a process of approximating a continuous signal with a discrete signal. For example, in the case of audio data, this audio data is sampled at appropriate intervals to obtain P samples, and P
The samples are put together to form one input vector in the P-dimensional space. Then, an appropriate number of representative vectors that are closest to this input vector are selected in advance, the input vector is approximated by the selected representative vector, and a number (ID) representing the selected representative vector.
Therefore, it is a method of encoding an input vector (that is, an input signal). Vector quantization is a very important technique in the field of data compression and coding.

Formally, vector quantization is a mapping Q from an n-dimensional Euclidean interval R ⁿ to a subset M of R ⁿ :

[0004]

[Equation 1] Can be regarded as That is, by vector quantization, each element (that is, input vector) x _i εR ⁿ of the input vector set V = {x _i | i = 1,2, ...} becomes a finite vector set (that is, a set of all representative vectors). ) M = {m _j |
It is converted into one element m _j εR ⁿ with 1 ≦ j ≦ L}.
This is because, as shown in FIG. 1, the n-dimensional space Ω in which the input vector x exists is divided into L regions (cells) {S _j | 1 ≦ j.
≦ L divided into}, equivalent to allocate one by one representative vector m _j in each cell S _j. That is, each cell is defined as follows.

[0005]

[Equation 2] Here, the finite set M is called a codebook, L is the size or number of levels of the codebook, and m _j is a representative vector or codeword.

When the original signal (input vector) x is approximated by the representative vector m, distortion d (x, m) occurs due to the approximation. Although various methods of evaluating the strain d can be considered, usually,
L ² norm:

[0007]

[Equation 3] Calculated by Therefore, in vector quantization, the average distortion D over the input vector set V:

[0008]

[Equation 4] Representative vector set (codebook) to minimize
It is important to design M = {m _j | 1 ≦ j ≦ L}. Codebook design is also called the learning process in vector quantization.

The codebook design method includes a batch method and an online method. In the batch method, a codebook is designed so as to minimize the average distortion D, given a finite number of learning data (input data used for learning). At present, there is no analytical method for minimizing the average distortion D, but as a known algorithm that guarantees a minimal solution, the LBG algorithm (Y. Lind) is used.
e, A. Buzo and RMGray, "An algorithm for vector
quantizer design ", IEEE Trans. Communication, Vol.
COM-28, pp. 84-95, 1980). Since the learning data is fixed in the batch method, it is not possible to follow the distribution variation of the signal, and a large capacity memory for storing a large amount of learning data is required.

On the other hand, in the online method, the learning data is sequentially presented, and the codebook is updated each time (therefore, the number of learning data may be infinite in principle). Therefore, unlike the batch method, it is not necessary to store all the learning data, and the codebook is updated in real time to adapt to changes in the properties of the data, so it is a more practical method. . This method is called Adaptive Vector Quantization (A).
VQ).

As a known algorithm for designing a codebook in adaptive vector quantization, competitive learning (Co
There is a learning algorithm based on mpetitive learning (CL) (D. Rumelhart and J. McClelland, "Parallel
distributed processing: Explorations in the micros
tructure of cognition ", chap. 5, MIT Press, 198
6). The general procedure of the competitive learning algorithm is as follows. In the following description, the learning data presented at the t-th is represented by x (t), the representative vector before updating at that time is represented by m (t), and the representative vector after updating is represented by m (t + 1). To do.

(Codebook Design Algorithm in Adaptive Vector Quantization) Step 1 First, the codebook is initialized by initializing the L representative vectors m _j with appropriate random numbers. That is,

[0013]

[Equation 5] And Step 2 Determine the nearest neighbor vector m _c by the following equation.

[0014]

[Equation 6] Step 3 Update the nearest neighbor vector according to the following equation.

[0015]

[Equation 7] -Step 4 Set t ← t + 1 and return to Step 2.

Here, α (t) is called a learning rate, and in order to guarantee the convergence of the learning result, a function that decreases with increasing t is usually used, and a positive number sufficiently smaller than 1 is set as an initial value. To be done.

As can be seen from this algorithm, in the competitive learning algorithm, only the nearest neighbor vector is learned, and the other representative vectors have the previous values. In the initialized state, the arrangement of representative vectors is not optimized,
It means that some of the representative vectors are not the nearest neighbor vectors, and thus this representative vector is not learned. This will be described with reference to FIG. The existence range of the n-dimensional input vector, that is, the signal existence area X is localized in the n-dimensional space Ω, and further, seven representative vectors C _{1 to} C ₇ are generated by the initialization, and four of them are represented. Only the vectors C _{4 to} C ₇ are in the signal existence region X.
It is supposed to be placed inside. Then, the probability that the representative vectors C _{1 to} C ₃ outside the signal existing region X become the nearest neighbor vectors in the learning process is extremely small, so that the learning is not performed and does not contribute to the quantization. Therefore, the quantization is performed only by the _four representative vectors C _{4 to} C ₇ in the signal existing region X, and the seven representative vectors C ₁ to C ₁
The average distortion becomes large as compared with the case where the quantization is performed using all of C ₇ .

Recently, in order to solve the problem that the average distortion in the competitive learning algorithm becomes large, a method of adding a "conscience mechanism" (CM) to the competitive learning algorithm has been proposed. CM is a kind of fine method, and is a mechanism that prevents a certain representative vector from becoming the nearest neighbor vector frequently. Several concrete methods of realizing CM have been proposed, and as a representative of them, Frequency Sensitive Competitive Learning Algoru
thm; FSCL (Frequency Sensitive Competitive Learning) algorithm (S.
C. Ahalt, AK Krishnamurthy, P. Chen and DE M
elton, "Competitive learning algorithms for vector
quantizaition ", Neural Networks, Vol.3, PP. 227-2
90, 1990). This algorithm modifies equation (5) in step 2 of the competitive learning algorithm described above as follows.

[0019]

[Equation 8] Where f _i (t) is the input (learning) vector x (0), ..., x
In the learning process by (t), it represents the frequency at which the representative vector m _i (t) becomes the nearest neighbor vector. Competitive learning method [expression
In (5)], the representative vector having the smallest distortion with the input vector x (t) is selected as the nearest neighbor vector.
In the FSCL method [Equation (7)], even a representative vector whose distortion is not minimum can be a nearest neighbor vector if the frequency is relatively small. Therefore, by adding CM, even a representative vector initialized in a signal-free area can be gradually attracted to the signal area in the learning process. As a result, all the representative vectors contribute to the quantization, and good quantization can be realized.

[0020]

However, the above-mentioned FS
The CM represented by the CL algorithm has a plurality of signal existing regions of the learning data in the n-dimensional Euclidean space where the learning data (input vector) exists, and these signal existing regions are arranged to be separated from each other. In some cases, there are limits. FIG. 3 is a diagram for explaining this limit, and shows a case where the input vector is two-dimensional for simplification.

In the figure, it is assumed that the learning data is uniformly generated in two square areas A and B in the two-dimensional space. These areas A and B are sufficiently separated. Here, the ten representative vectors C _{1 to} C ₁₀ are initialized and the area A
Two representative vectors C ₁ and C ₂ are arranged in the area B, and eight representative vectors C _{3 to} C ₁₀ are arranged in the area B. The necessary condition for minimizing the average distortion is that the same number (5) of representative vectors is assigned to each of the two areas A and B, assuming that the two areas A and B are uniformly distributed. . Therefore, in the learning process, three representative vectors must move from the area B to the area A.

However, in order for the representative vector to move from the area B to the area A, the representative vector in the area B needs to be the nearest vector to the signal x generated in the area A. This is because in the above formula (7), for all the representative vectors (m _{i, A} ) in the region A, the representative vectors (m _{j, B} ) satisfying the following formula are in the region B. Means that it must exist.

[0023]

[Equation 9] If there is no movement of the representative vector between the two areas A and B, the frequency ratio: f _{i, A} / f _{j, B} is the number of representative vectors in each area in the learning process from the assumption of uniform distribution. A reciprocal ratio of f _{i, A} / f _{j, B} ≈8 / 2 = 4. Therefore, the expression (8) is satisfied by d (x, m _{i, A} ) and d
Determined only by (x, m _{j, B} ). When the distance between the areas A and B is sufficiently larger than the size of the areas A and B, d (x, m _i ,
_{Since A} ) >> d (x, m _{j, B} ), it is impossible in principle to move the representative vector between regions in CM. Further, as can be easily inferred from the above consideration, even when the distance between the regions is relatively small, the number of movements of the representative vector between the regions where the learning data exists depends on the initial arrangement thereof, and therefore in CM, The initial placement of the representative vector greatly affects the quantization performance.

If all the learning data are given in advance, it is possible to obtain the optimum value of the initial arrangement number of representative vectors for each region by estimating the probability distribution of the learning data by some method. However, in the adaptive quantization vector, since the learning data is sequentially given,
The distribution of the learning data cannot be estimated in advance, and it is impossible to perform such ideal initial arrangement.

When handling a high-dimensional feature vector such as voice or image, the signal is localized in several regions in the high-dimensional vector space, regardless of the distribution of the feature vector. Is empirically known. Therefore, the above-mentioned problems in CM cannot be ignored in practical use.

According to the present invention, when designing a codebook in adaptive vector quantization, even if the training data is localized in several areas, it is possible to use a codeword to which an arbitrary initial value is assigned. Another object of the present invention is to provide an adaptive vector quantization method and apparatus which can reduce average distortion as much as possible.

[0027]

According to the adaptive vector quantization method of the present invention, a set consisting of a finite number of representative vectors is prepared in advance, and for each of the input vectors sequentially obtained as a time series, A representative vector having the smallest distortion of the set is selected from the set as a nearest neighbor vector, and a correction step of sequentially correcting the representative vector in the set based on the nearest neighbor vector and the distortion. In the adaptive vector quantization method having, in the selecting step, for each of the representative vectors, the frequency selected as the nearest neighbor vector is measured, and in the correcting step, the vector is calculated for each of the representative vectors. A first step of calculating the fitness in accordance with the selected frequency and stochastically generating a copy of the representative vector in accordance with the fitness; A second step of adding a minute random vector to each of the duplicates generated in the step to make a new representative vector, and making each of the new representative vectors the new set is added. The step 2 is executed when the presented cumulative number of input vectors reaches a predetermined number.

The adaptive vector quantizing device of the present invention has a set storing means for holding a set of a finite number of representative vectors, and input vectors are sequentially input as a time series so that distortion with the input vector is minimized. Selecting means for selecting a representative vector from the set as a nearest neighbor vector,
In an adaptive vector quantizer having a quantizing means for outputting a quantization result based on the nearest neighbor vector, and a modifying means for successively modifying a representative vector in the set based on the nearest neighbor vector and the distortion. For each of the representative vectors, a frequency storage unit that stores the frequency selected as the nearest neighbor vector is provided, and when the correction unit has reached a predetermined number of accumulated presentation numbers of the input vector, The fitness is calculated for each of the representative vectors by referring to the frequency storage means, the fitness is calculated according to the frequency, and a copy of the representative vector is stochastically generated according to the fitness, and each of the generated copies is generated. To a new representative vector by adding a minute random vector to the set storage means. It is intended to update the retained set.

[0029]

In the adaptive vector quantization method of the present invention, the update process based on "proliferation (generation)" and "disappearance" of the representative vector is added to the sequential update process of the representative vector in the conventional competitive learning algorithm. Is. That is, in the learning process, the frequency at which each representative vector becomes the nearest neighbor vector is stored, and when the total number of learning data (input vectors) reaches a certain value (T), the relative magnitude relation of the frequencies is shown. Based on this, "reproduction" and "disappearance" are performed to create a duplicate of each representative vector, and at the same time, a random vector with a small norm is added to the obtained duplicate representative vector. For a representative vector with a high frequency, a plurality of replicas corresponding to the frequency are created, and for a representative vector with a low frequency, no replica is generated. By introducing such "growth" and "disappearance" processes, even if the existence region of the input vector is localized in multiple spatial regions, as described below, any initialization is performed. This representative vector can be updated so that the average distortion can be made as stable and small as possible.

In this case, the representative vector copy creation process is performed stochastically. Also, in the normal case, the total number of representative vectors should not be changed before and after the duplication, and for the rearranged representative vectors, the conventional sequential update processing should be performed until the accumulated total number of input learning data reaches a certain value. Apply. Then, when the above-mentioned certain value is reached, the updating process based on multiplication / disappearance of the present invention is performed. After that, the above processing is repeatedly executed.

The present invention will be described in more detail below.

In general, the arbitrarily-represented representative vector is updated only when it becomes the nearest vector in the learning process. When the distribution of the input vector (signal distribution) stays within only one continuous region, each representative vector asymptotically approaches the optimal arrangement by sufficiently learning as described above. At this time, from the representative vector side,
Since a larger number (small number) of representative vectors are arranged in the dense (sparse) part of the signal distribution, it is considered that each representative vector becomes the nearest neighbor vector with equal frequency.

By the way, when the signal distribution is localized in a plurality of regions at distant locations, the number of representative vectors in each region is biased. As described above, it is difficult for the conventional algorithm to eliminate this bias.
On the other hand, in the present invention, when the cumulative total number of input vectors reaches a certain number, the representative vector is duplicated based on the relative magnitude relation of the frequencies by the multiplication / disappearance process.
When a representative vector becomes the nearest neighbor vector relatively frequently (rarely), a larger number (small number) of duplicates will be created in the vicinity of the representative vector. From a macro point of view, this replication process is equivalent to a large movement of the representative vector in the signal space instantaneously, and the representative vector is moved to the optimum position regardless of whether the signal distribution is continuous or discontinuous. It becomes possible.

In the present invention, the conventional sequential update process is applied to the rearranged representative vector until the cumulative total number of input learning data reaches a certain value, and the cumulative number of input data reaches a certain value. Since the update process based on the multiplication / disappearance is reached again, and this process is repeated thereafter, the variance of the frequency of being selected as the nearest neighbor vector to the representative vector gradually increases while the process is continuously executed. It becomes smaller, and the representative vector approaches the optimal arrangement.

In general, optimum quantization regardless of whether the distribution is continuous or discontinuous is equivalent to the fact that each representative vector becomes the nearest neighbor vector with equal frequency. Therefore, by performing the processing shown in the present invention, optimum or sufficiently close quantization is achieved.

[0036]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, embodiments of the present invention will be described with reference to the drawings. FIG. 4 is a block diagram showing the configuration of an adaptive vector quantization device according to an exemplary embodiment of the present invention. This adaptive vector quantization device includes a quantization unit 1 that quantizes an input vector that is input and outputs a result, a representative vector holding unit 2 that includes a codebook 3, a sample of the input vector that has been input at regular intervals, and as learning data. It is composed of a learning data creation unit 4 for outputting, a distortion calculation unit 5, and a correction unit 6 incorporating a frequency table 7. Codebook 3 here
Holds a predetermined number L of representative vectors, and the configuration thereof is as shown in FIG. That is, for each representative vector, its ID (identification number) and the vector value of that vector (the coordinate value of the vector, for example, n
(A _i , b _i , ..., N _i )) and are stored for the dimensional representative vector m _i . Further, the distortion calculation unit 5 calculates the distortion (for example, distance) between the input vector from the learning data creation unit 4 and each of the representative vectors stored in the codebook 3, and sets the representative vector with the smallest distortion as the nearest neighbor vector. , The ID of the vector is sent to the correction unit 6 and the quantization unit 1.

As shown in FIG. 6, the frequency table 7 stores the ID of each representative vector and the frequency selected as the nearest neighbor vector. When the ID of the representative vector is sent from the distortion calculation unit 5, the correction unit 6 counts (accumulates) the number of input vectors and adds 1 to the frequency corresponding to the ID in the frequency table 7. Is configured. Then, when the number of inputs reaches a certain value, the correction unit 6 calculates the fitness for each representative vector according to the frequency stored in the frequency table 7, and generates a duplicate of each representative vector according to the fitness. , A random vector with a small norm is added to the duplicated representative vector, the codebook 3 is rewritten with the obtained representative vector group, the frequency table 7 is initialized (each frequency is set to 0), and the duplicate is generated next time. The certain value is rewritten to the cumulative number (of the input vector) to be processed. Further, when the cumulative number of input vectors is other than the certain value, the correction unit 6 determines the representative vector that has become the nearest neighbor vector according to the conventional method [Equation (6) in the “Conventional Technology” column]. It is configured to update. Here, the fitness indicates a relative magnitude relationship between frequencies.

Next, the operation of this embodiment will be described with reference to the flowchart of FIG. This operation is based on the method of the present invention.

In the codebook 3, a predetermined number (for example, L) of representative vectors are initialized by random numbers and stored.

When an input vector is input, the input vector is sent to the quantizer 1 and the learning data generator 4,
The learning data creation unit 4 samples the data at regular intervals and sends it to the distortion calculation unit 5 as learning data x (t). The distortion calculation unit 5 uses the representative vector m _i (1
≦ i ≦ L) and the distortion of the input learning data x (t) are calculated, for example, by the method described in the formula (5) in the “Prior art” section, and the representative vector having the smallest distortion is It is determined as the side vector m _c (step 101),
The ID of the representative vector is sent to the quantization unit 1 and the correction unit 6. The quantizer 1 uses the sent ID to
Output the quantized data to the outside.

On the other hand, the correction section 6 counts the cumulative number t of input vectors, and first checks whether t matches a certain value T (step 102). When t ≠ T, 1 is added to the corresponding frequency f _c corresponding to the input ID in the frequency table 7 (step 103), and the method described in Formula (6) in “Prior Art” is used. The representative vector corresponding to the nearest neighbor vector m _c is modified / updated and stored in the codebook 3 (step 104).

On the other hand, if t = T in step 102, the correction unit 6 refers to the frequency table 7 and normalizes the frequency f _i (1 ≦ i ≦ L) for each representative vector (step 105). ). This normalization is performed according to the following equation.

[0043]

[Equation 10] Then, the correction unit 6 sets the interval [0, 1] consisting of real numbers to the normalization frequency obtained in step 105.

[0044]

[Outer 1] To divide. As a result, the section [0,1] has L subsections (s ₁ , s ₂ , ..., L) whose length is proportional to the frequency f _i .
s ₃ ). Then initialize the replication number (and g _i) for each representative vector (g _i = 0, _i =
1, ..., L). Then, L uniform random numbers are generated for the interval [0, 1], and for each random number (∈ [0, 1]),
It is determined to which of the above-mentioned partial sections (s _i = 0, i = 1, ..., L) belong. For example, if it belongs to s _j , 1 is added to the number of copies corresponding to the section (g _i ← g _i +
1). As a result, the number of representative vectors m _i after duplication is obtained in proportion to the frequency f _i .
clearly

[0045]

[Equation 11] Therefore, the total number of representative vectors after replication is the same as before replication. Then, for each representative vector m _i , the representative vectors of the duplications g _{i corresponding} to the representative vector m _i are created (step 106). At this point, the duplicate representative vector and the corresponding original representative vector are the same.

Next, for each duplicated representative vector,
A small norm random vector is added by the following formula to set a new representative vector (step 108). Here, ε _i (t) εR ⁿ is a small random vector of norm.

[0047]

[Equation 12]By the above steps 105 to 107,
In the vicinity of the table vector, the copy proportional to the frequency is stochastic.
The newly created duplicate m ^* _iSo
And update the entire codebook 3 (step 10
8).

Then, the correction section 6 initializes the frequency table 7 for all the representative vectors (step 109), substitutes 2T for the above-mentioned value T (step 110), and ends the processing. By substituting 2T for T, the above steps 105 to 110 are performed again when the cumulative number of input vectors reaches 2T.

Here, the duplication process will be described with reference to an actual example. Here, the process in the case where the number of representative vectors is 7 and the number of inputs of learning data is 100 and the fitness is calculated from the frequency to make a duplicate will be described. For example, at this point, seven representative vectors m _i (1 ≦ i ≦ 7)
It is assumed that the respective frequencies f _{i of the above} are shown in FIG. The new representative vector m ^* _{i for this} is shown in FIG.
Shown in (B). No replication is created for the representative vectors m ₃ , m ₄ , m ₇ whose frequency is 0, and the frequency is 40.
It can be seen that three replicas are created for the representative vector m ₁ which has the largest number of times.

Next, the result of comparison between the case of using the method of the present embodiment and the case of using the conventional method in the learning process of the representative vector for the same learning data will be described. FIG. 9 shows the learning process of the representative vector according to the present embodiment according to the cumulative number of input data, and FIG. 10 shows the same process for the conventional method. In the figure, each corner point represents a representative vector, and t represents the cumulative total number of input data. Here, the learning data is generated in a uniform distribution in two square regions A and B having the same area in the two-dimensional space, and the total number (L) of representative vectors is 18
I made it into an individual. Then, under the initial condition, each representative vector was set at a position in the area B on the lower side of the drawing. Therefore, in this case, in order to achieve optimum quantization, nine representative vectors should be arranged in each of the two areas A and B (from the lower area B in the figure to the upper area A in the figure,
The movement of individual representative vectors is a necessary condition.

In the conventional method shown in FIG. 10, t = 400.
After that, there is no movement of the representative vector between areas, and only movement (learning) within each area. The number of representative vectors between the regions is fixed at 5 and 13. On the other hand, in FIG. 9 according to the present embodiment, first, a certain representative vector moves to the upper area in the drawing (for example, from t = 20 to t).
= 80), a "duplicate" is generated in the upper area (eg t
= 300), and after that, "proliferation" and "disappearance" are repeated several times, and then nine representative vectors are placed in both areas, and finally the desired quantization is realized. There is.

[0052]

As described above, according to the present invention, the updating process based on the "proliferation (generation)" and "disappearance" of the representative vector is added to the sequential updating process of the representative vector in the conventional algorithm. , Even if the probability distribution of the input learning data is localized in multiple regions, the learning result is almost independent of the initial value, and the optimal or sufficiently close learning result is obtained. There is an effect that vector quantization can be realized.

[Brief description of drawings]

FIG. 1 is a schematic diagram for explaining the principle of vector quantization.

FIG. 2 is a diagram showing a relationship between a signal existing area and a representative vector.

FIG. 3 is a diagram showing a relationship between a region and a representative vector when there are two regions where a signal exists.

FIG. 4 is a block diagram showing a configuration of an adaptive vector quantization device according to an exemplary embodiment of the present invention.

FIG. 5 is a diagram showing a configuration example of a codebook.

FIG. 6 is a diagram showing a configuration example of a frequency table.

FIG. 7 is a flowchart illustrating the operation of the apparatus of FIG.

FIG. 8 is a diagram illustrating a replication process of a representative vector.

FIG. 9 is a diagram showing a situation of adaptation of a representative vector according to the method of the present invention.

FIG. 10 is a diagram showing a situation of adaptation of a representative vector by a conventional method.

[Explanation of symbols]

1 quantization part 2 representative vector holding part 3 codebook 4 learning data creation part 5 distortion calculation part 6 correction part 7 frequency table 101-110 steps A, B area C _i , m _i representative vector f _i frequency ε _i random vector Ω n-dimensional space

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁵ Identification code Office reference number FI technical display location H04N 7/13 Z

Claims (3)

[Claims] 1. A set of a finite number of representative vectors is prepared in advance, and for each of the input vectors that are sequentially obtained as a time series, a representative vector that minimizes distortion with the input vector is selected from the set. In an adaptive vector quantization method having a selecting step of selecting a nearest neighbor vector, and a modifying step of sequentially modifying a representative vector in the set based on the nearest neighbor vector and the distortion, in the selecting step, For each of the representative vectors, the frequency selected as the nearest neighbor vector is measured, and in the correction step, the fitness is calculated according to the frequency with which the vector is selected for each of the representative vectors. A first step of stochastically generating a copy of the representative vector according to the fitness, and a minute lander for each of the copies generated in the first step. A second step of adding the vectors to make a new representative vector and making each of the new representative vectors into a new set, the first and second steps being the accumulated cumulative number of input vectors. The adaptive vector quantization method is characterized in that the adaptive vector quantization method is executed when a predetermined number is reached. 2. The number of copies generated for each representative vector in the first step is probabilistically processed to a number proportional to the frequency with which the representative vector is selected.
The adaptive vector quantization method according to claim 1, wherein the total number of the duplicates is equal to the number of representative vectors before the first step. 3. A set storage means for holding a set consisting of a finite number of representative vectors, and a representative vector from which the input vector is sequentially input as a time series and distortion with the input vector is minimized. Selecting means for selecting the nearest neighbor vector, quantizing means for outputting a quantization result based on the nearest neighbor vector, and sequentially modifying the representative vector in the set based on the nearest neighbor vector and the distortion In the adaptive vector quantization device having a correction means for performing, for each of the respective representative vectors, a frequency storage means for storing the frequency selected as the nearest neighbor vector, and the correction means, When the number of accumulated presentations reaches a predetermined number, the frequency storage means is referred to and an appropriate value is selected for each of the representative vectors according to the selected frequency. Degree is calculated and a replica of the representative vector is stochastically generated according to the fitness, and a small random vector is added to each of the generated replicas to obtain a new representative vector. An adaptive vector quantizing device for updating the set held in the set storage means as a new set.