JP2002232296A - Method and device for generating code book for vector quantization - Google Patents

Abstract

PROBLEM TO BE SOLVED: To select the same representative vector as full search at high speed only by searching for the representative vectors of one part. SOLUTION: By repeatedly executing a sorting step (F102) for sorting representative vectors on the basis of the values of representative vectors, first arithmetic steps (F105 and F108) for calculating the distortions of input vector and representative vector and calculating the address of the representative vector to be next searched, decision steps (F106 and F109) for comparing two distortions provided by the first arithmetic steps and deciding the candidates of the representative vector to be encoded and second arithmetic step (F111) for calculating centroid/average distortion of representative vectors on the basis of the result decided by the decision step, the code book of vector quantization is generated.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a vector quantization technique, and more particularly to an apparatus and a method for generating a codebook for vector quantization in pattern recognition, computational geometry, image signals, and audio signals.

[0002]

2. Description of the Related Art Vector quantization is a technique for encoding an input space with a finite number of representative vectors. Examples of application of vector quantization include feature extraction of pattern recognition, post-placement problem of computational geometry, information compression, quantization of images and sounds, and the like.

FIG. 34 shows a conceptual diagram of vector quantization.

[0004] Vector quantization has two aspects: (1) encoding an input vector and (2) generating a codebook.

In the procedure of (1) encoding an input vector in vector quantization, a codebook W = Ｗw composed of N representative vectors (sometimes called a code vector) for an input vector x (1050). _i ∈R ^K : i =
The closest representative vector w _min is selected from 0,..., N−1｝ (1040) (1060), and the index min of the representative vector or the value of the representative vector is output (1070).

[0006] At that time, as a measure of when choosing the closest representative vectors, non-negative strain measure d (x, w _i) to use. That is, the expression for outputting the number of the representative vector can be expressed by the following expression (1).

[0007]

(Equation 1)

An expression for outputting the value of the representative vector can be expressed by the following expression (2).

[0009]

(Equation 2)

Further, in (2) codebook generation of vector quantization, an initial codebook (1010) is set, and an input vector set X = (x
^(T) : t = 0,..., T−1) (1030), and iteratively optimizes (1020) by repeatedly modifying the value of the representative vector so as to match the input vector set (1020). Obtain the optimal codebook (1040). The input vector for codebook generation may be different from the input vector input in (1) encoding of the input vector.

The evaluation of the codebook is performed using an expected value (average distortion) of distortion generated when a given set of input vectors is encoded with a corresponding representative vector. This average distortion is a non-convex function and has many local minima. The optimal codebook is the codebook that minimizes the average distortion. That is, a vector quantizer having an optimal codebook can perform encoding with the least distortion. The average strain is defined by the following equation (3).

[0012]

(Equation 3)

Where δ _ti is the Kronecker delta, which is 1 when the input vector x _t is encoded into the representative vector w _i , and 0 otherwise.

As an algorithm for generating a codebook, an adaptive correction type competitive learning algorithm and a batch correction type LBG algorithm [Linde, Y., Buzo, A., & Gray, R.
M .. “An algorithm for vector quantizer design.” IEE
E Transactions on Communications, 28, pp. 84-95, 198
0.] is well known.

[Ueda, N., & Nakano, R .. "A new co
mpetitive learning approach based on an equidistor
tion principle for designing optimal vector quanti
According to zers. "Neural Networks, 7, pp.1211-1227, 1994.], the competitive learning algorithm is an adaptation of the LBG algorithm. Here, only the codebook generation method by the LBG algorithm is taken up. It is also applicable to adaptive competitive learning algorithms.

The LBG algorithm is an algorithm that repeats the following steps (a) and (b) until a convergence condition is satisfied.

[0017] (a) for the input vector set of the T code book for generating, searching the closest representative vectors w _min in each input vector x ^(t) from the current codebook is encoded The representative vector number min is stored.

(B) Using the set of input vectors and the number of the representative vector into which each input vector is encoded, calculate the center of gravity of the representative vector and update the codebook.

Step (a) of the LBG algorithm, which is a codebook generation method, is a step of searching for a representative vector having the smallest distortion for each input vector and encoding the representative vector. This encoding step is very time-consuming because it is based on a full search that calculates the distortion with all the representative vectors and selects the closest representative vector.
Since this operation needs to be performed for all input vectors, it takes an enormous amount of time.

As one of the techniques for solving this problem, there is a vector quantization method and a vector quantization apparatus disclosed in Japanese Patent Application Laid-Open No. 6-45946. This vector quantization technique is described in the literature [Keiichi Nakano, Hironori Kasahara: "Fast Vector Quantization Using Sorted Codebooks (Fast Vecto
r Quantization Using Sorted Codebook: VQ-SC) ”Transactions of the Institute of Electronics, Information and Communication Engineers D-II Vol.J77-D-II No.10 pp.198
4-1992, 1994] and the VQ-SC
Is roughly the following algorithm:

(1) The representative vectors are sorted for each dimension, and the order is stored.

(2) A flag indicating whether or not the searched representative vector has been cleared is cleared.

(3) Calculate the strain that is the lower limit of the strain.

(4) On the basis of the principle of the branch and bound method, the search for a representative vector that does not improve the accuracy of the solution by using the upper limit of the distortion is omitted. The searched representative vector sets a flag. When the search for the representative vector used for the lower limit of the distortion is omitted, the lower limit of the distortion is updated. The search is continued until the upper limit of the strain and the lower limit of the strain match. The method of determining the next representative vector to be searched is heuristic.

By using the upper limit value of strain and the lower limit value of strain, if a part of the codebook is examined, the same representative vector as in the case of full search can be searched at high speed.

[0026]

However, the above method has the following problems.

(1) It is necessary to sort the representative vectors for each dimension. The higher the dimension, the longer the time required for sorting.

(2) It is necessary to store the order of the representative vectors sorted for each dimension, and the memory capacity to store becomes enormous.

(3) It is necessary to clear a flag indicating which representative vector has been searched for each input vector, and if there are many input vectors, the time for clearing will be enormous. Also, a memory capacity for flags is required for the number of representative vectors.

(4) Since the next representative vector to be searched may be determined heuristically, it is necessary to consider an efficient determination method.

Accordingly, it is an object of the present invention to provide a codebook generating method and a codebook generating apparatus for vector quantization, which can solve the above-mentioned problems.

[0032]

Here, for a K-dimensional input vector x, a representative vector w having the minimum distortion is set.
When searching for a _min, for a dimension k, it is not necessary to search for the representative vector w _i such that the following equation (4).

[0033]

(Equation 4)

However, equation (4) shows that the encoded w
_This is a condition when _min is obtained, and the equation (4) cannot be calculated. In addition, the values of the representative vectors for the k-dimension are not arranged regularly, and the order is in a random state, so that it is not possible to determine the search termination condition.

Therefore, the present invention employs the following method.

(1) The smallest value found in the search so far
Let the representative vector be w^cur _minAnd That is, search
Is completed, the representative vector w finally obtained^cur
_minIs the minimum strain d_min = D^cur _minAlso
One representative vector w_min = W ^cur _minBecomes

(2) Before starting the search, the representative vectors are sorted in ascending order (or descending order) according to the k-dimensional value. If there is a representative vector that satisfies the expression (4), it is sorted, so that there is no possibility that the further representative vector will become the representative vector having the minimum distortion. That is, by sorting the representative vectors, the search efficiency can be greatly increased.

Considering the calculation of the distortion in equation (4), the range in which the value of the representative vector in each dimension can be generally different. Therefore, choosing the one with the widest range should reduce the number of searches. In the gradation conversion of an image, R, G, and B take values of 0 to 255 in any dimension. Therefore, it is expected that the same number of times will be obtained even if any dimension is selected.

The method of terminating the search is determined by sorting, but it is desirable to terminate the search with a minimum number of times. For that purpose, it is important to determine the minimum representative vector w ^cur _min set first. It is expected that the search efficiency will greatly change depending on how to determine the initial representative vector.

For the method of determining the initial representative vector to be searched for in two cases, it is considered that a method of setting as follows is an efficient choice.

When designing a vector quantizer using a design algorithm such as the LPG algorithm, since the same set of input data is used, the cycle immediately before convergence is more likely to be encoded into the same representative vector as the previous cycle. . Therefore, the representative vector encoded in the previous cycle is set as an initial value.

When an input vector is encoded into a predetermined representative vector, a representative vector obtained by encoding adjacent input data is generally used as an initial value.

That is, in the method of generating a codebook for vector quantization according to the present invention, a sorting step of sorting the representative vectors based on the values of the representative vectors, a distortion of the input vector and the representative vector are calculated, and then a search is performed. A first operation step of calculating an address of a representative vector, a determination step of comparing two distortions obtained in the first operation step to determine a candidate of a representative vector to be encoded, and a determination result by the determination step And a second calculation step of calculating the center of gravity of the representative vector and calculating the average distortion based on the above.

Further, in the method of generating a codebook for vector quantization according to the present invention, a sorting step of sorting the representative vectors based on the values of the representative vectors, and calculating a distortion between the input vector and the plurality of representative vectors in parallel. A first calculating step of calculating an address of a next representative vector to be searched, a determining step of comparing a plurality of distortions obtained by the first calculating step to determine a candidate of a representative vector to be encoded, A second calculation step of calculating the center of gravity of the representative vector and calculating the average distortion based on the determination result of the determination step, wherein the above steps are repeated.

In the method of generating a codebook for vector quantization according to the present invention, a sorting step for sorting representative vectors based on the values of the representative vectors, a distortion of the input vectors and the representative vectors are calculated, and then a search is performed. A first operation step of calculating an address of a representative vector, a determination step of comparing two distortions obtained in the first operation step to determine a candidate of a representative vector to be encoded, and a determination result by the determination step A second calculation step of performing a center of gravity calculation, an average strain calculation, and a partial strain calculation based on the above, a partial strain sorting step of sorting on the basis of the partial strain obtained in the second calculation step, and a partial strain sorting step. A comparison method for comparing the difference between two sorted partial strains And a representative vector moving step of moving a representative vector having a small partial distortion to a vicinity of a representative vector having a large partial distortion based on the comparison result in the comparing step, and repeatedly performing each of the above steps. And

Further, in the method for generating a codebook for vector quantization according to the present invention, a sorting step for sorting representative vectors based on the values of representative vectors, and a distortion between an input vector and a plurality of representative vectors are calculated in parallel. A first operation step of calculating an address of a next representative vector to be searched, a determining step of comparing a plurality of distortions obtained in the first operation step to determine a candidate of a representative vector to be encoded, A second operation step of performing center-of-gravity calculation / average strain calculation / partial strain calculation based on a result of the determination step; a partial strain sorting step of sorting based on the partial strain obtained by the second operation step; The difference between the two partial strains sorted by the partial strain sorting step A comparing step of compare, on the basis of the comparison result of the comparing step, and a representative vector moving step of moving a small representative vectors of partial distortion in the vicinity of large representative vector strain portion,
Each of the above steps is repeatedly performed.

Further, in the method of generating a codebook for vector quantization according to the present invention, a sorting step of sorting representative vectors based on the values of representative vectors, a distortion of input vectors and representative vectors are calculated, and then a search is performed. A first operation step of calculating an address of a representative vector, a determination step of comparing two distortions obtained in the first operation step to determine a candidate of a representative vector to be encoded, and a determination result by the determination step A second calculation step of performing a center of gravity calculation, an average strain calculation, and a partial strain calculation based on the above, a partial strain sorting step of sorting on the basis of the partial strain obtained in the second calculation step, and a partial strain sorting step. Comparison comparing the disparity of sorted partial strains And a representative vector moving step of moving a plurality of representative vectors having small partial distortions to a vicinity of the representative vector having large partial distortion based on the comparison result in the comparing step, and repeatedly performing each of the above steps. Features.

Further, in the codebook generating method for vector quantization according to the present invention, a sorting step of sorting the representative vectors based on the values of the representative vectors, and calculating a distortion between the input vector and the plurality of representative vectors in parallel. A first operation step of calculating an address of a next representative vector to be searched, a determining step of comparing a plurality of distortions obtained in the first operation step to determine a candidate of a representative vector to be encoded, A second operation step of performing center-of-gravity calculation / average strain calculation / partial strain calculation based on a result of the determination step; a partial strain sorting step of sorting based on the partial strain obtained by the second operation step; Disparity of multiple partial strains sorted by partial strain sorting step A comparing step of comparing, and a representative vector moving step of moving a plurality of representative vectors having a small partial distortion to a vicinity of a representative vector having a large partial distortion based on the comparison result in the comparing step. It is characterized by performing.

Further, in the codebook generating method for vector quantization according to the present invention, a sorting step of sorting the representative vectors based on the values of the representative vectors, a distortion of the input vectors and the representative vectors of the codebook are calculated. A first calculating step of calculating an address of a representative vector to be searched for; a determining step of comparing two distortions obtained in the first calculating step to determine a candidate of a representative vector to be encoded; And a second operation step of adaptively calculating a representative vector and calculating an average distortion based on the determination result of the step, wherein the above steps are repeatedly performed.

Further, in the codebook generating method for vector quantization according to the present invention, the sorting step of sorting the representative vectors based on the values of the representative vectors and the distortion of the input vector and the plurality of representative vectors of the codebook are performed in parallel. A first operation step of calculating the address of a representative vector to be searched next, and a plurality of distortions obtained in the first operation step are compared to determine a candidate of a representative vector to be encoded. And a second operation step of adaptively calculating a representative vector and calculating an average distortion based on a result of the determination in the determination step, wherein the steps are repeatedly performed.

Further, in the codebook generating method for vector quantization according to the present invention, a sorting step of sorting the representative vectors based on the values of the representative vectors, a distortion between the input vectors and the representative vectors of the codebook are calculated. A first calculating step of calculating an address of a representative vector to be searched for; a determining step of comparing two distortions obtained in the first calculating step to determine a candidate of a representative vector to be encoded; A second operation step of performing adaptive representative vector calculation / average distortion calculation / partial distortion calculation based on the determination result of the step; and sorting based on the partial distortion obtained in the second operation step. A partial strain sorting step, and 2 sorted by the partial strain sorting step. And a representative vector moving step of moving a representative vector having a small partial strain to a vicinity of a representative vector having a large partial strain based on the comparison result in the comparing step. Each step is repeatedly performed.

Further, in the method of generating a codebook for vector quantization according to the present invention, a sorting step of sorting representative vectors based on the values of representative vectors, and a distortion between an input vector and a plurality of representative vectors of the codebook are performed in parallel. A first operation step of calculating the address of a representative vector to be searched next, and a plurality of distortions obtained in the first operation step are compared to determine a candidate of a representative vector to be encoded. A second operation step of performing adaptive representative vector calculation / average distortion calculation / partial distortion calculation based on a result of the determination in the above determination step; and a partial distortion obtained in the second operation step. By the partial strain sorting step of sorting on the basis of A comparing step of comparing the difference between the two divided partial strains, and a representative vector moving step of moving a representative vector having a small partial strain to a vicinity of a representative vector having a large partial strain based on the comparison result in the comparing step. And the above steps are repeatedly performed.

Further, in the method of generating a codebook for vector quantization according to the present invention, a sorting step for sorting the representative vectors based on the values of the representative vectors, a distortion between the input vectors and the representative vectors of the codebook are calculated. A first calculating step of calculating an address of a representative vector to be searched for; a determining step of comparing two distortions obtained in the first calculating step to determine a candidate of a representative vector to be encoded; A second operation step of performing adaptive representative vector calculation / average distortion calculation / partial distortion calculation based on the determination result of the step; and sorting based on the partial distortion obtained in the second operation step. A partial strain sorting step, and a compound sorted by the partial strain sorting step. And a representative vector moving step of moving a plurality of representative vectors having a small partial strain to a vicinity of a representative vector having a large partial strain based on the comparison result in the comparing step. Then, each of the above steps is repeatedly performed.

Further, in the codebook generating method for vector quantization according to the present invention, the sorting step of sorting the representative vectors based on the values of the representative vectors and the distortion of the input vector and the plurality of representative vectors of the codebook are performed in parallel. A first operation step of calculating the address of a representative vector to be searched next, and a plurality of distortions obtained in the first operation step are compared to determine a candidate of a representative vector to be encoded. A second operation step of performing adaptive representative vector calculation / average distortion calculation / partial distortion calculation based on a result of the determination in the above determination step; and a partial distortion obtained in the second operation step. A partial strain sorting step of sorting based on A comparing step of comparing the disparities of the plurality of sorted partial strains, and a representative vector movement for moving the plurality of representative vectors having a small partial strain to the vicinity of the representative vector having a large partial strain based on the comparison result in the comparing step And repeating the above steps.

The vector quantization codebook generating apparatus according to the present invention comprises: an input vector set storage means for storing an input vector set; a codebook storage means for storing a codebook; a sorting means for rearranging the codebook; A first calculating means for calculating the distortion of the input vector and the representative vector and calculating an address of a representative vector to be searched next;
Determining means for comparing the two distortions to determine a candidate for a representative vector to be encoded; second calculating means for calculating the center of gravity of the representative vector and calculating the average distortion based on the determination result by the determining means; It is characterized by comprising a sorting means, a first calculating means, a judging means and a control means for controlling a processing flow of the second calculating means.

The codebook generating apparatus for vector quantization according to the present invention comprises an input vector set storing means for storing an input vector set, a codebook storing means for storing a codebook, and a sorting means for rearranging the codebook. Comparing the distortion of the input vector with the plurality of representative vectors in parallel and calculating the addresses of the representative vectors to be searched next, and the plurality of distortions obtained by the first computing means. Determining means for determining a candidate for a representative vector to be encoded as a result of the calculation, and calculating and calculating the center of gravity of the representative vector based on the determination result by the determining means.
It is characterized by comprising a second calculating means for calculating the average strain, and a control means for controlling the processing flow of the sorting means, the first calculating means, the judging means and the second calculating means.

The codebook generating apparatus for vector quantization according to the present invention comprises an input vector set storing means for storing an input vector set, a codebook storing means for storing a codebook, and a sorting means for rearranging the codebook. And a first calculating means for calculating the distortion of the input vector and the representative vector and calculating the address of the next representative vector to be searched next, and comparing the two distortions obtained by the first calculating means with each other. Determining means for determining a representative vector candidate, second calculating means for calculating a center of gravity, average strain calculation, and partial strain calculation based on the determination result by the determining means, and partial distortion obtained by the second calculating means. The partial strain sorting means for sorting on the basis of Comparing means for comparing the difference between two partial strains, based on a comparison result by the comparing means, a representative vector moving means for moving a representative vector having a small partial strain to a vicinity of a representative vector having a large partial strain, and the sorting means, It is characterized by comprising a first calculating means, a determining means, a second calculating means, a partial strain sorting means, a partial strain comparing means, and a control means for controlling a processing flow of the representative vector moving means.

Further, the codebook generating apparatus for vector quantization according to the present invention includes an input vector set storing means for storing an input vector set, a codebook storing means for storing a codebook, and a sorting means for rearranging the codebook. And a first calculating means for calculating a distortion between an input vector and a plurality of representative vectors in parallel and calculating an address of a representative vector to be searched next, and a plurality of distortions obtained by the first calculating means. Determining means for determining a candidate of a representative vector to be encoded as a result, and calculating a center of gravity, calculating an average distortion,
A second calculating means for performing a partial strain calculation, a partial strain sorting means for sorting based on the partial strain obtained by the second calculating means, and a difference between the two partial strains sorted by the partial strain sorting means. A comparing means for comparing, a representative vector moving means for moving a representative vector having a small partial distortion to a vicinity of a representative vector having a large partial distortion based on a comparison result by the comparing means, the sorting means, the first calculating means, It is characterized by comprising a determining means, a second calculating means, a partial strain sorting means, a partial strain comparing means, and a control means for controlling the processing flow of the representative vector moving means.

The codebook generating apparatus for vector quantization according to the present invention comprises an input vector set storing means for storing an input vector set, a codebook storing means for storing a codebook, and a sorting means for rearranging the codebook. And a first calculating means for calculating the distortion of the input vector and the representative vector and calculating the address of the next representative vector to be searched next, and comparing the two distortions obtained by the first calculating means with each other. Determining means for determining a representative vector candidate, second calculating means for calculating a center of gravity, average strain calculation, and partial strain calculation based on the determination result by the determining means, and partial distortion obtained by the second calculating means. The partial strain sorting means for sorting on the basis of Comparing means for comparing the difference of several partial distortion, based on the comparison result by the comparison means,
A representative vector moving means for moving a plurality of representative vectors having a small partial strain to a vicinity of a representative vector having a large partial strain; and the sorting means, the first calculating means, the determining means, the second calculating means, the partial strain sorting means, It is characterized by comprising a control means for controlling the processing flow of the partial distortion comparing means and the representative vector moving means.

Further, the codebook generating apparatus for vector quantization according to the present invention comprises an input vector set storing means for storing an input vector set, a codebook storing means for storing a codebook, and a sorting means for rearranging the codebook. Comparing the distortion of the input vector with the plurality of representative vectors in parallel, and calculating the distortion of the representative vector to be searched next, and the plurality of distortions obtained by the first computing means. Determining means for determining a candidate of a representative vector to be encoded as a result, and calculating a center of gravity, calculating an average distortion,
A second calculating means for performing a partial strain calculation, a partial strain sorting means for sorting based on the partial strain obtained by the second calculating means, and a difference between a plurality of partial strains sorted by the partial strain sorting means A representative vector moving means for moving a plurality of representative vectors having a small partial distortion to a vicinity of a representative vector having a large partial distortion based on the result of the comparing means; Means, determination means, second calculation means for performing second calculation and partial strain calculation, partial strain sorting means, partial strain comparison means, and control means for controlling the processing flow of the representative vector moving means. It is characterized by.

Further, the codebook generating apparatus for vector quantization according to the present invention comprises: an input terminal to which an input vector is input; sorting means for rearranging the codebook;
Codebook storage means for storing the sorted codebooks, and the input vector input through the input terminal and the distortion of the representative vector of the codebook stored in the codebook storage means are calculated and then searched. First calculating means for calculating the address of the representative vector, determining means for comparing the two distortions obtained by the first calculating means to determine a representative vector candidate to be encoded, and determining by the determining means Based on the result, the second arithmetic means for adaptively calculating the representative vector and calculating the average distortion, and the processing flow of the sorting means, the first arithmetic means, the determination means, and the second arithmetic means are controlled. And control means.

Further, the codebook generating apparatus for vector quantization according to the present invention comprises: an input terminal to which an input vector is input; sorting means for rearranging the codebook;
Codebook storage means for storing the sorted codebooks, and calculating in parallel the distortion between the input vector input via the input terminal and a plurality of representative vectors of the codebook stored in the codebook storage means. A first calculating means for calculating an address of a representative vector to be searched next; a determining means for comparing a plurality of distortions obtained by the first calculating means to determine a candidate of a representative vector to be encoded; Based on the determination result by the determination means,
A second arithmetic means for adaptively calculating the representative vector and calculating the average distortion, and a control means for controlling the processing flow of the sorting means, the first arithmetic means, the judging means and the second arithmetic means. It is characterized by being performed.

In addition, the codebook generating apparatus for vector quantization according to the present invention comprises: an input terminal to which an input vector is input; sorting means for rearranging the codebook;
Codebook storage means for storing the sorted codebooks, and the input vector input through the input terminal and the distortion of the representative vector of the codebook stored in the codebook storage means are calculated and then searched. First calculating means for calculating the address of the representative vector, determining means for comparing the two distortions obtained by the first calculating means to determine a representative vector candidate to be encoded, and determining by the determining means Second calculating means for performing adaptive representative vector calculation, average strain calculation, and partial strain calculation based on the result, and partial strain sorting for sorting based on the partial strain obtained by the second calculating means Means for comparing the difference between two partial strains sorted by the partial strain sorting means; Based on the comparison result of the partial representative vector moving means for moving in the vicinity of large representative vector strain portions small representative vector strained, said sorting means, first calculating means, judging means, second
And control means for controlling the processing flow of the partial strain sorting means, the partial strain comparing means, and the representative vector moving means.

Also, the codebook generating apparatus for vector quantization according to the present invention comprises: an input terminal to which an input vector is inputted; sorting means for rearranging the codebook;
Codebook storage means for storing the sorted codebooks, and calculating in parallel the distortion between the input vector input via the input terminal and a plurality of representative vectors of the codebook stored in the codebook storage means. A first calculating means for calculating an address of a representative vector to be searched next; a determining means for comparing a plurality of distortions obtained by the first calculating means to determine a candidate of a representative vector to be encoded; Based on the determination result by the determination means,
Second arithmetic means for performing adaptive representative vector calculation / average strain calculation / partial strain calculation; partial strain sorting means for sorting on the basis of the partial strain obtained by the second arithmetic means; Comparing means for comparing the difference between the two partial strains sorted by the strain sorting means, and a representative vector for moving a representative vector having a small partial strain to a vicinity of a representative vector having a large partial strain based on a comparison result by the comparing means. Moving means, the sorting means, the first computing means,
It is characterized by comprising a determining means, a second calculating means, a partial strain sorting means, a partial strain comparing means, and a control means for controlling the processing flow of the representative vector moving means.

Further, the codebook generating apparatus for vector quantization according to the present invention comprises: an input terminal to which an input vector is inputted; sorting means for rearranging the codebook;
Codebook storage means for storing the sorted codebooks, and the input vector input through the input terminal and the distortion of the representative vector of the codebook stored in the codebook storage means are calculated and then searched. First calculating means for calculating the address of the representative vector, determining means for comparing the two distortions obtained by the first calculating means to determine a representative vector candidate to be encoded, and determining by the determining means Second calculating means for performing adaptive representative vector calculation, average strain calculation, and partial strain calculation based on the result, and partial strain sorting for sorting based on the partial strain obtained by the second calculating means Means for comparing the difference between a plurality of partial strains sorted by the partial strain sorting means; and A representative vector moving means for moving a plurality of representative vectors having a small partial distortion to a vicinity of a representative vector having a large partial distortion based on the comparison result by the step; the sorting means, the first calculating means, the judging means, and the second It is characterized by comprising a calculating means, a partial strain sorting means, a partial strain comparing means, and a control means for controlling the processing flow of the representative vector moving means.

Further, in the codebook generating apparatus for vector quantization according to the present invention, an input terminal to which an input vector is input, a sorting means for rearranging the codebook, and a codebook storage means for storing the sorted codebook. And, in parallel, calculate the distortion between the input vector input via the input terminal and a plurality of representative vectors of the codebook stored in the codebook storage means,
First calculating means for calculating an address of a representative vector to be searched next; determining means for comparing a plurality of distortions obtained by the first calculating means to determine candidates for a representative vector to be encoded; Based on the determination result by the determination means, adaptive representative vector calculation, average distortion calculation,
A second calculating means for performing a partial strain calculation, a partial strain sorting means for sorting based on the partial strain obtained by the second calculating means, and a plurality of partial strains sorted by the partial strain sorting means. Comparing means for comparing disparities; representative vector moving means for moving a plurality of representative vectors having a small partial strain to a vicinity of a representative vector having a large partial strain based on a comparison result by the comparing means;
, A control means for controlling the processing flow of the partial distortion sorting means, the partial distortion comparing means, and the representative vector moving means.

[0067]

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

The present invention is applied to, for example, a codebook generating apparatus 100 for a vector quantizer having a configuration as shown in FIG.

The codebook generating device 100 includes an input vector set memory 110, a sorting device 120,
Codebook memory 130, first computing device 140, determining device 150, second computing device 160, and control device 17
It consists of 0.

[0070] In this codebook generator 100, the input vector set memory 110, an input vector set X =: ^{Memory (x (t) ∈R K t} = 0, ···, T-1) stores the , The address is specified by the first arithmetic unit 140, and the corresponding input vector x ^(t)
And the number of the initial representative vector are passed to the first arithmetic unit 140. Further, the input vector set memory 110 receives the number min of the initial representative vector of the input vector x ^(t) from the determination device 150 and stores it.

The sorting device 120 includes the control device 17
Upon receiving a sorting instruction from 0 and an instruction on which dimension k is to be sorted, the code book is read from the code book memory 130 and the code book is sorted based on the specified dimension. Also, the sorting device 12
0 is a codebook memory 1 for storing sorted codebooks.
Write to the specified 30 addresses.

[0072] codebook memory 130, the codebook _{^{W (p) = {w i}} (p) ∈R K: i = 0, ···, N
-1} is a memory that receives an address from the first arithmetic unit 140 that performs distortion calculation and address calculation, and corresponding representative vector data (representative vector value w _i ^(p) , number i) To the first arithmetic unit 140. Further, the codebook memory 130 receives data (updated codebook W ^{(p + 1)} ) and an address to be written from the second arithmetic unit 160 that performs the calculation of the center of gravity and the average strain, and writes the codebook to the corresponding address. It is. Also, the codebook memory 130
The code book is passed to the sorting device 120. Further, the codebook memory 130 stores the sorting device 1
Receiving the code book sorted by 20 and the address to be written, and storing the code book at the specified address.

The first arithmetic unit 140 designates an address in the input vector set memory 110 and receives the corresponding input vector x ^(t) and the number of the initial representative vector.
Further, the first arithmetic unit 140 receives an instruction for calculating a distortion / address from the control unit 170, specifies the address of a representative vector to be read into the codebook memory 130, and specifies the corresponding data (representative vector value w _i). ^(P) ,
Number i). Also, the first arithmetic unit 140
Is the distortion d (x
_^(T), to calculate the ^{w i (p)).} The first computing device 140, passes the strain _{^{d (x (t), w}} i (p)) and data (number of the representative vector value) of the determination device 150. In addition, the first arithmetic unit 140 receives an instruction on the dimension k serving as a sorting criterion from the control device 170, and determines the distortion d (x _k ^(t) , w _ik ^{( p)} ) of only k dimensions in the determination device 150. Pass to. Further, the first arithmetic unit 140 receives, from the determination unit 150, up / down information indicating the direction of the representative vector to be searched based on the determination result, and calculates the address of the next representative vector to be searched. Note that, when receiving the information “finish” indicating the end of the search, the first arithmetic device 140 starts encoding the next input vector x ^{(t + 1)} .

The judgment device 150 calculates the current minimum distortion d.
^cur _min and the number min of the representative vector are held, and mi having the current minimum distortion d ^cur _min
For the n-th representative vector and the i-th representative vector calculated by the first arithmetic unit 140, the distortion d
Compares the _{^{(x (t), w i}} (p)), the strain is the smaller of the new min number, the current minimum distortion ^d _{cur min}
And Further, the determination device 150 receives a determination instruction from the control device 170 and receives a k-dimensional distortion d.
(X _k ^(t) , w _ik ^(p) ) and the current minimum strain d
^cur _min , d (x _k ^(t) ,
If w _ik ^(p) is larger, the information of up / down indicating the direction of the next representative vector to be searched is changed, and if both the up and down directions have been checked, the information finish indicating search has been completed To the arithmetic unit 140, and d
If (x _k ^(t) , w _ik ^(p) ) is smaller, information for examining the next representative vector in the same direction is sent to the first arithmetic unit 1.
Pass to 40. Further, the determination device 150 determines that the up / down
When the search ends in both directions and the search ends, the input vector x ^(t) , the number min of the initial representative vector having the current minimum distortion, and the minimum distortion d ^cur _min are output to the second arithmetic unit 160. Further, when the search is finished, the determination device 150 stores the number min of the initial representative vector having the current minimum distortion in the input vector set memory 110 as the initial representative vector of the input vector x ^(t) .

The second arithmetic unit 160 includes a judgment unit 150
From the input vector x^(T)The representative vector encoded with
When the input number min and the input number min are encoded
D ^cur _minReceive. And this second operation
The device 160 stores the data corresponding to the representative vector number min.
T (molecule w_minnumAnd denominator w_minden,further
Average strain D_sum) Is calculated by the following equation (5), (6) and
And (7).

W _minnum + = x ^(t) (5) w _minden + = 1 (6) D _sum + = d ^cur _min (7) Further, the second arithmetic unit 160 operates on T input vectors. When the controller 170 receives an instruction that the encoding has been completed and the above-described intermediate calculation has been completed, the calculation of the representative vector centroid w _i ^{(p + 1)} and the average distortion D are performed by the following equations (8) and (9). Perform according to.

[0077]

(Equation 5)

The second arithmetic unit 160 receives all data (numerator w) according to an instruction from the control unit 170.
_Inum , denominator w _iden , and average strain D _sum , ∀i) are cleared.

The second arithmetic unit 160 passes the data (the updated codebook W ^{(p + 1)} ) and the address to be written, and writes the updated codebook to the corresponding address.

The second arithmetic unit 160 includes the control unit 1
According to the instruction from 70, an arbitrary initial codebook is set in the codebook memory 130.

The control device 170 controls the sorting device 12
0, the processing flow of the first arithmetic device 140, the determination device 150, and the second arithmetic device 160 is controlled. In addition, the control device 170 specifies a dimension serving as a sorting reference to the sorting device 120. In addition, the control device 170 specifies the address (or number) of the initial representative vector to the first arithmetic device 140, and indicates the dimension used as the reference for sorting. Further, the control device 170 controls the second arithmetic device 16
Instruct 0 to set an arbitrary initial codebook.

The code book generating apparatus 1 having such a configuration
00, in the present invention, fast vector quantization (Vector) using the representative vectors sorted according to the codebook generation algorithm shown in the flowchart of FIG.
Quantization by Searching in Sirted Codebook: VQ
-SSC) is generated.

That is, the code book generating apparatus 10
In the case of 0, first, in step (F101), a representative vector initial arrangement is set. Also, a reference dimension k for sorting the representative vectors is specified.

In the next step (F102), the representative vectors are sorted based on the value of the representative vector of dimension k. The information in the sorted order can be considered in the following two ways.

(1) The order of the representative vectors is directly changed. If it is not desired to change the number of the representative vector, the structure is replaced with a structure having information on the number of the representative vector. Since the numbers are directly changed, the number lower of the representative vector lower than the own is the number-1 of the own array.
And the sequence number. Also, the number upper of the representative vector higher than the own is the array number obtained by adding +1 to the array number of the own.

(2) When preparing an array indicating the order of sorting, the information to be held by each representative vector is represented by the lower number of the representative vector lower than the own and the upper number of the representative vector higher than the own. is there.

In the next step (F103), the distortion between the input vector x ^(t) and the initial representative vector w _i ^(p) is calculated, and the distortion d (x ^(t) ,
and w _i ^(p)) of the current minimum distortion ^d _{cur min,} keep the number i as a candidate for coding.

[0088] _{^{d cur min = d (x (}} t), w i (p)) (10) min = i (11) In the next step (F104), the initial representative vector w
The number of the representative vector on both sides of _i ^(p) , that is, lower = the number of the next representative vector in the lower direction upper = the number of the next representative vector in the upper direction

In the next step (F105), the input vector
Torr x^(T)And representative vector w_{lo wer} ^(P)Tonohi
Z d (x^(T), W_lower ^(P)) Calculation
D (x_k ^(T),
w_lowerk ^(P)) Is saved. And this
In step (F105), the calculated strain d (x_k
^(T), W_lowerk ^(P)) Is the current minimum strain
d^cur _minIs smaller than, that is, d (x^(T), W_lower ^(P)) <D^cur _min It is determined whether or not (12) holds, and the equation (12) holds.
Then d^cur _min= D (x^(T), W_lower ^(P)(13) It is assumed that min = lower (14).

In the next step (F106), it is determined whether to continue the search in the lower direction. This step (F1
At 06), if d (x _k ^(t) , w _lowerk ^(p) )> d ^cur _min (15) is satisfied, the search in the lower direction is completed, and the process proceeds to step (F108). , Next step (F
Go to 107).

Step (F107) is a step of updating the representative vector number lower, and the current lower
The number of the representative vector next to the r-th representative vector in the lower direction is set as a new lower, and the process returns to step (F105).

In step (F108), the input vector
x^(T)And the representative vector w of the upper number_upper
^(P) And the strain d (x^(T),
w_upper ^(P)) Is calculated, and d (x
_k ^(T), W_upperk ^(P)) Is saved. This
In step (F108), the calculated strain d (x
^(T), W_upper ^(P)) Is the current minimum strain d
^cur _minIs smaller than, that is, d (x^(T), W_upper ^(P)) <D^cur _min It is determined whether or not (16) holds, and this (16) equation holds.
If d^cur _min= D (x^(T), W_upper ^(P)(17) min = upper (18)

In the next step (F109), it is determined whether to continue the search in the upper direction. This step (F1
In 09), if d (x _k ^(t) , w _upperk ^(p) )> d ^cur _min (19) is satisfied, the search in the upper direction is completed, and the process proceeds to step (F111). , Next step (F
Proceed to 110).

Step (F110) is a step of updating the representative vector number upper, and the current upper
The process returns to step (F108) with the number of the representative vector next to the numbered representative vector in the upper direction as a new upper.

In the next step (F111), the input vector is encoded, the center of gravity of the representative vector is calculated, and the average distortion is prepared.

In this step (F111), the input vector x ^(t) is encoded into a representative vector w _min ^(p) of the number min having the minimum distortion d ^cur _min .

In this step (F111), the step of updating the position of the encoded representative vector min (F111)
114), the numerator w _minnum and the denominator w
The calculation of _{min den} and the average strain D _sum is as follows (2
This is performed according to equations (0), (21) and (22).

W _minnum + = x ^(t) (20) w _minden + = 1 (21) D _sum + = d ^cur _min (22) In the next step (F112), encoding of all input vectors is completed. It is determined whether or not. In this step (F112), if t = T−1 holds, the process proceeds to a step (F114); otherwise, the process proceeds to the next step (F113).

In step (F113), t = t + 1 is set, and the process returns to step (F103) to search for the next input vector.

In step (F114), the center of gravity of the representative vector is calculated and the average distortion is calculated. This step (F114) is a step of updating the center of gravity of the representative vector in the LBG algorithm.
It can be represented by an equation.

[0101]

(Equation 6)

However, since the calculation of the numerator and denominator in equation (23) has been completed in the previous step (F111), the following (2)
4) The equation may be calculated.

[0103]

(Equation 7)

The average strain D is calculated by the following equation (25).

[0105]

(Equation 8)

In the next step (F115), convergence determination is performed. If the convergence determination condition is not satisfied in step (F115), the process returns to step (F102),
If the convergence determination condition is satisfied, the process ends.

Here, the convergence determination conditions in this step (F115) may be as follows. Others may be used.

(1) The process ends when the representative vector is no longer updated.

(2) When the average strain becomes equal to or less than the desired average strain, the process is terminated.

(3) Terminate at the number of repetitions.

Here, the coding algorithm (VQ-SS) for vector quantization in this codebook generating apparatus is used.
The operation of C) will be described using an example.

In this example, a square error distortion measure is used as the distortion measure.

D (x, w) = (x−w) ² (26) Naturally, the present invention is also effective when other strain measures are used. The present invention is also effective when a different strain measure is used for each dimension.

A three-dimensional input vector x = (12,16,
FIG. 3 shows an example in which the distortion is calculated for all the representative vectors as in the conventional method, and the representative vectors to be subjected to vector quantization (encoding) are obtained. In this case, 12
If distortion is not calculated for the representative vectors, the representative vector to be encoded cannot be determined. In this example,
The 9th representative vector w ₉ = (14,
18, 3).

On the other hand, in the coding algorithm of vector quantization in this codebook generating apparatus, k
The representative vectors are sorted for = 1. The representative vectors are numbered in the sorted order for easy understanding. Actually, it is not necessary to renumber the representative vectors, and the program may be programmed to save the numbers of the representative vectors on both sides (upper direction / lower direction) of the own representative vector.

If the representative vector to be encoded is known in advance, since the minimum distortion is 8, a range of 2 or less, which is the maximum integer of √8 or less, may be searched. FIG.
From this, the representative vector that becomes 12 ± 2 in one dimension is
Only the three representative vectors (7), (8) and (9) need to be examined.

However, since the representative vector to be encoded cannot be known in advance, the search is performed using the following algorithm (search 0 to search 6).

That is, in search 0, as shown in FIG. 4, the initial representative vector is set to the (7) th representative vector, and the distortion d (x, w ₇ ) is calculated.

D (x, w ₇ ) = 17 lower = 6 upper = 8 Further, the minimum distortion d ^cur _min = 17 and the number min = 7 of its representative vector are stored and stored.

In search 1, as shown in FIG.
The distortion d (x, w) of the representative vector of r (6)
_lower ).

D (x ₁ , w _{lower 1} ) = 9 d (x, w _lower ) = 189 Then, when d (x, w _lower ) is compared with the current minimum distortion d ^cur _min , d (x, w _lower )
Since w _lower ) is greater than d ^cur _min , the minimum distortion d ^cur _min retains its value, that is, d ^cur _min = 17 min = 7. _{Also, d (x 1, w lower1} ) because the current minimum of less than the strain ^{d cu} _{r min,} moves to explore 2 of the next lower = 5.

In the next search 2, as shown in FIG.
The distortion d (x, w) of the representative vector of the wer (5)
_lower ).

D (x ₁ , w _lower1 ) = 9 d (x, w _lower ) = 14 Then, when d (x, w _lower ) is compared with the current minimum distortion d ^cur _min , d (x, w _{lower 1} )
Since w _lower ) is smaller than d ^cur _min , the minimum distortion d ^cur _min is updated to d ^cur _min = 14 min = 5. D (x ₁ , w _lower1 ) is d ^cur
Since it is smaller than _min = 14, the process moves to the next search 3 of lower = 4.

In search 3, as shown in FIG.
The distortion of the representative vector of r (4) is calculated.

D (x ₁ , w _lower1 ) = 16 d (x, w _lower ) = 186 Then, when d (x, w _lower ) is compared with the current minimum distortion d ^cur _min , d (x, w _{lower 1} )
Since w _lower ) is greater than d ^cur _min , the minimum distortion d ^cur _min retains its value, that is, d ^cur _min = 14 min = 5. D (x ₁ , w _lower1 ) is d
Since it is larger than ^cur _min = 14, the search in the lower direction ends, and the process then proceeds to search 4 in the upper direction with upper = 8.

In search 4, as shown in FIG.
The distortion of the representative vector of r (8) is calculated.

[0127] _{d (x 1, w upper1)} = 1 d (x, w upper) and _{= 46, d (x, w} upper) if the comparing the current minimum distortion ^d _{cur min,} ^d (x,
Since w _lower ) is greater than d ^cur _min , the minimum distortion d ^cur _min retains its value, that is, d ^cur _min = 14 min = 5. D (x ₁ , w _upper1 ) is d
^{Since cur} _min = 14, the next upper
Move to search 5 where = 9.

In search 5, as shown in FIG.
The distortion of the r (9) -th representative vector is calculated.

D (x ₁ , w _{upper 1} ) = 4 d (x, w _upper ) = 8 Then, when d (x, w _upper ) is compared with the current minimum distortion d ^cur _min , d (x, w _upper )
Since w _lower ) is smaller than d ^cur _min , the minimum distortion d ^cur _min is updated so that d ^cur _min = 8 min = 9. D (x ₁ , w _upper1 ) is d ^cur
_{Since min} = 9, the search 6 for the next up0
Move on to

In search 6, as shown in FIG.
The distortion of the representative vector of er (10) is calculated.

D (x ₁ , w _{upper 1} ) = 9 d (x, w _upper ) = 41 Then, when d (x, w _upper ) is compared with the current minimum distortion d ^cur _min , d (x, w _upper )
Since w _lower ) is greater than d ^cur _min , the minimum distortion d ^cur _min retains its value, that is, d ^cur _min = 8 min = 9. D (x ₁ , w _upper1 ) is d
Since it is larger than ^cur _min = 9, the search in the upper direction ends.

The search in both the lower direction and the upper direction is completed in this way, and the finally obtained minimum distortion d
^The min = 9 representative vector with ^cur _min = 8 is the representative vector to be encoded for the input vector x = (12,16,13).

Here, from FIG. 11, if the representative vector to be coded is known in advance, it is sufficient to search for three representative vectors. However, since it is generally impossible, it is necessary to determine and search for an initial representative vector. There is. In this case, since the search was performed while updating the minimum distortion, a representative vector having the minimum distortion could be obtained by searching for seven representative vectors. In the gradation conversion, since the number of representative vectors is 256, it can be expected that the search range is greatly limited.

Further, the present invention is more effective when a high-speed technique using a partial sum of distortions is combined. In the high-speed technique using the partial sum of the distortion, the partial sum of the distortion represented by the following expression (27) is equal to the current minimum distortion d ^cur.
_If it is larger than _min , the next representative vector is searched. In other words, it is a method of increasing the efficiency of the search by comparing the strain in the middle of calculation with the current minimum strain, instead of finding the strain to the end.

[0135]

(Equation 9)

A high-speed technique using the partial sum of distortions is described in the literature [Bei CD. & Gray, RM.
of minimum distortion encoding algorithm for vecto
r quantization. ”IEEE Transactions on Communicati
ons, COM-33, pp. 1132-1133, 1985].

Here, only a difference will be described with reference to FIG. 12 as a modification of the above-described codebook generating apparatus shown in FIG.

In the code book generating apparatus shown in FIG. 12, the code book memory 130 passes the code book and the address to the sorting apparatus 120.

[0139] Sorting device 120 is controlled by control device 17.
Upon receiving a sort instruction from 0 and an instruction on which dimension k is to be sorted, the code book address and the code book are read from the code book memory 130, and the code book is sorted based on the designated dimension. Then, the addresses of the representative vectors are rearranged in the order of the sorted representative vectors, and the addresses arranged in the sorted order are written in the codebook sort table 180.

The first arithmetic unit 140 includes a judging device 150
, Receives up / down information indicating the direction of the representative vector to be searched based on the determination result, refers to the codebook sort table 180 for the address of the next representative vector to be searched, and receives the address. Then, the address is designated and the corresponding data (representative vector number i and value w _i ^(p), etc.) is read into the code book memory 130.

The codebook sort table 180 is
The address received from the sorting device 120 is stored. Further, the codebook sort table 180 passes the address of the representative vector to the first arithmetic unit 140 in accordance with the first instruction from the arithmetic unit 140 to refer to the next representative vector to be searched.

Here, [Gersho, A .. “Asy”
mptotically optimal block quantization. "IEEE Trans
actions on Information Theory, 25, pp. 373-380, 197
9.], [Yamada, Y., Tazaki, S. & Gray, RM.
totic performance of block quantizers with differe
nce distortion measures. "IEEE Transactions on Inf
ormation Theory}, 26, pp. 6-14, 1980.], [Ueda, N., &
Nakano, R .. “A new competitive learning approach
based on an equidistortion principle for designing
optimal vector quantizers. "Neural Networks, 7, p
p.1211-1227, 1994.], etc., from a theoretical study of vector quantization, the global optimum when the Euclidean (l ₂ ) norm represented by the following equation 28 is used as the r-th power is used for the distortion measure. The equal strain principle, which is a necessary condition of the solution, is shown.

[0143]

(Equation 10)

When the probability distribution of the input space is composed of a large number of disjoint clusters with respect to a large number of representative vectors, the arrangement of the representative vectors in the vector quantizer which has the minimum average distortion is as follows. Are arranged so that their partial strains are equal to each other.

The partial distortion D _i of the i-th representative vector
Is calculated using the following equation (29).

[0146]

[Equation 11]

Here, δ _it is Kronecker's delta, and is 1 when the representative vector i has the smallest distortion.
Becomes

FIG. 13 shows a simple example of the principle of equal distortion.
Black circles are representative vectors, and dotted lines are Voronoi regions. This is the optimal arrangement of representative vectors for a problem where the input vector arises uniformly from two square clusters. At that time, the partial distortion of each representative vector is in the same state.

As described above, since the LBG algorithm and the competitive learning are descent methods, the global minimum can be found when the codebook is arranged near the global minimum. In view of this, the present invention focuses on increasing the possibility that an optimal arrangement can be obtained by setting the arrangement of the representative vectors near the arrangement indicated by the equal distortion principle. Since it is difficult to place the representative vector in the vicinity of the optimal layout, the representative vector is moved so that the partial distortions become equal during updating.

Therefore, in the present invention, an LBG having a Min-Max Order Compare (MMOC) operation based on the equal distortion principle is added.
We propose an algorithm. The MMOC operation calculates the partial distortion during the update and rearranges the representative vectors to reduce the variation of the partial distortion. By the MMOC operation, it is possible to make the partial distortions of the representative vectors close to each other.

For example, the codebook generating apparatus 200 having a configuration as shown in FIG. 14 can execute an LBG algorithm to which an MMOC operation using VQ-SSC encoding is added.

A code book generating apparatus 200 shown in FIG.
Are the input vector set memory 210, the sorting device 220, the codebook memory 230, the first arithmetic device 2
40, determination device 250, second arithmetic device 260, partial strain sorting device 262, partial strain comparison device 26
4. It comprises a representative vector moving device 266 and a control device 270.

[0153] In this codebook generator 200, the input vector set memory 210, an input vector set X =: ^{Memory (x (t) ∈R K t} = 0, ···, T-1) stores the , The address is specified by the first arithmetic unit 240, and the corresponding input vector x ^(t)
And the number of the initial representative vector are passed to the first arithmetic unit 240. Further, the input vector set memory 210 receives the number min of the initial representative vector of the input vector x ^(t) from the determination device 250 and stores it.

The sorting device 220 includes the control device 27
Upon receiving a sorting instruction from 0 and an instruction on which dimension k is to be sorted, the code book is read from the code book memory 230, and the code book is sorted based on the specified dimension. Also, the sorting device 22
0 indicates the sorted code book is stored in the code book memory 2
Write to the specified 30 addresses.

[0155] codebook memory 230, the codebook _{^{W (p) = {w i}} (p) ∈R K: i = 0, ···, N
-1} is a memory that receives an address from the first arithmetic unit 240 that performs distortion calculation and address calculation, and corresponding representative vector data (representative vector value w _i ^(p) , number i) To the first arithmetic unit 240. In addition, the codebook memory 230 receives data (an address to be written with the updated codebook W ^{(p + 1))} from the representative vector moving device 266, and writes the updated codebook to the corresponding address. Further, the codebook memory 230 transfers the codebook to the sorting device 220. Further, the code book memory 230 receives the code book sorted by the sorting device 220 and the address to be written,
Store the codebook at the specified address.

The first arithmetic unit 240 has a set of input vectors.
Designate an address in the integration memory 210, and
Kutor x^(T)And the number of the initial representative vector.
Further, the first arithmetic unit 240 receives a signal from the control unit 270.
Receiving instructions for calculating addresses and addresses, codebook memo
Specify the address of the representative vector you want to read into the library 230
And the corresponding data (representative vector value w_i ^(P),
Number i). Also, the first arithmetic unit 240
Is the distortion d (x
^(T), W_i ^(P)) Is calculated. And the first operation
The device 240 has a strain d (x^(T), W_i ^(P)) And de
Data (representative vector number and value) to the determination device 250
You. Also, the first arithmetic unit 240
Received the instruction of dimension k, which was the reference for sorting, and k dimensions
Injury strain d (x_k ^(T) , W_ik ^(P)) Also judge
Hand over to device 250. Further, the first arithmetic unit 240
Representative vector to be searched from the determination device 250 based on the determination result.
The next search is received upon receiving up / down information indicating the direction of the file.
The address of the representative vector to be calculated is calculated. Note that the first
Arithmetic unit 240 receives information finish indicating the end of the search
Then the next input vector x^{(T + 1)}Encoding
Take it.

The determination device 250 determines the current minimum distortion d.
^cur _minAnd the representative vector number min
And the current minimum strain d^cur _minMi with
Calculated by the n-th representative vector and the first arithmetic unit 240
Distortion d for the i-th representative vector
(X^(T), W_i ^(P)), The smaller the strain
Is the new min number, the current minimum strain d^cur _min
And Also, the determination device 250
In response to the instruction of the judgment, distortion d of only k dimensions
(X_k ^(T), W_ik ^(P)) And the current minimum strain d
^cur _minAnd d (x_k ^{(T )} , W_ik
^(P)) Is larger, the next representative vector to search
Of up / down indicating the direction of
When both p / down directions have been checked, search ends
Is passed to the first computing device 240, and d
(X_k ^(T) , W_ik ^(P)) Is smaller if
Information for examining the next representative vector in the direction is stored in a first arithmetic unit
Pass it to 240. Further, the determination device 250 determines whether the up / down
n When both directions are checked and the search is finished,
Input vector x to second arithmetic unit 260^(T), Current
The initial representative vector numbers min and
Minimum strain d^cur _minIs output. Furthermore, judgment
The device 250, when the search end finish is reached,
Enter the number min of the initial representative vector with small distortion
Vector x^(T)Input vector as the initial representative vector of
Stored in the file set memory 210.

The second arithmetic unit 260 is provided with a judgment unit 250
Receives the input vector x ^(t) , the encoded representative vector number min, and the distortion d ^cur _min when the input vector is encoded. Then, the second arithmetic unit 260 calculates the data (numerator w _minnum and denominator w _minden , and partial strain D _sum ) corresponding to the representative vector number min by using the following equation (30):
This is performed according to equations (1) and (32).

W _minnum + = x ^(t) (30) w _minden + = 1 (31) D _minsum + = d ^cur _min (32) In addition, the second arithmetic unit 260 operates on T input vectors. Then, when the controller 270 receives an instruction that the encoding has been completed and the intermediate calculation has been completed, the encoding has been completed and the instruction has been received from the controller 270 that the intermediate calculation has been completed. _i ^{(p + 1)} , the partial strain D _i and the average strain D are calculated by the following equation (33):
This is performed according to the equations (34) and (35).

[0160]

(Equation 12)

[0161] Further, the second arithmetic unit 260, the control unit 270, all the data (molecular _{w inum} and denominator _{w iden,} further partial distortion _{D isum,} ∀i)
Clear

The partial distortion sorting device 262 receives data (updated codebook and partial distortion) from the second arithmetic unit 260. The partial strain sorting device 262 sorts the codebook and the partial strain based on the magnitude of the partial strain. Further, the partial strain sorting device 262 passes the data (the sorted codebook and the partial strain) to the partial strain comparison device 264.

The partial distortion comparing device 264 receives data (sorted codebook and partial distortion) from the partial distortion sorting device 262. This partial strain comparison device 264 compares a large partial strain with a small partial strain. When the disparity is large, information for moving the representative vector having the small partial distortion to the vicinity of the representative vector having the large partial distortion is passed to the representative vector moving device 266. Further, the data (the above-mentioned movement information and codebook) is passed to the representative vector moving device 266.

The representative vector moving device 266 receives, from the partial distortion comparing device 264, information and a codebook for moving the representative vector having the small partial distortion to the vicinity of the representative vector having the large partial distortion. The representative vector moving device 266 moves the representative vector having the small partial distortion to the vicinity of the representative vector having the large partial distortion based on the movement information. The representative vector moving device 266 stores the codebook memory 2
Data in 30 (updated codebook W ^{(p + 1)} )
And write the updated codebook to the corresponding address. Further, the representative vector moving device 266 sets an arbitrary initial codebook in the codebook memory 230 according to an instruction from the control device 270.

The control device 270 includes the sorting device 22
0, first computing device 240, determining device 250, second computing device 260, partial strain sorting device 262, partial strain comparing device 264, and representative vector moving device 26
6 is controlled. Also, the control device 270
A dimension serving as a reference for sorting is designated to the sorting device 220. Further, the control device 270 controls the first arithmetic device 24
The address (or number) of the initial representative vector is designated as 0, and the dimension used as the reference for sorting is designated. Further, control device 270 instructs second arithmetic device 260 to set an arbitrary initial codebook.

The code book generating apparatus 2 having such a configuration
In the present invention, in the present invention, Mi
A codebook is generated according to a flowchart shown in FIG. 15 by an LBG algorithm to which an n-Max Order Compare (MMOC) operation is added.

That is, in the vector quantizer 200, first, in step (F201), a representative vector initial arrangement is set. Also, a reference dimension k for sorting the representative vectors is specified. Further, parameters R and m of the MMOC operation are set.

In the next step (F202), the representative vectors are sorted based on the value of the representative vector of dimension k. The information in the sorted order can be considered in the following two ways.

(1) The order of the representative vectors is directly changed. If it is not desired to change the number of the representative vector, the structure is replaced with a structure having information on the number of the representative vector. Since the numbers are directly changed, the number lower of the representative vector lower than the own is the number-1 of the own array.
And the sequence number. Also, the number upper of the representative vector higher than the own is the array number obtained by adding +1 to the array number of the own.

(2) When an array indicating the order of sorting is prepared, the information to be held by each representative vector is represented by the lower number of the representative vector lower than the own and the upper number of the representative vector higher than the own. is there.

In the next step (F203), the distortion between the input vector x ^(t) and the initial representative vector w _i ^(p) is calculated, and the distortion d (x ^(t) ,
and w _i ^(p)) of the current minimum distortion ^d _{cur min,} keep the number i as a candidate for coding.

[0172] _{^{d cur min = d (x (}} t), w i (p)) (36) min = i (37) In the next step (F 204), the initial representative vectors w
The number of the representative vector on both sides of _i ^(p) , that is, lower = the number of the next representative vector in the lower direction upper = the number of the next representative vector in the upper direction

In the next step (F205), the input vector
Torr x^(T)And representative vector w_{lo wer} ^(P)Tonohi
Z d (x^(T), W_lower ^(P)) Calculation
D (x_k ^(T),
w_lowerk ^(P)) Is saved. And this
In step (F205), the calculated strain d (x_k
^(T), W_lowerk ^(P)) Is the current minimum strain
d^cur _minIs smaller than, that is, d (x^(T), W_lower ^(P)) <D^cur _min It is determined whether or not (38) holds, and the equation (38) holds.
Then d^cur _min= D (x^(T), W_lower ^(P)(39) Let min = lower (40).

In the next step (F206), it is determined whether to continue the search in the lower direction. This step (F2
At 06), if d (x _k ^(t) , w _lowerk ^(p) )> d ^cur _min (41) is satisfied, the search in the lower direction is completed, and the process proceeds to step (F208). , Next step (F
207).

The step (F207) is a step of updating the number “lower” of the representative vector.
The number of the representative vector next to the r-th representative vector in the lower direction is set as a new lower, and the process returns to step (F205).

In step (F208), the input vector x ^(t) and the representative vector w _{upper of the} upper number are input.
^(P) and the strain of _{^{d (x (t), w}} upper (p))
Is calculated, and d (x _k ^(t) , w
_{Save upperk} ^(p) ). In this step (F208), the calculated strain d (x ^(t) , w
_upper ^{( p)} is the current minimum distortion d ^cur
_min , that is, whether d (x ^(t) , w _upper ^(p) ) <d ^cur _min (42) holds, and if this equation (42) holds, then d ^cur _min = d (x ^(t) , w _upper ^(p) ) (43) ^Let min = upper (44).

In the next step (F209), it is determined whether to continue the search in the upper direction. This step (F2
09), if d (x _k ^(t) , w _upperk ^(p) )> d ^cur _min (45) is satisfied, the search in the upper direction is completed, and the process proceeds to step (F211). , Next step (F
Proceed to 210).

Step (F210) is a step of updating the representative vector number upper, and the current upper
The process returns to step (F208) with the number of the representative vector next to the numbered representative vector in the upper direction as a new upper.

In the next step (F211), encoding of the input vector, calculation of the center of gravity of the representative vector, and preparation of the average distortion are performed.

In this step (F211), the input vector x ^(t) is encoded into the representative vector w _min ^(p) of the number min having the minimum distortion d ^cur _min .

In this step (F211), the step of updating the position of the encoded representative vector min (F2)
214), the numerator w _minnum and the denominator w
The calculation of _{min den} and the partial strain D _minsum is performed according to the following equations (46), (47) and (48).

W _minnum + = x ^(t) (46) w _minden + = 1 (47) D _minsum + = d ^cur _min (48) In the next step (F212), encoding of all input vectors is completed. It is determined whether or not. In this step (F212), if t = T−1 holds, the process proceeds to step (F214), otherwise, the process proceeds to the next step (F213).

In step (F213), t = t + 1, and the flow returns to step (F203) to search for the next input vector.

In step (F214), the calculation of the center of gravity of the representative vector and the calculation of the average distortion and the partial distortion are performed. This step (F214) is a step of updating the center of gravity of the representative vector in the LBG algorithm, and can be represented by the following equation (49).

[0185]

(Equation 13)

However, since the calculation of the numerator and denominator in equation (49) has been completed in the previous step (F211), the following (511)
Equation (0) may be calculated.

[0187]

[Equation 14]

[0188] The partial distortion _{D i} is calculated by the following (51) below.

[0189]

(Equation 15)

Then, the average strain D is calculated by the following equation (52).

[0191]

(Equation 16)

In the next step (F215), convergence determination is performed. When the convergence determination condition in this step (F215) is not satisfied, the process proceeds to step (F216),
If the convergence determination condition is satisfied, the process ends.

Here, the convergence determination conditions in this step (F215) may be as follows. Others may be used.

(1) The process ends when the representative vector is no longer updated.

(2) When the average strain becomes equal to or less than the desired average strain, the process ends.

(3) Terminate at the number of repetitions.

In the next step (F216), the representative vectors and the partial distortions are sorted according to the partial distortions.

Then, in the next step (F217) and step (F218), the MMOC operation is executed.

That is, in the step (F217), as shown in FIG.
R partial strains D in ascending order of _max [r]
Compare _min [r].

As a result of the comparison in step (F217), if the following equation (53) does not hold, the process proceeds to step (F202). If the following equation (53) holds, the process proceeds to step (F218).

D _max [r]> m × D _min [r], (r = 0,..., R−1) (53) Further, in step (F218), the following equation (54) is used. In addition, the representative vector w _min ^{(p + 1)} [r] having a small partial strain D _min [r] is _replaced with the representative vector w _max having a large partial strain D _max [r].
^{(P + 1)} Rearranged in the vicinity of [r], and step (F2)
Return to 02).

W _min ^{(p + 1)} [r] ← w _max ^{(p + 1)} [r] + ε (54) where ε (≪1) is a perturbation term. Up to R movements are possible in one iteration.

In the MMOC operation, the MM in the current iteration is
The number of movements by the OC operation is MM in the previous iteration.
It hardly occurs that the number of movements R 'is increased by the OC operation. Therefore, the MMOC operation is always R
It is more efficient to apply the number R ′ in the previous iteration to the MMOC operation instead of applying the number to the MMOC operation.
Since R ′ eventually becomes zero, at least one MMOC operation is performed when R ′ = 0.

FIG. 17 is a conceptual diagram showing the operation of the LBG algorithm to which the MMOC operation is added in the simplest case where R = 1.
Shown in FIG. 18 shows the actual operation.

Here, the operation of the LBG algorithm to which the MMC operation is added will be described with reference to FIG.

In the first iteration, the second representative vector is not used for encoding at all, that is, the partial distortion D ₂ =
0. In contrast, it has been used very well to the encoding, the largest part distortion D ₄ representative vectors fourth. Therefore, the partial distortion D of the second representative vector
Portion of the representative vector of _2min and fourth strain _{D 4Max}
When comparing, since the following equation (55) _{D 4max> m · D 2min (} 55) holds and moves the representative vector number 2 in the vicinity of the representative vector of the fourth. This movement reduces the difference in partial strain.

Then, the arrangement of the representative vectors is corrected using the LBG algorithm.

In the second iteration, the partial distortion D ₁ of the _first representative vector is the smallest and the partial distortion D ₄ of the _fourth representative vector is the largest, as in the first iteration. Therefore, when comparing the partial distortion _{D 4 ma x} of the representative vector of the partial distortion _{D 1min} and No.4 No.1 representative vectors, when the next (56) Formula _{D 4max> m · D 1min (} 56) is satisfied, The first representative vector is moved to the vicinity of the fourth representative vector. This movement further reduces the difference in partial strain.

Then, the arrangement of the representative vectors is corrected using the LBG algorithm.

In the third iteration, the partial distortion D1 of the _first representative vector is the smallest, and the partial distortion D5 of the _fifth representative vector is the largest. However, when the partial distortion D _1min of the first representative vector is compared with the partial distortion D _4max of the fifth representative vector, D _4max > m · D _1min (57) does not hold because there is almost no difference in partial distortion. The arrangement of the representative vectors is an arrangement in the vicinity of the global optimal solution shown by the equal distortion principle.

Therefore, by repeating the LBG algorithm several times, an optimal arrangement of the representative vectors can be obtained.

Further, according to the present invention, the partial sum of the distortion represented by the above-mentioned equation (27) is equal to the current minimum distortion d ^cur.
_{When min} is larger than _min , a further effect is obtained by combining a high-speed technique using a partial sum of distortions for searching for the next representative vector.

Here, only a difference will be described with reference to FIG. 19 as a modification of the vector quantizer 200 shown in FIG.

In the vector quantizer 200 'shown in FIG. 19, the codebook memory 230 passes the codebook and the address to the sorting device 220.

The sorting device 220 includes the control device 27
Upon receiving a sorting instruction from 0 and an instruction on which dimension k is to be sorted, the code book address and code book are read from the code book memory 230, and the code book is sorted based on the specified dimension. Then, the addresses of the representative vectors are rearranged in the order of the sorted representative vectors, and the addresses arranged in the sorted order are written in the codebook sort table 280.

The first arithmetic unit 240 includes a judgment unit 250
, Receives up / down information indicating the direction of the representative vector to be searched based on the determination result, goes to the codebook sort table 280 for the address of the next representative vector to be searched, and receives the address. Then, the corresponding data (representative vector number i and value w _i ^(P), etc.) is read into the codebook memory 230 by designating the address.

Then, the code book sort table 280 is
The address received from the sorting device 220 is stored. Further, the codebook sort table 280 passes the address of the representative vector to the first arithmetic device 240 in accordance with the instruction from the first arithmetic device 240 to refer to the next representative vector to be searched.

Next, a description will be given of a simulation result of encoding of a still image by the vector quantizer 200.

As test images, tone conversion was performed on the 24 BMP files shown in FIGS. 20, 21, 22 and 23 included in Kodak PHOTOCD.

Full search by LBG algorithm with MM0C operation (R = 1) and each BMP by VQ-SSC
When the gradation conversion simulation was performed on the file, the result shown in FIG. 24 was obtained. FIG. 25A shows the results of this simulation, comparing the execution time (◆) of gradation conversion by full search and the execution time (●) of gradation conversion by VQ-SSC for each BMP file. FIG. 25 (b) shows the execution time of the operation (full search) as an improvement amount of the execution time (VQ-
Numerical values when normalized by the execution time of (SSC) are shown.

According to the simulation results of the gradation conversion shown in FIGS. 24 and 25, the LBG algorithm (full search) converges in 197 repetitions on average and about 2000 seconds, whereas the VQ-SSC If the encoding is performed, it converges in about 200 seconds on average. This execution time is VQ
-It is reduced to an average of 1 / 10.1 compared to the case without SSC.

A full search by the LBG algorithm to which the MMOC operation (R = 30) was added and a simulation of gradation conversion for each BMP file by VQ-SSC were performed, and the results shown in FIG. 26 were obtained. Regarding the simulation result, FIG.
Execution time of gradation conversion by full search for MP file (◆) and execution time of gradation conversion by VQ-SSC (●)
In addition, FIG. 27B shows a numerical value obtained by normalizing the execution time of the operation (full search) with the execution time of the operation (VQ-SSC) as the improvement amount of the execution time. Is shown.

According to the simulation results of the gradation conversion shown in FIGS. 26 and 27, the LBG algorithm (full search) to which the MMOC operation (R = 30) is applied converges in about 250 repetitions on average and about 250 seconds. VQ-
If encoding is performed by SSC, the average value converges in about 33.3 seconds. This execution time is reduced to an average of 1 / 7.6 as compared with the case without VQ-SSC.

Further, the present invention is implemented by a codebook generating apparatus 300 having a configuration as shown in FIG. 28, for example.

The code book generating device 300 includes an input terminal 310, a sorting device 320, a code book memory 330, a first arithmetic device 340, a determining device 350,
It comprises a second arithmetic unit 360 and a control unit 370.

This codebook generating apparatus 300 is a VQ
-Performs competitive learning using SSC encoding.
And input vector x via input terminal 310^(T)
∈R ^KIs input to the first arithmetic unit 340.

The sorting device 320 includes the control device 37
Upon receiving a sort instruction from 0 and an instruction on which dimension k is to be sorted, the code book is read from the code book memory 330, and the code book is sorted based on the specified dimension. Also, the sorting device 32
0 indicates that the sorted code book is stored in the code book memory 3.
Write to the specified 30 addresses.

[0228] codebook memory 330, the codebook _{^{W (t) = {w i}} (t) ∈R K: i = 0, ···, N
-1} is a memory that receives an address from the first arithmetic unit 340 that performs distortion calculation and address calculation, and corresponding representative vector data (representative vector value w _i ^(t) , number i) To the first arithmetic unit 340. Further, the codebook memory 330 receives the data (the updated representative vector value w _min ^{(t + 1)} ) and the write address from the second arithmetic unit 360 that performs the center of gravity calculation and the average strain calculation, and stores the corresponding address in the corresponding address. Codebook is written. Further, the codebook memory 330 transfers the codebook to the sorting device 320. Further, the code book memory 330 receives the code book sorted by the sorting device 320 and the address to be written, and stores the code book at the specified address.

The first arithmetic unit 340 has an input terminal 310
Via the input vector x ^(t) . The first arithmetic unit 340 receives the instruction of calculating the distortion / address from the control unit 370, specifies the address of the representative vector to be read into the codebook memory 330, and specifies the corresponding data (representative vector value w _i ^{( t)} , number i, etc.)
Read. Further, the first arithmetic unit 340 calculates the distortion d (x ^(t) , w _i between the input vector and the representative vector.
^(T) ) is calculated. Then, the first arithmetic unit 340
The strain _{^{d (x (t), w}} i (t)) and data pass (number and value of the representative vector) to the determination unit 350. Also,
The first arithmetic unit 340 receives an instruction of the dimension k serving as a sorting criterion from the control unit 370, and also transfers the k-dimensional distortion d (x _k ^(t) , w _ik ^(t) ) to the determination unit 350. . Further, the first arithmetic unit 340 includes a determination unit 350
, Receives up / down information indicating the direction of the representative vector to be searched based on the determination result, and calculates the address of the next representative vector to be searched. Note that the first arithmetic unit 34
When the information “finish” indicating the end of the search is received, the encoding of the next input vector x ^{(t + 1)} is started.

The judgment device 350 calculates the current minimum distortion d.
^cur _min and the number min of the representative vector are held, and mi having the current minimum distortion d ^cur _min
The distortion d for the n-th representative vector and the i-th representative vector calculated by the first arithmetic unit 340
Compares the _{^{(x (t), w i}} (t)), the strain is the smaller of the new min number, the current minimum distortion ^d _{cur min}
And Further, the determination device 350 receives a determination instruction from the control device 370 and receives a k-dimensional distortion d.
(X _k ^(t) , w _ik ^(t) ) and the current minimum strain d
^cur _min , d (x _k ^(t) ,
If w _ik ^(t) is larger, the information of up / down indicating the direction of the next representative vector to be searched is changed, and if both up / down directions have been checked, information indicating the end of search Is passed to the first arithmetic unit 340, and if d (x _k ^{(t )} , w _ik ^(t) ) is smaller, information for examining the next representative vector in the same direction is passed to the first arithmetic unit 340. Further, the determination device 350 determines that the up /
When the search is completed in both the down and search directions, the input vector x ^(t) is input to the second arithmetic unit 360.
Data of the representative vector having the current minimum distortion (number min and value w _min ^(t) ) and minimum distortion d
Output ^cur _min . Further, the determination device 350
When the search is finished, the number min of the representative vector having the current minimum distortion is set to the next input vector x
The second arithmetic unit 3 as an initial representative vector of ^{(t + 1)}
Pass to 60.

[0231] The second arithmetic unit 360 includes a judging device 350.
From the input vector x ^(t) and the data of the representative vector having the current minimum distortion (number min and value w _min
^{(T )} ) and the minimum distortion d ^cur _min . The second computing device 360 calculates the average distortion D
The calculation of _sum is performed according to the following equation (58).

D _sum + = d ^cur _min (58) Further, the second arithmetic unit 360 updates the value of the representative vector according to the following equation (59).

W _min ^{(t + 1)} = w _min ^(t) + η ^(t) [x ^(t) −w _min ^{(t )} ] (59) The second arithmetic unit 360 is a representative of the codebook memory 330. The updated representative vector value w _min ^{(t + 1)} is written to the address of the vector number min.

[0234] When the arithmetic unit 360 receives an instruction from the control unit 370 that the encoding has been completed for the T input vectors and the intermediate calculation has been completed, the arithmetic unit 360 calculates the average distortion D as follows. This is performed according to equation (60).

[0235]

[Equation 17]

The second arithmetic unit 360 receives data (average distortion D) according to an instruction from control unit 370.
_sum ) is cleared.

Further, the second arithmetic unit 360 sets an arbitrary initial code book in the code book memory 330 according to an instruction from the control unit 370.

The control device 370 controls the sorting device 32
0, the processing flow of the first arithmetic unit 340, the determination unit 350, and the second arithmetic unit 360 is controlled. Further, the control device 370 specifies a dimension serving as a sorting reference to the sorting device 320. Further, the control device 370 specifies the address (or number) of the initial representative vector to the first arithmetic device 340, and indicates the dimension used as the reference for sorting. Further, the control device 370 controls the second arithmetic device 36
Instruct 0 to set an arbitrary initial codebook.

The code book generating apparatus 3 having such a configuration
At 00, the present invention generates a codebook according to the encoding algorithm shown in the flowchart of FIG.

That is, the code book generating device 30
In the case of 0, first, in step (F301), a representative vector initial arrangement is set. Also, a reference dimension k for sorting the representative vectors is specified.

In the next step (F302), the representative vectors are sorted based on the value of the representative vector of dimension k. From the second time on, the updated representative vector w
Only ^{(t + 1)} is corrected to enter the corresponding location.
Regarding the information of the sorted order, the following two methods can be considered.

(1) The order of the representative vectors is directly changed. If it is not desired to change the number of the representative vector, the structure is replaced with a structure having information on the number of the representative vector. Since the numbers are directly changed, the number lower of the representative vector lower than the own is the number-1 of the own array.
And the sequence number. Also, the number upper of the representative vector higher than the own is the array number obtained by adding +1 to the array number of the own.

(2) In the case of preparing an array indicating the order of sorting, the information to be held by each representative vector is represented by the lower number of the representative vector lower than the own and the upper number of the representative vector higher than the own. is there.

In the next step (F303), the distortion between the input vector x ^(t) and the initial representative vector w _i ^(t) is calculated, and the distortion d (x ^(t) ,
and w _i ^(t)) of the current minimum distortion ^d _{cur min,} keep the number i as a candidate for coding.

[0245] _{^{d cur min = d (x (}} t), w i (t)) (61) min = i (62) In the next step (F304), the initial representative vector w
The number of the representative vector on both sides of _i ^(t) , that is, lower = the number of the next representative vector in the lower direction, and upper = the number of the next representative vector in the upper direction.

At the next step (F305), the input vector
Torr x^(T)And representative vector w_{lo wer} ^(T)Tonohi
Z d (x^(T), W_lower ^(T)) Calculation
D (x_k ^(T),
w_lowerk ^(T)) Is saved. And this
In step (F305), the calculated strain d (x_k
^(T), W_lowerk ^(T)) Is the current minimum strain
d^cur _minIs smaller than, that is, d (x^(T), W_lower ^(T)) <D^cur _min It is determined whether or not (63) holds, and the equation (63) holds.
Then d^cur _min= D (x^(T), W_lower ^(T)) (64) min = lower (65)

In the next step (F306), it is determined whether to continue the search in the lower direction. This step (F3
At 06), if d (x _k ^(t) , w _lowerk ^(t) )> d ^cur _min (66) is satisfied, the search in the lower direction is completed, and the process proceeds to step (F308). , Next step (F
Proceed to 307).

The step (F307) is a step of updating the number “lower” of the representative vector.
The number of the representative vector next to the r-th representative vector in the lower direction is set as a new lower, and the process returns to step (F305).

In the step (F308), the input vector x ^(t) and the representative vector w _{upper of the} upper number are determined.
^(T) and the strain of _{^{d (x (t), w}} upper (t))
Is calculated, and d (x _k ^(t) , w
_upperk ^(t) ) is saved. In this step (F308), the calculated strain d (x ^(t) , w
_upper ^{( t)} ) is the current minimum distortion d ^cur
_{It is} determined whether or not d (x ^(t) , w _upper ^(t) ) <d ^cur _min (67). If this equation (67) holds, then d ^cur _min = d (x ^(t) , w _upper ^(t) ) (68) min = upper (69)

In the next step (F309), it is determined whether to continue the search in the upper direction. This step (F3
In 09), if d (x _k ^(t) , w _upperk ^(t) )> d ^cur _min (70) is satisfied, the search in the upper direction is completed, and the process proceeds to step (F311). , Next step (F
Proceed to 310).

Step (F310) is a step of updating the representative vector number upper, and the current upper
The number of the representative vector next to the numbered representative vector in the upper direction is set as a new upper, and the process returns to step (F308).

In the next step (F311), the input vector is encoded, the center of gravity of the representative vector is calculated, and the average distortion is prepared. In this step (F311), the input vector x ^(t) is encoded into the representative vector w _min ^(t) of the number min having the minimum distortion d ^cur _min . In this step (F311), the value of the encoded representative vector is updated according to the following equation (71).

W _min ^{(t + 1)} = w _min ^(t) + η ^(t) [x ^(t) −w _min ^{(t )} ] (71) In this step (F311), the average strain D
The calculation of _sum is performed according to the following equation (72).

D _sum + = d ^cur _min (72) When t = T−1, the average strain D is calculated according to the following equation (73).

[0255]

(Equation 18)

In the next step (F312), it is determined whether or not encoding of all input vectors has been completed. In this step (F312), if the data is completed, the processing is terminated; otherwise, the next step (F312)
Proceed to 13).

In step (F313), t = t + 1 is set, and the process returns to step (F302) to search for the next input vector.

Further, according to the present invention, the partial sum of the distortion represented by the above-mentioned equation (27) is equal to the current minimum distortion d ^cur.
When the value is larger than _min , a further effect is obtained by combining a high-speed method using a partial sum of distortions for searching for the next representative vector.

Here, only a difference will be described with reference to FIG. 30 as a modification of the codebook generating apparatus 300 shown in FIG. 28 described above.

The code book generating device 30 shown in FIG.
At 0 ', codebook memory 330 passes the codebook and address to sorting device 320.

The sorting device 320 includes the control device 37
Upon receiving a sort instruction from 0 and an instruction on which dimension k to sort on, the code book address and code book are read from the code book memory 330, and the code book is sorted based on the specified dimension. Then, the addresses of the representative vectors are rearranged in the order of the sorted representative vectors, and the addresses arranged in the sorted order are written in the codebook sort table 380.

The first arithmetic unit 340 includes a judging device 350
, Receives up / down information indicating the direction of the representative vector to be searched based on the determination result, refers to the codebook sort table 380 for the address of the next representative vector to be searched, and receives the address. Then, the address is specified and the corresponding data (representative vector number i and value w _i ^(t), etc.) is read into the code book memory 330.

The code book sort table 380 is
The address received from the sorting device 320 is stored. Further, the codebook sort table 380 passes the address of the representative vector to the first arithmetic device 340 in accordance with the instruction from the first arithmetic device 340 to refer to the next representative vector to be searched.

Further, the present invention is implemented by a codebook generating apparatus 400 having a configuration as shown in FIG. 31, for example.

The code book generation device 400 includes an input terminal 410, a sorting device 420, a code book memory 430, a first arithmetic device 440, a determination device 450,
Second computing device 460, partial strain sorting device 4
62, a partial distortion comparing device 464, a representative vector moving device 466, and a control device 470.

This codebook generating apparatus 400 is a VQ
A performs a competitive learning plus MMOC operation using the encoding by -SSC, the input vector ^{x (t)} ∈R ^K through the input terminal 410 is input to the first arithmetic unit 440.

The sorting device 420 includes the control device 47
Upon receiving a sorting instruction from 0 and an instruction on which dimension k is to be sorted, the code book is read from the code book memory 430 and the code book is sorted based on the designated dimension. Also, the sorting device 42
0 indicates that the sorted code book is stored in the code book memory 4.
Write to the specified 30 addresses.

[0268] codebook memory 430, the codebook _{^{W (t) = {w i}} (t) ∈R K: i = 0, ···, N
-1} is a memory that receives an address from the first arithmetic unit 440 that performs distortion calculation and address calculation, and corresponding representative vector data (representative vector value w _i ^(t) , number i) To the first arithmetic unit 440. Further, the codebook memory 430 receives the data (updated codebook W ^(T) ) and the write address from the representative vector moving device 466, and writes the codebook to the corresponding address. In addition, the codebook memory 430 receives data (updated representative vector value w _min) from the second arithmetic unit 460 that performs representative vector calculation, average strain calculation, and partial strain calculation.
^{(T + 1)} ) and the write address are received, and the codebook is written to the corresponding address. Also, the codebook memory 430 transfers the codebook to the sorting device 420. Further, the codebook memory 430
Receives the code book sorted by the sorting device 420 and the address to be written, and stores the code book at the specified address.

The first arithmetic unit 440 has an input terminal 410
Via the input vector x ^(t) . The first arithmetic unit 440 receives an instruction for calculating a distortion / address from the control unit 470, specifies the address of a representative vector to be read into the codebook memory 430, and specifies the corresponding data (representative vector value w _i ^{( t)} , number i, etc.)
Read. The first arithmetic unit 440, distortion ^d between the input vector the representative vector ^{(x (t),} _{w i}
^(T) ) is calculated. Then, the first arithmetic unit 440
The strain _{^{d (x (t), w}} i (t)) and data pass (number and value of the representative vector) to the determination unit 450. Also,
The first arithmetic unit 440 receives an instruction of the dimension k serving as a sorting criterion from the control unit 470, and also passes the distortion d (x _k ^(t) , w _ik ^(t) ) of only k dimensions to the determination unit 450. . Further, the first arithmetic unit 440 includes a determination unit 450
, Receives up / down information indicating the direction of the representative vector to be searched based on the determination result, and calculates the address of the next representative vector to be searched. Note that the first arithmetic unit 44
When the information “finish” indicating the end of the search is received, the encoding of the next input vector x ^{(t + 1)} is started.

The judgment device 450 determines the current minimum distortion d.
^cur _min and the number min of the representative vector are held, and mi having the current minimum distortion d ^cur _min
The distortion d for the n-th representative vector and the i-th representative vector calculated by the first arithmetic unit 440
Compares the _{^{(x (t), w i}} (t)), the strain is the smaller of the new min number, the current minimum distortion ^d _{cur min}
And Further, the determination device 450 receives a determination instruction from the control device 470 and receives a k-dimensional distortion d.
(X _k ^(t) , w _ik ^(t) ) and the current minimum strain d
^cur _min , d (x _k ^(t) ,
If w _ik ^(t) is larger, the information of up / down indicating the direction of the next representative vector to be searched is changed, and if both up / down directions have been checked, information indicating the end of search Is passed to the first arithmetic unit 440, and if d (x _k ^{(t )} , w _ik ^(t) ) is smaller, information for examining the next representative vector in the same direction is passed to the first arithmetic unit 440. Further, the determination device 450 determines whether the up /
When the search is finished in both the down and search directions, the input vector x ^(t) ,
Data of the representative vector having the current minimum distortion (number min and value w _min ^(t) ) and minimum distortion d
Output ^cur _min . Further, the determination device 450
When the search is finished, the number min of the representative vector having the current minimum distortion is set to the next input vector x
The second arithmetic unit 4 as an initial representative vector of ^{(t + 1)}
Pass to 60.

The second arithmetic unit 460 includes a judgment unit 450
From the input vector x ^(t) and the data of the representative vector having the current minimum distortion (number min and value w _min
^{(T )} ) and the minimum distortion d ^cur _min . Then, the second arithmetic unit 460 _generates data (partial distortion D _minsum ) corresponding to the representative vector number min.
Is calculated according to the following equation (74).

D _minsum + = d ^cur _min (74) The second arithmetic unit 460 updates the value of the representative vector according to the following equation (75).

W _min ^{(t + 1)} = w _min ^(t) + η ^(t) [x ^(t) −w _min ^{(t )} ] (75) The second arithmetic unit 460 is a representative of the codebook memory 330. The updated representative vector value w _min ^{(t + 1)} is written to the address of the vector number min.

When the arithmetic unit 460 receives an instruction from the control unit 470 that the encoding has been completed for the T input vectors and the intermediate calculation has been completed, the following (7)
6) Calculate the partial strain D _i and the average strain D according to the equations (77) and (77).

[0275]

[Equation 19]

Further, the second arithmetic unit 460 controls all data (partial distortion D
clear _isum , ∀i).

The partial distortion sorting device 462 receives data (updated codebook and partial distortion) from the second arithmetic unit 460. Further, the partial strain sorting device 462 specifies an address from the codebook memory 430 and stores the corresponding data (codebook W).
^(T) ). This partial strain sorting device 4
62 sorts the codebook and the partial distortion based on the magnitude of the partial distortion. Further, the partial strain sorting device 462 includes a partial strain comparison device 464.
To the data (sorted codebook and partial distortion)
give.

The partial distortion comparing device 464 receives data (sorted codebook and partial distortion) from the partial distortion sorting device 462. This partial strain comparison device 464 compares a large partial strain with a small partial strain. When the disparity is large, information for moving a representative vector having a small partial strain to a representative vector having a large partial strain is passed to the representative vector moving device 466. Further, the data (the movement information and the codebook W ^(T) ) are passed to the representative vector movement device 466.

The representative vector moving unit 466 receives, from the partial distortion comparing unit 464, information for moving the representative vector having the small partial distortion to the vicinity of the representative vector having the large partial distortion and the codebook W ^(T) . The representative vector moving device 466 moves the representative vector having the small partial distortion to the vicinity of the representative vector having the large partial distortion based on the movement information. Then, the representative vector moving device 466 stores the data (the updated codebook W) in the codebook memory 430.
^(T) ), and the updated codebook is written to the corresponding address. Further, the representative vector moving device 466 stores an arbitrary initial codebook in the codebook memory 4 according to an instruction from the control device 470.
Set to 30.

The control device 470 controls the sorting device 42
0, first computing device 440, determining device 450, second computing device 460, partial strain sorting device 462, partial strain comparing device 464, and representative vector moving device 46
6 is controlled. Also, the control device 470
Instruct the representative vector moving device 466 to set an arbitrary initial codebook. In addition, control device 470 designates sorting device 420 as a dimension serving as a reference for sorting.
Further, the control device 470 specifies the address (or number) of the initial representative vector to the first arithmetic device 440, and indicates the dimension used as the reference for sorting.

The code book generating apparatus 4 having such a configuration is described.
In the present invention, in the present invention, Mi
A codebook is generated according to the flowchart shown in FIG. 32 by competitive learning with an n-Max Order Compare (MMOC) operation.

That is, the code book generating device 40
In the case of 0, first, in step (F401), a representative vector initial arrangement is set. Also, a reference dimension k for sorting the representative vectors is specified. In addition, MMO
The parameters R and m of the C operation are set.

In the next step (F402), the representative vectors are sorted based on the value of the representative vector of dimension k. From the second time on, the updated representative vector w
Only ^{(t + 1)} is corrected to enter the corresponding location.
Regarding the information of the sorted order, the following two methods can be considered.

(1) The order of the representative vectors is directly changed. If it is not desired to change the number of the representative vector, the structure is replaced with a structure having information on the number of the representative vector. Since the numbers are directly changed, the number lower of the representative vector lower than the own is the number-1 of the own array.
And the sequence number. Also, the number upper of the representative vector higher than the own is the array number obtained by adding +1 to the array number of the own.

(2) In the case of preparing an array indicating the order of sorting, the information to be held by each representative vector is represented by the lower number of the representative vector lower than the own and the upper number of the representative vector higher than the own. is there.

In the next step (F403), the distortion between the input vector x ^(t) and the initial representative vector w _i ^(t) is calculated, and the distortion d (x ^(t) ,
and w _i ^(t)) of the current minimum distortion ^d _{cur min,} keep the number i as a candidate for coding.

[0287] _{^{d cur min = d (x (}} t), w i (t)) (78) min = i (79) In the next step (F404), the initial representative vector w
The number of the representative vector on both sides of _i ^(t) , that is, lower = the number of the next representative vector in the lower direction, and upper = the number of the next representative vector in the upper direction.

In the next step (F405), the input vector
Torr x^(T)And representative vector w_{lo wer} ^(T)Tonohi
Z d (x^(T), W_lower ^(T)) Calculation
D (x_k ^(T),
w_lowerk ^(T)) Is saved. And this
In step (F405), the calculated strain d (x_k
^(T), W_lowerk ^(T)) Is the current minimum strain
d^cur _minIs smaller than, that is, d (x^(T), W_lower ^(T)) <D^cur _min It is determined whether or not (80) holds, and this (80) equation holds.
Then d^cur _min= D (x^(T), W_lower ^(T)) (81) min = lower (82)

In the next step (F406), it is determined whether to continue the search in the lower direction. This step (F4
At 06), if d (x _k ^(t) , w _lowerk ^(t) )> d ^cur _min (83) is satisfied, the search in the lower direction is completed, and the process proceeds to step (F408). , Next step (F
Go to 407).

The step (F407) is a step of updating the lower number of the representative vector, and the current lower
The number of the representative vector next to the r-th representative vector in the lower direction is set as a new lower, and the process returns to step (F405).

In the step (F408), the input vector x ^(t) and the representative vector w _{upper of the} upper number are determined.
^(T) and the strain of _{^{d (x (t), w}} upper (t))
Is calculated, and d (x _k ^(t) , w
_upperk ^(t) ) is saved. In this step (F408), the calculated strain d (x ^(t) , w
_upper ^{( t)} ) is the current minimum distortion d ^cur
_{It is} determined whether or not d (x ^(t) , w _upper ^(t) ) <d ^cur _min (84) holds. If this equation (84) holds, then d ^cur _min = d (x ^(t) , w _upper ^(t) ) (85) min = upper (86)

In the next step (F409), it is determined whether to continue the search in the upper direction. This step (F4
In 09), if d (x _k ^(t) , w _upperk ^(t) )> d ^cur _min (87) is satisfied, the search in the upper direction is completed, and the process proceeds to step (F411). , Next step (F
Go to 410).

Step (F410) is a step of updating the representative vector number upper, and the current upper
The number of the representative vector next to the numbered representative vector in the upper direction is set as a new upper, and the process returns to step (F408).

At the next step (F411), the input vector
Encoding and updating representative vectors and preparing mean distortion
I do. In this step (F411), the input
Kutor x^(T)Is the minimum strain d^cur _minhave
Representative vector w of number min _min ^(T)Encoded into
You. In this step (F411), the encoded representative
The value of the vector is updated according to the following equation (88).

W _min ^{(t + 1)} = w _min ^(t) + η ^(t) [x ^(t) −w _min ^{(t )} ] (88) In this step (F411), the partial distortion D
The calculation of _minsum is performed according to the following equation (89).

[0296] ^{_{D minsum + = d cur min (}} 89) Then, when t = T-1 is satisfied, the calculation of partial strain _{D i} and the average strain D. Also, the partial strain D
_i is calculated by the following equation (90).

[0297]

(Equation 20)

The average strain D is calculated by the following equation (91).

[0299]

(Equation 21)

In the next step (F412), it is determined whether or not encoding of all input vectors has been completed. In this step (F412), if the data is completed, the processing is terminated; otherwise, the next step (F4)
Proceed to 13).

Step (F413) is a step for determining whether or not to enter the MMOC mode. If t = T-1 does not hold, the flow advances to step (F414).
When t = T−1 holds, step (F415)
Proceed to.

In step (F414), t = t + 1, and the flow returns to step (F402) to search for the next input vector.

At step (F415), the representative vectors and the partial distortions are sorted according to the partial distortions.

Then, in the next step (F416) and step (F417), the MMOC operation is executed.

That is, in the step (F416), as shown in FIG.
R partial strains D in ascending order of _max [r]
Compare _min [r].

As a result of the comparison in step (F306), when the following equation (92) does not hold, the procedure returns to step (F402), and when the equation (92) holds, the procedure proceeds to step (F417).

D _max [r]> mxD _min [r], (r = 0,..., R−1) (92) Further, in step (F417), as shown in the following equation (93): In addition, the representative vector w _min ^(T) [r] having the partial distortion D _min [r] in the smaller order is replaced with the representative vector w _max having the partial distortion D _max [r] in the larger order.
^(T) Rearranged near [r], and step (F40)
Return to 2).

W _min ^(T) [r] ← w _max ^(T) [r] + ε (93) where ε (≪1) is a perturbation term. Up to R movements are possible in one iteration.

Also, according to the present invention, the partial sum of the distortion represented by the above equation (27) is equal to the current minimum distortion d ^cur.
_{When min} is larger than _min , a further effect is obtained by combining a high-speed technique using a partial sum of distortions for searching for the next representative vector.

Here, only a difference will be described with reference to FIG. 33 as a modified example of the code book generating apparatus 400 shown in FIG.

The code book generating device 40 shown in FIG.
At 0 ', codebook memory 430 passes the codebook and address to sorting device 420.

[0312] The sorting device 420 includes the control device 47.
Upon receiving a sort instruction from 0 and an instruction on which dimension k is to be sorted, the code book address and code book are read from the code book memory 430, and the code book is sorted based on the specified dimension. Then, the addresses of the representative vectors are rearranged in the order of the sorted representative vectors, and the addresses arranged in the sorted order are written in the codebook sort table 480.

[0313] The first arithmetic unit 440 includes a judging unit 450.
, Receives up / down information indicating the direction of the representative vector to be searched based on the determination result, goes to the codebook sort table 480 for the address of the next representative vector to be searched, and receives the address. Then, the address is passed to the codebook memory 430, and the corresponding data (representative vector number i and value w _i ^(t), etc.) is read.

The codebook sort table 480 is
The address received from sorting device 420 is stored. In addition, the codebook sort table 480 passes the address of the representative vector to the first arithmetic device 440 according to the instruction from the first arithmetic device 440 to refer to the next representative vector to be searched.

[0315]

As described above, according to the present invention,
By searching only a part of the representative vectors, the same representative vector as in the full search can be selected at high speed. Further, according to the present invention, the number of memories to be stored can be reduced as compared with the conventional method. Further, according to the present invention, as compared with the conventional method, since the sorting only needs to be performed for any one dimension, the time can be reduced. Further, according to the present invention, as compared with the conventional method, there is no need for a memory for storing the order of the sorted representative vectors. According to the present invention,
Compared with the conventional method, a memory for a flag indicating whether or not the search for the representative vector has been completed is not required, and no time is required to clear the flag for each input vector. Further, according to the present invention, the method of determining the next representative vector to be searched is uniquely determined.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a codebook generation device to which the present invention has been applied.

FIG. 2 is a flowchart showing an LBG algorithm process using VQ-SSC, which is executed in the codebook generation device.

FIG. 3 is a diagram showing an example of a search for a representative vector by a full search.

FIG. 4 is a diagram illustrating a search example (search 0) of a representative vector by the encoding algorithm (VQ-SSC).

FIG. 5 is a diagram showing an example of searching for a representative vector (search 1).

FIG. 6 is a diagram showing a search example (search 2) of a representative vector.

FIG. 7 is a diagram showing an example of searching for a representative vector (search 3).

FIG. 8 is a diagram showing an example of searching for a representative vector (search 4).

FIG. 9 is a diagram showing an example of searching for a representative vector (search 5).

FIG. 10 is a diagram showing an example of searching for a representative vector (search 6).

FIG. 11 is a diagram showing a search range of the encoding algorithm (VQ-SSC).

FIG. 12 is a block diagram showing a configuration of a modified example of the codebook generation device shown in FIG.

FIG. 13 is a diagram illustrating the equal distortion principle.

FIG. 14 is a block diagram illustrating a configuration of a codebook generation device to which the present invention has been applied.

FIG. 15 shows a VQ-SS in the codebook generating apparatus.
Min-Max Order Comp based on equal strain principle using C
It is a flowchart which shows the LBG algorithm which added the are (MMOC) operation.

FIG. 16 is a diagram schematically illustrating a comparison process of sorted partial distortion in ascending order and partial distortion in descending order in the LBG algorithm.

FIG. 17 is a diagram showing L obtained by adding an MMOC operation when R = 1.
It is a conceptual diagram of operation | movement of a BG algorithm.

FIG. 18 is a view showing L obtained by adding an MMOC operation when R = 1;
It is a figure which shows the actual operation | movement of a BG algorithm typically.

FIG. 19 is a block diagram showing a configuration of a modification of the vector quantizer shown in FIG.

FIG. 20 is a diagram showing six Kodak PHOTOCDs (24).

FIG. 21 is a diagram showing six Kodak PHOTOCDs (24).

FIG. 22 is a diagram showing six Kodak PHOTOCDs (24).

FIG. 23 is a diagram showing six Kodak PHOTOCDs (24).

FIG. 24: Full search by LBG algorithm with MMOC operation (R = 1) and Kodak PHOT by VQ-SSC
FIG. 14 is a diagram illustrating a simulation result of gradation conversion for OCD.

FIG. 25 is a diagram showing a comparison between the execution time (◆) of gradation conversion by full search and the execution time (●) of gradation conversion by VQ-SSC and the amount of improvement in execution time for the simulation results.

FIG. 26: Full search by LBG algorithm with MMOC operation (R = 30) and Kodak PH by VQ-SSC
FIG. 14 is a diagram illustrating a simulation result of gradation conversion for OTOCD.

FIG. 27 is a diagram showing, for the simulation results, a comparison between the execution time (◆) of gradation conversion by full search and the execution time (●) of gradation conversion by VQ-SSC and the amount of improvement in execution time.

FIG. 28 is a block diagram illustrating a codebook generation device to which the present invention has been applied.

FIG. 29 is a flowchart showing a competitive learning process using VQ-SSC executed in the codebook generation device.

FIG. 30 is a block diagram showing a configuration of a modification of the codebook generation device shown in FIG. 29.

FIG. 31 is a block diagram showing a codebook generation device to which the present invention is applied.

FIG. 32 is a flowchart showing a competitive learning process using VQ-SSC executed in the codebook generating device.

FIG. 33 is a block diagram showing a configuration of a modified example of the codebook generation device shown in FIG. 31.

FIG. 34 is a conceptual diagram of vector quantization.

[Explanation of symbols]

100, 100 ', 200, 200' Codebook generation device, 110, 210 Input vector set memory, 1
20, 220 sorting device, 130, 230 codebook memory, 140, 240 first computing device,
150, 250 determining device, 160, 260 second computing device, 170, 270 control device, 180, 280 codebook sort table, 262 partial strain sorting device, 264 partial strain comparing device, 266 representative vector moving device, 300, 300 ', 400,400'
Codebook generation device, 110, 210 Input vector set memory, 120, 220 Sorting device, 31
0,410 input terminal, 320,420 sorting device, 330,430 codebook memory, 340,
440 first arithmetic unit, 350, 450 determination unit,
360, 460 second arithmetic unit, 370, 470 control unit, 380, 480 codebook sort table, 462
Partial strain sorting device, 464 Partial strain comparison device, 466 Representative vector moving device

 ──────────────────────────────────────────────────続 き Continued on the front page F term (reference) 5D015 GG00 5D045 DA11 5J064 AA03 BA13 BB14 BC01 BC14 BC21 BD01

Claims (32)

[Claims] A sorting step of sorting the representative vectors based on the values of the representative vectors; a first calculating step of calculating a distortion of the input vector and the representative vector, and calculating an address of a representative vector to be searched next; A determining step of comparing two distortions obtained in the first operation step to determine a candidate for a representative vector to be encoded; and calculating a center of gravity of the representative vector and calculating an average distortion based on the determination result in the determining step And a second calculation step of performing the above-mentioned steps, and repeatedly performing the above steps. 2. The method according to claim 1, wherein in the determining step, a candidate of a representative vector to be encoded is determined by comparing two distortions obtained in the first operation step, and a partial sum of the distortions is compared. The method of generating a codebook for vector quantization according to claim 1. 3. The method according to claim 1, further comprising a storage step of storing an arrangement order of the codebooks sorted in the sorting step. 4. The codebook for vector quantization according to claim 1, wherein a representative vector in which a previous input vector is encoded is designated as an initial representative vector. Generation method. 5. The vector quantization code according to claim 1, wherein a representative vector coded at the time of the previous iteration is designated as the initial representative vector. Book generation method. 6. A sorting step for sorting the representative vectors based on the values of the representative vectors, calculating a distortion between the input vector and the plurality of representative vectors in parallel, and calculating an address of a representative vector to be searched next. And a determining step of comparing a plurality of distortions obtained in the first calculating step to determine a candidate for a representative vector to be encoded; and calculating a center of gravity of the representative vector based on the determination result in the determining step. A second calculation step for calculating an average distortion; and a codebook generation method for vector quantization, wherein the above steps are repeatedly performed. 7. A sorting step of sorting the representative vectors based on the values of the representative vectors, a first operation step of calculating a distortion of the input vector and the representative vector, and calculating an address of a representative vector to be searched next; A determination step of comparing two distortions obtained in the first operation step to determine a candidate of a representative vector to be encoded; and calculating a center of gravity based on a determination result in the determination step.
A second operation step of performing average strain calculation / partial strain calculation, a partial strain sorting step of sorting based on the partial strain obtained by the second operation step, and two parts sorted by the partial strain sorting step A comparison step of comparing the disparity of distortion; and a representative vector moving step of moving a representative vector having a small partial strain to a vicinity of a representative vector having a large partial distortion based on a comparison result in the comparing step. , And a codebook generation method for vector quantization. 8. A sorting step for sorting the representative vectors based on the values of the representative vectors, calculating a distortion between the input vector and the plurality of representative vectors in parallel, and calculating an address of a representative vector to be searched next. A determination step of comparing a plurality of distortions obtained in the first calculation step to determine a candidate for a representative vector to be encoded; and calculating a center of gravity based on a determination result of the determination step.
A second operation step of performing average strain calculation / partial strain calculation, a partial strain sorting step of sorting based on the partial strain obtained by the second operation step, and two parts sorted by the partial strain sorting step A comparison step of comparing the disparity of distortion; and a representative vector moving step of moving a representative vector having a small partial strain to a vicinity of a representative vector having a large partial distortion based on a comparison result in the comparing step. , And a codebook generation method for vector quantization. 9. A sorting step for sorting the representative vectors based on the values of the representative vectors, a first operation step for calculating a distortion of the input vector and the representative vector, and calculating an address of the representative vector to be searched next; A determination step of comparing two distortions obtained in the first operation step to determine a candidate of a representative vector to be encoded; and calculating a center of gravity based on a determination result in the determination step.
A second operation step of performing average strain calculation / partial strain calculation, a partial strain sorting step of sorting on the basis of the partial strain obtained in the second operation step, and a plurality of pieces sorted by the partial strain sorting step A comparison step of comparing differences in partial strains, and a representative vector moving step of moving a plurality of representative vectors having small partial strains to a vicinity of a representative vector having large partial strains based on a comparison result in the comparison step, A method of generating a codebook for vector quantization, comprising repeating the above steps. 10. A sorting step for sorting a representative vector based on a value of a representative vector, calculating a distortion between an input vector and a plurality of representative vectors in parallel, and calculating an address of a representative vector to be searched next. A determination step of comparing a plurality of distortions obtained in the first calculation step to determine a candidate for a representative vector to be encoded; and calculating a center of gravity based on a determination result of the determination step.
A second operation step of performing average strain calculation / partial strain calculation, a partial strain sorting step of sorting on the basis of the partial strain obtained in the second operation step, and a plurality of pieces sorted by the partial strain sorting step A comparison step of comparing differences in partial strains, and a representative vector moving step of moving a plurality of representative vectors having small partial strains to a vicinity of a representative vector having large partial strains based on a comparison result in the comparison step, A method of generating a codebook for vector quantization, comprising repeating the above steps. 11. A sorting step for sorting the representative vectors based on the values of the representative vectors, and a first operation for calculating the distortion of the input vector and the representative vector of the codebook and calculating the address of the representative vector to be searched next. A step of comparing the two distortions obtained in the first operation step to determine a candidate for a representative vector to be encoded; and an adaptive representative vector A second calculation step of calculating and calculating average distortion, wherein the above steps are repeatedly performed, and a codebook generation method for vector quantization is provided. 12. A sorting step of sorting the representative vectors based on the values of the representative vectors, calculating a distortion between the input vector and a plurality of representative vectors of the codebook in parallel, and calculating an address of a representative vector to be searched next. A first operation step of performing a determination, a plurality of distortions obtained in the first operation step are compared to determine a candidate of a representative vector to be encoded, and a determination is made based on a result of the determination step. A second operation step of calculating a typical representative vector and calculating an average distortion, and repeatedly performing each of the above steps. 13. A sorting step for sorting the representative vectors based on the values of the representative vectors, a first operation for calculating a distortion of the input vector and the representative vector of the codebook, and calculating an address of a representative vector to be searched next. A step of comparing the two distortions obtained in the first operation step to determine a candidate for a representative vector to be encoded; and an adaptive representative vector A second operation step of performing calculation / average strain calculation / partial strain calculation; a partial strain sorting step of sorting on the basis of the partial strain obtained by the second operation step; Comparing the difference between the two partial strains, A representative vector moving step of moving a representative vector having a small partial distortion to a vicinity of a representative vector having a large partial distortion based on a comparison result in the comparing step, wherein the above steps are repeatedly performed. Codebook generation method. 14. A sorting step for sorting the representative vectors based on the values of the representative vectors, calculating a distortion between the input vector and a plurality of representative vectors of the codebook in parallel, and calculating an address of a representative vector to be searched next. A first operation step of performing a determination, a plurality of distortions obtained in the first operation step are compared to determine a candidate of a representative vector to be encoded, and a determination is made based on a result of the determination step. A second calculation step of calculating a typical representative vector, calculating an average strain, and calculating a partial strain; a partial strain sorting step of sorting on the basis of the partial strain obtained in the second calculation step; A comparison step for comparing the difference between two partial strains sorted by the sorting step; And a representative vector moving step of moving a representative vector having a small partial distortion to a vicinity of a representative vector having a large partial distortion based on a comparison result in the comparing step, wherein the steps are repeatedly performed. To generate a codebook for vector quantization. 15. A sorting step for sorting the representative vectors based on the values of the representative vectors, a first operation for calculating a distortion of the input vector and the representative vector of the codebook, and calculating an address of a representative vector to be searched next. A step of comparing the two distortions obtained in the first operation step to determine a candidate for a representative vector to be encoded; and an adaptive representative vector A second operation step of performing calculation, calculation of average strain, and partial strain calculation; a partial strain sorting step of sorting based on the partial strain obtained in the second operation step; Comparing the disparity of the plurality of partial strains, A representative vector moving step of moving a plurality of representative vectors having a small partial distortion to the vicinity of the representative vector having a large partial distortion based on the comparison result in the comparing step, wherein the above steps are repeatedly performed. Codebook generation method for quantization. 16. A sorting step for sorting the representative vectors based on the values of the representative vectors, calculating a distortion between the input vector and a plurality of representative vectors of the codebook in parallel, and calculating an address of a representative vector to be searched next. A first operation step of performing a determination, a plurality of distortions obtained in the first operation step are compared to determine a candidate of a representative vector to be encoded, and a determination is made based on a result of the determination step. A second calculation step of calculating a typical representative vector, calculating an average strain, and calculating a partial strain; a partial strain sorting step of sorting on the basis of the partial strain obtained in the second calculation step; A comparison step for comparing the disparities of a plurality of partial strains sorted by the sorting step. And a representative vector moving step of moving a plurality of representative vectors having a small partial distortion to a vicinity of a representative vector having a large partial distortion based on the comparison result in the comparing step. A codebook generation method for vector quantization characterized by the following. 17. An input vector set storing means for storing an input vector set, a code book storing means for storing a code book, a sorting means for rearranging a code book, and calculating a distortion of an input vector and a representative vector. First calculating means for calculating an address of a representative vector to be searched for; determining means for comparing two distortions obtained by the first calculating means to determine a candidate for a representative vector to be encoded; A second calculating means for calculating the center of gravity of the representative vector and calculating the average distortion based on the determination result by the above, and controlling the processing flow of the sorting means, the first calculating means, the determining means and the second calculating means. A vector quantization codebook generating apparatus including a control unit. 18. The method according to claim 18, wherein the determination unit compares two distortions obtained by the first operation unit to determine a candidate of a representative vector to be encoded, and compares partial sums of the distortions. The codebook generation apparatus for vector quantization according to claim 17. 19. The codebook generation apparatus according to claim 17, further comprising a storage unit configured to store an arrangement order of the codebooks sorted by the sorting unit. 20. The codebook for vector quantization according to claim 17, wherein a representative vector in which a previous input vector is encoded is designated as an initial representative vector. Generator. 21. The vector quantization code according to claim 17, wherein a representative vector coded in a previous iteration is designated as an initial representative vector. Book generation device. 22. An input vector set storage means for storing an input vector set, a codebook storage means for storing a codebook, a sorting means for rearranging a codebook, and a method for parallelizing distortion between an input vector and a plurality of representative vectors. Calculating means for calculating the address of a representative vector to be searched next, and determining means for comparing a plurality of distortions obtained by the first calculating means to determine a candidate for a representative vector to be encoded. And a second calculating means for calculating the center of gravity of the representative vector and calculating the average distortion based on the result of the determination by the determining means; and the sorting means, the first calculating means, the determining means, and the second calculating means. A vector quantization codebook generating apparatus, comprising: a control unit that controls a processing flow. 23. An input vector set storage means for storing an input vector set, a codebook storage means for storing a codebook, a sorting means for rearranging a codebook, and calculating a distortion of an input vector and a representative vector. First calculating means for calculating an address of a representative vector to be searched for; determining means for comparing two distortions obtained by the first calculating means to determine a candidate for a representative vector to be encoded; A second calculating means for performing a center of gravity calculation, an average strain calculation, and a partial strain calculation based on the determination result according to the above, a partial strain sorting means for sorting based on the partial strain obtained by the second calculating means, 2 sorted by sorting means
Comparing means for comparing a difference between two partial strains; a representative vector moving means for moving a representative vector having a small partial strain to a vicinity of a representative vector having a large partial strain based on a comparison result by the comparing means; A vector quantization codebook generating apparatus comprising: a first calculating means, a determining means, a second calculating means, a partial strain sorting means, a partial strain comparing means, and a control means for controlling a processing flow of a representative vector moving means. . 24. An input vector set storage means for storing an input vector set, a codebook storage means for storing a codebook, a sorting means for rearranging a codebook, and a method for parallelizing distortion between an input vector and a plurality of representative vectors. Calculating means for calculating the address of a representative vector to be searched next, and determining means for comparing a plurality of distortions obtained by the first calculating means to determine candidates for a representative vector to be encoded. And second calculating means for calculating the center of gravity, average strain calculation, and partial strain calculation based on the determination result by the determining means, and partial strain sorting means for sorting based on the partial strain obtained by the second calculating means. And 2 sorted by the partial strain sorting means.
Comparing means for comparing a difference between two partial strains; a representative vector moving means for moving a representative vector having a small partial strain to a vicinity of a representative vector having a large partial strain based on a comparison result by the comparing means; A vector quantization codebook generating apparatus comprising: a first calculating means, a determining means, a second calculating means, a partial strain sorting means, a partial strain comparing means, and a control means for controlling a processing flow of a representative vector moving means. . 25. An input vector set storing means for storing an input vector set, a code book storing means for storing a code book, a sorting means for rearranging a code book, and calculating a distortion of an input vector and a representative vector. First calculating means for calculating an address of a representative vector to be searched for; determining means for comparing two distortions obtained by the first calculating means to determine a candidate for a representative vector to be encoded; A second calculating means for performing a center of gravity calculation, an average strain calculation, and a partial strain calculation based on the determination result according to the above, a partial strain sorting means for sorting based on the partial strain obtained by the second calculating means, Comparing means for comparing a difference between a plurality of partial strains sorted by sorting means, A representative vector moving means for moving a plurality of representative vectors having a small partial distortion to a vicinity of a representative vector having a large partial distortion, based on a comparison result by the comparing means, the sorting means, the first calculating means, the determining means, the second And a control means for controlling the processing flow of the partial distortion sorting means, the partial distortion comparison means, and the representative vector moving means. 26. An input vector set storage means for storing an input vector set, a codebook storage means for storing a codebook, a sorting means for rearranging a codebook, and a method for parallelizing distortion between an input vector and a plurality of representative vectors. Calculating means for calculating the address of a representative vector to be searched next, and determining means for comparing a plurality of distortions obtained by the first calculating means to determine a candidate for a representative vector to be encoded. And second calculating means for calculating the center of gravity, average strain calculation, and partial strain calculation based on the determination result by the determining means, and partial strain sorting means for sorting based on the partial strain obtained by the second calculating means. And a ratio comparing the disparities of a plurality of partial strains sorted by the partial strain sorting means. Means, a representative vector moving means for moving a plurality of representative vectors having a small partial strain to a vicinity of a representative vector having a large partial strain based on the result of the comparing means, the sorting means, the first calculating means, and the judging means. , A second calculation means for performing a second calculation / partial distortion calculation, a partial distortion sorting means, a partial distortion comparing means, and a control means for controlling a processing flow of the representative vector moving means, a vector quantization code. Book generation device. 27. An input terminal to which an input vector is input, sorting means for rearranging codebooks, codebook storage means for storing sorted codebooks, and an input vector input via the input terminal. A first calculating means for calculating the distortion of the representative vector of the codebook stored in the codebook storing means and calculating an address of a representative vector to be searched next; Determining means for comparing the distortion to determine a candidate of a representative vector to be encoded; second calculating means for performing adaptive representative vector calculation / average distortion calculation based on the determination result by the determining means; A vector comprising a sorting unit, a first computing unit, a determining unit, and a control unit for controlling a processing flow of the second computing unit. Codebook generator for quantization. 28. An input terminal to which an input vector is input, sorting means for rearranging a codebook, codebook storage means for storing a sorted codebook, and an input vector input via the input terminal. A first calculating means for calculating a distortion with a plurality of representative vectors of the codebook stored in the codebook storing means in parallel and calculating an address of a representative vector to be searched next; A determining means for comparing the obtained plurality of distortions to determine a candidate of a representative vector to be encoded; and a second calculating means for adaptively calculating a representative vector and calculating an average distortion based on the determination result by the determining means. And a control means for controlling the processing flow of the sorting means, the first calculation means, the determination means, and the second calculation means. A codebook generator for vector quantization to be generated. 29. An input terminal for inputting an input vector, sorting means for rearranging a codebook, codebook storage means for storing a sorted codebook, and an input vector input via the input terminal. A first calculating means for calculating the distortion of the representative vector of the codebook stored in the codebook storing means and calculating an address of a representative vector to be searched next; A determining means for comparing the distortion to determine a candidate for a representative vector to be encoded; and a second means for adaptively calculating a representative vector, calculating an average distortion, and calculating a partial distortion based on the determination result by the determining means. Calculating means; partial strain sorting means for sorting based on the partial strain obtained by the second calculating means; Sorted by partial strain sorting means 2
Comparing means for comparing the difference between two partial strains; representative vector moving means for moving a representative vector having a small partial strain to a vicinity of a representative vector having a large partial strain based on a comparison result by the comparing means; A vector quantization codebook generating apparatus comprising: a first calculating means, a determining means, a second calculating means, a partial strain sorting means, a partial strain comparing means, and a control means for controlling a processing flow of a representative vector moving means. . 30. An input terminal to which an input vector is input, sorting means for rearranging a codebook, codebook storage means for storing a sorted codebook, and an input vector input via the input terminal. A first calculating means for calculating in parallel a distortion with a plurality of representative vectors of the codebook stored in the codebook storing means and calculating an address of a representative vector to be searched next; Determining means for determining a candidate for a representative vector to be encoded by comparing a plurality of obtained distortions; adaptive representative vector calculation / average distortion calculation / partial distortion calculation based on the determination result by the determining means And a partial strain sorter that sorts on the basis of the partial strain obtained by the second arithmetic means. 2 sorted by the partial strain sorting means
Comparing means for comparing a difference between two partial strains; a representative vector moving means for moving a representative vector having a small partial strain to a vicinity of a representative vector having a large partial strain based on a comparison result by the comparing means; A vector quantization codebook generating apparatus comprising: a first calculating means, a determining means, a second calculating means, a partial strain sorting means, a partial strain comparing means, and a control means for controlling a processing flow of a representative vector moving means. . 31. An input terminal to which an input vector is input, sorting means for rearranging a codebook, codebook storage means for storing a sorted codebook, and an input vector input via the input terminal. A first calculating means for calculating the distortion of the representative vector of the codebook stored in the codebook storing means and calculating an address of a representative vector to be searched next; A determining means for comparing the distortion to determine a candidate for a representative vector to be encoded; and a second means for adaptively calculating a representative vector, calculating an average distortion, and calculating a partial distortion based on the determination result by the determining means. Calculating means; partial strain sorting means for sorting based on the partial strain obtained by the second calculating means; Comparing means for comparing a difference between a plurality of partial strains sorted by the partial strain sorting means, and a plurality of representative vectors having small partial strains in the vicinity of a representative vector having large partial strains based on a comparison result by the comparing means. A representative vector moving means to be moved; and a control means for controlling a processing flow of the sorting means, the first calculating means, the determining means, the second calculating means, the partial strain sorting means, the partial strain comparing means, and the representative vector moving means. A vector quantization codebook generating apparatus comprising: 32. An input terminal to which an input vector is input, sorting means for rearranging a codebook, codebook storage means for storing a sorted codebook, and an input vector input via the input terminal. A first calculating means for calculating a distortion with a plurality of representative vectors of the codebook stored in the codebook storing means in parallel and calculating an address of a representative vector to be searched next; Determining means for comparing the obtained plurality of distortions to determine a candidate for a representative vector to be encoded; and adaptively calculating representative vectors, calculating average distortion, calculating partial distortion based on the determination result by the determining means. And a partial strain sorter that sorts on the basis of the partial strain obtained by the second arithmetic means. A comparing means for comparing a difference between a plurality of partial strains sorted by the partial strain sorting means; and a representative vector having a small partial strain having a large partial strain based on a comparison result by the comparing means. The representative vector moving means for moving to the vicinity of the representative vector, and the processing flow of the sorting means, the first calculating means, the determining means, the second calculating means, the partial strain sorting means, the partial strain comparing means, and the representative vector moving means. A vector quantization codebook generating apparatus, comprising: a control unit for controlling the codebook.