JP2001251192A - Conjugate structure vector quantization method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform preliminary selection in conjugate structure vector quantization. SOLUTION: In a conjugate structure vector quantization method, the synthesized vector of two code vectors taken out from respective code books storing the two code vectors and a vector whose quantization error is minimum with respect to a target vector are selected. At the time of retrieving only a preliminary selected number of candidate vectors from each of the two code books, the code books are divided into a number identical to the preliminary selected number of vectors. Only one vector whose distance with the target vector is the minimum is retrieved and the preliminary selected number of code vectors are set to be the candidate vectors.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a vector quantization method widely used in the field of high-efficiency compression coding of voice / audio and images.

[0002]

2. Description of the Related Art In vector coding, a plurality of input samples are combined into a target vector in signal encoding.
This is an encoding method in which a code vector closest to a target vector is selected from a code book which is a list of code vectors prepared in advance, and an input sample is represented by the code vector number. In this method, when the distribution of input samples has a certain tendency, coding can be performed with high efficiency by preparing a code book having the same distribution tendency as that of the input samples. In vector quantization, if the amount of information that can be used for encoding is equal,
It is more efficient to increase the number of quantization bits in vector quantization and increase the number of code vector samples (code vector length).

However, the codebook scale is proportional to the length of the code vector and proportional to the power of the number of quantization bits of 2. Therefore, to increase the quantization efficiency, it is necessary to store the codebook. The disadvantage is that the memory becomes too large. To solve this problem, a conjugate structure vector quantization method as shown in FIG. 1 has been proposed. In this method, a code signal from two small codebooks 10 and 11 is added by an adder 12 to obtain a reproduced signal. The distance between the reproduced signal and the target vector is determined by the distance calculation unit 13, and the optimum code vector search unit 14 selects each code vector in the code books 10 and 11 so that the distance is minimized. In this way, the efficiency of vector quantization can be increased without increasing the amount of memory.

[0004] Another disadvantage of large-scale vector quantization is that it takes a large amount of computation to find an optimal code vector. In the conjugate structure vector quantization, in order to reduce the amount of calculation, as shown in FIG.
The small codebooks 17 and 18 are configured by independently searching the candidate vectors 5 and 16 and preliminarily selecting a plurality of plausible candidate vectors.
7 and the small codebooks 1:18 are searched for the optimum combination, and the selected two code vector numbers are output as the optimum code vector numbers.

[0005] Of these, in the same way as the selection method shown in FIG. 1, the main selection unit 19 performs small code book combinations for all combinations of code vectors included in the small codebooks 0:17 and 1:18. Code vector 0 from book 0 and code vector 1 from small code book 1 are added, the distance between the obtained reproduced signal and the target vector is calculated, and the codes of code vector 0 and code vector 1 that give the minimum distance A method of outputting a vector number 0 and a code vector number 1 is general. A method for performing this actual selection at a higher speed has been filed in Japanese Patent Application No. 9-357159 (refer to Japanese Patent Application Laid-Open No. 11-191939).

[0006]

When the method described above is applied to the main selection, the ratio of the operation amount of the preliminary selection to the operation amount of the entire vector quantization relatively increases, and the vector quantization operates at a high speed. In such a case, the calculation amount of the preliminary selection becomes a problem. The preliminary selection of the vector quantization is performed by the codebooks 0, 1:10, and 11 in FIG.
6. Using the small codebooks 17 and 18, the operation is performed as shown in the flowchart of FIG. In this process, for all the code vectors, the distance d ² between the target vector and r times (r is a constant) and the code vector is calculated. If the distance d ² is smaller than the end of the small code book, A position where distances are arranged in ascending order is searched, and a new code vector is inserted there. For the operation of updating the small codebook, operations of searching for the nearest point and shifting the small codebook are required, and these operations impose a burden on the amount of calculation. This burden increases as the number of preliminary selections (the size of the small codebook) increases.

Accordingly, an object of the present invention is to provide a method for performing preselection at high speed in conjugate structure vector quantization.

[0008]

According to the present invention, when M candidate vectors are preselected from each of two codebooks of a conjugate structure vector quantizer, the codebooks are preselected as shown in FIG. Divide into the same number as the number M,
After obtaining the sub-codebooks, for each sub-codebook, only one code vector having the smallest distance from the target vector is searched, and the selected M code vectors are set as candidate vectors. When dividing the codebook, the code vectors may be rearranged in advance.

[0009]

The flow of the procedure of the preliminary selection when the preliminary selection is performed as described above is as shown in the flowchart of FIG. Comparing this processing with the flow of the conventional preliminary selection shown in FIG. 3, the amount of calculation for distance calculation in steps S4 and S22 is equal, and the amount of calculation for condition determination in steps S5 and S23 is also equal. Step S24 or Step S
Comparing the process A of 26 with the process B of the steps S6 and S7, the amount of calculation per process is much smaller in the process B. This makes it possible to reduce the amount of computation for preliminary selection as compared with the conventional method.

Further, when this method is used, the efficiency may decrease due to the division of the codebook. However, by rearranging the codebook in advance, the decrease in the efficiency due to the division can be minimized.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS As shown in FIG. 2, the configuration of vector quantization in the embodiment of the present invention is a conjugate structure vector quantizer, and codebooks 0, 1:10, 11, preselection units 0, 1 : 15, 16, small codebook 0, 1: 1
7, 18 and a main selection unit 19, which is the same as the conventional configuration of conjugate structure vector quantization using preliminary selection. A part newly introduced by the present invention is a method of preliminary selection in the preliminary selection units 0, 1:15, and 16.

In this embodiment, codebook 0:10,
The number of code vectors constituting codebook 1:11 is N in both cases. The number of code vectors to be preselected and stored in the small codebook, that is, the preselected number is M. The code vectors in the codebooks 0, 1:10, and 11 are numbered in the order of storage.
The values are represented by idx0 and idx1, respectively. The preliminary selection unit 0:15 shown in FIG.
1 and M optimal code vector search units, input a code vector from code book 0:10, select an optimal code vector candidate, and select a small code book 0: 1.
7 is output. The configuration of the preliminary selection unit 1:16 is the same as the configuration of the preliminary selection unit 0:15. However, input is performed from the codebook 1:11, and output is performed to the small codebook 1:18.

The codebook dividing section 21 divides the codebook into M subcodebooks of size N / M. In other words, the code vector in the codebook is vector cv [idxs], and the number of the divided sub-codebook is
ub, if the code vector in each sub-codebook is represented by a vector cvs [isub] [idxs], the codebook is divided according to the following rules.

[0014]

(Equation 1)

For the code vector cvs [isub] [idxs], code vectors having the same value of isub belong to the same isub-th sub-codebook. As another division method, the code book may be rearranged using a predefined mapping table and then divided. In this case, the codebook division rules are as follows.

[0016]

(Equation 2)

As an example of the definition of the mapping table,
A good result can be obtained by preparing the same number of mapping tables as the code vector length and defining the mapping tables so that the values of arbitrary elements of each code vector are sorted. Each code vector of the code book thus divided is sent to the optimum code vector search unit. Of the M sub-codebooks, the isub-th optimal code vector search unit searches the isub-th sub-codebook for a code vector with the smallest distance to the target vector. If the elements of each code vector cvs [isub] [idxs] are expressed as cvs [isub] [idxs] [ielem] and each element of the target vector is expressed as x [ielem], the distance d ² for all idxs is as follows. Calculate [isub] [idxs].

[0018]

(Equation 3)

Here, r is a constant, and good results can be obtained by selecting about 0.5. The value of len_d may be the vector length length of the target vector or the minimum value satisfying the following condition.

[0020]

(Equation 4)

Here, k is a constant from 0 to 1, and a good result can be obtained by selecting about 0.9. In vector quantization in which code information is added to a code vector,
Instead of performing the distance calculation as in equation (3),
The combination of idxs and isign is performed as follows.

[0022]

(Equation 5)

The sign [i] represents code information, and the index i takes a value of 0 or 1. The value is sign [0] =
1, sign [1] = − 1. D calculated in this way
^{2 Of} [isub] [idxs], the code vector cvs [isub] [idxs] that gives the minimum value is selected as the isub-th candidate vector, and is stored in the isub-th small codebook. In vector quantization in which code information is added to a code vector, codes are also stored in a small codebook.

The above procedure of the optimum code vector search section is performed for all isubs. Preliminary selection section 0:
The flow of the 15 steps is shown in the flowchart of FIG. FIG.
In the above, there are only M optimal code vector searching units 22
May be performed sequentially or in parallel, but here, an example of the case of performing sequentially will be described. In step S1, the subcodebook number is initialized by setting isub to 0, and the process proceeds to step S2. In step S2, codebook 0 is divided into the isub-th sub-codebook in accordance with the division rules given in the examples of equations (1) and (2), and the process proceeds to step S3. In step S3, the optimal code vector search in the isub-th sub-codebook is performed by setting the number idxs of the code vector in the sub-codebook to 0 and setting it to the maximum value that can take the value of the minimum distance d ² _min. Is initialized, and the procedure goes to step S4. • In Step S4, the idxs th code vector in the sub-codebook, the distance d ² between r times the target signal, equation (3)
After the calculation according to the calculation method described in the examples such as and Equation (5), the process proceeds to step S5. • In Step S5, the distance d ² compares the minimum distance d ² min, proceed to smaller steps S6, otherwise, the process proceeds to step S8. In step S6, the idxs-th code vector is stored in the small codebook 0, and the process proceeds to step S7. • In Step S7, by setting the value of d ² as the value of the minimum distance d ² min, and updates the minimum distance d ² min, flow proceeds to step S8. -In step S8, add 1 to the code vector number idxs,
Proceed to step S9. In step S9, the value of idxs is checked. If the value is equal to the sub-codebook size N / M, the process proceeds to step S10; otherwise, the process proceeds to step S4. In step S10, 1 is added to the subcodebook number isub, and the process proceeds to step S11. In step S11, the value of isub is checked, and if it is equal to the number M of subcodebooks, the processing of the preliminary selection unit 0:15 is terminated,
Otherwise, go to step S2.

The flow of the procedure of the preliminary selection unit 1:16 is the same as the flow of the procedure of the preliminary selection unit 0:15. However, the number of 0 assigned to the codebook and the small codebook is 1.

[0026]

According to the present invention, in conjugate structure vector quantization, preliminary selection can be performed with a small amount of operation, and as a result, the amount of operation of conjugate structure vector quantization can be reduced.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a conjugate vector quantizer.

FIG. 2 is a block diagram showing a configuration of a conjugate structure vector quantizer with preliminary selection.

FIG. 3 is a flowchart showing a conventional method of preliminary selection in a conjugate structure vector quantizer.

FIG. 4 is a block diagram showing first and third embodiments of the present invention.

FIG. 5 is a flowchart illustrating the method of the present invention of preselection in a conjugate structure vector quantizer.

[Explanation of symbols]

 10, 11 codebook 12 addition unit 13 distance calculation unit 14 optimal code vector search unit 15, 16 preliminary selection unit 17, 18 small codebook 21 codebook division unit 22 optimal codebook search unit

 ──────────────────────────────────────────────────続 き Continued on the front page F term (reference) 5D045 DA11 5J064 AA02 AA03 BA13 BC01 BC08 BD02 BD03 5K041 AA05 AA08 AA09 CC01 CC02 EE38 HH11 9A001 EE05 FF01 GG08 GG16

Claims (4)

[Claims] 1. A first codebook storing N1 code vectors and a second codebook storing N2 code vectors.
In a conjugate structure vector quantization method for selecting a combination vector of two code vectors extracted from each of the codebooks and a quantization error with respect to a target vector, the first codebook is divided into N1 / M1 codes. For the first stage of obtaining M1 subcodebooks storing vectors, and for each of the divided M1 subcodebooks, one optimal code vector is selected, and M1 is selected.
A second step of dividing the second codebook into three candidate vectors, a third step of dividing the second codebook and obtaining M2 subcodebooks containing N2 / M2 codevectors, and a step of dividing the M2 subcodes For each book, select the optimal code vector one by one,
A fourth step of determining the number of candidate vectors, and a target of the combined vector among combinations of M1 candidate vectors selected from the first codebook and M2 candidate vectors selected from the second codebook. 5. A conjugate structure vector quantization method comprising a fifth step of selecting a vector having a minimum quantization error with respect to a vector. 2. The conjugate structure vector quantization method according to claim 1, wherein the first step is a step of rearranging the code vectors constituting the first codebook, and the step of rearranging the rearranged codebook. The method includes a seventh step of dividing by dividing N1 / M1 pieces, and the fourth step includes an eighth step of rearranging the code vectors constituting the second codebook, and a step of rearranging the rearranged codebooks by N2 / M1. 9. A conjugate structure vector quantization method, comprising a ninth step of dividing by dividing into M2 pieces. 3. The conjugate structure vector quantization method according to claim 1, wherein said second and fourth steps are performed by r times (r) times a target vector for all code vectors in the sub-codebook.
A conjugate structure vector quantization method, comprising: a tenth step of calculating a distance to a constant vector, and an eleventh step of selecting a code vector having a minimum distance calculated in the tenth step. 4. The conjugate structure vector quantization method according to claim 3, wherein in the tenth step, a distance is calculated using a code vector in a sub-codebook and a part of elements of a target vector. A conjugate structure vector quantization method characterized by the following.