JP3056339B2 - Code Bit Allocation Method for Reference Vector in Vector Quantization - Google Patents

Abstract

PURPOSE:To improve the anti-bit error performance by the gray coding for the vector quantization of an information source. CONSTITUTION:An evaluation function of an input vector is calculated for each reference vector stored in a vector quantization table (201). Then the exchange of code bit trains is repeated between the adjacent reference vectors until the change value DELTAG of the grosses G of correlative coefficients Rij caused by the exchange of code bit trains between the adjacent vectors is not positive (202). Then the vector quantization table is turned into a gray code.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to an information source with respect to a reference vector used when the vector quantization, which code bit
Vector quantization criterion vector that determines whether
The present invention relates to a method for allocating code bits to a tor .

[0002]

2. Description of the Related Art Methods of quantizing several information sources that are correlated with each other include scalar quantization for performing quantization for each information source and vector quantization for quantizing a combination of information source values. Can be roughly divided into

[0003] When comparing the two methods at the same bit rate, vector quantization is superior to scalar quantization in terms of quantization quality. Vector quantization has been used in many fields.

[0004]

However, in the case of the vector quantization, the following 2 is compared with the scalar quantization.
There are two disadvantages.

[0005] First, the amount of calculation required for quantization is large, which increases the memory capacity and processing time required for processing. Second, it is difficult to improve the bit error resistance performance (the ability to suppress the quality of the dequantized (decoded) information from deteriorating due to the incorporation of a bit error into the code bit string).

The bit error resistance performance will be described in detail.

In the scalar quantization, gray coding of a code bit string (that is, reducing the Hamming distance between bit patterns having close inverse quantization values) has already been put into practical use. The bit error resistance performance can be relatively easily improved by performing the gray coding without depending on the increase of the bit error.

However, in the case of vector quantization, the method of Gray coding itself has not yet reached the level of practical use, which makes it difficult to improve bit error resistance.

[0009] The present invention has been made in view of the above circumstances, the hand upon vector quantization of information sources, if possible occupy vector quantization to realize an improvement in resistance to bit error performance
It is an object of the present invention to provide a method of allocating a code bit to a reference vector for multiplexing .

[0010]

SUMMARY OF THE INVENTION Vector quantization of the present invention
The method of allocating the sign bit to the reference vector of
Corresponds to each of multiple reference vectors used for child transformation
Reference vector for vector quantization to determine sign bit
, Wherein an evaluation function for each of a plurality of input vectors is obtained for each of the reference vectors, and two reference vectors whose sign bits are different from each other by one bit.
Based on the definition of the adjacent reference vectors, adjacent
Multiple evaluations of one of the reference vectors
Function and multiple evaluation functions of the other reference vector.
The correlation coefficient between the evaluation functions
The sum G of all the correlation coefficients of the quasi-vector is obtained,
Multiple reference vectors until the sum of coefficients G is the largest
The exchange of code bits is repeated between
You.

[0011]

To the reference vector of vector quantization according to the present invention
In the code bit allocation method of
Coefficient obtained from multiple evaluation functions
Indicates whether those reference vectors are approximate or not.
are doing. In other words, the two coefficients whose correlation coefficients are positively correlated
For quasi-vectors, code bits with short Hamming distance
Should be assigned. Of these two reference vectors
Extends the relationship to the relationship between all two adjacent reference vectors.
The value of the extended evaluation is the value G, and this value G is the largest.
Exchange sign bits between multiple reference vectors
I do. Thereby, when vector quantization of the information source,
Improvement of bit error resistance performance can be realized.

[0012]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS First, a basic concept for realizing gray coding according to the present invention will be described. here,
Gray coding is used for vector quantization of information sources.
To improve the bit error resistance performance.
This means associating a torque with a sign bit. The quantizer vector quantization when the input vector (input information source was set) is given, for all reference vectors are stored in the vector quantization table for use in quantization of the input vector Then, for example, an evaluation function as shown in the following equation (1) is obtained, and a reference vector that is most similar to the input vector is selected from all the reference vectors. And one-to-one correspondence with the selected reference vector
The amount of the sign bit sequence (hereinafter referred to as sign bit)
Output as child output.

(Equation 1) Here, which coded bit is associated with which reference vector
Bit error in vector quantization
Performance changes. Below, the highest bit error resistance performance
Method for associating reference vectors with coded bits
The concept of (Gray coding method) will be described. Many (here
In this case, as a statistic for a learning sample composed of S input vectors, a total of S sets of correlation coefficients Rij of the evaluation function Ei and the evaluation function Ej obtained for each input vector are considered. Considering the physical meaning of the correlation coefficient Rij, the following can be said.

[0013] [1] standard base when the correlation coefficient Rij is a good evaluation function Ei according to when the reference vector xi is close to 1.0
The evaluation function Ej related to the vector xj is also good, and the evaluation function E
When i is bad, the evaluation function Ej is also bad. Paraphrase
Then, the reference vector xi and the reference vector xj are vector
This indicates that there is an approximate relationship from the viewpoint of quantization.
You. [2] reference base when the evaluation function Ei is good correlation coefficient Rij is according to when the reference vector xi is close to -1.0
The evaluation function Ej relating to the vector xj is bad, and the evaluation function E
When i is bad, the evaluation function Ej is good. Paraphrase
Then, the reference vector xi and the reference vector xj are vector
From the point of view of quantization
ing. [3] When the correlation coefficient Rij is close to 0.0 The evaluation function Ei and the reference vector xj related to the reference vector xi
In the evaluation of the evaluation function Ej according to indicate that the correlation is little <br/> exist. In other words, the reference vector x
i and the reference vector xj are close to each other in terms of vector quantization.
It is difficult to say that they are similar or that there are large differences
It is shown that. In each evaluation using the correlation coefficient Rij described above, from the viewpoint of improving the bit error resistance performance,
In the case of [1], the sign bit for the reference vector xi is
And bets, as the Hamming distance between the code bit decreases with respect to the reference vector xj, it is desirable to assign <br/> sign bit of the reference vectors xi and xj. Also, [2]
In the case of, the reference vectors xi and xi and
It is desirable to assign the sign bit of xj .

In other words, if the magnitude of the correlation coefficient Rij is determined and the result is reflected in the Hamming distance between code bits , gray coding of each reference vector on the vector quantization table is achieved. Thus, the bit error resistance performance can be improved by Gray coding.

Next, the magnitude of the correlation coefficient Rij is determined by each reference vector.
A method of reflecting the value on the Hamming distance between code bits of a torque will be described.

[0016] In vector quantization table stored in which each reference vector, the reference vector each other differ only 1 bit in the code bit is defined as a vector that is adjacent.

FIG. 2 shows that the number of reference vectors stored in the vector quantization table is eight, and three
If the sign bit of the bit is associated with code-bi
This shows the adjacency between the units .

As shown in the following equation (2), all adjacent relations
Two reference vectors xi in engagement, the correlation coefficient between xj Rij
Is G. One certain correlation coefficient Rij is as described above.
Between the two adjacent reference vectors xi and xj
Is evaluated, but the value G is
Between two reference vectors xi and xj that are
Since it is the sum of the correlations (correlation coefficients), all the reference vectors
To evaluate the sign bit allocation for
I have. G = ΣRij (2) The larger this value G is, the smaller the quality deterioration at the time of bit error mixing is. Therefore, this value G represents the bit error resistance performance of the vector quantization table. Also, by modifying the adjacent relationship between the reference vectors (allocation of code bits) so that the value G becomes as large as possible, it is possible to achieve gray coding that minimizes quality deterioration when bit errors are mixed. .

As a method of deforming the adjacent relation between the reference vectors so as to increase the value G, the following method can be considered. Note that the adjacent relationship between the reference vectors
Is defined as above, so that the
Deforming the adjacency is done by using the reference vector and the sign bit.
Means to change the correspondence relationship with. Also, before deformation
After that, the assignment of the sign bit of each reference vector is
It is a candidate, and deforming is the latter candidate
It means good. As shown in FIG. 3A, four reference vectors x out of many reference vectors
Focus on i, xj, xk, and xl. Among these four reference vectors xi, xj, xk, and xl, reference vector xi and reference vector xj, reference vector xj and reference vector x
k, the reference vector xl and the reference vector xl, and the reference vector xl and the reference vector xi are adjacent to each other.
You. In addition, the reference vector xi, the reference vector xi1,
xi2,... xiM are adjacent, and the reference vector xj contains the reference
Vector xj1, xj2 ,,, xjM is adjacent to the reference vector xk, adjacent reference vectors XK1, XK2 ,,, XKM, the reference vector xl is the reference vector xl1, xl
It is assumed that 2,... XlM are adjacent.

In such an adjacency relationship, code bit exchange processing for exchanging code bits to be assigned to each of the reference vectors xi, xj, xk, xl between these reference vectors xi, xj, xk, xl. (Circular exchange) as appropriate
However, these reference vectors xi, xj, xk, xl
Search for an optimal code bit to be performed. It will be described later
Thus, the four reference vectors xi, xj,
The search is performed while changing xk and xl. In the case of FIG.
From, as shown in (b) of FIG. 3, each of the four reference vectors xi, xj, xk, so as to maintain the adjacency with respect to the reference vector sign bit of xl are adjacent code bits by the so-called cyclic permutation Exchange processing. That is,
Mark associated with reference vector xi before cyclic permutation
No. bits are associated with the reference vector xj before the cyclic permutation,
Similarly, reference vectors xj, xk, xl and sign bits
Is changed. According to such a cyclic permutation, the above-described reference vector xi and reference vectors xi1, xi
2. Adjacency relationship between, xiM, reference vector xj and reference vector
Torr XJ1, Xj2 ,,, adjacencies with XjM, reference vector xk and the reference vector XK1, adjacencies with XK2 ,,, XKM, reference vectors xl and the reference vector xl1, xl2 ,,,
The adjacency with xlM is eliminated. Then, instead, a new reference vector xl and a new reference vector xi
1, xi2,... XiM, adjacency relation, reference vector xi and base
Adjacent relationship between quasi-vectors xj1, xj2, xjM, adjacency between reference vector xj and reference vectors xk1, xk2, xkM, reference vector xk and reference vectors xl1, xl
The adjacency with 2,... XlM is obtained.

These four reference vectors xi, xj,
When the bit error resistance performance after executing the code bit exchange processing between xk and xl is G1, and the bit error resistance performance before executing the above code bit exchange processing is G0, the bit error resistance performance G1 Is larger than the bit error resistance performance G0, it means that the bit error resistance performance is improved by the code bit exchange processing, and the gray code conversion is promoted by the code bit exchange processing.

Therefore, it is necessary to determine the magnitude relationship between the bit error resistance G1 and the bit error resistance G0 after the exchange processing. The difference between the bit error resistance G1 and the bit error resistance G0 is required. Are the reference vectors xi, xj, xk,
Since it is caused by a change in the code bit adjacent to x l , the following equation (3) is used to determine the magnitude relation between the bit error resistance performance G1 and the bit error resistance performance G0.
As shown in the above, the difference ΔG between the bit error resistance performance G1 and the bit error resistance performance G0 may be used .

[0023] In ΔG = G1-G0 ... (3 ) before mentioned equation (3), .DELTA.G is positive if each reference vector xi, xj, xk, improved resistance to the bit error performance by the exchange processing of the sign bit of xl As a result, it means that the Gray coding has been advanced. Deformation of the adjacency relationship due to the sign bit exchange process is defined as i, j, k, l
By repeating until there are no more pairs, the bit error resistance performance is maximally improved, and as a result, Gray coding of the vector quantization table is achieved. When ΔG is 0 or negative, each of the reference vectors xi, xj, x
The k / xl code bit exchange processing means that the bit error resistance performance is not improved, which means that the bit error resistance performance has already reached its limit and Gray coding has been completed.

The embodiment of the present invention is based on the above-described concept. FIG. 1 shows a processing procedure of a method of gray-coding a vector quantization table according to an embodiment of the present invention. .

In one embodiment of the present invention, first, an evaluation function for an input vector is obtained for each reference vector stored on the vector quantization table (step 201).

[0026] Then, the Hamming distance of the code bit is based on the definition of the reference vector to each other is 1 adjacent to each other, reference vector xi which are adjacent on the vector quantization table, xj mutual of the evaluation function Ei, assuming the sum G of the correlation coefficient Rij of ej, the correlation coefficient Rij that between reference vectors each other caused me by exchanging processing code bicycloalkyl <br/> Tsu DOO
Sum up the change amount ΔG of G is no longer positive, repeat the replacement process of the code bit between adjacent reference vectors, at the time of replacement processing immediately before the change amount ΔG positive ones no longer occurs in the reference vector The assigned sign bit is determined as the sign bit of each reference vector (step 202). Note that the exchange processing is the exchange of code bits.
Is a process for determining whether or not is appropriate, and the value ΔG is
If it is positive, the exchange of the sign bit is confirmed, and the value ΔG is negative or
If 0, no sign bit exchange is performed. Sign bit assigned to each reference vector by repeating the exchange of the sign bit in the step 202,
This is to maximize the sum G of correlation coefficients Rij between adjacent reference vectors, and as a result, gray coding that minimizes deterioration in quality when bit errors are mixed is achieved.

Next, a specific processing procedure in step 202 will be described with reference to FIG. In this embodiment, four reference vectors that focuses as shown in FIG. 3 is in the chain state, number <br/> bits of the code bits of each reference vector is assumed to be M bits. Therefore, the reference vector
There are 2 M powers. The parameters i, j,
k and l are decimal numbers ranging from 0 to (2 to the power of M−1).
And represents the sign bit. Reference vector
Xi, xj, xk, and xl are the sign bits at that time.
Assume that i, j, k, and l are associated with each other. What
Before entering the processing of FIG.
Code bits are assigned as initial candidates. First, reference vectors xi, xj, x adjacent in a chain state
i in k, xl is set to 0, p is set to 0, and q is set to 1 (step 101). Thereby, the reference vectors xi and
Then, the reference vector x0 is set. Next, j is p of i
It is assumed that the bit is inverted (step 102).
That is, as the reference vector xj,
For the sign bit obtained by inverting the p-th bit of the sign bit i,
It shall be assigned. Thus, p is in M bits
Is specified. Next, l is the q of i
It is assumed that the bit is inverted (step 103).
That is, the reference vector xl is used as the reference vector xl.
For the sign bit obtained by inverting the q-th bit of the sign bit i,
It shall be assigned. Thus, q is also in M bits
Is specified. Next, k is the p of i
The bit and the q-th bit are inverted (step 104). That is, as the reference vector xk,
The p-th bit and the q-th bit of the sign bit i of the quasi-vector xi
Shall be associated with the sign bit with its eyes inverted.
You. Next, ΔG is obtained from the above set of i, j, k, l by equation (3) (step 105). Sand
The four reference vectors xi, x set at this time
ΔG is determined for j, xk, and xl. In addition, four criteria
When the vectors xi, xj, xk and xl are set for the first time
Goes from step 104 to step 108 immediately.
You. It is determined whether or not ΔG is positive. If ΔG is positive, the code bits of each of the reference vectors xi, xj, xk, and xl are cyclically replaced (fixed bit exchange is determined) , and step 10 is performed.
Returning to 1 (steps 106 and 107). However, if ΔG is not positive, the value of q is increased by 1 (step 1).
06,108).

If the value of q is increased by 1 in step 108, it is determined whether or not the value of q is greater than M. If q is less than M, the process returns to step 102 (step 109). .

If it is determined in step 109 that q> M, p is incremented by 1 and q is set to p + 1 (step 110), and it is determined whether or not p> M.
If p is equal to or smaller than M, the process returns to step 102 (step 111).

If it is determined in step 11 that p> M, the value of i is increased by one, p is set to 0, and q is set to 1 (step 112). 2 M
It is determined whether a power of above, i is in the case of more than 2 to the power M to terminate the process, i is smaller than the power of 2 M, the process returns to step 102 (step 113 ).

As is clear from the above, the sign bit i,
References for changing j, k, l and corresponding to them
The vector is also changed sequentially (reference vector and code
While changing the candidate of the correspondence relationship with the
Based on the relationship between the optimal reference vector and the sign bit,
Since the relationship is searched for, it is possible to realize appropriate Gray coding that can maximize bit error resistance performance in quantization.

FIG. 5 shows experimental data for confirming the effect of the Gray coding method of the vector quantization table of the above-described embodiment. In the LPC (linear predictive coding) type speech encoder, the LPC A comparison is made between the SN ratio of the decoded speech when the Gray quantization is applied to the parameter vector quantization table by the method of the above-described embodiment and the SN ratio of the decoded speech when the Gray encoding is not performed according to the embodiment. It is measured.

[0033] In FIG. 5, base curve A to which the sign bit after performing an exemplary embodiment Gray coded
Shows the SN ratio of the audio signal when decoding the vector quantization output, the curve B the audio signal when decoding the vector quantization output of applying the sign bit before performing the Gray coded according to an embodiment SN The ratio is shown.

The SN ratio between the curves A and B is better in the case of gray coding, and the tendency becomes more prominent as the bit error rate increases. That is, it can be confirmed that the bit error resistance performance is satisfactorily improved by the Gray coding.

Note that, in the above-described embodiment, a case where a set in which four reference vectors xi, xj, xk, and xl are adjacent to each other in a chain state is subjected to the exchange processing. The exchange process of the ット is realized by cyclic permutation. But,
When other sets are used, for example, when two adjacent reference vectors xi and xj are to be replaced,
The sign bit of each reference vector may be exchanged with the partner to be contiguous.

FIGS. 6 (a) and 6 (b) show how adjacent relations are transformed by exchanging code bit strings when two reference vectors xi and xj are adjacent to each other.

Initially, as shown in FIG. 6A, two reference vectors xi and xj are adjacent to each other, and reference vectors xi and xi are included in the reference vector xi.
2,, xiM are adjacent and the reference vector xj is the reference vector
Le xj1, xj2 ,,, xjM is to be adjacent.

In such an adjacency relationship, the reference vectors xi1, xi adjacent to the respective reference vectors xi, xj
2, xiM and xj1, xj2, xjM are exchanged between adjacent reference vectors xi and xj . According to the code bit exchange processing, the reference vector xi and the reference vectors xi1, xi2,.
, The reference vector xj and the reference vector xj1,
The adjacency with xj2,. Then, instead, a new reference vector xj and the reference base
Vector xi1, adjacencies with xi2 ,,, XIM, adjacency of the reference vector xi and the reference vector xj1, xj2 ,,, xjM is obtained.

[0039] Such a code replacement processing resistance bit error performance after execution of bit G3, when the resistance bit error performance before executing the exchange process of the sign bit of the aforementioned G2, the above-described embodiment Similarly to the case, as shown in the following equation (4), the bit error resistance performance G3 and the bit error resistance performance G2
Is calculated from the difference ΔG.

ΔG = G 3 −G 2 (4) In the above equation (4), if ΔG is positive, the bit error resistance performance is improved by exchanging the code bits of the reference vectors xi and xj, and as a result, gray It means that coding has been advanced. By repeating the modification of the adjacency relationship due to the code bit exchange process until there is no more i, j pair with ΔG being positive, the bit error resistance performance is maximized, and as a result, the gray level of the vector quantization table is reduced. Coding is achieved. If ΔG is 0 or negative, it means that the bit error resistance performance is not improved by the code bit exchange processing of each of the reference vectors xi and xj, and the improvement of the bit error resistance performance has already reached its limit. , Means that Gray coding has been completed.

FIG. 7 shows a specific processing procedure of the code bit exchange processing shown in FIG. It is assumed that the number of bits of the code bit string of each reference vector is M bits as in the case of the embodiment.

First, the adjacent reference vectors xi, xj
Is set to 0 and p is set to 1 (step 30).
1).

Next, j is obtained by inverting the p-th bit of i (step 302).

Next, ΔG is obtained from the above set of i and j by equation (4) (step 303).

It is determined whether ΔG is positive. If ΔG is positive, the sign bits of the reference vectors xi and xj are exchanged, and
Return to step 301 (steps 304 and 305). However, if ΔG is not positive, the value of p is increased by 1 (steps 304 and 306).

If the value of p is increased by 1 in step 306, it is then determined whether or not the value of p is greater than M. If p is less than M, the flow returns to step 302 (step 307). .

If it is determined in step 307 that p> M, i is incremented by 1 and p is set to 1 (step 308), and it is determined whether or not i is equal to or larger than 2 to the power of M. If i is greater than or equal to 2M , the process is terminated. If i is smaller than 2M , the process returns to step 302 to repeat the process (step 309).

Also in the procedure shown in this figure, gray coding of the vector quantization table can be realized in the same manner as in the above-described embodiment, and the number of bits of the code bit string is increased at the time of vector quantization of the information source. Without using such a measure, it is possible to improve bit error resistance performance by gray coding.

[0049]

As is clear from the above description, according to the present invention, when the vector quantization of the information source is performed,
The reference vector is encoded to improve the error performance.
Signal bits.

[Brief description of the drawings]

FIG. 1 is a schematic explanatory diagram of a processing procedure according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram of an adjacency relationship between three code bits.

FIG. 3 is an explanatory diagram of a modification of the adjacency relationship due to the exchange of code bits.

FIG. 4 is an explanatory diagram of gray coding by exchanging code bit strings in one embodiment of the present invention.

FIG. 5 is an explanatory diagram of an effect of one embodiment of the present invention.

FIG. 6 is an explanatory diagram of a modification of the adjacency relationship according to another embodiment of the present invention.

FIG. 7 is an explanatory diagram of Gray coding by exchanging code bit strings in another embodiment of the present invention.

[Description of Signs] xi, xj, xk, xl Reference Vector

──────────────────────────────────────────────────続 き Continuation of front page (72) Inventor Katsutoshi Ito 1-7-12 Toranomon, Minato-ku, Tokyo Oki Electric Industry Co., Ltd. (58) Field surveyed (Int. Cl. ⁷ , DB name) H04B 14 / 00-14/06 G10L 19/00 H04N 7/24

Claims (1)

(57) [Claims] 1. A plurality of reference vectors used for vector quantization.
Vector that determines the sign bit associated with each
A sign bit allocation method to the reference vector of the torque quantization, for each reference vector, obtains the respective <br/> against evaluation function of a plurality of input vectors, the two reference vectors sign bit 1 bit different Adjacent
Based on the definition of the reference vectors, adjacent
Multiple evaluation functions of one of the reference vectors
And multiple evaluation functions of the other reference vector
Then, a correlation coefficient between these evaluation functions is obtained, and a sum G of all correlation coefficients of adjacent reference vectors is obtained.
Therefore, until a total sum G of the correlation coefficients becomes maximum,
It is characterized by repeated sign bit exchanges between quasi-vectors.
Sign bit to reference vector for vector quantization
Assignment method.