JP3412118B2 - Conjugate structure vector quantization method, apparatus therefor, and program recording medium - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is widely used in the field of high-efficiency compression coding of voice, audio, image, etc., and a target vector is obtained by combining code vectors of two codebooks having a conjugate structure relationship with each other. The present invention relates to a quantization method, an apparatus therefor, and a program recording medium.

[0002]

2. Description of the Related Art In vector quantization, a plurality of input samples are collectively set as a target vector in signal coding.
This is a coding method in which a code vector closest to the target vector is selected from a code book that is a list of code vectors prepared in advance, and a plurality of input samples are represented by the code vector numbers.

In this method, when the distribution of input samples has a certain tendency, it is possible to perform coding with high efficiency by preparing a codebook having the same distribution tendency as the distribution tendency of input samples. In vector quantization, if the amount of information that can be used for encoding is the same, increase the number of quantization bits in vector quantization, that is, the number of code vectors, and increase the number of code vector samples, that is, the code vector length. It is better to do it.

However, the scale of the codebook is proportional to the length of the code vector and is proportional to the square of the number of quantization bits. Therefore, if the quantization efficiency is to be increased, it is necessary to store the codebook. It has the disadvantage that the memory becomes too large. In order to solve this problem, FIG.
A conjugated structure vector quantization method such as A has been proposed. This method uses two small codebooks 10,
Each code vector from 11 is added by an adder 13 to obtain a reproduced signal. The distance calculator 13 calculates the distance between the reproduced signal and a target signal, and the optimum code vector searcher 14 is used to minimize the distance. In the codebook 10,1
Each code vector of 1 is selected. This way,
It is possible to improve the efficiency of vector quantization without increasing the memory amount too much.

Another drawback of large-scale vector quantization is that searching for an optimum code vector requires a large amount of calculation. In the conjugate structure vector quantization, in order to reduce the calculation amount, the code vectors of the codebooks 10 and 11 are made to be conjugate structure vectors to each other as shown in FIG.
0 and 11 are independently searched by the preselection units 15 and 16, respectively, and some plausible candidate vectors are preselected to form the small codebooks 17 and 18, and the optimal combination is selected from the small codebooks 17 and 18. Is used.

In this method, the smaller the number of candidates at the time of preliminary selection is, the higher the effect of reducing the amount of calculation can be obtained. However, on the other hand, since the information on the opposite side to be combined is not used in the preliminary selection,
There is a problem that the efficiency of quantization deteriorates as the number of candidates for preliminary selection decreases.

[0007]

DISCLOSURE OF THE INVENTION The present invention provides a method for quantizing a conjugated structure vector with a small amount of calculation while maintaining a high quantization efficiency, an apparatus therefor, and a program recording medium.

[0008]

According to the present invention, each code vector in one codebook, the code vector in the other codebook, and a plurality of valid ones are searched for in the combination search of conjugate structure vector quantization. , And the quantization error between the target vector and the combination of the code vectors of both codebooks in each group is minimized.

In order to select a group of combinations of each code vector of the one codebook and a plurality of valid code vectors in the other codebook, for example, the code vector is set close to the target vector. Is converted to a one-dimensional evaluation value that can be evaluated, and using this value,
Conjugate vector (code vector of the other codebook)
Find a valid combination with. It is highly effective to use the magnitude of the parallel component of the code vector as the one-dimensional evaluation value with the target vector. The action conjugated structure vector quantization, as shown in FIG. 4 B, and calculating the combination of the code vectors v1 in the code vectors v0 and codebook 11 in the codebook 10, the distance R between the target vector t, A combination having the minimum value R _min is selected. In the conventional method, distance calculation and comparison with the minimum value are performed for all possible combinations, but in the present invention, as shown in FIG. 1, for example, the parallel component C0 of the code vectors v0 and v1 with each target vector t
Error R _p with the C1 synthesis C0 + C1 and the target vector t of
The distance calculation is performed only for the combinations in which the current minimum error range R _{p min} falls. This way,
This can be omitted for combinations that do not require distance calculation and comparison with the minimum value, and the amount of calculation can be reduced. Since the one-dimensional value can be sorted in ascending or descending order and the value search can be performed with a small amount of calculation, the amount of preprocessing for omitting the distance calculation can be reduced.

Further, in the method of the present invention, since the method of excluding the case where the quantization error always exceeds the present minimum value is taken, there is no reduction in the quantization efficiency by applying this method.

[0011]

Embodiment 1 FIG. 2 shows Embodiment 1 of the present invention. In the first embodiment, in order to select an effective combination of the code vectors in the code book 11 for each code vector in the code book 10, the closeness to the target vector is determined as the magnitude of the parallel component of the code vector to the target vector. This is a case where the determination is performed using. The number of code vectors forming the codebook 10 and the codebook 11 is N in both cases.

In the procedure 1, the magnitude C ^a (n) of the parallel component of each code vector v _a (n) in the code books 10 and 11 with the target vector t is calculated as in the equation (1). _{^{C a (n) = v a}} (n) · t / T T = ‖t‖ (1) n is the index number of the codebook code vectors v, a represents the identification number of the code book 10 and 11 , The value is 0 or 1. That is, the inner product of v _a (n) and t is divided by the size (amplitude) of t to obtain the size of the parallel component. These are stored in the parallel component storage units 21a and 21b (FIG. 3A). Also, the size of the target vector ‖t‖
Is stored in the storage unit 22.

In procedure 2, one of the N codebooks 11 is rearranged so that the magnitude C ¹ of the parallel component is in ascending order. Various algorithms have been proposed for performing this rearrangement at high speed, and heap sorting is an example. The contents of the storage unit 21b are arranged in this order. In procedure 3, the magnitude C ⁰ of the parallel component of each code vector in the other N codebooks 0 is based on the target vector and the minimum quantization error up to the present.
For each (n), the range (group) of code vectors v ¹ (n) to be combined in the codebook 11 is determined. Therefore, the present minimum quantization error storage unit 23 is provided ( FIG. 3A).

First, the magnitude C ⁰ of the parallel component of the designated code vector v ₀ (n) in the codebook 10
For (n), the range of the parallel component of the effective code vector in the codebook 11 is determined by the combination. C ¹ _min (n) = ‖t ‖-C ⁰ (n) -√d ² _min C ¹ _max (n) = ‖t ‖-C ⁰ (n) + √d ² _min (2) where C ¹ _min is the lower limit value of the parallel components of the code vector in the codebook 11, which is effective in combination, C ¹
_max is an effective upper limit of the parallel components of the code vector in the codebook 11, and d ² _min is the minimum value of the quantization error. That is, from the size ‖t‖ of the target vector t to the size C ⁰ (n) of the parallel component of the code vector in the codebook 10 and the minimum value d ² _min of the quantization error.
The value obtained by subtracting the square root of is not less than C ¹ _min (n), and √d ² _min is added to the value obtained by subtracting the magnitude C ⁰ (n) of the parallel component of the code vector of the codebook 10 from ‖t‖. It is shown that the code vector in the code book 11 should be searched for each code vector v ₀ (n) for the value of C ¹ _max (n) or less .

Therefore, C¹ _min(N) <C¹(M) <C
¹ _maxThe range of m that is (n) is searched. m is in ascending order
C given¹The numbers in the order of (n) are shown. Code book
11 is C in step 2¹Sort in ascending order
Because it is obtained, C¹But C¹ _minMinimum value m above
_minAnd C¹ _maxMaximum value m below_maxThe value of m
If you search each, the range of m (m_min(N), m
_max(N)) can be specified. For example in Figure 3B
As shown, there is a C⁰About (n) C¹(N) is large
C in order¹ ₁(N), C¹ ₂(N), ... C¹ _N(N) and
Lined up, C¹ ₂(N)> C¹ _maxC in (n)¹ ₃(N)
<C¹ _maxIf (n), then m_max= 3 and C ¹ _i
(N)> C¹ _minC in (n)¹ _{i + 1}(N) <C
¹ _minIf (m), then m_min= I.

This procedure is performed for each value of n. That is, the process is performed for each code vector v ⁰ (n) of the codebook 10, the value of n starts from 0, and m _min (n) and m _max
After (n) is determined, n is added one by one and updating is performed, but the procedure proceeds to step 4 before updating. After step 4, the process proceeds to step 5, but after step 5 ends, the process returns to step 3. In the procedure 3, for each code vector in the codebook 10, the candidate of the code vector which is a valid combination in the codebook 11 is selected.

In step 4, the effective combination obtained in step 3
Each code vector v within the range ₁(M) and code
Vector v₀For the target vector t of the composition of (n)
Calculate the quantization error.     d²(M, n) = | t− {v₀(N) + v₁(M)} ‖²，           m_min(N)<m<m_max(N) (3)         n is the value used in step 3 In step 5, the quantization error d found in step 4²(M, n)
And the minimum value of the currently recorded quantization error d² _minAnd
Compare and d²If (m, n) is smaller, remember it
D of part 23² _minIs updated and the code in the storage unit 24 is updated.
The values of the indexes m and n are updated (5a). This update hand
The procedure is performed every time the quantization error is calculated in the procedure 4. Step 4
・ After the quantization error calculation / comparison is completed in 5, increment n by 1.
(5b), codebook 0 code that should still be processed
Check if index n remains (5c), and leave
If so, go to step 3. If not,
Quantize index of currently recorded code index
(5d). Where codebook 0
The index n of is output as is, but the codebook
The index m of 1 is the index before the rearrangement in step 2.
Output after correcting it. Example 2 The second embodiment is an extension of the first embodiment. Enter in step 1
Instead of using codebook 10 and codebook 11,
Small code obtained by preselecting each codebook
Dobook 0 and small codebook 1 are used. This preliminary selection
Can be performed in the same manner as the conventional method shown in FIG.
Do not decrease the quantization efficiency by increasing the number of candidates for preselection
To

As is clear from the above description, as the quantizer according to the present invention, the codebooks 10 and 11, the target vector amplitude calculating means, the storage unit 22 for storing the calculation result ‖t‖, and the respective code vectors are stored. Parallel vector amplitude calculating means for calculating the magnitudes of the components parallel to the target vector and the storage unit 21 for storing the magnitudes of those parallel components.
a and 21b, a rearrangement means for rearranging the storage contents of the storage section 21b in ascending order, a means for obtaining the effective lower limit value and the effective upper limit value by the equation (2) for the specified code vector in the codebook 10, and A means for obtaining code vector candidates in the codebook 11 which are effective combinations according to the effective lower limit value and the effective upper limit value, and a quantization error between these code vector candidates and the designated code vector in the codebook 10 A means for calculating, a means for comparing the calculated quantization error with the current minimum quantization error in the storage unit 23, and if the quantization error calculated by the comparison is smaller, Updating the minimum quantization error in the storage unit 23 and generating the minimum quantization error;
A means for updating the code indexes n and m of the storage unit 24 with the code indexes of both code vectors and whether or not all the code vectors of the codebook 10 have been designated are determined. Unless there are remaining specifications for all code vectors in the codebook 10, a means for obtaining an effective upper limit value, a means for obtaining an effective lower limit value, a means for obtaining a code vector candidate, a means for obtaining a quantization error, a repeating means for performing a comparing means, and It will be apparent that the apparatus is provided with a means for outputting the code index in the storage unit 24 at that time as a quantized output, and a control means for controlling each of the above means and executing the processing.

The processing of the quantizer is generally performed by decoding and executing a program. Further, in the above description, the magnitudes of the parallel components of the code vector 11 are rearranged in ascending order, but without doing so, the point is that the magnitude of the parallel components of the designated code vector and the magnitude of the target vector. And the current minimum quantization error, a candidate that is a valid combination of the designated code vector and the code vector of the codebook 11 may be obtained. For each code vector in co Dobukku 10,
A valid one is selected as a candidate in combination with this and a code vector in the codebook 11, the code vector in the codebook 10 is combined with each candidate of a valid combination corresponding to this, and a target vector is obtained. The code index of the combination of code vectors that minimizes the quantization error of is used as the quantized output.

[0020]

As described above, according to the present invention, useless combination of the code vector 10 and the code vector 11 can be predicted with a small amount of calculation in the conjugate structure vector quantization. When the number of code vectors in the codebook is about 64 and the length of the code vectors is about 20, it is expected that the amount of calculation will be reduced by about 1/10.

[Brief description of drawings]

FIG. 1 is a diagram for explaining whether an effective combination can be obtained depending on the size of a target vector in the conjugate structure vector quantization of the present invention and the components of two conjugate structure vectors parallel to the target vector.

FIG. 2 is a diagram showing an example of a functional configuration and a procedure to which the method of the present invention is applied.

3A is a diagram showing each storage unit used in the device of the present invention, and FIG. 3B is a diagram for explaining a valid lower limit value, a valid upper limit value, and a candidate range of valid combinations.

FIG. 4A is a diagram showing a functional configuration of conventional conjugate structure vector quantization, and B is a schematic diagram of distance calculation of ordinary conjugate structure vector quantization.

FIG. 5 is a diagram showing a functional configuration in which a calculation amount of a conventional conjugate structure vector quantization is reduced.

Continuation of the front page (56) Reference JP-A-8-123493 (JP, A) JP-A-6-282298 (JP, A) JP-A-9-120299 (JP, A) JP-A-10-154000 (JP , A) (58) Fields investigated (Int.Cl. ⁷ , DB name) H03M 7/30

Claims (8)

(57) [Claims] 1. A conjugate structure vector quantum in which synthesis of two code vectors respectively extracted from first and second codebooks storing mutually conjugate structure code vectors has a minimum quantization error with respect to a target vector. In the coding method, for each code vector in the first and second codebooks,
The step of obtaining the magnitude of the component parallel to the target vector, and the parallel composition for each code vector of the first codebook.
Min and the parallel component of the code vector of the second codebook
Of the combinations, the combined vector of the combination and the target vector
A valid combination with an error within the specified range
A step of selecting as the cause, and the features and synthesis of valid combinations between the singled out code vector, the quantization error between the target vector composed of the stage you find the code index indicating the minimum combination Conjugate structure vector quantization method. 2. A step of searching the code index.
Obtains a code index indicating the current minimum quantization error for each combination of code vectors, the predetermined range is above
Serial is the current minimum quantization error, the second co long as there remains a code vector in the first codebook to be searched
Valid combinations with the code vectors in the
Obtaining process continues, remaining If no conjugate structure vector quantization method according to claim 1, wherein the output code index and to Turkey the code index indicating the current minimum quantization error. 3. The first and second codebooks each include a small vector consisting of a set of code vectors preselected from first and second large codebooks each having more code vectors than the first and second codebooks. 3. The conjugated structure vector quantization method according to claim 1, which is a codebook. 4. A vector in which the combination of two code vectors respectively taken out from the first and second codebooks that store code vectors having a conjugate structure to each other is a target vector and a combination having a minimum quantization error is used as a code index. In the quantizer, a target vector amplitude calculating means for calculating the magnitude of the target vector, a target vector magnitude storing means for storing the magnitude of the target vector, and a target of each code vector in the first codebook. First parallel vector amplitude calculating means for obtaining the magnitude of the component parallel to the vector as the first parallel component, and the magnitude of the component parallel to the target vector of each code vector in the second codebook as the second parallel component. Second parallel vector amplitude calculating means to be obtained, the first parallel component, and the second parallel component are stored respectively. First and second parallel component storage means, error storage means for storing the current minimum quantization error, the first parallel component and the second parallel component for each code vector of the first codebook. The candidate selecting means for obtaining a code vector candidate that is an effective combination in the second codebook from the current minimum quantization error and the magnitude of the target vector, and the code vector of the first codebook correspond to this. An error calculation means for calculating a quantization error with each code vector candidate of the second codebook selected by the candidate selection means, and a quantization error calculated by the error calculation means with the current quantization error. Comparing means, an index storage means for storing a combination index of the code vector for which the current minimum quantization error is obtained, and the comparison result is calculated. If the quantization error is smaller, updating means for updating the current minimum quantization error to the calculated quantization error and updating the index of the index storage means, the candidate selecting means, the error Calculation means, the comparison means,
The updating means includes an iterative means for executing all the code vectors of the first codebook, and an output means for outputting the index of the index storage means as a quantized output when the execution of the iterative means is completed. A conjugated structure quantizer characterized by the following. 5. A rearrangement means for rearranging the second parallel components in order of size thereof, wherein the second parallel component storage means is a means for storing the rearranged ones, and the candidate selection. And a means for calculating an effective upper limit value and an effective lower limit value by adding and subtracting a square root of a current minimum quantization error to and from a value obtained by subtracting the target first parallel component from the size of the target vector, and the second The number m _max (n) of the largest one of the second parallel components smaller than the effective upper limit value in the parallel component storage means,
The number m _min (n) of the smallest one of the second parallel components larger than the above effective lower limit value is obtained, and the second codebook corresponding to each number in the range of number m _max (n) and number _mmin (n) 5. The conjugate structure vector quantization device according to claim 4, further comprising: candidate range determining means that uses the code vector of 1. as the code vector candidate. 6. Vector quantization in which the synthesis of two code vectors respectively taken out from the first and second codebooks that store code vectors having a conjugate structure to each other is selected with respect to a target vector and a quantization error is the smallest. A recording medium for recording a program for performing processing of an apparatus, the program comprising: a first step of obtaining a magnitude of a target vector; and a parallel process with the target vector of each code vector in the first and second codebooks. The second step of obtaining the magnitudes of different components, the magnitude of the parallel component, the magnitude of the target vector, and the second codebook for the code vectors of the first codebook that are sequentially designated. The magnitude of the parallel component of each code vector in
A third step of obtaining a candidate code vector of an effective combination with the specified code vector from the code vector in the second codebook based on the current minimum quantization error, and the specified code A fourth step of calculating a quantization error between the vector and each of the candidate code vectors, a fifth step of comparing the quantization error obtained in the fourth step with the current minimum quantization error, and a fifth step If the quantization error obtained in the fourth step is smaller in the process comparison, the current minimum quantization error is updated to that quantization error, and the holding code index is obtained in the fourth step. Is updated to the code indexes of both code vectors obtained, and a seventh step of determining whether all the code vectors have been specified in the third step, and a seventh step thereof If the designation has not ended, after the sixth step, or if the quantization error obtained in the fourth step in the comparison of the fifth step is larger, the designated code vector of the third step is changed to An eighth step of returning to the third step, and a ninth step of outputting the held code index as a quantization result when it is determined in the seventh step that the designation for all code vectors is completed, A computer-readable recording medium characterized by being executed. 7. The program executes a tenth step of rearranging the magnitudes of the parallel components of the second codebook obtained in the second step in ascending order, and the third step is the magnitude of the target vector. Therefore, the eleventh step of calculating the effective upper limit value and the effective lower limit value by adding and subtracting the square root of the current minimum quantization error to and from the value obtained by subtracting the parallel component of the designated code vector, From the size of the parallel components arranged in ascending order, the number m of the largest parallel components smaller than the above effective upper limit value
and _max (n), a twelfth step of obtaining the smallest of the numbers m _min in the effective lower limit is greater than the parallel component (n), each ranging from the number m _max (n), the number m _min (n) 7. The recording medium according to claim 6, further comprising a thirteenth step of setting the code vector of the second codebook corresponding to the number as the candidate code vector. 8. The first and second codebooks are preselected from the first and second large codebooks each having a larger number of code vectors than the first and second codebooks.
The recording medium according to claim 6 or 7, wherein the fourteenth step of obtaining the second codebook is executed.