JPH04334206A - Method of generating code book for vector quantization - Google Patents

Abstract

PURPOSE:To reduce the influence of an initial value so as to generate a code whose quantization distortion is small in a method of generating a code book for vector quantization such as sound encoding and the like. CONSTITUTION:A code book initial value given at a step 1 is updated using conventional code book generating algorithm at steps 2 and 3. The code book is analyzed and an ineffective code vector and a code book whose quantization amount is large are decided at a step 4. The ineffective vector is eliminated and the code vector whose quantization distortion is large is divided into two at a step 5. The same processing as that at the step 2 is executed at a step 6 and the reduction of quantization distortion is decided at a step 7. The code book is updated at a step 8 in the case of reduction and a processing operation is finished in other cases.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for creating a codebook for vector quantization used in audio coding, image coding, etc.

0002

[Prior Art] Figure 4 shows a conventional codebook creation method (L
BG algorithm) is shown. In FIG. 4, first, in step 11, initial values of the codebook are given. Then, in step 12, the training vector is classified using the distance of the code vector, and in step 13
In step 14, the centroid of the training vector classified into each code vector is calculated to obtain a new code vector, and in step 14, the quantization distortion is calculated using the training vector. Then, in step 15, the quantization distortion is compared, and if the quantization distortion has decreased, the codebook is updated in step 16, and in other cases, the codebook creation procedure is ended.

Next, the operation of the above conventional example will be explained in more detail. Assuming that the data is two-dimensional, in step 11, the code vector (y0, ·
. . , y3 ) are given as initial values, and then the training vectors (x0 , . . . , x7 ) are similarly given as shown in FIG. 5(a). In step 12, the training vectors x0,...,x7 are the code vectors y0, y1, y with the smallest distance.
Classified as 2. That is, x0, x1 are classified into y0, x2, x3 are classified into y1, and x4, x5, x6, x7 are classified into y2. Here, there is no training vector classified as code vector y3, and y3 is left out. Next, in step 13, the centroid of each classified training vector is calculated, and the code vectors y0, y1,
Obtain y2 and y3.

Next, in step 14, quantization distortion D is determined.

[0005]

[Math 1]

Next, in step 15, the quantization distortions are compared, and if the quantization distortion of the newly calculated codebook has decreased, the codebook is updated in step 16, and if the quantization distortion has not decreased, the codebook is updated. Finish the processing operation.

[0007] As described above, even with the conventional codebook creation method described above, a codebook with small quantization distortion can be created by calculating the centroid of the training vector.

[0008]

[Problems to be Solved by the Invention] However, in the conventional codebook creation method described above, the influence of the codebook given as an initial value remains until the end, resulting in a bias in the distribution of codevectors, such as codevector y3. However, there was a problem in that code vectors could not be used efficiently during quantization, making it difficult to minimize quantization distortion.

The present invention is intended to solve these conventional problems, and provides an excellent codebook creation method that can reduce the influence of the codebook initial value and create a codebook with small quantization distortion. The purpose is to provide.

0010

[Means for Solving the Problems] In order to achieve the above object, the present invention adds codebook distribution, deletion, and division functions to the conventional codebook creation algorithm to create code vectors and code vectors with a small quantization distortion reduction effect. The quantization distortion is reduced by determining, deleting, and dividing code vectors with large quantization distortion.

[0011]

[Operation] Therefore, according to the present invention, code vectors with a small quantization distortion reduction effect are deleted and code vectors with a large quantization distortion are divided to reduce the quantization distortion, thereby quantizing each code vector. Distortion can be averaged, the influence of the initial value of the codebook can be reduced, and a codebook with small quantization distortion can be created.

0012

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is a flowchart showing the processing procedure of a codebook creation method according to an embodiment of the present invention. In FIG. 1, first, in step 1, initial values of the codebook are provided. Next, in step 2, a new codebook is obtained using a conventional codebook creation algorithm such as the LBG method, and in step 3, the codebook is updated. In step 4, the updated codebook is analyzed, and in step 5, code vectors are deleted and divided. In step 6, a new codebook is obtained using the same method as in step 2, and in step 7, the quantization distortion is compared. If the quantization distortion decreases, the codebook is updated in step 8, and in other cases. ends the processing operation.

Next, details of the operation of the above embodiment will be explained. FIG. 2A shows an example of two-dimensional vectors as in the conventional example, where x0 to x7 are training vectors, and y0 to y3 are code vectors given as initial values. First, by steps 1, 2, and 3, y0 is x0, x1, y1 is x2,
The centroids of x3 and y2 are updated to the centroids of x4, x5, x6, and x7, and are distributed as shown in FIG. 2(b). Here, since there is no training vector classified as code vector y3, code vector y3 is not updated.

Next, in step 4, N (in FIG. 2, N
=4) code vectors y0,...,y3, n training vectors x0 (n=8 in Figure 2)
,...,x7, and the sum of quantization distortions Dm (m=0,...,N-1) for each code vector.
and the importance level Vm of the code vector is defined, and Dm and Vm are calculated for each code vector.

[0015]

[Math 2]

[0016]

[Math 3]

Next, in step 5, the third code vector y3 indicating the minimum Vm is deleted and the maximum Dm
The second code vector y2 indicating y2 is divided into two in the vicinity of y2 to create new y2 and y3. As a result, a code vector as shown in FIG. 3(a) is created.

In step 6, the codebook is updated using a conventional codebook creation algorithm, and in step 7, the quantization distortion D is compared.

[0019]

[Math 4]

If the quantization distortion has decreased, the codebook is updated in step 8, and the process returns to step 4. If the quantization distortion does not decrease, the processing operation ends. As a result, as shown in FIG. 3(b), a codebook is created in which the newly divided code vectors y2, y3 are updated to the centroids of the training vectors x4, x5 and x6, x7.

As described above, according to the above embodiment, the code vector y3, which is of low importance in reducing quantization distortion, which was left behind during classification in the conventional codebook creation algorithm, is deleted, and the quantization distortion is reduced. Divide the large code vector y2 into two to create a new code vector y2
, y3, it is possible to average the contribution of each code vector to the quantization distortion, and it is possible to reduce the influence of the initial value and create a codebook with small quantization distortion.

[0022]

Effects of the Invention As is clear from the above embodiments, the present invention eliminates code vectors that have a small quantization distortion reduction effect, divides code vectors that have a large quantization distortion, and shares the quantization distortion of the code vectors. Since the values are averaged, it is possible to reduce the influence of the initial value and create a codebook with small quantization distortion.

[Brief explanation of the drawing]

FIG. 1 is a flowchart showing the processing procedure of a codebook creation method in an embodiment of the present invention.

[Figure 2] (a) is a schematic diagram showing the initial values of the codebook in the same example. (b) is a schematic diagram of the codebook obtained by calculating the centroid of the training vector in the same example.

[
Figure 3 (a) is a schematic diagram of a codebook obtained by deleting unnecessary code vectors in the same example and dividing the code vector with large quantization distortion into two. Schematic diagram of the codebook obtained by calculating Lloyd's

[Fig. 4] Flowchart showing the processing procedure of the codebook creation method in the conventional example.

[Figure 5] (a) is a schematic diagram showing the initial values of the codebook in the same example. (b) is a schematic diagram of the codebook obtained by calculating the centroid of the training vector in the same example.

[Explanation of symbols]

1 Codebook initialization step 2 Step for codebook creation algorithm 3 Codebook update step 4 Codebook analysis step 5 Code vector deletion and division step 6 Step for codebook creation algorithm 7 Quantization distortion reduction determination step 8 Code Book update step

Claims (1)

[Claims] Claim 1: The code vector is updated by calculating the centroid using the training vector from the initial value of the code vector in the codebook, and then the distance between the code vector and the training vector is analyzed to determine the quantization distortion reduction effect. A method for creating a codebook for vector quantization, in which codevectors with small quantization distortion are deleted and codevectors with large quantization distortion are divided and the codebook is updated.