JPH0262167A - Optimum output vector searching method for vector quantization - Google Patents

Abstract

PURPOSE:To contrive the high speed of a searching operation by outputting either one of an optimum output vector candidate or the positive-negative symmetrical output vector candidate as an optimum vector according to the sign of an inner product value. CONSTITUTION:The inner product operation of an output vector candidate Yi<+> and an input vector X in a code book is operated, and the value having a maximum absolute value is selected as an optimum output vector candidate Y0<+>. When the sign of the inner product is positive, it is outputted as the optimum output vector as it is. When the sign of the inner product is negative, a vector Y0<-> obtained by inverting the sings of the respective elements of an optimum output vector candidate Y0 is outputted as the optimum vector. Thus, the number of the output vector candidates held in the code book can be reduced by half.

Description

[Detailed Description of the Invention] [Kaname Taka] This invention relates to an optimal output vector search method for vector quantization used in high-efficiency encoding of image signals or audio signals.

In order to speed up the search process by reducing the number of calculations required to search for an optimal output vector, a codebook is prepared that stores multiple output vectors with no positive/negative symmetry of elements, and Ijk'A output vector candidates are searched from the codebook based on the absolute value of the inner product of the output vector candidates, and depending on the sign of the inner product value, the optimal output vector candidate or the output vector candidate whose elements are symmetrical with it is searched. Either one is configured to be output as an HA output vector.

[Industrial application field]

The present invention is a vector ninetizer used for high-efficiency encoding of 3 image signals or audio signals, and in particular, a GAIN/5II APF vector 9 in which the average value of the component elements of the output vector is normalized to zero and the standard deviation is normalized to a constant value. I went through childbirth. This paper relates to an optimal output vector search method for vector quantization.

[Conventional technology]

GAIN/5IIAPE vector quantization is a type of vector quantization used for high-efficiency encoding of image signals or audio signals. This GAIN/Sll^PE vector quantization subtracts the average value of its elements from the input signal (input vector) and normalizes it by the standard deviation.

In the output vectors in the prepared codebook, the average value of the elements is all zero, and the standard deviation σ is normalized to σ=1, for example.

The search method for the bk output vector in vector quantization is to compare the input vector with all output vectors in the codebook, find the output vector with the least distance (error), and use it as the optimal output vector. be.

[Problem to be solved by the invention]

In -a vector quantization, when searching for an optimal output vector, it is necessary to search all output vectors in the codebook and calculate an error. For this reason, if the number of output vectors in the codebook is large, the amount of calculation becomes large, and there is a problem that the search takes a long time.

Therefore, an object of the present invention is to reduce the amount of calculation for searching for an optimal output vector and speed up the search process.

[Means to solve the problem]

FIG. 1 is a diagram explaining the principle of the present invention. In the optimal output vector search method for vector quantization according to the present invention, a plurality of output vector candidates y without positive/negative symmetry of elements,
A codebook Y+ in which tens is stored is prepared (step Sl). Each output vector candidate y in this codebook
, 10 and the input vector (Step 33).

Next, the optimal output vector candidate 7 is determined depending on the sign of the inner product value.
o +, or an output vector candidate yo− that is symmetrical with this output vector candidate yo+, in which the positive and negative of the constituent elements are reversed.
Either one of them is output as the optimal output vector (step 34).

[Effect]

An inner product operation is performed between the output vector candidates yI in the codebook and the input vector X, and the one with the maximum absolute value is selected as the optimal output vector candidate yo÷. For example, when the sign of this inner product is positive, its iJ
Output the output vector candidates yo ten as they are as optimal output vectors.

On the other hand, when the sign of the inner product is negative, a vector yo− with the sign of each element of the optimal output vector candidate yo reversed
is output as the optimal output vector.

As a result, the number of output vector candidates held in the codebook can be halved, and therefore the number of calculations for searching can be halved, thereby speeding up the optimal output vector search process.

〔Example〕

The present invention will be described in detail below with reference to one drawing. In this embodiment, the method of the present invention is applied to a motion compensated interframe predictive coding circuit.

In FIG. 2, l is a subtracter that calculates the difference between one block of input images and a predicted value P, and outputs a prediction error X. 2 is GAIN/SH^PF, vector spermization, where one block of prediction error X is input as input vector value m'
, gain f1. Output vector index i. Here, the input vector X consists of elements (pixels).

It is expressed as x= (xl, X2---xk).

3 is an adder which adds the quantized value X' and the predicted value P to generate a locally decoded signal A'. 4 is about 1 input signal
A frame memory 5 stores the previous screen by delaying the frame, and 5 compares the previous screen in the frame memory 4 with the input image A of the current screen to determine the input image.
6 is a motion compensator that detects the motion vector V, and 6 is a variable delay that shifts the previous screen of the frame memory 4 according to the motion vector V from motion compensation'/:t5 to generate a motion Tt+m predicted value P. It is a type.

GAIN/
The detailed configuration of the 5IIAPE vector quantizer 2 is shown in FIG. In FIG. 3, 20 is a codebook, which is made up of a set Y of a plurality (nl&1) of output vector candidates VI (i=l to n) that have been windowed in advance. Each output vector candidate y consists of '! = (y + I, y + □, ---yIk
), and the set Y of these is expressed as .

Y = (71, y2---yn l It is expressed as

This output vector candidate y has an average value of zero and is a normalized vector, so the average value m is.

k j=1 So, the standard deviation σ is.

becomes. These output vectors generally lie on a (k-1) dimensional hypersphere.

The output vector family M 7 + in the codebook 20 does not include elements whose elements are symmetrical with respect to each other. In other words, if a certain output vector group is Y0, and a group of output vectors whose elements are symmetrical with each other is Y-, then the output vector candidates included in the codebook 2o are Y+ or Y-. It's one or the other. Y hit output vector candidate.

7+ ”= (yl 1 ”ryi 2”+ ”-
``'-y1 is +1, this output vector candidate y,
The sign of each element of 10 is inverted and becomes the output vector candidate in Y-.

Y+-=()'ll", YI2"----3'ik")
= (yl+-, yt 2-+ "-""yi k-
). Therefore, the number of output vector candidates in codebook 2o is only for either Y0 or Y-, compared to the case where both Y0 and Y- are used (Y = Y + UY-). The number will be halved.

21 is a gain calculation and output vector search circuit,
Input vector X and output vector family in codebook 25? ! A gain (inner product: g+) with 31'+ is calculated to search for an optimal output vector, and the gain gi and vector index i are output.

22 is an average value calculation circuit for calculating the average value m' of the elements of the input vector a multiplier that multiplies 24
is the output gI'! from the multiplier 23. This is an adder that adds J and the average value m' from the average value calculation circuit 22 to reproduce the original signal.

The detailed configuration of the gain calculation and output vector search circuit 21 is shown in FIG. In Fig. 4.

211 is the key point Nxj of the input vector X and the key point of the output vector candidate y! 212 is an adder that performs a summation operation of the multiplied values of each element, 213 is a register, 214 is an absolute value calculation circuit that takes the absolute value, 215 is a maximum value detection circuit that detects the maximum value, 216 is a register, and the maximum value detection circuit 215 outputs the vector index i of the output vector candidate yI with the maximum gain gl.

The register 216 outputs the gain gi at that time.

The operation of the example device will be described below.

GAIN/Sll^PE vector quantization base 2, output vector group closest to 2X in terms of human quantization if 7
To search for r in the codebook 20, find the distortion d between the input vector Let this be the optimal output vector y. I mean.

The search for the optimal output vector (candidate yo is defined as

By searching for i that satisfies sin d I = sin xyl (where i = 1-n). Here, the distortion d is.

It is expressed as dl= x 71 =Σ(xjj・1 yi 1 )2. In this equation, the value of the first term depends only on the input vector X, and the third term has a constant value regardless of the output vector because the codebook has been normalized. , is the inner product (gain) of the input vector X of the second term and the output vector candidate y1.

It can be seen that i is the value that maximizes 2ΣXj')'IJ j=1. Therefore, by finding i that maximizes this inner product value, the optimal output vector y is obtained. can be explored.

Here, it is assumed that the codebook 20 stores only Y1 of YJ or Y-. Gathering Y
Absolute value hI of the inner product value of the output vector quantization and the input vector X in 10.

hI=x-I! ”, and from among them, find i for which the absolute value hi is maximum. At this time.

If gl=x −71+ ≧ O, YJ+ is selected as the optimal output vector yo.

gl〈0 If so, -y,+=y,- is selected.

As a result, the number of inner product operations can be reduced to half compared to the case where the output vector set Y includes both YJ and Y-. In this way, by utilizing the symmetry of the vectors in the codebook, the number of operations can be halved.

Looking at this from a different perspective, we search for the vector of i that maximizes hi from a codebook in which only Y, which has no positive/negative symmetry, is prepared in advance, and if hI≧0, that vector is output, and hi<Q Then, by outputting a vector with the signs of all elements of the selected vector inverted, it is equivalent to having twice as many output vectors in the codebook.

In addition, regarding the reproduction of the input vector, assuming that the average value of the input vector X is k j = 1, the reproduction value J f is as follows.

It can be determined by YJ'""m' + g i 'Yj.

〔Effect of the invention〕

According to the present invention, by performing the inner product operation N times, it is equivalent to performing a total of 2N searches for N output vector candidates and N output vector candidates obtained by reversing the sign of each element. . Also, from another perspective, for the same number of output vector candidates, the method of the present invention requires half the number of operations compared to the conventional method.

It is possible to speed up the search calculation for the optimal output vector.

[Brief explanation of the drawing]

FIG. 1 is a diagram explaining the principle of the optimal output vector search method for vector quantization according to the present invention. FIG. 2 is a block diagram showing a motion-compensated interframe predictive coding circuit that performs a vector quantization optimal output vector search method as an embodiment of the present invention. Figure 3 shows the GAIN/5IIA in the example circuit of Figure 2.
A block diagram showing a detailed configuration example of the PII' vector quantizer 2; FIG. 4 is a block diagram showing a detailed configuration example of the gain calculation and output vector search circuit 21 in the circuit shown in FIG. In fig. 1--1 subtractor 2-G^IN/5II API vector! Seven child converter 3.2
4,212 - Adder 4 - Frame memory 5 - Motion compensator 6 - Variable delay unit 20 - Codebook gain calculation and output vector average value calculation circuit 211 - Multiplier 216 - Register absolute value calculation circuit Maximum value detection circuit 23゜213゜tor search circuit

Claims (1)

[Claims] A codebook is prepared in which a plurality of output vector candidates without positive/negative symmetry of elements are stored, and the optimal output vector candidate is selected based on the absolute value of the inner product of the input vector and the output vector candidate. A vector quantization optimum vector searched from the book and configured so that either the optimal output vector candidate or an output vector candidate whose polarity is symmetrical to the optimal output vector candidate is output as the optimal output vector, depending on the sign of the inner product value. How to search for output vector.