JPS59201170A - Vector quantizer - Google Patents

Abstract

PURPOSE:To attain a coding system having high efficiency of quantization by providing a DPCM decoder and an amplitude calculator at the input side of a quantizer and therefore applying the predictive coding to the block average. CONSTITUTION:A mean value calculator 2 of a vector quantizer obtains the block average value of an input vector 1, and the correlation between blocks of mean value is used to compress data through a DPCM encoder 31. Thus a DPCM signal 35 and a DPCM decoding signal 37 are delivered. The signal 37 is subtracted from the vector 1 by a subtractor 4 to obtain a mean value separation vector 40. An amplitude calculator 6 calculates an intra-block amplitude 7 from the vector 40. At the same time, the vector 40 is divided by the amplitude by a divider 8 and a mean value separation normalizing vector 41 is applied to a register 10 to obtain a mean value separation normalizing input vector x'. Then the processing is performed through a code table memory 12, a parallel absolute value operator 15, a maximum element distortion detector, a code address counter 11, a latch 19, etc. As a result, the efficiency of quantization is improved.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to vector quantization in which a plurality of input signal sequences are grouped into blocks and quantized in a multidimensional signal space.

Let me briefly explain the principle of vector quantization.

Now, the information source input signal sequence is divided into pieces and the input vector
−CXi, X2...+XK]. At this time.

N representative points of the K-dimensional Euclidean signal space RK (XERK) (i.e. output - mactor) χ1 - [A □1 + y
42+..., y ]'s Sesera is Y"= [71+ 7
2 +・−・, C] and set it to 1. If R1, R2, .

Q:RK-)Y Here, R1==Q'(ヱ,)=(OtsuERK:Q(t)
)=ヱ,)UR1=RK,R,nRj-φ (1~j)
i=1 The above vector quantization Q is expressed as a cascade of encoding C and decoding. The encoding C is the mapping of the RK output vector to the index set I=[1,2,...,N], and the decoding DFi This is mapping from engineering to Y. Namely.

C: RK → I, D: Engineering → Y, and Q = D - C.

In vector quantization, the encoded output (2) is transmitted or recorded, resulting in extremely high encoding efficiency.

Vector quantization is to hematize an input vector to an output vector x1 that has the shortest distance (results in the least distortion). Specifically, the distance (distortion) between the input and output vectors is expressed as d(
``q + 3'1), it becomes as follows.

1f d (L V,) (d(X, 7 j)
for all j −1 XERi, that is, attraction→ヱ, indicates a relationship.

The set Y of output vectors B, as shown in Figure 1, is obtained by clustering using the information source input signal sequence that becomes the training model (selection of representative points and division of the signal space are performed in a manner that minimizes the sum of distortions). (repeat) can be found.

The configuration of a conventional vector quantizer will be described below with reference to FIGS. 2, 3, and 4.

FIG. 2 is an example of an encoder.

In the figure, (1) is an input vector, and (2) is an average value calculator.

(3) is the block average value calculated by the average value calculator (2), (4) is the subtracter, and (5) is the subtracter (
4) is the mean value separation vector subtracted by the block mean value (3), (6) is the amplitude calculator that calculates the amplitude within the block, (7) is the intra-block amplitude, (8) is the divider, (
9) converts the mean value separation vector (5) into the intra-block amplitude (7
), the mean value separation normalized vector obtained by dividing by (1
1 is a register.

αυ is the code table address counter, 02 is the code table memory, a3 is the code table output vector register, (14) is the parallel subtractor, α is the parallel absolute value calculator, αQ is the maximum element distortion detector, aη is the minimum distortion Vector detector, a barrel is an index signal, (11 is a latch.

I21 is the encoder output signal.

Further, FIG. 3 shows an example of the configuration of the amplitude calculator (6).

In the figure, (2υ is an absolute value calculator, and (2υ is an average value calculator.

Figure 4 shows an example of the configuration of a decoder.

In the figure, (c) is a latch, +2→ is an average value separation normalized output vector, (c) is a multiplier, @ is an average value separation output vector, (27) is an adder, (2) is an output 4 vector, @ is the output vector register, and (to) is the decoding fixture output signal.

Next, the operation will be explained.

First, the operation of the encoder will be explained according to FIGS. 2 and 3. The input signal sequence of the encoder is divided into blocks into input vectors X ``'' (Xl, X2
,...strong). For the input vector (1), the mean value calculator (2) calculates the block mean value m(3) by m-(xj). The subtractor (4) inputs the input vector (
Subtract the block average value m from 1) to obtain the average value separation vector (5) represented by xj-=m. Amplitude calculator (6
) calculates the intra-block amplitude σ from the mean separation vector. There are various methods for calculating σ, and 2 to 30 examples are shown below.

■ σ−E (Ixjml) ■ σ−[E(x−m)2]1/2 ■ a=mgX Find the absolute value of the mean value separation vector (5) and use the mean value calculator (
c) to find σ.

The divider (8) normalizes the input vector by dividing the mean value separation vector (5) by the intra-block amplitude σ,
Obtain the mean value separation normalized long sword vector X'.

That is, x'-(xjm)/σ By the mean value separation normalization, the vector is mapped onto a spherical surface centered on the origin of the dimensional signal space, and the quantization efficiency is increased. In other words, it is a mixture with scalar quantization. The relationship between pectonori' and quantization vector / at this time is shown in FIG.

The code table address counter 0υ sequentially calculates the average value separation normalized output vector y/from the code table a2.
, is read and latched into register α east.

The maximum element distortion detector αQ calculates the distortion d□ of y/1 from the parallel subtractor ◯ and the parallel absolute value calculator σ9 as follows.

di=d (X', yQ) -maX l x'j-
y/□j1 Next, the minimum distortion detector aη is sequentially read out y/.

The minimum value of the distortion d1 between 1 and 5' is detected. That is, the minimum distortion is d-mindl. The vector with the minimum distortion is the vector quantization output of vector winter'. When the minimum distortion vector is detected, a strobe signal is sent to the latch member.

Take in the index signal 0, which is the address of the vector. The latch OI also takes in the block average value (3) and intra-block amplitude (7) and outputs them as the encoder output signal (A).

Next, the operation of the decoder shown in FIG. 4 will be explained.

The encoder output signal (1) is taken into the latch (c) of the decoder, and the average value separated normalized output vector y'i read out according to the address indicated by the index signal (1υ) is latched to the register. □041 is multiplied by the intra-block amplitude (7) in the multiplier (c), and
The block average value (3) is added by the adder (2η), resulting in an output vector yc! (to), which is latched in the register (c) to obtain the decoder output signal 00.

That is, y, = σ・y', j 10m In this case, the encoding efficiency η is η if the code table has N vectors and a pit and b bit are assigned to the block average value and intra-block amplitude, respectively. = (18g2N+a + b)/K (bits/sample)
It is.

Since the conventional vector quantizer is constructed as described above, it has the following drawbacks. In other words, since mean value separation normalization does not perform any data compression processing on scalar quantities such as the mean value and amplitude, the compression effect of the vector quantizer is not fully utilized as a whole system. be.

This invention was made to eliminate these drawbacks by applying predictive coding to block average values that have strong correlations even between blocks.

The purpose is to provide a vector quantizer with high quantization efficiency.

An embodiment of the present invention will be described below with reference to the drawings.

FIG. 6 shows an example of the configuration of an encoder. In the figure, (1) is an input vector, and (2) is an average value calculator.

OD is the DPMC encoder 2oz is the subtracter, (to) is the prediction error signal, 04) is the quantizer, C351 is the quantized prediction error signal, (g) is the adder, GD is ])P OM
decoded signal, (to) is the predictor, C31 is the predicted signal, (4
) is the subtractor, (41 is the mean value separation vector, (6) is the amplitude calculator, (force is the intra-block amplitude, (8) is the divider, (
4υ is the mean value separation normalized vector, 0 is the register, α
υ is a code table address counter, Q2 is a code table memory, (1■ is a code table output vector register, 04 is a parallel subtracter, (151 is a parallel absolute value calculator, (1Q is a maximum element distortion detector, (1η is a Minimum distortion vector detector, 08 is index signal, ■ is latch, (
4) is the encoder output signal.

In fact, FIG. 7 shows an example of the configuration of the DPOM encoder 0υ.

In the figure, (4z is a delay element, (4 is a multiplier.)

FIG. 8 shows an example of the configuration of a decoder.

Next, the operation will be explained.

The operation of the encoder will now be described with reference to FIGS. 6 and 7. The input signal sequence of the encoder is grouped into professional,′,:! The input vector winter −(Xl, X2
．． ..., XK). For input vector (1),
The mean value calculator (2) calculates the block mean value m(3) by m = E (Xj). This block average value m
Since the data rate will be higher if it is transmitted using POM, D
A PCM encoder C31) performs data compression processing using the correlation between blocks. Here, for the block average value in (3), C31 is the prediction signal Pm, [phase] is the prediction error signal em, C1 is the DPCM signal △ ◇, C37+]) p is the CM decoded signal, and the quantizer If Oa is above the quantization noise generated to reduce the number of quantization levels, △ △ −1Qm=
==m-p□”-ga・prn-8m゛m°z△
At Δm 0 m pm , a DPCM signal (up to) with a reduced number of quantization levels can be formed. The predicted signal pm is obtained in a quadratic manner. That is, as the input of the predictor (to) D P I:! The M decoded signal clη is applied, a signal with an appropriate time delay (for example, 2 represents a delay of one block) is obtained by two delay elements, and a predetermined coefficient a is obtained by the 2 multiplier 03.
It is multiplied by 1n and becomes the prediction signal C31.

Although only one delay element is used here, a prediction signal may be formed by using several delay elements and linearly combining the delayed signals obtained from them.

A subtracter (4) subtracts the block average value DPOM decoded signal C171 from the input vector (1) to obtain an average value separation vector. An amplitude calculator (6) calculates the intra-block amplitude σ from the mean value separation vector. The divider (8) divides the mean value separation vector (41) by the intra-block amplitude (7) to obtain the mean value separation normalized input vector each', i.e. △X'=(Xj-m)/σ Code table address counter 01) sequentially reads the average value separated normalized output vector N' from the code table circle and latches it into register 0. The maximum element distortion detector OD obtains the distortion d1 of the capital and X'□ from the parallel subtracter Q4) and the parallel absolute value calculator 00 as follows.

d, -d (winter's'1) -m negative X1x/l-y'□j1
Although the maximum element type is used here as the definition of strain, a square bundle or absolute value strain may also be used.

Next, the minimum distortion detector aη detects the minimum value of the distortion d1 between 6' and 6' when read out sequentially. That is, the minimum distortion is d = +Qn 6.1. The vector with the minimum distortion is the vector quantization output of the vector. When the minimum distortion vector is detected, a strobe signal is sent to the latch (11).

The index signal 08, which is the address of the vector, is taken in. In latch α1, the block average value DPOM signal 0
9. It also takes in the intra-block amplitude (force) and outputs these as the encoder output signal.

Next, the operation of the decoder shown in FIG. 8 will be explained.

The encoder output signal ■ is taken into the latch 123 of the decoder, and the average value separated normalized output vector Y' read out according to the address indicated by the index signal a& is latched into the register a:i. On the other hand, the block average value D:
A D POM decoded signal 0D is obtained from the POM signal G9 using the following equation. Namely.

△　／＼ -0mpm The prediction signal pmCHI is obtained by a predictor.

The vector y/, 04 is multiplied by the intra-block amplitude (7) f: M'rt in the multiplier (C), and added to the adder Q7).
At , the DPCM decoded signal 67) of the block average value is added, resulting in an output vector yeυ, which is latched in the register to obtain the -C decoder output signal (to). Namely.

△y−0σ゛y′・10m J ]LI In this case, the coding efficiency η is the code table having N vectors, and the quantizers C3 and Il, which perform DPCM quantization of the block average value, perform quantization of small bits. If a level is given and b bits are allocated to the intra-block amplitude, Δη=(1ag2N +a, +b)/K (bits/sample).

Note that in this example, the block average value subtracted from the input vector on the encoding side is a DPOM decoded signal, but the block average values before being subjected to DPOM encoding and decoding may be directly arrayed. .

As described above, according to the present invention, coding with high quantization efficiency can be realized by performing predictive coding on block averages.

[Brief explanation of drawings]

Figure 1 is an explanatory diagram explaining the principle of vector quantization, Figure 2
This figure is a block diagram of a conventional vector quantization encoder, and FIG. 3 is a block diagram showing an example of an amplitude calculator. Fig. 4 is a block diagram showing an example of a conventional vector quantization decoder, Fig. 5 is an explanatory diagram showing the arrangement of output vectors in mean separation normalization vector quantization, and Fig. 6 is a diagram showing the arrangement of output vectors in mean separation normalized vector quantization. A block diagram showing one embodiment of such a vector quantization encoder, FIG. 7 is the DPC! FIG. 8 is a block diagram showing an example of a vector quantization decoder according to the present invention. In the figure, (1) is an input vector, and (2) is a block diagram showing an example of an M encoder. Average value calculator. (3) is the block average value, (4) is the subtractor, (5) is the average value separation vector, (6) is the amplitude calculator, (the sword is the temperature width within the block, (8) is the divider, (9) is an average value separation normalization vector. OQ is a register, +Il+ is a code table address counter, 02 is a code table memory, 03 is a code table output vector register, (1a is a parallel subtractor, 04 is a parallel absolute value calculator, arbitrary e is the maximum element distortion detector. 07) is the minimum distortion vector detector, αB is the index signal, Kanji is the latch, C2 is the encoder output signal, (2D is the absolute value calculator, (2 is the average value calculator) C2 is a latch, C24) is a mean value separation normalized output vector, (c)
is a multiplier. (a) is the average value separation output vector, (2η is the adder, (
2) is the output vector, (2!H- is the output vector register, (7) is the decoding fixture output signal, 0υ is the DrCM encoder, 0 is the subtractor, (G) is the prediction error signal, (B) is Quantizer 209 is a DPCM signal, (sub)) is an adder, 07
) is DPOM (jJ signal, (to) is DPOM predictor,
CI! J is DPC! M prediction signal. f4G is a mean value separation vector, (4υ is a mean value separation normalization vector, (42 is a delay element, (43 is a multiplier. Yes. Agent Masuo Oiwa Figure 1 y Figure 3 Figure 4 Figure 5 Figure 3 2

Claims (2)

[Claims] (1) Separate the intra-block average value from the input vector formed by dividing the input signal sequence into blocks (K is plural), and further divide the standard deviation component of the elements within the block or the difference between the maximum element and the minimum element 2 -1. an average value separation and normalization circuit that obtains an absolute value average ring of the height deviation as an intra-block amplitude, normalizes a vector obtained by separating the intra-block child devices by the intra-block amplitude, and forms an average value separation normalized vector; A predictor that predicts the current value from the past block average value in order to predictively encode the average value. a subtractor for obtaining a prediction error signal between the average value predicted by the predictor and the current average value; a quantizer for scalar quantizing the prediction error signal; and a quantizer for the value predicted by the predictor and the quantization. an adder that calculates a predicted decoded signal of the intra-block average value by adding the prediction error signal whose output level has been reduced by the adder, and an adder that calculates the predicted decoded signal of the average value within the block; a first code table memory that obtains and stores a set of output vectors; a code table address counter that sequentially reads the mean value separated normalized output vector from the first code table memory; and a mean value separated normalized input vector; A distortion calculator that calculates and compares the square bundle, maximum element distortion, or absolute value distortion with the sequentially read vectors to detect an average separation normalized output vector that is the minimum distortion, and a vector that provides the maximum /J% distortion. )/an encoder that encodes the code table address of the code table, the intra-block average value prediction error signal, and the intra-block imaging width; a decoder that reads out the mean-separated normalized output vector giving the minimum distortion from the same second code table memory; a multiplier that processes and combines the intra-block amplitude with the mean-separated normalized output vector; A vector quantizer characterized by comprising an adder that adds a predictively decoded intra-block average value to a vector obtained by combining amplitudes to obtain an output vector. (2) In the vector quantizer, when performing mean value separation and normalization of a vector, a mean value separation and normalization circuit that uses a DPCM decoded signal of the intra-block average value instead of the intra-block average value, and a code table address and the block Claim 1 is characterized in that it is equipped with an encoder that encodes the DPOM signal of the intra-block average value and the intra-block amplitude.
Vector quantizer described in ).