JPH03228433A - Multistage vector quantizing system - Google Patents

Abstract

PURPOSE:To execute the embedded encoding by containing each orthogonalized code vector in a code book of a quantizer of the post-stage, and selecting the code vector from this code book so that the norm of an error signal becomes minimum. CONSTITUTION:Between each quantizer 1A and quantizer 2A, an intermediate code book 27 in which many code vectors are contained is provided. Also, a coefficient corresponding to an amplification factor of the code vector outputted from a code book 13 of the quantizer 1A of the pre-stage is derived, quasi-orthogonalization for synthesizing this coefficient and each code vector outputted from the intermediate code book 27 is executed, and a quasi- orthogonalization arithmetic means 28 for containing it in a code book 23 of the quantizer 2A of the post-stage is provided. In such a state, from the code book 23, the code vector is selected so that the norm of an error signal becomes minimum by an arithmetic means 22 of the same stage, and also, amplified by an amplifying means 22, by which a quantizing code is obtained. In such a way, embedded encoding is executed, and also, an optimal vector being free from a loss is obtained by the quantizer of a second stage and thereafter.

Description

[Detailed Description of the Invention] Table of Contents Overview Industrial Field of Application Conventional Technology Problems to be Solved by the Invention Means for Solving the Problems Implementation Examples Outline of Effects of the Invention The present invention provides multiple gain/shape effects. Regarding a multi-stage vector quantization method in which a quantizer is connected to quantize and encode an input signal, and in particular, a multi-stage vector quantization method that can be applied to embedded coding, the method can perform embedded coding, and 2 In order to obtain optimal vectors without loss in the quantizers in the subsequent stages, we use a codebook that accommodates a large number of codevectors and an amplification device that amplifies the codevectors output from this codebook with a desired amplification factor. means for calculating an error signal by calculating an error between the code vector output from the amplifying means and the input vector; and selecting a code vector of the codebook such that the norm of the error signal is minimized. In a multi-stage vector quantization method in which a plurality of quantizers each having a calculation means for quantizing and encoding an input signal such as audio and image, that is, the input vector, is connected, each of the quantizers and the quantizer Between the quantizer and the quantizer, there is an intermediate codebook that stores a large number of code vectors, and each code vector output from this intermediate codebook is connected to the code output from the codebook of the previous stage quantizer via an amplification means. orthogonalization calculation means for orthogonalizing the vector and storing each of the orthogonalized code vectors in a codebook of a subsequent quantizer; The code vector is selected by the arithmetic means in the stage so that the norm of the error signal is minimized, and is amplified by the amplification means in the same stage to obtain a quantized code.

Industrial Application Field The present invention relates to a multi-stage vector quantization method that connects a plurality of gain/shape quantizers to quantize and encode an input signal, and in particular a multi-stage vector quantization method that can be applied to embedded coding. It is related to.

A multistage vector quantization method is already known as a band compression technique aimed at highly efficient transmission of audio and image signals, but this method uses a codebook consisting of multiple types of code vectors to encode the target. This method selects the pattern closest to the input vector (input audio/image signal), connects a desired number of quantizers that use the pattern number as encoded information, and quantizes the input signal. In this method, it is possible to perform embedded encoding in which reproduction is performed using any of the encoded information obtained for each quantizer connected in multiple stages.

This embedded encoding is, for example, discarding the encoded information from the second stage onward and performing reproduction using only the encoded information from the first stage. However, if the encoded information in the first stage is quantized into information that can withstand playback, and then the encoded information is quantized again in the second stage, the encoded information obtained in the second stage is , the information deviates from the input vector, that is, it becomes lossy information.

Therefore, it is necessary to eliminate this loss and also be able to implement embedded encoding.

BACKGROUND OF THE INVENTION FIG. 6 is a diagram for explaining a conventional multi-stage vector quantization method.

In this figure, 1 is the first stage quantizer.

This quantizer 1 quantizes and encodes a signal e0, which is an analog signal such as a voice sampled at a desired number of samples, that is, an input vector X. 2 is a second stage quantizer having the same function as the first stage.

The first stage quantizer 1 includes an error calculation section 11 and an operation section 12.
, a codebook 13 , and an amplifier 14 . The error calculation unit 11 calculates the input vector X and the amplifier 14.
It calculates the error with the vector 81all output from the vector 81all and outputs an error signal e. The codebook 13 contains all code vectors of codebook length L1 (i=L2i...L, , L, ≧1
〉. The calculation unit 12 includes an error calculation unit 11
Vectors a1 . At the same time, the amplifier 14 amplifies the vector a output from the codebook 13 with a gain g1 that minimizes the norm of the error signal e1, and outputs the vector g1a1i. Further, the second-stage quantizer 2 also includes an error calculation section 21, an operation section 22, a codebook 23, and the like, as in the first stage.
It is composed of an amplifier 24.

According to such a configuration, as shown in the vector diagram of FIG.
is amplified by the amplifier 14 to obtain the vector Laz, which is the optimal vector for the first stage.

In addition, the optimal vector in the second stage is 1 as shown in the figure.
The vector g1a1i obtained in the second stage and the vector g2az + output from the second stage amplifier 24 become a composite vector X1. Also, the illustrated vector g o
p i + and g. pt2 will be explained later.

Further, in this vector diagram, the error signal e1 becomes a broken line e1 connecting the tips of the input vector X and the vector gnat +, and the error signal e2 becomes a broken line e2 connecting the tips of the input vector X and the vector g2az.

In other words, according to the multi-stage vector quantization method in which the quantizers 1 and 2 are connected, the first stage data alone can withstand reproduction to some extent, even if the second stage data (quantized signal) is not the second stage data (quantized signal). This means that quantization has been achieved. Therefore, this method is called L A N (Local Area Network
For example, when the transmission line becomes busy, even if the second-stage quantized information is discarded by a switch, the first-stage information can be used even if the second-stage information is not present. Although the quality of the signal is degraded compared to the original signal, the receiving side can reproduce the original signal to some extent from the first stage information.

Therefore, it is possible to perform embedded encoding in which the information in the second stage is arbitrarily omitted and the information in the first stage is used for reproduction, so that flexible services can be provided.

Problems to be Solved by the Invention By the way, in the multi-stage vector quantization method described above, even if embedded encoding can be performed, as shown in the vector diagram in FIG. 7, the optimal vector obtained in the second stage is There was a problem in that x1 was separated from the input vector X by the distance of the broken line e2, and a loss occurred accordingly, making it impossible to obtain an optimal vector in the second stage.

To solve this problem, the first stage vector a is amplified to become the vector gopt+, and the second stage vector a21 becomes the vector g. pt2, and these vectors g a p t r and g. p
t2 may be combined to obtain an optimal vector that is substantially the same as the input vector X. However, in this case, the vector g obtained in the first stage. Since pt+ is different from the optimal vector gnat + originally obtained in the first stage, it becomes impossible to perform reproduction based on the information in the first stage. Therefore, in this method to solve the above-mentioned problem, even if the optimal vector can be obtained in the second stage, the information in the second stage is arbitrarily discarded and the embedded data is reproduced using the information in the first stage. There is a drawback that encoding cannot be performed.

The present invention has been made in view of these points,
It is an object of the present invention to provide a multistage vector quantization method that can perform embedded encoding and can obtain lossless optimal vectors in the second and subsequent quantizers.

Means to solve problems FIG. 1 is a diagram showing the principle of the present invention.

According to this figure, a codebook 13 that accommodates a large number of codevectors, and an amplification means 1 that amplifies the codevectors output from the codebook 13 at a desired amplification factor.
4, an error calculation means 11 for calculating the error between the code vector output from the amplification means 14 and the input vector to obtain an error signal, and a code in the code book 13 that minimizes the norm of the error signal. In a multi-stage vector quantization method in which a plurality of quantizers 1A each having a calculation means 12 for selecting a vector are connected to quantize and encode input signals such as audio and images, that is, the input vectors, each of the quantizers is An intermediate codebook 27 containing a large number of code vectors is located between the quantizer 1A and the quantizer 2A,
Each code vector outputted from this intermediate codebook 27 is made orthogonal to the codevector outputted from the codebook 13 of the preceding stage quantizer 1A via the amplifying means 14, and each of the orthogonalized code vectors is then outputted from the subsequent stage. Orthogonalization calculation means 2 stored in the codebook 23 of the quantizer 2A of
8, a code vector is selected from the code book 23 in which the orthogonalized code vectors are stored so that the norm of the error signal is minimized by the arithmetic means 22 of the same stage, and the amplification of the same stage is performed. The quantization code is obtained by amplification by the means 24.

In addition to such components, another codebook storing a desired code vector and a linear prediction means for linearly predicting and synthesizing each code vector of the other codebook are provided, The code vector obtained by the prediction means may be stored in the intermediate code book 27.

Furthermore, according to another aspect of the present invention, in the multi-stage vector quantization method described above, each quantizer 1A and quantizer 2A
An intermediate codebook 27 storing a large number of code vectors and a codebook 1 of the preceding stage quantizer 1A
3, a coefficient corresponding to the amplification factor of the code vector outputted via the amplifying means 14 is obtained, and quasi-orthogonalization is performed to synthesize this coefficient and each code vector outputted from the intermediate codebook 27, and this quasi-orthogonalization is performed. A quasi-orthogonalization calculation means 28 is provided for storing each orthogonalized code vector in the codebook 23 of the subsequent stage quantizer 2A, and a The code vector is selected by the arithmetic means 22 in the stage so that the norm of the error signal is minimized, and is amplified by the amplification means 22 in the same stage to obtain a quantized code.

In addition to these components, another codebook storing a desired codevector and a linear prediction means for linearly predicting and synthesizing each codevector of the other codebook are provided, The code vector obtained by the prediction means may be stored in the intermediate code book 27.

According to the present invention, each code vector output from the intermediate code book provided between each quantizer is converted from the code book of the preceding quantizer by the orthogonalization calculation means. The code vectors output through the amplification means are combined and orthogonalized, and each of the orthogonalized code vectors is stored in a codebook of a subsequent quantizer.

Then, from the codebook containing the orthogonalized code vectors, a code vector is selected by the calculation means at the same stage so that the norm of the error signal is minimized, and is amplified by the amplification means at the same stage. A quantized code is obtained.

Therefore, since it is possible to obtain a quantization code for each quantizer in each stage, it is possible to reproduce information for each quantizer in each stage. Embedded encoding for selective reproduction can be performed.

In addition, the quantized code obtained for each stage of quantizer has substantially the same vector as the input vector because the above-described orthogonalization operation has been performed in the signal processing process.

Also, each code vector obtained by linear prediction is
Since the code vectors are stored in an intermediate codebook and the stored code vectors are used for orthogonalization and quantization, the amount of information can be reduced, thereby increasing information processing efficiency.

Furthermore, according to the quasi-orthogonalization method of the present invention, a coefficient corresponding to the amplification factor of the code vector outputted through the amplifying means is obtained from the codebook of the quantizer in the previous stage, and the coefficient is Since the code vectors output from the amplifying means include quantization errors, the errors can be eliminated. Therefore, as mentioned above, even if there is a quantization error, the quantization code obtained by the quantizer at each stage has almost the same vector as the input vector.In other words, the information obtained by performing embedded coding is It becomes highly reliable.

Also, in this method of performing quasi-orthogonalization, as in the method of performing orthogonalization described above, each code vector obtained by linear prediction is stored in an intermediate codebook, and each stored code vector is used to perform standardization. Since orthogonalization and quantization are performed, the amount of information can be reduced, thereby increasing information efficiency.

Embodiments Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 2 is a diagram for explaining the multistage vector quantization method according to the first embodiment of the present invention, and in this figure, parts corresponding to the parts of the conventional example shown in FIG. 6 are given the same reference numerals. It is.

In this figure, 1A is a first-stage quantizer, 2A is a second-stage quantizer, and 3A is a third-stage quantizer.

The first stage quantizer 1A is different from the conventional quantizer 1 shown in FIG. a residual codebook 15 in which a plurality of white noises are stored;
This configuration includes a predictive reproduction filter 16 that reproduces the code of the residual codebook 15 and outputs it to the codebook 13. That is, according to such a configuration, the codes of the residual codebook 15 are transferred to the codebook 13 via the predictive reproduction filter 16, and the codes are transferred to the codebook 13 through the predictive reproduction filter 16.

Nocode vector al l (1"l+2i "'
+LL*Ll≧1). Further, by the calculation control of the calculation unit 12, the norm of the error signal e outputted from the error calculation unit 11 is minimized.
Vector a from code vector a in codebook 13.

is selected, and the amplifier 14 further outputs the error signal e1
The vector a1 output from the codebook 13 is amplified by a gain g1 that minimizes the norm of Ru.

In addition to the components of the first-stage quantizer 1A, the second-stage quantizer 2A has a codebook 27 and an arithmetic unit 28 connected between the predictive reproduction filter 26 and the codebook 23. This is what I composed.

Codebook 27 is similar to codebook 13 in the first stage,
A code coded with white noise output from the residual codebook 25 is taken in via the predictive reproduction filter 26, and a code vector a=i of codebook length L1 is obtained.
(+=1.2.-random2iL1≧1). Furthermore, the arithmetic unit 28 converts each vector a21 (see FIG. 3) output from the codebook 27 into the vector g1a1i output from the first stage amplifier 14.
The vector b21 obtained by the orthogonalization is outputted to the codebook 23 and the codebook 37 of the third stage quantizer 3A.

In addition, the vector b21 output from the arithmetic unit 28 is
In the codebook 23, the code vector b2t (+=1.2..., L, , L
, ≧1). Under the arithmetic control of the arithmetic unit 22, the vector b2 is selected from the code vector b21 and output to the amplifier 24. Also,
The amplifier 24 outputs the vector Ihb21 amplified with the desired gain g2 to the error calculation unit 21 and the arithmetic unit 38 of the third stage quantizer 3A. According to such an operation, the final vector obtained by the second stage quantizer 2A is the vector g1a1, as shown in FIG.
A vector X2 is obtained by combining i and the vector Ihb2i, which is the same vector as the input vector X input to the first stage.

The third-stage quantizer 3A includes a codebook 37 that temporarily stores the vector b2I output from the second-stage arithmetic device 28, and an arithmetic device 38 having the same function as the second-stage arithmetic device 28. The operation is similar to that of the second stage quantizer 2A. That is, each vector b output from the codebook 37 by the arithmetic unit 38
21 is orthogonalized at 90 degrees to the vector gzb2i output from the second stage amplifier 24, and the vector b3i obtained by the orthogonalization is output to the codebook 33 and stored therein. Then, under the arithmetic control of the arithmetic unit 32, the vector b3° is selected from the codebook 33 and output to the amplifier 34, where it is amplified by a desired gain g, and becomes the vector g3bz +
is output to the error calculation section 31. According to such an operation, the final vector obtained by the third stage quantizer 3A is the vector gzaa + and the vector gx
bx + and is a composite vector, which is 1
This is the same vector as the input vector X input to the first stage.

Therefore, according to the multistage vector quantization method in which such quantizers 1A, 2A, and 3A are connected, the data obtained by being quantized and encoded at each stage is data that can withstand reproduction, and the input data The data will have almost no error.

Next, a multi-stage vector quantization method according to a second embodiment of the present invention will be described. The structural elements of this second embodiment differ from those of the first embodiment only in the functions of the arithmetic units 28 and 38, so the diagram will be explained with reference to FIG.

The difference between the arithmetic units 28 and 38 used in the second embodiment from that of the first embodiment is that the arithmetic units 28 and 38 amplify the vector output from the amplifier of the quantizer in the previous stage. It has a function of detecting the amplification factor and performing quasi-orthogonalization that combines two vectors according to the detected amplification factor.

This function will be explained with reference to FIGS. 2 and 4. For example, the vector g multiplied by 783 by the first stage amplifier 14
1a1i actually contains an error of 6g1, but if this error is large, the vector gnat +
Even if the vector X2 is orthogonalized as in the first embodiment and the optimal vector X2 is obtained in the second stage, an error will occur between this vector X2 and the input vector X. Further, since the vector obtained by the orthogonalization is only on the plane indicated by the broken line S in the figure, the error δg1 cannot be corrected twice.

Therefore, in the arithmetic unit 28, first, the error δg included in the first vector gnat 1 of the vectors gnat + sequentially output from the amplifier 14 is calculated.
.

and the error δg1 included in the second vector g+a+2
Detect the midpoint of the second vector g+a+2 and the error δg1 of the third vector g+a13.
Find the midpoint between `` and the larger one of these midpoints, and from this, find the distribution coefficient h l/g +, which is the ratio of and g+. Then, quasi-orthogonalization is performed to combine each vector a21 output from the codebook 27 with the distribution coefficient h + /g + to obtain each vector b21. The same applies to the arithmetic unit 38.

Vector gna+ + based on such quasi-orthogonalization
The distribution of the composite vector of and vector g2az + is like the ellipse 40 shown in FIG. In addition, a straight line 41 (diameter of a circle 42) shown in the same figure shows the distribution of the composite vector of the vector g+a+ + and the vector g2az+ based on orthogonalization, and the circle 42 shows the distribution of the composite vector of the vector g+a+ + and the vector g2az+ based on the orthogonalization.
Vector gzaa based on code vector a21 of
The distribution of the composite vector of t and vector g+a+t is shown.

As can be seen from the figure, for example, when the vector g+a+ t output from the amplifier 14 contains an error -6g1, if the vector a2i is orthogonalized to this vector La+ +, the vector on the straight line 41 b21, and the vector gz obtained by amplifying the vector b21 selected from this vector b2i
The composite vector X2 of az and the vector g1a1i produces an error of -D with respect to the input vector X. However, if the vector gnat□ is made quasi-orthogonal to the vectors a and i, it becomes a vector b21 on the ellipse 40, and a vector b2 selected from this vector b21
Vector gzaz and vector L obtained by amplifying 1
The composite vector X2' with a+ + becomes almost the same vector as the input vector X.

Also, even if the vector La+ + contains an error +δg1, a composite vector X2 can be obtained by orthogonalization, but this vector X2 causes an error of +D with respect to the input vector X' in this case. .

However, according to the quasi-orthogonalization, the input vector
It is possible to obtain a composite vector X2' that is almost the same as '.

As mentioned above, according to the multi-stage vector quantization method as described above, it is possible to obtain a composite vector that is almost the same as the input vector at each stage, so information from the second and subsequent stages is arbitrarily discarded, and the
Embedded encoding can be performed in which reproduction is performed using the information of the rows. Therefore, if such a multi-stage vector quantization method is used in a communication device such as a LAN, flexible services can be provided.

As described in detail, the present invention has the following effects.

(2) That is, according to the present invention, each code vector outputted from the intermediate codebook between each quantizer and the quantizer is processed by the orthogonalization calculation means from the codebook of the preceding quantizer by the amplification means. The code vectors outputted through the quantizer are combined and orthogonalized, and each orthogonalized code vector is stored in a codebook of a subsequent quantizer. Then, from the codebook containing the orthogonalized code vectors, a code vector is selected by the calculation means at the same stage so that the norm of the error signal is minimized,
The signal is then amplified by an amplifying means at the same stage to obtain a quantized code.

Therefore, since it is possible to obtain a quantization code for each quantizer in each stage, it is possible to reproduce information for each quantizer in each stage. This has the advantage of being able to perform embedded encoding for selective reproduction.

Moreover, the quantization code vector obtained for each stage of quantizer is the one that has undergone the above-mentioned orthogonalization operation during the signal processing process, so it is almost the same vector as the input vector, and an optimal vector without loss is obtained. There is an effect that can be done.

■Also, in addition to the orthogonalization mentioned above, each code vector obtained by linear prediction is stored in an intermediate codebook,
Since each stored code vector is used for orthogonalization and quantization, the amount of information is small.
This has the effect of increasing informatization efficiency.

■According to another aspect of the present invention, a coefficient corresponding to the amplification factor of the code vector outputted via the amplifier is obtained from the codebook of the previous stage quantizer, and this coefficient and each of the coefficients outputted from the intermediate codebook are calculated. Since quasi-orthogonalization is performed to synthesize the code vector, even if the code vector output from the amplifier contains a quantization error, that error can be eliminated. This has the effect that the vector of the quantization code to be input can be made almost the same as the input vector. This has the effect that the information obtained by performing embedded encoding becomes highly reliable.

■In the quasi-orthogonalization of (■) above, similarly to (■), each code vector obtained by linear prediction is stored in an intermediate codebook, and each stored code vector is used to perform quasi-orthogonalization and quantization. Since this is done, the amount of information can be reduced, which has the effect of increasing informatization efficiency.

[Brief explanation of drawings]

FIG. 1 is a diagram of the principle of the present invention; FIG. 2 is a diagram for explaining a multi-stage vector quantization method according to an embodiment of the present invention; FIG. 3 is a vector diagram for explaining the first embodiment; Figure 4 is a vector diagram for explaining the second embodiment, Figure 5 is a vector diagram for explaining the second embodiment, and Figure 6 is a diagram for explaining the conventional multi-stage vector quantization method. , FIG. 7 is a vector diagram for explaining a conventional example. 1A, 2A... Quantizer, 11.21... Error calculating means, 12.22... Arithmetic means, 13.23... Codebook, 14.24... Amplifying means, 27... - Intermediate code book, 28... orthogonalization calculation means, quasi-orthogonalization calculation means, 12.
22...Arithmetic means, e0=X...Input signal (input vector), eIn e
2...Error signal, al+ La++ b2. g2b2* a2
1+ b2t・-code spectrum, L,! h...Amplification rate.

Claims (1)

[Claims] 1. A codebook containing a large number of code vectors (
13) and the code vector (a_1_i) output from this codebook (13) to a desired amplification factor (g_1).
an amplifying means (14) for amplifying the amplifying means (14);
The error signal (e_
1), and an error calculation means (11) for obtaining the error signal (e
The codebook (
A plurality of quantizers (1A) equipped with arithmetic means (12) for selecting the code vector (13) are connected, and input signals (e_0) such as audio and images, that is, the input vector (
In the multi-stage vector quantization method for quantizing and encoding X), an intermediate codebook (27) storing a large number of code vectors is located between each quantizer (1A) and the quantizer (2A). And, each code vector (a_2_i) output from this intermediate code book (27) is converted into a code vector (a_2_i) output from the code book (13) of the preceding stage quantizer (1A) via the amplification means (14). g_1a_1_i), and each orthogonalized code vector (b_2_i)
and an orthogonalization calculation means (28) for storing the orthogonalized code vectors (b_2_i) in the codebook (23) of the subsequent quantizer (2A). , the same stage calculation means (
22), the code vector (b_2_i) is selected such that the norm of the error signal (e_2) is minimized, and the code vector (b_2_i) is amplified by the amplification means (24) of the same stage, thereby obtaining a quantization code. Vector quantization method. 2. A codebook containing a large number of codevectors (
13) and the code vector (a_1_i) output from this codebook (13) to a desired amplification factor (g_1).
an amplifying means (14) for amplifying the amplifying means (14);
The error signal (e_
1), and an error calculation means (11) for obtaining the error signal (e
The codebook (
A plurality of quantizers (1A) equipped with arithmetic means (12) for selecting the code vector (13) are connected, and input signals (e_0) such as audio and images, that is, the input vector (
In the multi-stage vector quantization method for quantizing and encoding X), an intermediate codebook (27) storing a large number of code vectors is located between each quantizer (1A) and the quantizer (2A). From the codebook of the previous stage quantizer (1A), the amplification means (
14) The amplification factor (g
_1), and combine this coefficient with each code vector (a) output from the intermediate codebook (27).
_2_i), and stores each of the semi-orthogonalized code vectors (b_2_i) in the codebook (23) of the subsequent quantizer (2A). From the codebook (23) in which each of the quasi-orthogonalized code vectors (b_2_i) is stored, the code is calculated by the calculation means (22) at the same stage so that the norm of the error signal (e_2) is minimized. A multi-stage vector quantization method characterized in that a quantization code is obtained by selecting a vector (b_2_i) and amplifying it by an amplifying means (24) in the same stage. 3. Provide another codebook in which a desired code vector is stored, and linear prediction means for linearly predicting and synthesizing each code vector of the other codebook, and using the code vector obtained by the linear prediction means as described above. The multi-stage vector quantization method according to claim 1 or 2, wherein the multi-stage vector quantization method is stored in an intermediate codebook (27).