CN101198041A - Vector quantization method and device - Google Patents

Abstract

The invention discloses a vector quantitative method which comprises the following steps: input vectors are quantized by utilization of N codebooks and adjustment quantity of each codebook, and codebooks and adjustment quantity of quantized input vectors are obtained; N is a positive integer which is larger than or equal to 1. The invention also discloses a vector quantitative device on the basis of the method. The technical proposal provided by the invention is that: by utilization of participation of adjustment quantity of the codebooks in quantization of the input vectors, memory capacity of the codebooks is effectively reduced and operation quantity is only calculation amount needed for traversing N codebooks, thereby complexity of vector quantization can be effectively reduced.

Description

Vector quantization method and device

Technical field

The present invention relates to the signal processing technology field, relate in particular to a kind of vector quantization method and device.

Background technology

In that sound signal or vision signal are carried out in the digitalization compressed encoding process process,,, can reduce code rate and guarantee coding quality carrying out vector quantization by the synthetic vector of a plurality of input signal sample values according to predetermined code book.

Vector quantization is a kind of data compression technique efficiently, and this technology is divided into one group with every K continuous sample value of input signal sequence earlier, forms a vector in the K dimension theorem in Euclid space by this group sample value, afterwards, again each such vector is carried out quantification treatment.The complexity of vector quantization is mainly from two aspects: amount of calculation and memory space, and the complexity of vector quantization can increase along with the increase exponentially of vector dimension; And as a kind of data compression technique that diminishes, also there is the code rate problem in vector quantization except exist quantizing problem of dtmf distortion DTMF, and vector dimension and code rate are inversely proportional to.Under the network condition rugged environment, often avoid packet loss phenomenon in the packet transmission course by reducing code rate, guaranteeing certain coding quality, but correspondingly, also can increase the complexity of vector quantization.

Be example with existing encoding compression below, the vector quantization technology that is adopted in the existing voice cataloged procedure is described voice signal.

Existing the vector of leading spectral frequency (ISF) coefficient composition is carried out in the technology of vector quantization, adopt wideband adaptive audio coder ﹠ decoder (codec) (AMR-WB+) that the ISF vector is carried out quantification treatment usually.Wherein, the ISF coefficient is converted to by the linear predictor coefficient (LPC) of voice signal.AMR-WB+ adopts the mode of uniting division vector quantization (SVQ) and multi-stage vector quantization (MSVQ) to quantize the ISF vector, this vector quantization process mainly is: adopt single order moving average (MA) predicted method earlier, obtain the ISF prediction residual vector r (n) of current n frame, utilize multistage code book that r (n) is carried out quantification treatment then.Table 1 is the Bit Allocation in Discrete and the code book use table of existing ISF vector quantization process.

Table 1

Referring to table 1, AMR-WB+ to the process of ISF vector quantization is:

If the ISF vector is 16 n dimensional vector ns, r (n) also is one 16 n dimensional vector n so.Make r=[r ₁, r ₂..., r ₁₆], for reducing the memory space of code book in the vector quantization, AMR-WB+ divides multi-stage vector quantization (S-MSVQ) to the r vector:

Earlier 16 dimension r vectors are carried out first order division, obtain 2 sub-vectors, i.e. the r of 9 dimensions ₁ ⁽¹⁾The r of vector and 7 dimensions ₂ ⁽¹⁾Vector is used 8 bit quantization r respectively ₁ ⁽¹⁾And r ₂ ⁽¹⁾, corresponding code book number is respectively 256,256.

Afterwards, to r ₁ ⁽¹⁾And r ₂ ⁽¹⁾Carry out second level division respectively, obtain 5 sub-vectors altogether, wherein,

9 dimension r ₁ ⁽¹⁾Split into 3 sub-vectors, that is, and the r of 3 dimensions ₁₁ ⁽²⁾, use 6 bit quantizations, corresponding code book number is 64; The r of 3 dimensions ₁₂ ⁽²⁾, use 7 bit quantizations, corresponding code book number is 128; The r of 3 dimensions ₁₃ ⁽²⁾, use 7 bit quantizations, corresponding code book number is 128;

7 dimension r ₂ ⁽¹⁾Split into 2 sub-vectors, that is, and the r of 3 dimensions ₂₁ ⁽²⁾, use 5 bit quantizations, corresponding code book number is 32; The r of 3 dimensions ₂₂ ⁽²⁾, use 5 bit quantizations, corresponding code book number is 32.

Above-mentioned fission process obtains 7 sub-vectors altogether, is divided into to have joined 46 bits and be used for vector quantization.

2 sub-vectors that above-mentioned first order division is obtained carry out the one-level quantification, and 5 sub-vectors that second level division obtains carry out secondary and quantize.One-level quantizing process and secondary quantizing process are similar, therefore, can make r _xMake a general reference in 7 sub-vectors, the x of certain sub-vector ties up component, and x represents the dimension numbering of sub-vector; Be in the code book corresponding with this sub-vector, the x of y code word dimension component.Calculate r _xWith Between error ${\mathrm{\Δ}}_x=r_x-{\hat{r}}_x^y,$ And calculate Δ _xQuadratic sum:

$E=\underset{x=m}{\overset{n}{\mathrm{\Σ}}}{\left[{\mathrm{\Δ}}_x\right]}^2,$

With the distortion factor of E as vector quantization, from the code book corresponding, select the corresponding y of code word that makes the E minimum with this sub-vector, and with the call number output of y as this code word.

The vector quantization technology of existing AMR-WB+ adopts multistage code book quantization vector, can obtain higher coding quality, but also therefore increases the complexity of vector quantization.The complexity of vector quantization comes from memory space and these two aspects of amount of calculation, and code book memory space of the prior art is bigger, has also therefore increased amount of calculation, need travel through all code books when calculating.Therefore, prior art is still waiting further improvement on reduction vector quantization complexity.

Summary of the invention

In order to solve the problem of vector quantization complexity in the prior art, the invention provides a kind of better simply vector quantization method.

The present invention also provides a kind of generation code book adjustment amount method.

The present invention provides a kind of better simply vector quantization apparatus again.

A kind of method of vector quantization comprises:

Utilize the adjustment amount of N code book and each code book that input vector is carried out quantification treatment, draw the code book and the adjustment amount that quantize input vector, N is the positive integer more than or equal to 1.

A kind of method that produces the adjustment amount of code book comprises:

Calculate the mean vector of N code book, N is the positive integer more than or equal to 1;

Calculate the distance of each code book to this mean vector;

According to the distance of each code book, produce the adjustment amount of each code book to this mean vector.

A kind of vector quantization apparatus comprises: code book provides unit, adjustment amount that unit and quantifying unit are provided, wherein,

Code book provides the unit, is used for sending the N corresponding with an input vector code book to quantifying unit, and N is the positive integer more than or equal to 1;

Adjustment amount provides the unit, is used for sending the adjustment amount of each code book to quantifying unit;

Quantifying unit, be used to receive from adjustment amount provide the unit each code book adjustment amount and provide the unit from code book, utilize the adjustment amount of N code book and each code book that input vector is carried out quantification treatment, draw the code book and the adjustment amount that quantize input vector.

Above-mentioned vector quantization method carries out quantification treatment by the adjustment amount that utilizes N code book and each code book to input vector, be about in the quantification of adjustment amount participation to input vector of code book provided by the invention, thereby effectively reduce the code book memory space, and operand just travels through N the needed amount of calculation of code book.Therefore, utilize above-mentioned vector quantization method, can effectively reduce the complexity of vector quantization.

In the above-mentioned vector quantization apparatus, the quantifying unit utilization provides unit and adjustment amount to provide N the code book that the unit sends and the adjustment amount of each code book that input vector is done quantification treatment by code book, used code book is less, and operand just travels through N the needed amount of calculation of code book.Therefore, use above-mentioned vector quantization apparatus, can effectively reduce the complexity of vector quantization.

Description of drawings

Fig. 1 is the overview flow chart of the vector quantization method that provides of the embodiment of the invention;

Fig. 2 is the method flow diagram of the adjustment amount of the generation code book that provides of the embodiment of the invention;

Fig. 3 is the flow chart that produces ISF residual error code book and adjustment amount thereof in the embodiment of the invention;

Fig. 4 is the flow chart that adopts the vector quantization method quantification ISF residual error of the embodiment of the invention;

Fig. 5 is the structural representation of a preferred embodiment of vector quantization apparatus of the present invention;

Fig. 6 is the structural representation of another preferred embodiment of vector quantization apparatus of the present invention;

Fig. 7 is the structural representation of another vector quantization apparatus preferred embodiment of the present invention;

Fig. 8 is the structural representation of quantification treatment unit preferred embodiment among Fig. 7;

Fig. 9 is the structural representation of another vector quantization apparatus preferred embodiment of the present invention.

Embodiment

For the purpose, technical scheme and the advantage that make the embodiment of the invention is clearer, describe in further detail below in conjunction with accompanying drawing.

For reducing the complexity of vector quantization, the present invention provides a kind of vector quantization method by embodiment.

Fig. 1 is the overview flow chart of the vector quantization method that provides of the embodiment of the invention, and this flow process may further comprise the steps:

Step 101, utilize the adjustment amount of N code book and each code book that input vector is carried out quantification treatment.

Step 102, draw the code book and the adjustment amount that quantize input vector.

Wherein, N is the positive integer more than or equal to 1.The adjustment amount of code book is used to adjust code book so that to the quantized result of input vector near with this input vector.

The vector quantization method that provides around the embodiment of the invention shown in Figure 1 is done and is launched explanation below.

In concrete enforcement, before the adjustment amount that utilizes N code book and each code book carries out quantification treatment to input vector, also need to produce the adjustment amount of N code book.

Fig. 2 is a kind of method flow diagram that produces the adjustment amount of code book that the embodiment of the invention provides, and wherein, the dimension of establishing input vector is a dimension, and this flow process may further comprise the steps:

The mean vector of step 201, N code book of calculating.

In this step, the mean vector of code book also is a n dimensional vector n.In the present embodiment, the way of calculating the mean vector of N code book can be: after the same dimension component addition with all N code book, divided by N, just obtain this dimension component of mean vector.Making the i dimension component of j code book is Y _j(i), making the i dimension component of mean vector is AV (i), and then the calculating formula of AV (i) is:

$\mathrm{AV}\left(i\right)=\frac{\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{Y}_j\left(i\right)}{N},i=\mathrm{1,2},\·\·\·,a.$

Mean vector AV (a) be exactly a n dimensional vector n of forming by these AV (i) component: AV (a)=AV (i), i=1,2 ..., a}.

Step 202, calculate each code book to the distance of this mean vector.

In the present embodiment, this computational process is specifically: calculate the difference between the same dimension component of each the dimension component of j code book and mean vector, obtain a component difference, obtain the quadratic sum of this a component difference again, with the quadratic sum that obtains divided by a, again to the result of division evolution, just obtain j code book and be to the distance A VS (j) of this mean vector:

$\mathrm{AVS}\left(j\right)=\sqrt{\frac{\underset{i=1}{\overset{a}{\mathrm{\Σ}}}{[Y_j\left(i\right)-\mathrm{AV}\left(i\right)]}^2}{a}},j=\mathrm{1,2},\·\·\·,N,$

Step 203, according to each code book to the distance of this mean vector, produce the adjustment amount of each code book.

In the present embodiment, can produce the adjustment amount of this code book by producing the adjustment amount of each dimension component of each code book, specific practice is: this code book is set respectively ties up 1 of the adjustment amount correspondence of component or more than 1 scale factor; When the adjustment amount of each the dimension component that produces this code book, this code book is multiplied each other to each corresponding with the adjustment amount of this dimension component respectively scale factor of the distance of described mean vector, with each product of obtaining each adjustment amount as this dimension component correspondence of this code book; Sort to obtaining each product corresponding with this dimension component of this code book, the sequence number that each product is obtained is as the call number of the adjustment amount of this product correspondence.

With j code book is example, launches the process that i that explanation produces this j code book ties up the adjustment amount of component to be: with the base value of AVS (j) as j code book adjustment amount, scale factor is set, the base value of j code book adjustment amount is adjusted being used for further.Be made as the i dimension component Y of j code book _j(i) scale factor of She Zhiing be this set 0, A1 ,-A1, A2, the element the among-A2}, Y so _j(i) adjustment amount be the set 0, AVS (j) * A1 ,-AVS (j) * A1, AVS (j) * A2, the element among-AVS (j) * A2}; If A1=A2=0.3, then Y _j(i) adjustment amount be the set 0, AVS (j) * 0.3, the element among-AVS (j) * 0.3}.Simultaneously, the also call number of corresponding adjustment amount of the sequence number of each element in the set.

Embodiment illustrated in fig. 2ly provide a kind of method that produces the adjustment amount of code book, in the practical application, also can produce the adjustment amount of code book by other ways, as the simplest way is exactly the adjustment amount that produces each code book at random, but the adjustment amount that obtains with this way, residual error vector between code book and the input vector is carried out quantification treatment, the coding effect that is obtained, the coding effect stability that does not obviously adopt the middle adjustment amount that produces embodiment illustrated in fig. 2 to bring.

Carrying out vector quantization with the ISF vector that the ISF coefficient is formed below is example, describes the vector quantization method shown in Figure 1 that the embodiment of the invention provides in detail.And, coding effect and quantization complexity for ease of the vector quantization method of the vector quantization method of the embodiment of the invention relatively and AMR-WB+, in the embodiment of the invention, also adopt 46 bit numbers to encode, and establishing the ISF vector is 16 n dimensional vector ns, and the code book that is trained by this 16 ISF residual error of tieing up ISF vector correspondence is 256.

The method of the adjustment amount of the generation code book that provides based on Fig. 2, Fig. 3 is the flow chart that produces the adjustment amount of ISF residual error code book and each code book in the embodiment of the invention, this flow process may further comprise the steps:

Step 301, employing inter prediction calculate the ISF residual error of ISF vector.

This step adopts and has now in the quantification ISF vector, calculates the identical way of ISF residual error, calculates the ISF residual error, and this computational process of brief description is as follows:

Definition Z _nFor removing the n frame ISF vector isf behind the average mean_isf _n, then Yu Ce ISF residual error is R _n, its expression formula is:

Z _n＝isf _n-mean_isf

R _n＝Z _n-P _n

Wherein, P _nBe the prediction ISF vector of n frame:

$P_n=\mathrm{MA}\×R_{n-1}^q$

R wherein _N-1 ^qBe the ISF residual error after the former frame quantification, MA is an empirical value, generally gets 1/3.

Step 302, train 256 code books of 16 dimension ISF residual errors.

In this step, as known quantity, adopt the LBG algorithm to train a N corresponding code book input vector with input vector.Wherein, the LBG algorithm is a kind of algorithm that is used to train code book as known technology, and the relevant detailed process of this algorithm training code book that adopts can be with reference to data of literatures.In the practical application, also can be by other algorithm training code books.

Each code book that trains also is a n dimensional vector n, and a is the positive integer more than or equal to 1.

The mean vector of step 303,256 code books of calculating.

To the description of this step, can be referring to the description in the above-mentioned steps 202, wherein, N=256, a=16.

In step 304,256 code books of calculating, each code book is to the distance of this mean vector.

To the description of this step, can be referring to the description in the above-mentioned steps 203, obtain j code book and be to the distance A VS (j) of this mean vector:

$\mathrm{AVS}\left(j\right)=\sqrt{\frac{\underset{i=1}{\overset{16}{\mathrm{\Σ}}}{[Y_j\left(i\right)-\mathrm{AV}\left(i\right)]}^2}{16}},j=\mathrm{1,2},\·\·\·,256$

In the present embodiment, the corresponding distance value AVS (j) of code book, this distance value will in the practical application, also can calculate a plurality of base values of a code book as the base value of the adjustment amount of code book, still, will increase the amount of calculation that subsequent quantizatiion is handled.

In step 305,256 code books of generation, the adjustment amount of each dimension component of each code book.

In the present embodiment, for the 1st to 6 dimension component of each code book is provided with 7 adjustment amounts, corresponding scale factor is 0, A1 ,-A1, B1 ,-B1, C1 ,-C1; Be that the 7th to 13 dimension component is provided with 5 adjustment amounts, corresponding scale factor is 0, A2 ,-A2, B2 ,-B2; Be that the 14th to 16 dimension is provided with 3 adjustment amounts, corresponding scale factor is 0, A3 ,-A3.So, be example with j code book, with the AVS (j) that obtains in the step 305 adjustment amount base value as every dimension component of j code book, resulting j code book respectively tieed up the set that the adjustment amount V (k) of component forms and is:

For the 1st to 6 dimension component, k=1,2 ..., 7, i=1,2 ..., 6,

V(i，k)＝{0，AVS(j)×A ₁，-AVS(j)×A ₁，AVS(j)×B ₁，-AVS(j)×B ₁，AVS(j)×C ₁，-AVS(j)×C ₁}；

Wherein, k represents V (i, k) in the set, the sequence number of each element also is the call number of the adjustment amount of the 1st to 6 dimension component correspondence.

For the 7th to 13 dimension component, k '=1,2 ..., 5, i=7 ..., 13,

V(i，k′)＝{0，AVS(j)×A ₂，-AVS(j)×A ₂，AVS(j)×B ₂，-AVS(j)×B ₂}；

Wherein, k ' represents in V (i, the k ') set, and the sequence number of each element also is the call number of each adjustment amount of the 7th to 13 dimension component correspondence.

For the 14th to 16 dimension component, k "=1,2,3, i=14 ..., 16,

V(i，k″)＝{0，AVS(j)×A ₃，-AVS(j)×A ₃}；

Wherein, " (in i, the k ") set, the sequence number of each element also is the call number of each adjustment amount of the 14th to 16 dimension component correspondence to k to represent V.

In addition, in the present embodiment, if all proportions factor is 1, the adjustment amount of the every dimension component of the code book that obtains so is: V (i, k)={ 0, AVS (j) ,-AVS (j) }, k=1,2,3, i=1 ..., 16.

So far, just obtain 256 code books and adjustment amount thereof, afterwards,, quantize the ISF residual error utilizing these code books and corresponding adjustment amount.

The generation ISF residual error code book that provides based on Fig. 3 and the adjustment amount of each code book, Fig. 4 is the flow chart that adopts the vector quantization method quantification ISF residual error of the embodiment of the invention, wherein, utilize the adjustment amount of N code book and each code book that input vector is carried out quantification treatment, draw the code book of quantification input vector and the flow process of adjustment amount and may further comprise the steps:

Step 401, calculate in input vector and 256 code books difference vector between the adjustment amount of each code book and this code book.

In the present embodiment, input vector is the ISF residual error, calculates in input vector and 256 code books, and the difference vector between the adjustment amount of each code book and this code book is, in calculating ISF residual error and 256 code books, and the residual error vector between each code book.

The residual error vector that obtains in this step also is 16 n dimensional vector ns, and such residual error vector has 256, with the residual error vector representation is: res _{J, i}, j=1,2 ..., 256, i=1,2 ..., 16.

Step 402, according to the pattern of the adjustment amount that code book is set, 16 dimension residual error vectors are split into 3 sub-vectors.

In the present embodiment, in the description according to step 305, the adjustment amount of the 1st to 6 dimension component correspondence of code book, the adjustment amount of the 7th to 13 dimension component correspondence, and the adjustment amount of the 14th to 16 dimension component correspondence, 16 of j code book correspondence tieed up the residual error vectors is split into 3 sub-vectors and is:

res _j，i，i＝1，2，…，6；res _j，i，i＝7，…，13；res _j，i，i＝14，…，16。

Step 403, utilize the adjustment amount of each code book, the residual error vector is carried out quantification treatment.

In the present embodiment, utilize the adjustment amount of N code book and each code book that input vector is carried out quantification treatment and comprise: utilize the adjustment amount of each code book that the residual error vector between input vector and this code book is carried out quantification treatment.Quantizing process to the residual error vector is: the absolute difference between each adjustment amount of each the dimension component that calculates the residual error vector and the same dimension component of corresponding code book; Select call number with each the corresponding minimum absolute difference value of dimension component of residual error vector and corresponding adjustment amount.To quantize res _{J, i}, i=1,2 ..., 6 is example, this quantizing process is specially:

Calculate res _{J, i}The 1st dimension component and V (i, k) middle i=1, k=1,2 ..., the absolute difference in 7 the set between each element:

E _1，k＝|res _j，1-V(1，k)|，

V (1, k)=0, AVS (j) * A ₁,-AVS (j) * A ₁, AVS (j) * B ₁,-AVS (j) * B ₁, AVS (j) * C ₁,-AVS (j) * C ₁, obtain 7 absolute difference E _{1, k}, select minimum absolute difference value minE wherein _{1, k}, and the corresponding k value of record.

According to above-mentioned quantizing process, obtain res _{J, i}In, the minimum absolute difference value minE of the 1st to 6 dimension _{I, k}, i=1,2 ..., 6; The minimum absolute difference value minE of the 7th to 13 dimension _{I, k '}, i=7 ..., 13; The minimum absolute difference value minE of the 14th to 16 dimension _{I, k "}, i=14 ..., 16.

Step 404, calculate the distortion factor between each residual error vector adjustment amount corresponding, obtain N the distortion factor with the minimum absolute difference value.

In the present embodiment, the process of calculating the distortion factor between each residual error vector adjustment amount corresponding with the minimum absolute difference value is: to each the residual error vector in 256 residual error vectors, obtain the square value of the minimum absolute difference value of respectively tieing up the component correspondence of this residual error vector earlier, again each square value of tieing up the minimum absolute difference value of component correspondence is sued for peace, with the summed result that obtains as the residual error vector corresponding and the distortion factor between the adjustment amount with same code book, the residual error vector is corresponding one by one with the distortion factor, and the distortion factor that obtains has N.To calculate res _{J, i}, i=1,2 ..., 16 and adjustment amount between distortion factor E _jBe example, this calculating formula is:

$E_j=\underset{i=1}{\overset{6}{\mathrm{\Σ}}}{\left(\min E_{i,k}\right)}^2+\underset{i=7}{\overset{13}{\mathrm{\Σ}}}{\left(\min E_{{i,k}^{\′}}\right)}^2+\underset{i=14}{\overset{16}{\mathrm{\Σ}}}{\left(\min E_{{i,k}^{\′\′}}\right)}^2,j=\mathrm{1,2},\·\·\·,256$

The distortion factor that obtains has 256.

If establish $E1=\underset{i=1}{\overset{6}{\mathrm{\Σ}}}{\left(\min E_{i,k}\right)}^2,$ $E2=\underset{i=7}{\overset{13}{\mathrm{\Σ}}}{\left(\min E_{{i,k}^{\′}}\right)}^2,$ $E3=\underset{i=14}{\overset{16}{\mathrm{\Σ}}}{\left(\min E_{{i,k}^{\′\′}}\right)}^2,$ So, also can do weighted to E1, E2 and E3, this square value be multiply by corresponding weight coefficient, obtain the distortion factor after the weighting: E according to the difference of ISF residual error to the error sensitivity _j=E1+E2/ α+E3/ β, α and β are weight coefficients, are real number.

Step 405, select the minimum distortion degree from 256 distortion factors, code book that will be corresponding with this minimum distortion degree and adjustment amount are as the code book and the adjustment amount of the quantification input vector that draws.

In the present embodiment, establish E _j=E1+E2/ α+E3/ β is the minimum distortion degree, and j code book is exactly the code book that quantizes input vector so, is best code book, establishes the 1st minE among the E1 _{1, k}Corresponding k is 2, and the quantification adjustment amount of residual error vector the 1st dimension is exactly AVS (j) * A so ₁, be the best adjustment amount of the 1st dimension of this residual error vector, and the like, the best adjustment amount that residual error vector the 1st to 6 is tieed up obtained according to E1; According to E2, obtain the best adjustment amount of residual error vector the 7th to 13 dimension component; According to E3, obtain the best adjustment amount of residual error vector the 14th to 16 dimension, obtain E2 and E3 best adjustment amount and corresponding k value according to the 1st to the 6 dimension component correspondence that obtains residual error vector wherein, quantification adjustment amount and corresponding k ' value, the k of the 14th to 16 quantification adjustment amount of tieing up and correspondence ".

Step 406, the call number of adjustment amount and codebook index number output that will be corresponding with this minimum distortion degree.

In the present embodiment, obtaining E _jAfter, so correspondingly, with this E _jJ corresponding code book is best code book, and its call number is j; Mention in the description as step 405 that the call number of the quantification adjustment amount of residual error vector the 1st dimension is exactly the 1st minE that obtains according to E1 _{1, k}Corresponding k=2, and the like, each ties up the call number of the best adjustment amount correspondence of component to obtain the residual error vector.

So far, the process that adopts vector quantization method of the present invention to quantize the ISF residual error finishes, and afterwards, the call number that finally obtains as coding result, is sent to decoding end.According to the encoding and decoding agreement, decoding end finds and input vector corresponding quantitative vector according to call number, draws synthetic speech signal.

Need to prove, when concrete execution in step 401, also can sue for peace respectively to each dimension component each adjustment amount corresponding of each code book earlier with this dimension component, calculate the absolute difference between the same dimension component of input vector and each and the value then, like this, just can directly obtain the absolute difference in step 403 description, can continue execution in step 403 operation afterwards then.

Owing to when image/video signals is encoded, also can sample to signal, obtain burst, and can form vector by these sequences, therefore, the above-mentioned vector quantization method that the embodiment of the invention provides also is applicable to the image/video signals cataloged procedure.

First kind of vector quantization method that the embodiment of the invention also provides based on the invention described above embodiment provides a kind of vector quantization apparatus.Fig. 5 is the structural representation of a preferred embodiment of this device, and this device comprises: code book provides unit 501, adjustment amount that unit 502 and quantifying unit 503 are provided, wherein,

Code book provides unit 501, is used for sending the N corresponding with an input vector code book to quantifying unit 503, and N is the positive integer more than or equal to 1;

Adjustment amount provides unit 502, is used for sending the adjustment amount of each code book to quantifying unit 503;

Quantifying unit 503, be used to receive the adjustment amount that each code book of unit 502 is provided from adjustment amount, with unit 501 is provided from code book, utilize the adjustment amount of N code book and each code book that input vector is carried out quantification treatment, draw the code book and the adjustment amount that quantize input vector.

Because the adjustment amount of N code book and each code book can produce before vector quantization, therefore, when carrying out vector quantization, can send quantifying unit 503 to by code book and the adjustment amount that code book provides unit 501 and adjustment amount to provide unit 502 that vector quantization need be used.

Fig. 6 is another preferred embodiment of vector quantization apparatus of the present invention, and this device further comprises: adjustment amount generation unit 601, be used to calculate the mean vector of N code book, and N is the positive integer more than or equal to 1; Calculate the distance of each code book to this mean vector; To the distance of this mean vector, produce the adjustment amount more than 1 or 1 of each code book according to each code book, send the adjustment amount of each code book of producing to adjustment amount unit 502 is provided.

Fig. 7 is the structural representation of another preferred embodiment of vector quantization apparatus of the present invention, this device can further comprise based on Fig. 6 device: residual error vector computing unit 701, be used for calculating an input vector and N code book, residual error vector between each code book sends the residual error vector between each code book to quantifying unit 503.

Fig. 8 is a preferred embodiment of Fig. 7 quantifying unit 503, and quantifying unit 503 comprises: quantize subelement 801, computation subunit 802 and chooser unit 803, wherein,

Quantize subelement 801, be used to receive the adjustment amount that each code book of unit 502 is provided from adjustment amount, with residual error vector from residual error vector computing unit, the absolute difference between each the dimension component that calculates each residual error vector and each adjustment amount of the same dimension component of corresponding code book; Select each the dimension component corresponding minimum absolute difference value with the residual error vector, the result of calculation of minimum absolute difference value that will be corresponding with each dimension component of residual error vector sends computation subunit 802 to;

Computation subunit 802, be used to receive result of calculation from quantizing the corresponding minimum absolute difference value of each dimension component subelement 801 and each residual error vector, calculate the square value of the minimum absolute difference value of respectively tieing up the component correspondence of this residual error vector, again each square value of tieing up the minimum absolute difference value of component correspondence is sued for peace, with the summed result that obtains as the distortion factor between this residual error vector and the adjustment amount, calculate N the distortion factor of N residual error vector, send N the distortion factor to chooser unit 803;

Chooser unit 803 is used to receive N the distortion factor from computation subunit 802, selects the minimum distortion degree, and code book and the adjustment amount corresponding with the minimum distortion degree, code book that will be corresponding with the minimum distortion degree and adjustment amount output.

Fig. 9 is the structural representation of another preferred embodiment of vector quantization apparatus, this device can further comprise based on Fig. 7: output unit 901, be used to receive code book corresponding and adjustment amount, codebook index that will be corresponding number and export with the call number of the corresponding adjustment amount of minimum distortion degree with the minimum distortion degree with the minimum distortion degree from chooser unit 803.

The performance of the vector quantization method shown in Figure 1 that provides for the further check embodiment of the invention, AMR-WB+ and the inventive method have been carried out the voice quality objective evaluating, evaluating software is P.862 objective evaluating standard broadband section's perceptual speech quality (WB-PESQ) of International Telecommunications Union's telecommunication standards tissue (ITU-T), also has the averaging spectrum distortion.Cycle tests is the cycle tests of the AMR-WB+ that provides of third generation partner program (3GPP) tissue and the cycle tests that video compression standard (AVS) provides, form by 71 cycle testss, comprising voice, music, voice music, music background voice etc.Test result is as follows:

But 1 perceptual speech quality (PESQ) average, AMR-WB+ is 3.801, the present invention is 3.874.2 code checks (10.4kbps and 24kbps) are in totally 142 test items in 71 cycle testss, and AMR-WB+ obtains winning (the PESQ average is the highest) winning and accounts for 47, and winning of this programme accounts for 95.

2, averaging spectrum distortion, the averaging spectrum distortion of AMR-WB+ is 0.960156, and wherein 2-4dB accounts for 3.84%, and＞4db accounts for 0.095%; And averaging spectrum distortion of the present invention is 0.878902, and wherein 2-4dB accounts for 0.963%, and＞4dB accounts for 0.00158%

Referring to table 1, the vector quantization scheme one-level of AMR-WB+ quantizes to use 256,256 code books, and memory space is 256 * (9+7)=4096 floating numbers; Secondary quantizes to use 64,128,128,32,32 code books, and memory space is 64 * 3+128 * 3+128 * 3+32 * 3+32 * 4=1184 floating number, thereby total storage capacity is 4096+1184=5280 floating number.Corresponding multiplying amount is 256 code books of traversal search that one-level quantizes, therefrom find out the one-level code book of 4 optimums, search for the secondary code book then, will travel through the one-level code book of 4 optimums and 64,128,128,32,32 code books that secondary quantizes in the process of search secondary code book.

Compare with the vector quantization scheme of AMR-WB+, when adopting vector quantization method of the present invention to quantize the ISF residual error, 256 code books are encoded with 8 bits, the adjustment amount of code book is encoded with 38 bits.Aspect memory space, in the embodiment of the invention, required storage be the memory space of 256 code books, be 256 * 16=4096 floating-point, add 256 adjustment amount base values, memory space is total up to 4096+256=4352 floating number.Amount of calculation is that multiplication calculates, for traveling through 256 needed amounts of calculation of code book.

As seen, the technical scheme that the embodiment of the invention provides all has remarkable advantages on coding quality and vector quantization complexity.

In sum, the technique scheme that the embodiment of the invention provides is by utilizing the code book corresponding with input vector, and the adjustment amount of these code books, input vector is carried out quantification treatment, adopt the way of combined calculation, when finding out best adjustment amount, find the best code book corresponding with this best adjustment amount, make the distortion factor minimum of whole quantizing process, thereby guarantee higher coding quality.Wherein, best adjustment amount is exactly the adjustment amount corresponding with the minimum distortion degree, and best code book is exactly the code book corresponding with this adjustment amount.Simultaneously, the technical scheme that the embodiment of the invention provides utilizes the adjustment amount of code book that the residual error vector between code book and the input vector is done quantification treatment, be different from the existing way that quantizes the residual error vector with multistage code book, therefore can reduce the code book memory space, and operand just travels through N the needed amount of calculation of code book, after N one-level code book of the existing traversal of comparing, need to travel through a plurality of secondary code books again, reduce amount of calculation.Therefore, the technical scheme that the embodiment of the invention provides when guaranteeing higher coding quality, reduces the complexity of vector quantization.

Claims (25)

1. a vector quantization method is characterized in that, comprising:

Utilize the adjustment amount of N code book and each code book that input vector is carried out quantification treatment, draw the code book and the adjustment amount that quantize input vector, N is the positive integer more than or equal to 1.

2. method according to claim 1 is characterized in that, before the adjustment amount that utilizes N code book and each code book carried out quantification treatment to input vector, this method further comprised: the adjustment amount that produces N code book is:

Calculate the mean vector of N code book;

Calculate the distance of each code book to this mean vector;

According to the distance of each code book, produce the adjustment amount of each code book to this mean vector.

3. method according to claim 2 is characterized in that, the adjustment amount of described each code book of generation is: the adjustment amount that produces each dimension component of each code book.

4. method according to claim 3 is characterized in that, the adjustment amount of each dimension component of each code book of described generation is:

Each code book is set respectively ties up 1 of the adjustment amount correspondence of component or more than 1 scale factor;

This code book is multiplied each other to each corresponding with the adjustment amount of this dimension component respectively scale factor of the distance of described mean vector, obtain each product corresponding, this each product should be tieed up each adjustment amount of component correspondence as this code book with this dimension component of this code book.

5. method according to claim 2 is characterized in that, this method further comprises: the call number that produces each adjustment amount.

6. method according to claim 5, it is characterized in that, the call number of described each adjustment amount of generation is: sort to obtaining each product corresponding with this dimension component of this code book, the sequence number that each product is obtained is as the call number of the adjustment amount of this product correspondence.

7. according to each described method in the claim 2 to 6, it is characterized in that the dimension of described mean vector is a dimension, a is the positive integer more than or equal to 1, and the mean vector of a described calculating N code book is: making the i dimension component of j code book is Y _j(i), making the i dimension component of described mean vector is AV (i), then:

$\mathrm{AV}\left(i\right)=\frac{\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{Y}_j\left(i\right)}{N},$ i＝1，2，…，a，

Described mean vector AV (a)=and AV (i), i=1,2 ..., a}.

8. method according to claim 7 is characterized in that, each code book of described calculating to the distance of this mean vector is:

$\mathrm{AVS}\left(j\right)=\sqrt{\frac{\underset{i=1}{\overset{a}{\mathrm{\Σ}}}{[Y_j\left(i\right)-\mathrm{AV}\left(i\right)]}^2}{a}},$ j＝1，2，…，N，

Wherein, AVS (j) is the distance of j code book to described mean vector.

9. method according to claim 1 is characterized in that, the adjustment amount of the described N of a utilization code book and each code book carries out before the quantification treatment input vector, and this method further comprises:

Calculate in input vector and N the code book difference vector between the adjustment amount of each code book and this code book.

10. method according to claim 9 is characterized in that, calculates in input vector and N the code book, and the difference vector between the adjustment amount of each code book and this code book comprises:

Calculate in input vector and N the code book residual error vector between each code book.

11. method according to claim 10, it is characterized in that the adjustment amount of the described N of a utilization code book and each code book carries out quantification treatment to input vector and comprises: utilize the adjustment amount of each code book that the residual error vector between input vector and this code book is carried out quantification treatment.

12. method according to claim 11, it is characterized in that, the adjustment amount of each dimension component correspondence of each code book has more than 1 or 1, the described adjustment amount that utilizes each code book carries out quantification treatment to input vector and residual error vector between this code book: the absolute difference between each the dimension component that calculates this residual error vector and each corresponding adjustment amount of the same dimension component of this code book, draw with this residual error vector respectively tie up each corresponding minimum absolute difference value of component, and the adjustment amount corresponding with each minimum absolute difference value.

13. method according to claim 12, it is characterized in that, described draw with this residual error vector respectively tie up each corresponding minimum absolute difference value of component after, this method further comprises: calculate the distortion factor between this residual error vector and the adjustment amount corresponding with the minimum absolute difference value, the described distortion factor of calculating between this residual error vector adjustment amount corresponding with the minimum absolute difference value is: the square value of respectively tieing up the corresponding minimum absolute difference value of component of obtaining this residual error vector earlier, again square value of the minimum absolute difference value of each dimension component correspondence is sued for peace, with the summed result that obtains as the distortion factor between described this residual error vector adjustment amount corresponding with the minimum absolute difference value.

14. method according to claim 13, it is characterized in that, behind the square value of the described minimum absolute difference value of respectively tieing up the component correspondence of obtaining this residual error vector, this method further comprises: this square value be multiply by corresponding weight coefficient, and described weight coefficient is a real number.

15. method according to claim 13, it is characterized in that, the residual error vector is corresponding one by one with the distortion factor, the described distortion factor has N, this method further comprises: select the minimum distortion degree from N the distortion factor, code book that will be corresponding with described minimum distortion degree and adjustment amount are as the code book and the adjustment amount of the described quantification input vector that draws.

16. method according to claim 15 is characterized in that, described method further comprises: the call number of code book that will be corresponding with described minimum distortion degree, and the output of the call number of the adjustment amount corresponding with described minimum distortion degree.

17. a method that produces the adjustment amount of code book is characterized in that, comprising:

Calculate the mean vector of N code book, N is the positive integer more than or equal to 1;

Calculate the distance of each code book to this mean vector;

According to the distance of each code book, produce the adjustment amount of each code book to this mean vector.

18. method according to claim 17 is characterized in that, the adjustment amount of described each code book of generation is: the adjustment amount that produces each dimension component of each code book.

19. method according to claim 18 is characterized in that, this method further comprises: the call number that produces each adjustment amount.

20. method according to claim 19, it is characterized in that, the call number of described each adjustment amount of generation is: sort to obtaining each product corresponding with this dimension component of this code book, the sequence number that each product is obtained is as the call number of the adjustment amount of this product correspondence.

21. a vector quantization apparatus is characterized in that, comprising: code book provides unit, adjustment amount that unit and quantifying unit are provided, wherein,

Code book provides the unit, is used for sending the N corresponding with an input vector code book to quantifying unit, and N is the positive integer more than or equal to 1;

Adjustment amount provides the unit, is used for sending the adjustment amount of each code book to quantifying unit;

Quantifying unit, be used to receive from adjustment amount provide the unit each code book adjustment amount and provide the unit from code book, utilize the adjustment amount of N code book and each code book that input vector is carried out quantification treatment, draw the code book and the adjustment amount that quantize input vector.

22. device according to claim 21 is characterized in that, described device further comprises: the adjustment amount generation unit, be used to calculate the mean vector of N code book, and N is the positive integer more than or equal to 1; Calculate the distance of each code book to this mean vector; To the distance of this mean vector, produce the adjustment amount more than 1 or 1 of each code book according to each code book, send the adjustment amount of each code book of producing to adjustment amount the unit is provided.

23. according to claim 21 or 22 described devices, it is characterized in that described device further comprises: residual error vector computing unit is used for calculating an input vector and N code book, residual error vector between each code book sends the residual error vector between each code book to quantifying unit.

24. device according to claim 23 is characterized in that, described quantifying unit comprises: quantize subelement, computation subunit and selected cell, wherein,

Quantize subelement, be used to receive the adjustment amount that each code book of unit is provided from adjustment amount, with residual error vector from residual error vector computing unit, the absolute difference between each the dimension component that calculates each residual error vector and each adjustment amount of the same dimension component of corresponding code book; Select each the dimension component corresponding minimum absolute difference value with the residual error vector, the result of calculation of minimum absolute difference value that will be corresponding with each dimension component of residual error vector sends computation subunit to;

Computation subunit, be used to receive result of calculation from quantizing the corresponding minimum absolute difference value of each dimension component subelement and each residual error vector, calculate this residual error vector the minimum absolute difference value of respectively tieing up the component correspondence square, square suing for peace the minimum absolute difference value of each dimension component correspondence again, with the summed result that obtains as the distortion factor between this residual error vector and the adjustment amount, calculate N the distortion factor of N residual error vector, send N the distortion factor to selected cell;

The chooser unit is used to receive N the distortion factor from computation subunit, selects the minimum distortion degree, and code book and the adjustment amount corresponding with the minimum distortion degree, code book that will be corresponding with the minimum distortion degree and adjustment amount output.

25. device according to claim 24, it is characterized in that, described device further comprises: output unit, be used to receive code book corresponding and adjustment amount, codebook index that will be corresponding number and export with the call number of the corresponding adjustment amount of minimum distortion degree with the minimum distortion degree with the minimum distortion degree from the chooser unit.

Images