CN105205838A - Vector quantization codebook construction method based on chaotic particle swarm algorithm - Google Patents

Abstract

The invention requests to protect a vector quantization codebook construction method based on a chaotic particle swarm algorithm, and particularly relates to a vector quantizer based on a particle swarm, wherein the vector quantizer can be used for a voice compression coding and image compression system. According to the method, a particle swarm algorithm and a chaotic algorithm are combined to be used for construction of a vector quantization codebook, and the algorithm can be extended to a multistage vector quantizer. Chaotic optimization is integrated in the particle swarm optimization algorithm so that speed of the particle swarm algorithm for being away from the local extreme point can be accelerated, convergence precision of the algorithm can be enhanced and performance of the vector quantizer is enabled to be improved.

Description

A kind of vector quantization code book building method based on Chaos particle swarm optimization algorithm

Technical field

The invention belongs to field of data compression, be specifically related to a kind of vector quantization method and vector quantizer, have a wide range of applications in voice compression coding and image compression system.

Background technology

Vector quantization is the new development of Shannon information theory in source coding theory, several scalar datas is formed a vector, in vector space, then gives overall quantification, reach and not only have compressed data but also do not lose too many information.Therefore, vector quantization is a kind of data compression technique efficiently, has ratio of compression large, and the advantage such as to decode simple, has a wide range of applications at voice and image code domain.The basic theories of vector quantization has people as far back as twentieth century six the seventies and pays close attention to, and starts gradual perfection in the 1980s.Realize a practical vector quantizer, what first need to solve is the design of code book.If code book performance is not good, by the coding quality of impact vector quantization system, as caused the decline of decoding end phonetic synthesis quality or image reconstruction quality.LBG algorithm is a kind of code book building method of classics proposed in 1980 by Linde, Buzo and Gray, because of its have theoretical tight, be easy to implement, restrain feature fast, achieve in vector quantization field and apply widely.But the performance of LBG algorithm too relies on inceptive code book, and is easily absorbed in local optimum.For this reason, scholars, on the basis of LBG algorithm, propose a lot of code book construction algorithm improved, as simulated annealing code book algorithm, PNN code book algorithm, neural network code book algorithm, immune cat group optimize code book algorithm etc.Particle cluster algorithm combines with chaos algorithm by the present invention, proposes a kind of code book building method based on Chaos particle swarm optimization algorithm.The method can increase the diversity of solution space, improves ability of searching optimum and convergence precision, also can process the problem in ghost chamber simultaneously.

Chaos is a kind of non-linear phenomena that nature extensively exists, and Chaos Variable seems mixed and disorderly, but its change procedure contains inherent regularity, is described as last century Mo maximum discovery, for the progress of promotion modern science plays an important role.There is no strict definition to chaos at present, generally the motion state with randomness obtained by determinacy equation is called chaos, the variable presenting chaos state is called Chaos Variable.Chaotic motion is seemingly random, but imply exquisite immanent structure, there is ergodicity, randomness and the feature to starting condition susceptibility, can repeatedly not travel through all states by himself rule within the specific limits, therefore, these character of chaotic motion are utilized to be optimized search.Following logic equation describes a most typical chaos system.

Z _n+1＝μZ _n(1-Z _n),n＝0,1,2,…

μ in above formula is controling parameter, after this value is determined, by any initial value Z ₀(0,1) iteration can go out a time series Z determined ₀, Z ₁..., Z _n.

Particle group optimizing (ParticleSwarmOptimization, PSO) algorithm is the inspiration by Kennedy and Eberhart in nineteen ninety-five by artificial life result of study at first, a kind of computing technique of evolving based on swarm intelligence of migrating with proposing during collective behaviour of looking for food simulation flock of birds in process.Because this algorithm is simple, fast convergence rate, and features such as less (as without the need to gradient information) is required to objective function, therefore develop very rapid, and to succeed application at numerous areas.In PSO model, a bird in the corresponding search volume of each solution of optimization problem, be somebody's turn to do " bird " also referred to as particle, each particle has a speed, determines its direction of circling in the air and distance.PSO is initialized as a group random particles, and then particle starts to follow current optimal particle motion, until optimum solution is found in search in whole solution space.In each iteration, particle upgrades oneself by following the trail of two extreme values, and an optimum solution being particle oneself and finding, is called individual extreme value p _best; Another is the optimum solution that whole population finds, and is called global extremum g _best.Suppose to use represent i-th particle, wherein n is the dimension of particle, and the desired positions of its experience is expressed as and the desired positions of whole colony experience is expressed as the speed of particle i is by searching current optimal particle principle, particle i changes speed and position by (1) formula and (2) formula.

$v_i^j(t+1)={\mathrm{wv}}_i^j\left(t\right)+c_1{r}_1\left(p_i^j\right(t)-x_i^j(t\left)\right)+c_2{r}_2\left(g_i^j\right(t)-x_i^j(t\left)\right)---\left(1\right)$

$x_i^j(t+1)=x_i^j\left(t\right)+v_i^j(t+1)---\left(2\right)$

Wherein, t is current iterations; c ₁, c ₂for Studying factors; r ₁, r ₂for obeying the random number that (0,1) distributes, w is the inertia weight factor.Research shows: larger w value is conducive to jumping out local minizing point, and less w value is conducive to algorithm convergence and improves the precision of separating.In a lot of algorithm, the normal linear decrease value adopting (3) formula to calculate w.

$w=w_{max}-(w_{max}-w_{min})\frac{t}{t_{max}}---\left(3\right)$

T _maxfor the maximum iteration time of setting; w _maxfor maximum inertia weight; w _minfor minimum inertia weight.In addition, for making particle rapidity be unlikely to excessive, can setting speed higher limit V _max.When the speed of certain one dimension exceedes this setting speed, be V with regard to making the speed of this one dimension _max.

Although it is simple that particle swarm optimization algorithm has principle, be easy to the advantage applied, also have and be easily absorbed in Local Extremum, later stage of evolution speed of convergence is slow, the shortcomings such as precision is poor.

Summary of the invention

For the deficiency of particle swarm optimization algorithm, propose a kind of vector quantization code book building method based on Chaos particle swarm optimization algorithm.Technical scheme of the present invention is as follows: a kind of vector quantization code book building method based on Chaos particle swarm optimization algorithm, and it comprises the following steps:

101, extract the characteristic ginseng value of voice signal or picture signal, the characteristic ginseng value of extraction is as the trained vector constructing code book;

102, be stored in calculator memory by the whole trained vector X generated needed for vector quantization code book, the set of whole X represents with S;

103, initialization of population, arranges the maximum iteration time MaxLoop of iterative algorithm, the number K of particle, vector quantization progression M, maximal rate V _max, minimum speed V _min; Initial velocity p [k] .v, the k=1 of particle are set, 2 ..., K; Fitness initial value p [k] .f, k=1 are set, 2 ..., K; Iterative initial value m=1 is set, progression initial value kkk=1;

104, in the whole trained vector X from step 102, the N number of trained vector of random extraction forms an inceptive code book repeat K time, obtain K particle p [k]; According to the most contiguous criterion, often input a particle, be divided into N number of subset in, namely when time, following formula is set up:

$d(X,y_l^{m-1})\≤d(X,y_i^{m-1}),\∀i,i\≠l$

$d(X,y_l^{m-1})=\left|\right|X-y_l^{m-1}|{|}^2$

105, new code word is calculated $p\[k\].Y=(p\[k\].y_1^m,p\[k\].y_2^m,...,p\[k\].y_N^m),$ Wherein be barycenter, computing formula is as follows, num _iit is set the number of middle trained vector;

$p\[k\].y_i^m=\frac{1}{{\mathrm{num}}_i}\underset{X\∈S_i^m}{\Σ}X$

Add up each cell, i.e. each subset the trained vector number comprised, by the distortion of each trained vector by sorting from big to small.For ghost chamber, then cell trained vector number is not 1 and the maximum trained vector of distortion replaces ghost chamber as new code word, the trained vector number in this cell is subtracted 1 simultaneously;

106, the fitness of each particle is calculated

$p\[k\].f^{\left(m\right)}=\underset{i=1}{\overset{N}{\Σ}}\underset{X\∈S_i^m}{\Σ}d(X,y_i^m)$

Relatively fitness p [k] .f ^(m)with p [k] .f _bestif, p [k] .f ^(m)be less than fitness p [k] .f of optimal location _best, then p [k] .f is upgraded _bestwith fitness p [k] .f of more each particle _bestwith the fitness f of the optimal location of group _gbestif, p [k] .f _bestbe less than f _gbest, then f is upgraded _gbestwith colony's optimal particle $p.Y_{gbest}=(p\[k\].y_1^m,p\[k\].y_2^m,...,p\[k\].y_N^m);$

If 107 iterations m aliquots 3, then carry out chaos optimization to optimal particle.In following formula, β _iand α _ibe maximal value and the minimum value of characteristic parameter i-th dimension, μ is controling parameter;

$z_i=\frac{p\[k\].y_i-{\mathrm{\α}}_i}{{\mathrm{\β}}_i-{\mathrm{\α}}_i}$

z _i＝μ(1-z _i)

p[k].y _i＝α _i+(β _i-α _i)z _i

Circulate three times, produce 3 particles, retain the particle that fitness is minimum, and replace a particle of random selecting in current group; Judge whether iterations is greater than MaxLoop, if be greater than MaxLoop, then jumps to step 109, otherwise forwards step 108 to, continue to perform;

108, the speed of more new particle and particle;

109, ask trained vector to quantize residual error, output codons, progression kkk adds 1, if kkk>M, terminates; If kkk≤M, reset initial value, and trained vector is changed into the quantification residual error vector of upper level.

Further, if vector quantization progression M is set to 1, what design is single-stage vector quantizer code book; If M>1, what design is M level vector quantizer code book.

Further, the computing formula of inertia weight in particle swarm optimization factor w is:

$w=w_{\min }+(w_{max}-w_{min})\frac{1+cos\[\frac{f_g(m-1)}{f_i(MaxLoop-1)}\mathrm{\π}\]}{2}$

In above formula, f _gglobal optimum's fitness, w _maxrepresent maximum inertia weight, w _minrepresent minimum inertia weight, f _ibe current particle fitness, MaxLoop is maximum iteration time, and m is current iteration number of times.

Advantage of the present invention and beneficial effect as follows:

The present invention takes full advantage of the wide and chaos algorithm in particle cluster algorithm hunting zone and has ergodic feature, particle cluster algorithm is combined with chaos algorithm, increase the diversity of solution space, improve ability of searching optimum and convergence precision, effective process is also done to the problem in ghost chamber simultaneously, optimized the performance of quantizer.The present invention can be used for, in voice and image compression encoding, having a good application prospect and practical value.

Accompanying drawing explanation

Fig. 1 the invention provides the code book construction algorithm process flow diagram of preferred embodiment based on Chaos-Particle Swarm Optimization.

Embodiment

Below in conjunction with accompanying drawing, the invention will be further described:

Chaos optimization thought is incorporated in particle swarm optimization algorithm, proposes a kind of based on Chaos particle swarm optimization algorithm.This algorithm makes solution space have diversity, can improve ability of searching optimum and convergence precision.Adopt the step of Chaos particle swarm optimization algorithm structure vector quantization code book as shown in Figure 1.

In particle cluster algorithm, the size of inertia weight factor w have impact on the search capability of algorithm.When w is larger, be conducive to the ability of searching optimum improving algorithm, but local search ability is more weak; When w is less, the local search ability of particle can be strengthened, but the ability exploring new region is more weak.In standard particle group algorithm, w is definite value, so just cannot make full use of the exploring ability of particle.In Linear recurring series particle cluster algorithm, w is not fixing, is to upgrade according to formula (3), makes w relevant with iterations, and along with iterations increases, w linearly reduces.Relative to standard particle group algorithm, although this algorithm improves to some extent in performance, premature problem does not still obtain essence and solves, and especially the diversity in particle later stage is deteriorated, and causes speed of convergence slack-off, is very easily absorbed in local optimum.

The fitness f of particle is a parameter of reflection particle current location quality, certain is had to the particle of higher fitness, and the regional area at its place may exist the some p that can obtain upgrading global optimum _best, i.e. p _bestrepresented solution is better than global optimum g _best.In order to enable global optimum be upgraded rapidly, p can be found rapidly _best, the inertia weight factor of particle should be reduced to strengthen its local search ability.And for the lower particle of fitness, current location is poor, there is the probability being better than globally optimal solution in region lower, in order to jump out current region, the inertia weight factor should be increased, to strengthen ability of searching optimum.So the inertia weight factor is not only relevant with iterations, also closely related with fitness f, in the present invention, inertia weight factor w is calculated by formula below.

$w=w_{min}+(w_{max}-w_{min})\frac{1+cos\[\frac{f_g(m-1)}{f_i(MaxLoop-1)}\mathrm{\π}\]}{2}$

In above formula, f _gglobal optimum's fitness, f _ibe current particle fitness, MaxLoop is maximum iteration time, and m is current iteration number of times.In order to strengthen ability of searching optimum and the convergence precision of particle further, chaos optimization is carried out to the optimal value searched out at every turn.Utilize the ergodicity of chaotic motion, based on the optimal location that current whole population searches, produce chaos sequence, the particle of the optimal location in the chaos sequence produced is substituted at random the position of a particle in current particle group.In order to reduce the complexity of algorithm, the present invention carries out chaos optimization at every turn, but takes every Q iteration to carry out a chaos optimization, effectively can reduce algorithm complex like this, little to the quality influence of last acquisition particle.

The structure of optimum vector quantizer needs to follow following two principle:

1, arest neighbors condition

If given code book C={y ₀, y ₁..., y _n-1, size is N, then optimal dividing { the R in input vector space ₀, R ₁..., R _n-1should meet $R_i=\left\{x\right|d(x,y_j)=\underset{0\≤j\≤N-1}{min}d(x,y_j)\}.$

2, barycenter condition

To given division { R ₀, R ₁..., R _n-1, optimum code word y _jmust be corresponding cell R _j" barycenter ", i.e. y _j=cent (R _j).

In example of the present invention, intend carrying out three grades of vector quantizations to speech characteristic parameter line spectrum pair frequency LSF, construct three grades of LSF code books, its size is respectively 256,64,32, trained vector used is the LSF parameter adopting MELP (Mixed Excitation Linear Prediction) (MELP) scrambler to extract from a large amount of Chinese and English speech database, and vector dimension is 10.Concrete step is as follows:

1) be stored in calculator memory by the whole trained vector X generated needed for vector quantization code book, the set of whole X represents with S;

2) the maximum iteration time MaxLoop=20 of iterative algorithm is set, the number K=15 of particle, vector quantization progression M=3, maximal rate V _max=150, minimum speed V _min=-150, progression initial value kkk=1;

3) initial velocity p [k] .v, the k=1 of particle are set, 2 ..., K;

4) fitness initial value p [k] .f, k=1 are set, 2 ..., K;

5) iterative initial value m=1 is set;

6) from trained vector, the N number of trained vector of random extraction forms an inceptive code book $p\[k\].Y=(p\[k\].y_1^0,p\[k\].y_2^0,...,p\[k\].y_N^0),$ Repeat K time, obtain K particle p [k];

7) according to the most contiguous criterion, often input a particle, be divided into N number of subset in, namely when time, following formula is set up:

$d(X,y_l^{m-1})\≤d(X,y_i^{m-1}),\∀i,i\≠l$

$d(X,y_l^{m-1})=\left|\right|X-y_l^{m-1}|{|}^2$

8) new code word is calculated $p\[k\].Y=(p\[k\].y_1^m,p\[k\].y_2^m,...,p\[k\].y_N^m),$ Wherein be barycenter, computing formula is as follows, num _iit is set the number of middle trained vector.

$p\[k\].y_i^m=\frac{1}{{\mathrm{num}}_i}\underset{X\∈S_i^m}{\Σ}X$

9) the trained vector number that each cell comprises is added up, by the distortion of each trained vector by sorting from big to small.For ghost chamber, then cell trained vector number is not 1 and the maximum trained vector of distortion replaces ghost chamber as new code word, the trained vector number in this cell is subtracted 1 simultaneously;

10) fitness of each particle is calculated

$p\[k\].f^{\left(m\right)}=\underset{i=1}{\overset{N}{\Σ}}\underset{X\∈S_i^m}{\Σ}d(X,y_i^m)$

11) fitness p [k] .f is compared ^(m)with p [k] .f _bestif, p [k] .f ^(m)be less than p [k] .f _best, then p [k] .f is upgraded _bestwith $p\[k\].Y_{best}=(p\[k\].y_1^m,p\[k\].y_2^m,...,p\[k\].y_N^m);$

12) fitness p [k] .f is compared _bestwith f _gbestif, p [k] .f _bestbe less than f _gbest, then f is upgraded _gbestwith colony's optimal particle $p.Y_{gbest}=(p\[k\].y_1^m,p\[k\].y_2^m,...,p\[k\].y_N^m)$

13) if m can divide exactly 3, then chaos optimization is carried out to optimal particle, β in following formula _ithe maximal value that LSF i-th ties up, α _ibe the minimum value that LSF i-th ties up, μ is controling parameter;

$z_i=\frac{p\[k\].y_i-{\mathrm{\α}}_i}{{\mathrm{\β}}_i-{\mathrm{\α}}_i}$

z _i＝μ(1-z _i)

p[k].y _i＝α _i+(β _i-α _i)z _i

Circulate three times, produce 3 particles, retain the particle that fitness is minimum, and replace a particle of random selecting in current group.

14) judge whether iterations is greater than MaxLoop, if be greater than MaxLoop, then jumps to step 17, otherwise forwards step 15 to, continue to perform.

15) according to the speed of following formula more new particle and particle

$p\[k\].v_i^{m+1}=wp\[k\].v_i^m+c_1{r}_1(p\[k\].Y_{best}-p\[k\].y_i^m)+c_2{r}_2(p.Y_{gbest}-p\[k\].y_i^m)$

$p\[k\].y_i^{m+1}=p\[k\].y_i^m+p\[k\].v_i^{m+1}$

m＝m+1

16) jump to the 7th step, continue to perform;

17) try to achieve trained vector and quantize residual error, output codons, progression kkk adds 1.If kkk>M, jump to step 18; If kkk≤M, reset initial value, and trained vector is changed into the residual error of upper level quantification.Initial value arranges as follows:

if(kkk＝2)

N＝64；

V _max＝f _gbest；V _min＝-f _gbest；

if(kkk＝3)

N＝32；

V _max＝f _gbest；V _min＝-f _gbest；

18) terminate

These embodiments are interpreted as only being not used in for illustration of the present invention limiting the scope of the invention above.After the content of reading record of the present invention, technician can make various changes or modifications the present invention, and these equivalence changes and modification fall into the scope of the claims in the present invention equally.

Claims (3)

1., based on a vector quantization code book building method for Chaos particle swarm optimization algorithm, it is characterized in that, comprise the following steps:

101, extract the characteristic ginseng value of voice signal or picture signal, the characteristic ginseng value of extraction is as the trained vector constructing code book;

102, be stored in calculator memory by the whole trained vector X generated needed for vector quantization code book, the set of whole X represents with S;

103, initialization of population, arranges the maximum iteration time MaxLoop of iterative algorithm, the number K of particle, vector quantization progression M, maximal rate V _max, minimum speed V _min; Initial velocity p [k] .v, the k=1 of particle are set, 2 ..., K; Fitness initial value p [k] .f, k=1 are set, 2 ..., K; Iterative initial value m=1 is set, progression initial value kkk=1;

104, in the whole trained vector X from step 102, the N number of trained vector of random extraction forms an inceptive code book repeat K time, obtain K particle p [k]; According to the most contiguous criterion, often input a particle, be divided into N number of subset in, namely when time, following formula is set up:

$d(X,y_l^{m-1})\≤d(X,y_i^{m-1}),\∀i,i\≠l$

$d(X,y_l^{m-1})=\left|\right|X-y_l^{m-1}|{|}^2$

105, new code word is calculated $p\[k\].Y=(p\[k\].y_1^m,p\[k\].y_2^m,...,p\[k\].y_N^m),$ Wherein be barycenter, computing formula is as follows, num _iit is set the number of middle trained vector;

$p\[k\].y_i^m=\frac{1}{{\mathrm{num}}_i}\underset{X\∈S_i^m}{\Σ}X$

Add up each cell, i.e. each subset the trained vector number comprised, by the distortion of each trained vector by sorting from big to small.For ghost chamber, then cell trained vector number is not 1 and the maximum trained vector of distortion replaces ghost chamber as new code word, the trained vector number in this cell is subtracted 1 simultaneously;

106, the fitness of each particle is calculated

$p\[k\].f^{\left(m\right)}=\underset{i=1}{\overset{N}{\Σ}}\underset{X\∈S_i^m}{\Σ}d(X,y_i^m)$

Relatively fitness p [k] .f ^(m)with p [k] .f _bestif, p [k] .f ^(m)be less than fitness p [k] .f of optimal location _best, then p [k] .f is upgraded _bestwith fitness p [k] .f of more each particle _bestwith the fitness f of the optimal location of group _gbestif, p [k] .f _bestbe less than f _gbest, then f is upgraded _gbestwith colony's optimal particle $p.Y_{gbest}=(p\[k\].y_1^m,p\[k\].y_2^m,...,p\[k\].y_N^m);$

If 107 iterations m aliquots 3, then carry out chaos optimization to optimal particle, in following formula, β _iand α _ibe maximal value and the minimum value of characteristic parameter i-th dimension, μ is controling parameter;

$z_i=\frac{p\[k\].y_i-{\mathrm{\α}}_i}{{\mathrm{\β}}_i-{\mathrm{\α}}_i}$

z _i＝μ(1-z _i)

p[k].y _i＝α _i+(β _i-α _i)z _i

Circulate three times, produce 3 particles, retain the particle that fitness is minimum, and replace a particle of random selecting in current group; Judge whether iterations is greater than MaxLoop, if be greater than MaxLoop, then jumps to step 109, otherwise forwards step 108 to, continue to perform;

108, the speed of more new particle and particle;

109, ask trained vector to quantize residual error, output codons, progression kkk adds 1, if kkk>M, terminates; If kkk≤M, reset initial value, and trained vector is changed into the quantification residual error vector of upper level.

2. a kind of vector quantization code book building method based on Chaos particle swarm optimization algorithm according to claim 1, is characterized in that, if vector quantization progression M is set to 1, what design is single-stage vector quantizer code book; If M>1, what design is M level vector quantizer code book.

3. a kind of vector quantization code book building method based on Chaos particle swarm optimization algorithm according to claim 1, it is characterized in that, the computing formula of inertia weight in particle swarm optimization factor w is:

$w=w_{\min }+(w_{max}-w_{\min })\frac{1+cos\[\frac{f_g(m-1)}{f_i(MaxLoop-1)}\mathrm{\π}\]}{2}$

In above formula, f _gglobal optimum's fitness, w _maxrepresent maximum inertia weight, w _minrepresent minimum inertia weight, f _ibe current particle fitness, MaxLoop is maximum iteration time, and m is current iteration number of times.