JSci.maths.wavelet.cdf3_5.CDF3_5 Maven / Gradle / Ivy
Go to download
Show more of this group Show more artifacts with this name
Show all versions of jsci Show documentation
Show all versions of jsci Show documentation
JSci is a set of open source Java packages. The aim is to encapsulate scientific methods/principles in the most natural way possible. As such they should greatly aid the development of scientific based software.
It offers: abstract math interfaces, linear algebra (support for various matrix and vector types), statistics (including probability distributions), wavelets, newtonian mechanics, chart/graph components (AWT and Swing), MathML DOM implementation, ...
Note: some packages, like javax.comm, for the astro and instruments package aren't listed as dependencies (not available).
The newest version!
package JSci.maths.wavelet.cdf3_5;
import JSci.maths.wavelet.*;
import JSci.maths.*;
/******************************************
* Cohen-Daubechies-Feauveau
* with N=3 and
* Ntilde=5 adapted to the interval
* by Deslauriers-Dubuc-Lemire
* @author Daniel Lemire
*****************************************/
public final class CDF3_5 extends Multiresolution implements Filter {
protected final static int filtretype=2;
protected final static int minlength=12;
/****************************************
* This method is used to compute
* how the number of scaling functions
* changes from on scale to the other.
* Basically, if you have k scaling
* function and a Filter of type t, you'll
* have 2*k+t scaling functions at the
* next scale (dyadic case).
* Notice that this method assumes
* that one is working with the dyadic
* grid while the method "previousDimension"
* define in the interface "Filter" doesn't.
******************************************/
public int getFilterType () {
return(filtretype);
}
public MultiscaleFunction primaryScaling(int n0, int k) {
return(MultiSpline3_5.scaling(n0,k));
}
public MultiscaleFunction dualScaling(int n0, int k) {
return(new DualScaling3_5(n0,k));
}
public MultiscaleFunction primaryWavelet(int n0, int k) {
return(MultiSpline3_5.wavelet(n0,k));
}
public MultiscaleFunction dualWavelet(int n0, int k) {
return(new DualWavelet3_5(n0,k));
}
/********************************************
*********************************************/
final static double[] v0={8703/5120d, 14851/5120d, -7497/2560d, -4137/2560d, 4703/2560d, 5799/2560d, -3315/1024d, 1105/1024d};
final static double[] v1={-983/5120d, 2949/5120d, 4977/2560d, 1617/2560d, -2423/2560d, -2559/2560d, 1515/1024d, -505/1024d};
final static double[] v2={2897/51200d, -8691/51200d, -3863/25600d, 23177/25600d, 116611/76800d, 22521/25600d, -3153/2048d, 2833/6144d, 15/256d, -5/256d};
final static double[] v3={-789/51200d, 2367/51200d, 531/25600d, -4749/25600d, -869/25600d, 16323/25600d, 4429/2048d, -751/2048d, -97/256d, 19/256d, 15/256d, -5/256d};
final static double[] v4={7/3200d, -21/3200d, -3/1600d, 37/1600d, -3/1600d, -99/1600d, -55/128d, 189/128d, 175/128d, -13/128d, -97/256d, 19/256d, 15/256d, -5/256d};
final static double[] v5={0d, 0d, 0d, 0d, 0d, 0d, 17/128d, -51/128d, -13/128d, 175/128d, 175/128d, -13/128d, -97/256d, 19/256d, 15/256d, -5/256d};
final static double[] vg={-5/256d, 15/256d, 19/256d, -97/256d, -13/128d, 175/128d, 175/128d, -13/128d, -97/256d, 19/256d, 15/256d, -5/256d};
final static double[] vd0=ArrayMath.invert(v0);
final static double[] vd1=ArrayMath.invert(v1);
final static double[] vd2=ArrayMath.invert(v2);
final static double[] vd3=ArrayMath.invert(v3);
final static double[] vd4=ArrayMath.invert(v4);
final static double[] vd5=ArrayMath.invert(v5);
/********************************************
*********************************************/
final static double[] phvg={-1/2d, 3/2d, -3/2d, 1/2d};
/****************************************
* This method return the number of "scaling"
* functions at the previous scale given a
* number of scaling functions. The answer
* is always smaller than the provided value
* (about half since this is a dyadic
* implementation). This relates to the same idea
* as the "Filter type". It is used by
* the interface "Filter".
*****************************************/
public int previousDimension (int k) {
return(Cascades.previousDimension(filtretype,k));
}
public CDF3_5 () {}
/****************************************
* This is the implementation of the lowpass
* Filter. It is used by the interface
* "Filter". Lowpass filters are normalized
* so that they preserve constants away from
* the boundaries.
*****************************************/
public double[] lowpass (double[] v, double[] param) {
return(lowpass(v));
}
/****************************************
* This is the implementation of the highpass
* Filter. It is used by the interface
* "Filter". Highpass filters are normalized
* in order to get L2 orthonormality of the
* resulting wavelets (when it applies).
* See the class DiscreteHilbertSpace for
* an implementation of the L2 integration.
*****************************************/
public double[] highpass (double[] v, double[] param) {
return(highpass(v));
}
/****************************************
* This is the implementation of the lowpass
* Filter. It is used by the interface
* "Filter". Lowpass filters are normalized
* so that they preserve constants away from
* the boundaries.
*****************************************/
public double[] lowpass (double[] gete) {
if(gete.length<12) {
throw new IllegalScalingException("The array is not long enough : "+gete.length+" < 12");
}
double[] sortie=new double[2*gete.length-2];
int dl0=gete.length-1;
for(int k=6;k<=dl0-6;k++) {
for(int L=-6;L<=5;L++){
sortie[2*k+L]+=vg[L+6]*gete[k];
}
}
sortie=ArrayMath.add(sortie,gete[0],v0,0);
sortie=ArrayMath.add(sortie,gete[1],v1,0);
sortie=ArrayMath.add(sortie,gete[2],v2,0);
sortie=ArrayMath.add(sortie,gete[3],v3,0);
sortie=ArrayMath.add(sortie,gete[4],v4,0);
sortie=ArrayMath.add(sortie,gete[5],v5,0);
int p0=sortie.length-vd0.length;
int p1=sortie.length-vd1.length;
int p2=sortie.length-vd2.length;
int p3=sortie.length-vd3.length;
int p4=sortie.length-vd4.length;
int p5=sortie.length-vd5.length;
sortie=ArrayMath.add(sortie,gete[dl0],vd0,p0);
sortie=ArrayMath.add(sortie,gete[dl0-1],vd1,p1);
sortie=ArrayMath.add(sortie,gete[dl0-2],vd2,p2);
sortie=ArrayMath.add(sortie,gete[dl0-3],vd3,p3);
sortie=ArrayMath.add(sortie,gete[dl0-4],vd4,p4);
sortie=ArrayMath.add(sortie,gete[dl0-5],vd5,p5);
return(sortie);
}
/****************************************
* This is the implementation of the highpass
* Filter. It is used by the interface
* "Filter". Highpass filters are normalized
* in order to get L2 orthonormality of the
* resulting wavelets (when it applies).
* See the class DiscreteHilbertSpace for
* an implementation of the L2 integration.
*****************************************/
public double[] highpass(double[] v) {
if(v.length<4) {
throw new IllegalScalingException("The array is not long enough : "+v.length+" < 4");
}
double[] ans=new double[2*v.length+2];
for(int k=0;k