A Class for performing the Clark Clean Algorithm on Arrays. More...

#include <ClarkCleanModel.h>

Inheritance diagram for casa::ClarkCleanModel:

Public Member Functions
	ClarkCleanModel ()
	The default constructor does nothing more than initialise a zero length array to hold the deconvolved model.
	ClarkCleanModel (Array< Float > &model)
	Construct the ClarkCleanModel object and initialise the model.
	ClarkCleanModel (Array< Float > &model, Array< Float > &mask)
	Construct the ClarkCleanModel object and initialise the model ans mask.
void	getMask (Array< Float > &mask) const
void	setMask (const Array< Float > &mask)
void	setMask (Array< Float > &mask)
void	getModel (Array< Float > &model) const
void	setModel (const Array< Float > &model)
	Set the current model.
void	setModel (Array< Float > &model)
virtual void	setProgress (ClarkCleanProgress &ccp)
	Set/get the progress display.
virtual ClarkCleanProgress &	getProgress ()
Bool	solve (ConvolutionEquation &eqn)
	Using a Clark clean deconvolution proceedure solve for an improved estimate of the deconvolved object.
Bool	singleSolve (ConvolutionEquation &eqn, Array< Float > &residual)
virtual void	setPsfPatchSize (const IPosition &psfPatchSize)
	These functions set various "knobs" that the user can tweak and are specific to the Clark clean algorithm.
virtual IPosition	getPsfPatchSize ()
virtual void	setHistLength (const uInt HistBins)
	Set the size of the histogram used to determine how many pixels are "active" in a minor iteration.
virtual uInt	getHistLength ()
virtual void	setMaxNumberMinorIterations (const uInt maxNumMinorIterations)
	Set the maximum number of minor iterations to perform for each major cycle.
virtual uInt	getMaxNumberMinorIterations ()
virtual void	setInitialNumberIterations (const uInt initialNumberIterations)
	Set and get the initial number of iterations.
virtual uInt	getInitialNumberIterations ()
virtual void	setMaxNumberMajorCycles (const uInt maxNumMajorCycles)
	Set the maximum number of major cycles to perform.
virtual uInt	getMaxNumberMajorCycles ()
virtual void	setMaxNumPix (const uInt maxNumPix)
	Set the maximum number of active pixels to use in the minor iterations.
virtual uInt	getMaxNumPix ()
virtual void	setMaxExtPsf (const Float maxExtPsf)
	Set the maximum exterior psf value.
virtual Float	getMaxExtPsf ()
virtual void	setSpeedup (const Float speedup)
	An exponent on the F(m,n) factor (see Clark[1980]) which influences how quickly active pixels are treated as unreliable.
virtual Float	getSpeedup ()
virtual void	setCycleSpeedup (const Float speedup)
	The structure of various AIPS++ algorithms creates major cycles around the Clark Clean (or other deconvolution algrithms.
virtual Float	getCycleSpeedup ()
virtual const Float	threshold ()
	We are overwriting Iterate's threshold() method to put out speedup in it.
virtual void	setChoose (const Bool askForChoice)
	The user can be asked whether to stop after every minor cycle.
virtual Bool	getChoose ()
Private Member Functions
void	doMinorIterations (Array< Float > &model, Matrix< Float > &pixelValue, const Matrix< Int > &pixelPos, const Int numPix, Matrix< Float > &psfPatch, Float fluxLimit, uInt &numberIterations, Float Fmn, const uInt totalIterations, Float &totalflux)

Int	cacheActivePixels (Matrix< Float > &pixVal, Matrix< Int > &pixPos, const Array< Float > &data, const Float fluxLimit)
	Find all the pixels in the residual that are greater than fluxlimit, and store the values in the pixelsValue Matrix, and their positions in the pixelPos Matrix.
void	absHistogram (Vector< Int > &hist, Float &minVal, Float &maxVal, const Array< Float > &data)
	make histogram of absolute values in array
Float	biggestResiduals (Float &maxRes, const uInt maxNumPix, const Float fluxLimit, const Array< Float > &residual)
	Determine the flux limit if we only select the maxNumPix biggest residuals.
Float	getPsfPatch (Array< Float > &psfPatch, ConvolutionEquation &eqn)
	Work out the size of the Psf patch to use.
Float	maxResidual (const Array< Float > &residual)
	The maximum residual is the absolute maximum.
void	maxVect (Vector< Float > &maxVal, Float &absVal, Int &offset, const Matrix< Float > &pixVal, const Int numPix)
void	subtractComponent (Matrix< Float > &pixVal, const Matrix< Int > &pixPos, const Int numPix, const Vector< Float > &maxVal, const Vector< Int > &maxPos, const Matrix< Float > &psf)
Float	absMaxBeyondDist (const IPosition &maxDist, const IPosition &centre, const Array< Float > &array)
Bool	stopnow ()
Private Attributes
uInt	theHistBins
Float	theMaxExtPsf
uInt	theMaxNumberMinorIterations
uInt	theInitialNumberIterations
Int	theMaxNumberMajorCycles
uInt	theMaxNumPix
IPosition	thePsfPatchSize
Float	theSpeedup
Float	theCycleSpeedup
Bool	theChoose
Array< Float >	theMask
LogIO	theLog
Int	theIterCounter
	There are too many iterations counters.
ClarkCleanProgress *	itsProgressPtr
Bool	itsJustStarting

Detailed Description

A Class for performing the Clark Clean Algorithm on Arrays.

Intended use:

Public interface

Review Status

Date Reviewed:: yyyy/mm/dd

Prerequisite

ResidualEquation/ConvolutionEquation
LinearModel/LinearEquation Paradigm

Etymology

This class is called ClarkCleanModel because thats the algorithm it uses deconvolve the model.

Synopsis

This class is used to perform the Clark Clean Algorithm on an Array. Only the deconvolved model of the sky are directly stored by this class. The point spread function (psf) and convolved (dirty) image are stored in a companion class which is must be derived from ResidualEquation.

The cleaning works like this. The user constructs a ClarkCleanModel by specifying an initial model of the sky. This can by be one,two,three... dimensional depending on the dimension of the psf (see below). The user then constructs a class which implements the forward equation between the model and the dirty image. Typically this will be the ConvolutionEquation class, although any class which has a ResidualEquation interface will be work (but perhaps very slowly, as the ConvolutionEquation class has member functions optimised for cleaning)

The user then calls the solve() function (with the appropriate equation class as an arguement), and this class will perform the Clark clean. The various clean parameters are set (prior to calling solve) using the functions derived from the Iterate class, in particular setGain(), setNumberIterations() & setThreshold() (to set a flux limit).

The solve() function does not return either the deconvolved model or the residuals. The solved model can be obtained using the getModel() function (derived from ArrayModel()) and the residual can be obtained using the residual() member function of the Convolution/Residual Equation Class.

The size and shape of the model used in this class MUST be the same as the convolved data (Dirty Image), stored in the companion ResidualEquation Class. However the model (and convolved data) can have more dimensions than the psf, as well as a different size (either larger or smaller). When the dimensionality is different the cleaning is done independendtly in each "plane" of the model. (Note this has not been implemented yet but is relatively simple to do if necessary).

This multi-dimensionalty is exploited when cleaning arrays of StokesVectors. Here the Array of StokesVectors is decomposed into a stack of 4 Floating point arrays and the cleaning is done on all the the arrays simultaneosly. The criterion for choosing the brightest pixel has been generalised by using the "length" of the Stokesvector in 4 dimensional space.

A companion class to this one is MaskedClarkCleanModel. This provides the same functionality but is used with MaskedArrays which indicate which regions of the model to search for clean components.

Example

    Matrix<Float> psf(12,12), dirty(10,10), initialModel(10,10);
    ...put appropriate values into psf, dirty, & initialModel...\.
    ClarkCleanModel<Float> deconvolvedModel(initialModel); 
    ConvolutionEquation convEqn(psf, dirty);
    deconvolvedModel.setGain(0.2); 
    deconvolvedModel.setNumberIterations(1000);
    Bool convWorked = deconvolvedModel.solve(convEqn);
    Array<Float> finalModel, residuals;
    if (convWorked){
      finalModel = deconvolvedModel.getModel();
      ConvEqn.residual(deconvolvedModel, finalResidual);
    }

Motivation

This class is needed to deconvolve images.

Template Type Argument Requirements (T)

have tested this class with Arrays of Float StokesVector

To Do

Make changes so that multidimensions work as advertised
compare timing with other clean implementations (ie, Mark's CleanTools, SDE, AIPS & miriad)

Definition at line 146 of file ClarkCleanModel.h.

Constructor & Destructor Documentation

casa::ClarkCleanModel::ClarkCleanModel ( )

The default constructor does nothing more than initialise a zero length array to hold the deconvolved model.

If this constructor is used then the actual model must be set using the setModel() function of the ArrayModel class.

casa::ClarkCleanModel::ClarkCleanModel ( Array< Float > & model )

Construct the ClarkCleanModel object and initialise the model.

casa::ClarkCleanModel::ClarkCleanModel	(	Array< Float > &	model,
		Array< Float > &	mask
	)

Construct the ClarkCleanModel object and initialise the model ans mask.

Member Function Documentation

void casa::ClarkCleanModel::absHistogram	(	Vector< Int > &	hist,
		Float &	minVal,
		Float &	maxVal,
		const Array< Float > &	data
	)		`[private]`

make histogram of absolute values in array

Float casa::ClarkCleanModel::absMaxBeyondDist	(	const IPosition &	maxDist,
		const IPosition &	centre,
		const Array< Float > &	array
	)		`[private]`

Float casa::ClarkCleanModel::biggestResiduals	(	Float &	maxRes,
		const uInt	maxNumPix,
		const Float	fluxLimit,
		const Array< Float > &	residual
	)		`[private]`

Determine the flux limit if we only select the maxNumPix biggest residuals.

Flux limit is not exact due to quantising by the histogram

Int casa::ClarkCleanModel::cacheActivePixels	(	Matrix< Float > &	pixVal,
		Matrix< Int > &	pixPos,
		const Array< Float > &	data,
		const Float	fluxLimit
	)		`[private]`

Find all the pixels in the residual that are greater than fluxlimit, and store the values in the pixelsValue Matrix, and their positions in the pixelPos Matrix.

Increases the size of the output matrices as necessary, but does not decrease them. So the actual number of "active" pixels is returned. This will always be less than (or equal to) the matrix size.

void casa::ClarkCleanModel::doMinorIterations	(	Array< Float > &	model,
		Matrix< Float > &	pixelValue,
		const Matrix< Int > &	pixelPos,
		const Int	numPix,
		Matrix< Float > &	psfPatch,
		Float	fluxLimit,
		uInt &	numberIterations,
		Float	Fmn,
		const uInt	totalIterations,
		Float &	totalflux
	)		`[private]`

Do all the minor iterations for one major cycle. Cleaning stops when the flux or iteration limit is reached.

virtual Bool casa::ClarkCleanModel::getChoose ( ) [virtual]

virtual Float casa::ClarkCleanModel::getCycleSpeedup ( ) [virtual]

virtual uInt casa::ClarkCleanModel::getHistLength ( ) [virtual]

virtual uInt casa::ClarkCleanModel::getInitialNumberIterations ( ) [virtual]

void casa::ClarkCleanModel::getMask ( Array< Float > & mask ) const

virtual Float casa::ClarkCleanModel::getMaxExtPsf ( ) [virtual]

virtual uInt casa::ClarkCleanModel::getMaxNumberMajorCycles ( ) [virtual]

virtual uInt casa::ClarkCleanModel::getMaxNumberMinorIterations ( ) [virtual]

virtual uInt casa::ClarkCleanModel::getMaxNumPix ( ) [virtual]

void casa::ClarkCleanModel::getModel ( Array< Float > & model ) const [virtual]

Reimplemented from casa::ArrayModel< Float >.

virtual ClarkCleanProgress& casa::ClarkCleanModel::getProgress ( ) [inline, virtual]

Definition at line 172 of file ClarkCleanModel.h.

References itsProgressPtr.

Float casa::ClarkCleanModel::getPsfPatch	(	Array< Float > &	psfPatch,
		ConvolutionEquation &	eqn
	)		`[private]`

Work out the size of the Psf patch to use.

virtual IPosition casa::ClarkCleanModel::getPsfPatchSize ( ) [virtual]

virtual Float casa::ClarkCleanModel::getSpeedup ( ) [virtual]

Float casa::ClarkCleanModel::maxResidual ( const Array< Float > & residual ) [private]

The maximum residual is the absolute maximum.

void casa::ClarkCleanModel::maxVect	(	Vector< Float > &	maxVal,
		Float &	absVal,
		Int &	offset,
		const Matrix< Float > &	pixVal,
		const Int	numPix
	)		`[private]`

virtual void casa::ClarkCleanModel::setChoose ( const Bool askForChoice ) [virtual]

The user can be asked whether to stop after every minor cycle.

virtual void casa::ClarkCleanModel::setCycleSpeedup ( const Float speedup ) [virtual]

The structure of various AIPS++ algorithms creates major cycles around the Clark Clean (or other deconvolution algrithms.

The cycleSpeedup parameter causes the threshold to edge up as thresh = thresh_0 * 2^( iter/cycleSpeedup ); ignored if cycleSpeedup <= 0.

virtual void casa::ClarkCleanModel::setHistLength ( const uInt HistBins ) [virtual]

Set the size of the histogram used to determine how many pixels are "active" in a minor iteration.

Default value is 1000 is OK for everything except very small cleans.

virtual void casa::ClarkCleanModel::setInitialNumberIterations ( const uInt initialNumberIterations ) [virtual]

Set and get the initial number of iterations.

void casa::ClarkCleanModel::setMask ( const Array< Float > & mask )

void casa::ClarkCleanModel::setMask ( Array< Float > & mask )

virtual void casa::ClarkCleanModel::setMaxExtPsf ( const Float maxExtPsf ) [virtual]

Set the maximum exterior psf value.

This is used to determine when to stop the minor itartions. This is normally determined from the psf and the number set here is only used if this cannot be determined. The default is zero.

virtual void casa::ClarkCleanModel::setMaxNumberMajorCycles ( const uInt maxNumMajorCycles ) [virtual]

Set the maximum number of major cycles to perform.

virtual void casa::ClarkCleanModel::setMaxNumberMinorIterations ( const uInt maxNumMinorIterations ) [virtual]

Set the maximum number of minor iterations to perform for each major cycle.

virtual void casa::ClarkCleanModel::setMaxNumPix ( const uInt maxNumPix ) [virtual]

Set the maximum number of active pixels to use in the minor iterations.

The specified number can be exceeded if the topmost bin of the histogram contains more pixels than specified here. The default is 10,000 which is suitable for images of 512by512 pixels. Reduce this for smaller images and increase it for larger ones.

void casa::ClarkCleanModel::setModel ( const Array< Float > & model ) [virtual]

Set the current model.

Reimplemented from casa::ArrayModel< Float >.

void casa::ClarkCleanModel::setModel ( Array< Float > & model ) [virtual]

Reimplemented from casa::ArrayModel< Float >.

virtual void casa::ClarkCleanModel::setProgress ( ClarkCleanProgress & ccp ) [inline, virtual]

Set/get the progress display.

Definition at line 171 of file ClarkCleanModel.h.

References itsProgressPtr.

virtual void casa::ClarkCleanModel::setPsfPatchSize ( const IPosition & psfPatchSize ) [virtual]

These functions set various "knobs" that the user can tweak and are specific to the Clark clean algorithm.

The more generic parameters ie. clean gain, and maximum residual fluxlimit, are set using functions in the Iterate base class. The get functions return the value that was actually used after the cleaning was done.

set the size of the PSF used in the minor iterations. If not set it defaults to the largest useful Psf (ie. min(modelSize*2, psfSize))

virtual void casa::ClarkCleanModel::setSpeedup ( const Float speedup ) [virtual]

An exponent on the F(m,n) factor (see Clark[1980]) which influences how quickly active pixels are treated as unreliable.

Larger values mean more major iterations. The default is zero. I have no experience on when to use this factor.

Bool casa::ClarkCleanModel::singleSolve	(	ConvolutionEquation &	eqn,
		Array< Float > &	residual
	)

Bool casa::ClarkCleanModel::solve ( ConvolutionEquation & eqn )

Using a Clark clean deconvolution proceedure solve for an improved estimate of the deconvolved object.

The convolution/residual equation contains the psf and dirty image. When called with a ResidualEquation arguement a quite general interface is used that is slow. The convolution equation contains functions that speed things up. The functions return False if the deconvolution could not be done.

Bool casa::ClarkCleanModel::stopnow ( ) [private]

void casa::ClarkCleanModel::subtractComponent	(	Matrix< Float > &	pixVal,
		const Matrix< Int > &	pixPos,
		const Int	numPix,
		const Vector< Float > &	maxVal,
		const Vector< Int > &	maxPos,
		const Matrix< Float > &	psf
	)		`[private]`