Base class for ImageKmeansModelEstimator object. More...

#include <itkImageKmeansModelEstimator.h>

Inheritance diagram for itk::ImageKmeansModelEstimator< TInputImage, TMembershipFunction >:

Collaboration diagram for itk::ImageKmeansModelEstimator< TInputImage, TMembershipFunction >:

Public Types
typedef vnl_matrix< double >	CodebookMatrixOfDoubleType
typedef vnl_matrix< int >	CodebookMatrixOfIntegerType
typedef SmartPointer< const Self >	ConstPointer
typedef ImageRegionConstIterator < TInputImage >	InputImageConstIterator
typedef TInputImage::ConstPointer	InputImageConstPointer
typedef ImageRegionIterator < TInputImage >	InputImageIterator
typedef TInputImage::PixelType	InputImagePixelType
typedef TInputImage::Pointer	InputImagePointer
typedef TInputImage	InputImageType
typedef TInputImage::PixelType::VectorType	InputImageVectorType
typedef TMembershipFunction::Pointer	MembershipFunctionPointer
typedef SmartPointer< Self >	Pointer
typedef ImageKmeansModelEstimator	Self
typedef ImageModelEstimatorBase < TInputImage, TMembershipFunction >	Superclass
Public Member Functions
virtual ::itk::LightObject::Pointer	CreateAnother (void) const
virtual CodebookMatrixOfDoubleType	GetCodebook () const
virtual int	GetMaxSplitAttempts () const
virtual const char *	GetNameOfClass () const
virtual double	GetOffsetAdd () const
virtual double	GetOffsetMultiply () const
CodebookMatrixOfDoubleType	GetOutCodebook ()
virtual double	GetThreshold () const
void	SetCodebook (CodebookMatrixOfDoubleType InCodebook)
virtual void	SetMaxSplitAttempts (int _arg)
virtual void	SetOffsetAdd (double _arg)
virtual void	SetOffsetMultiply (double _arg)
virtual void	SetThreshold (double _arg)
Static Public Member Functions
static Pointer	New ()
Protected Member Functions
void	Allocate ()
void	GenerateData ()
void	PrintKmeansAlgorithmResults ()
Private Types
typedef TInputImage::SizeType	ImageSizeType
typedef TInputImage::PixelType::VectorType	InputPixelVectorType
Private Member Functions
void	EstimateKmeansModelParameters ()
virtual void	EstimateModels ()
	ImageKmeansModelEstimator (const Self &)
void	NearestNeighborSearchBasic (double *distortion)
void	operator= (const Self &)
void	Perturb (double oldCodeword, int scale, double newCodeword)
void	Reallocate (int oldSize, int newSize)
void	SplitCodewords (int currentSize, int numDesired, int scale)
int	WithCodebookUseGLA ()
int	WithoutCodebookUseLBG ()
Private Attributes
CodebookMatrixOfDoubleType	m_Centroid
CodebookMatrixOfDoubleType	m_Codebook
CodebookMatrixOfDoubleType	m_CodewordDistortion
CodebookMatrixOfIntegerType	m_CodewordHistogram
SizeValueType	m_CurrentNumberOfCodewords
double	m_DoubleMaximum
int	m_MaxSplitAttempts
SizeValueType	m_NumberOfCodewords
double	m_OffsetAdd
double	m_OffsetMultiply
double	m_OutputDistortion
int	m_OutputNumberOfEmptyCells
double	m_Threshold
bool	m_ValidInCodebook
SizeValueType	m_VectorDimension
CodebookMatrixOfDoubleType	GetKmeansResults (void)
	ImageKmeansModelEstimator ()
	~ImageKmeansModelEstimator ()
virtual void	PrintSelf (std::ostream &os, Indent indent) const

Detailed Description

template<class TInputImage, class TMembershipFunction>
class itk::ImageKmeansModelEstimator< TInputImage, TMembershipFunction >

Base class for ImageKmeansModelEstimator object.

itkImageKmeansModelEstimator generates the kmeans model (cluster centers). This object performs clustering of data sets into different clusters, either using user-provided seed points as an initial guess or generating the clusters using a recursive approach when the user provides the number of desired clusters. Each cluster is represented by its cluster center. The two algorithms used are the generalized Lloyd algorithm (GLA) and the Linde-Buzo-Gray algorithm. The cluster centers are also referred to as codewords and a table of cluster centers is referred as a codebook.

As required by the GLA algorithm, the initial seed cluster should contain approximate centers of clusters. The GLA algorithm genrates updated cluster centers that result in a lower distortion than the input seed cluster when the input vectors are mapped/classified/labelled using the given codebooks.

If no codebook is provided, the Linde-Buzo-Gray algorithm is used. This algorithm uses the GLA algorithm at its core to generate the centroids of the input vectors (data). However, since there is no initial codebook, LBG first creates a one word codebook (or centroid of one cluster comprising of all the input training vectors). The LBG uses codeword or centroid splitting to create an increasing number of clusters. Each new set of clusters are optimized using the GLA algorithm. The number of clusters increases as $2^{n}$ n= 0, 1, ... The codebook is expected to be in the form of a vnl matrix, where there are N rows, each row representing the cluster mean of a given cluster. The number of columns in the codebook should be equal to the input image vector dimension.

The threshold parameter controls the ''optimality'' of the returned codebook, where optimality is related to the least possible mean-squared error distortion that can be found by the algorithm. For larger thresholds, the result will be less optimal. For smaller thresholds, the result will be more optimal. If a more optimal result is desired, then the algorithm will take longer to complete. A reasonable threshold value is 0.01.

If, during the operation of the algorithm, there are any unused clusters or cells, the m_OffsetAdd and m_OffsetMultiply parameters are used to split the cells with the highest distortion. This function will attempt to fill empty cells up to 10 times (unless the overall distortion is zero). Using 0.01 is a reasonable default values for the m_OffsetAdd and m_OffsetMultiply parameters.

If the GLA is unable to resolve the data into the desired number of clusters or cells, only the codewords which were used will be returned.

In terms of clustering, codewords are cluster centers, and a codebook is a table containing all cluster centers. The GLA produces results that are equivalent to the K-means clustering algorithm.

For more information about the algorithms, see A. Gersho and R. M. Gray, {Vector Quantization and Signal Compression}, Kluwer Academic Publishers, Boston, MA, 1992.

This object supports data handling of multiband images. The object accepts the input image in vector format only, where each pixel is a vector and each element of the vector corresponds to an entry from 1 particular band of a multiband dataset. A single band image is treated as a vector image with a single element for every vector.

This function is templated over the type of input image. In addition, a second parameter for the MembershipFunction needs to be specified. In this case a Membership function that store cluster centroids models needs to be specified.

The Update() function enables the calculation of the various models, creates the membership function objects and populates them.

Note: There is a second implementation of k-means algorithm in ITK under the itk::statistics namespace. While this algorithm (GLA/LBG based algorithm) is memory efficient, the other algorithm is time efficient.

See also:: KdTreeBasedKmeansEstimator, WeightedCentroidKdTreeGenerator, KdTree; ScalarImageKmeansImageFilter

Wiki Examples:

All Examples

Compute kmeans clusters of pixels in an image

Definition at line 131 of file itkImageKmeansModelEstimator.h.