WDSS-IIw2algs: kmeans::KMeansSegmenter< T


Public Types
typedef Cluster< SingleFeature >	RegionProperty
	The simplest region property is a cluster of a single feature.
typedef std::vector< int >	Replacement
	Given a region with label i, Replacement[i] provides the new label, usually at a particular scale.
Public Member Functions
virtual LImage< int >	segment (const LImage< int > &orig)
virtual	~KMeansSegmenter ()
	KMeansSegmenter (int min_val=0, int max_val=100, size_t number_of_clusters=16)
	Pass in the range of values to consider in thresholding.
void	setOutputImageDir (const std::string &newDir)
	Pass in an empty string to turn off output.
void	writeGraph (const std::string &filename) const
	Write out a DOT output file (in a form that the Graphviz utility dot can read).
LImage< int >	getScale (size_t scale, bool rawlabels=false) const
	The different scales.
const Replacement &	getReplacement (size_t scale) const
	Gets the replacement at different scales.
size_t	getNumScales () const
	Returns the total number of scales in the segmentation.
Static Public Member Functions
static void	setMarkovWeight (float lambda)
	Set the weight of the Markov energy relative to the Discontiguity measure.
static void	computeRegionProperty (const LImage< int > &orig, const LImage< int > &label, vector< RegionProperty > *fillme)
	Fills in the passed in vector with a vector of computed region properties.
static void	computeRegionProperty (const LImage< int > &label, vector< RegionProperty > *fillme)
	Fills in the passed in vector with a vector of computed region properties.
Protected Member Functions
LImage< int >	doKMeans (const ImageAttr< T > &attr, LImage< int > label, vector< Cluster< T > > clusters, int iter_no=0) const
	Do K means segmentation, filling in the cluster array.
LImage< int >	initialize (const LImage< int > &orig, const ImageAttr< T > &attr, vector< Cluster< T > > *clusters) const
	Get an initial cut of clusters, somehow.
void	computeClusterAttr (const ImageAttr< T > &attr, const LImage< int > &label, vector< Cluster< T > > *clusters, bool has_K_levels=true) const
	Compute cluster attributes from the image attribute and labeling.
int	findBestCluster (const T &current, const vector< Cluster< T > > &clusters, const LImage< int > &labels, int x, int y) const
	Return label corresponding to best cluster available.
float	computeMeasureDistance (const Cluster< T > &cluster, const T &current, int x, int y) const
	Way to compute distance between a point (value,x,y) and a cluster's mean.
float	computeMeasureDistance (const T &other, const T &current) const
	Finds distance between two statistics.
float	computeDiscontiguityDistance (const LImage< int > &label, int x, int y, int candidate_label) const
	Way to compute discontiguity if the current label were to be changed to the candidate.
virtual LImage< int >	computeDetailImage (const LImage< int > &label, const ImageAttr< T > &attr, std::vector< Replacement > *replacements_list)
	This function may be over-ridden to provide a different way of selecting among the regions produced by K-means segmentation.
virtual void	computeReplacements (const LImage< int > &label, const ImageAttr< T > &attr, std::vector< Replacement > *replacements_list)
	This function may be over-ridden to provide a different way of choosing how to combine/replace regions at higher levels.
Static Protected Member Functions
static void	expand_segment (int x, int y, const LImage< int > &orig, LImage< int > &segimg, int label)
	Set the segimg at x,y to be label and then call this function.
static LImage< int >	relabel (const LImage< int > &label_img)
	Will relabel the image, removing unused labels.
static LImage< int >	pruneRegions (const LImage< int > &label, RegionPruner &)
	This method removes all regions that meet certain criteria.
Static Protected Attributes
static const int	Unset = -1000
static const int	Invalid = -2000
Friends
class	Pruner
Classes
class	Pruner
	Controls pruning of regions based on size. More...

Member Typedef Documentation

typedef Cluster< SingleFeature > kmeans::Segmenter::RegionProperty [inherited]

The simplest region property is a cluster of a single feature.

typedef std::vector<int> kmeans::Segmenter::Replacement [inherited]

Given a region with label i, Replacement[i] provides the new label, usually at a particular scale.

The Replacement structure is such that replacement[i] provides the replacement label for the current label i. This replacement may be 0 or negative, which means that no replacement exists.

Reimplemented in kmeans::MergingKMSegmenter< T >.

Constructor & Destructor Documentation

template<class T>

virtual kmeans::KMeansSegmenter< T >::~KMeansSegmenter ( ) [inline, virtual]

template<class T>

kmeans::KMeansSegmenter< T >::KMeansSegmenter	(	int	min_val = `0`,
		int	max_val = `100`,
		size_t	number_of_clusters = `16`
	)

Pass in the range of values to consider in thresholding.

Member Function Documentation

template<class T>

void kmeans::KMeansSegmenter< T >::computeClusterAttr	(	const ImageAttr< T > &	attr,
		const LImage< int > &	label,
		vector< Cluster< T > > *	clusters,
		bool	has_K_levels = `true`
	)			const `[protected]`

Compute cluster attributes from the image attribute and labeling.

During K-means, the computation is always done with K levels, but pass in false if you are using this after K-means and relabeling.

template<class T>

virtual LImage<int> kmeans::KMeansSegmenter< T >::computeDetailImage	(	const LImage< int > &	label,
		const ImageAttr< T > &	attr,
		std::vector< Replacement > *	replacements_list
	)			`[protected, virtual]`

This function may be over-ridden to provide a different way of selecting among the regions produced by K-means segmentation.

Parameters:

	label	The image with the K-means segmentation done and labels computed based on 8-neighborhood, but no selection among the regions performed.
	attr	The attributes computed at every pixel of the image to be segmented.
	replacements_list	The list of replacements; the first element (most detail) will be filled in here.

Returns:: The most detail image will be returned. Caller should call relabel() on it in order to get a segmented image.

template<class T>

float kmeans::KMeansSegmenter< T >::computeDiscontiguityDistance	(	const LImage< int > &	label,
		int	x,
		int	y,
		int	candidate_label
	)			const `[protected]`

Way to compute discontiguity if the current label were to be changed to the candidate.

Returns a number between 0 and 1.

template<class T>

float kmeans::KMeansSegmenter< T >::computeMeasureDistance	(	const T &	other,
		const T &	current
	)			const `[protected]`

Finds distance between two statistics.

template<class T>

float kmeans::KMeansSegmenter< T >::computeMeasureDistance	(	const Cluster< T > &	cluster,
		const T &	current,
		int	x,
		int	y
	)			const `[protected]`

Way to compute distance between a point (value,x,y) and a cluster's mean.

The base class does not use the location information, but derived classes may want to. Between 0 and 1.

static void kmeans::Segmenter::computeRegionProperty	(	const LImage< int > &	label,
		vector< RegionProperty > *	fillme
	)			`[static, inherited]`

Fills in the passed in vector with a vector of computed region properties.

The input label image should be a labeled image such as that returned by the segmenter. WARNING: The resulting properties will not have any of the value elements set because the original image was not passed in.

static void kmeans::Segmenter::computeRegionProperty	(	const LImage< int > &	orig,
		const LImage< int > &	label,
		vector< RegionProperty > *	fillme
	)			`[static, inherited]`

Fills in the passed in vector with a vector of computed region properties.

The input label image should be a labeled image such as that returned by the segmenter.

template<class T>

virtual void kmeans::KMeansSegmenter< T >::computeReplacements	(	const LImage< int > &	label,
		const ImageAttr< T > &	attr,
		std::vector< Replacement > *	replacements_list
	)			`[protected, virtual]`

This function may be over-ridden to provide a different way of choosing how to combine/replace regions at higher levels.

Parameters:

	label	The image with the K-means segmentation done and labels computed based on 8-neighborhood, but no selection among the regions performed.
	attr	The attributes computed at every pixel of the image to be segmented.
	replacements_list	The list of replacements; the first element (most detail) should be filled already. Coarser detail will be filled in here.

template<class T>

LImage<int> kmeans::KMeansSegmenter< T >::doKMeans	(	const ImageAttr< T > &	attr,
		LImage< int > *	label,
		vector< Cluster< T > > *	clusters,
		int	iter_no = `0`
	)			const `[protected]`

Do K means segmentation, filling in the cluster array.

static void kmeans::Segmenter::expand_segment	(	int	x,
		int	y,
		const LImage< int > &	orig,
		LImage< int > &	segimg,
		int	label
	)			`[static, protected, inherited]`

Set the segimg at x,y to be label and then call this function.

It will expand out segimg to all contiguous values that match

template<class T>

int kmeans::KMeansSegmenter< T >::findBestCluster	(	const T &	current,
		const vector< Cluster< T > > &	clusters,
		const LImage< int > &	labels,
		int	x,
		int	y
	)			const `[protected]`

Return label corresponding to best cluster available.

template<class T>

size_t kmeans::KMeansSegmenter< T >::getNumScales ( ) const

Returns the total number of scales in the segmentation.

The returned values correspond to the previous call to segment()

template<class T>

const Replacement& kmeans::KMeansSegmenter< T >::getReplacement ( size_t scale ) const

Gets the replacement at different scales.

0 is the most detail, 1 is less and so on. If the scale passed in exceeeds the number of scales in the segmentation, an invalid Replacement ( .size() == 0 ) is returned.

The returned values correspond to the previous call to segment()

template<class T>

LImage<int> kmeans::KMeansSegmenter< T >::getScale	(	size_t	scale,
		bool	rawlabels = `false`
	)			const

The different scales.

The returned values correspond to the previous call to segment()

Parameters:

	scale	The scale to return. 0 is the most detail, 1 is less ...
	rawlabels	If true, then no relabeling of the original image is done. If false, the labeling corresponds to that of raw_segimg.

Returns:: image at scale. If the scale passed in exceeeds the number of scales in the segmentation, an invalid image (dim_x()*dim_y() <= 0) is returned.

template<class T>

LImage<int> kmeans::KMeansSegmenter< T >::initialize	(	const LImage< int > &	orig,
		const ImageAttr< T > &	attr,
		vector< Cluster< T > > *	clusters
	)			const `[protected]`

Get an initial cut of clusters, somehow.

The implementation here is to divide the measurement space (max-min) into equal parts.

static LImage<int> kmeans::Segmenter::pruneRegions	(	const LImage< int > &	label,
		RegionPruner &
	)			`[static, protected, inherited]`

This method removes all regions that meet certain criteria.

The function object RegionPruner should be provided to target the appropriate regions.

Usually, you'd want to relabel the image that you get after pruning.

static LImage<int> kmeans::Segmenter::relabel ( const LImage< int > & label_img ) [static, protected, inherited]

Will relabel the image, removing unused labels.

template<class T>

virtual LImage<int> kmeans::KMeansSegmenter< T >::segment ( const LImage< int > & orig ) [virtual]

Implements kmeans::Segmenter.

template<class T>

static void kmeans::KMeansSegmenter< T >::setMarkovWeight ( float lambda ) [static]

Set the weight of the Markov energy relative to the Discontiguity measure.

The default value is 0.6 and should be left alone unless you know what you are doing. If wish to change it, provide any number in the range [0,1].

This is more that "just" a static function -- you are setting it for all values of the template parameter T as well!

template<class T>

void kmeans::KMeansSegmenter< T >::setOutputImageDir ( const std::string & newDir )

Pass in an empty string to turn off output.

template<class T>

void kmeans::KMeansSegmenter< T >::writeGraph ( const std::string & filename ) const

Write out a DOT output file (in a form that the Graphviz utility dot can read).

This is a directed graph.

Friends And Related Function Documentation

template<class T>

friend class Pruner [friend]

Member Data Documentation

const int kmeans::Segmenter::Invalid = -2000 [static, protected, inherited]

const int kmeans::Segmenter::Unset = -1000 [static, protected, inherited]

kmeans::KMeansSegmenter< T > Class Template Reference

Detailed Description

template<class T>
class kmeans::KMeansSegmenter< T >

Public Types

Public Member Functions

Static Public Member Functions

Protected Member Functions

Static Protected Member Functions

Static Protected Attributes

Friends

Classes

Member Typedef Documentation

Constructor & Destructor Documentation

Member Function Documentation

Friends And Related Function Documentation

Member Data Documentation

kmeans::KMeansSegmenter< T > Class Template Reference

Detailed Description

template<class T> class kmeans::KMeansSegmenter< T >

Public Types

Public Member Functions

Static Public Member Functions

Protected Member Functions

Static Protected Member Functions

Static Protected Attributes

Friends

Classes

Member Typedef Documentation

Constructor & Destructor Documentation

Member Function Documentation

Friends And Related Function Documentation

Member Data Documentation

template<class T>
class kmeans::KMeansSegmenter< T >