max-normalization algorithms

Functions
template<class T >
Image< T >	maxNormalize (const Image< T > &src, const T mi, const T ma, const MaxNormType normtyp=VCXNORM_DEFAULT, int nbiter=FANCYITER, const Image< float > *lrexcit=0)
	Generic implementation (select algorithm with normtyp).
template<class T >
Image< T >	maxNormalizeNone (const Image< T > &src, const T mi, const T ma)
	Specialized implementation (corresponds to VCXNORM_NONE).
template<class T >
Image< T >	maxNormalizeStd (const Image< T > &src, const T mi, const T ma)
	Specialized implementation (corresponds to VCXNORM_MAXNORM).
template<class T >
Image< T >	maxNormalizeFancyFast (const Image< T > &src, const T mi, const T ma, const int nbiter=FANCYITER)
	Specialized implementation (corresponds to VCXNORM_FANCYFAST).
template<class T >
Image< T >	maxNormalizeFancy (const Image< T > &src, const T mi, const T ma, const int nbiter=FANCYITER, const double weakness=1.0, const Image< float > *lrexcit=0)
	Specialized implementation (corresponds to VCXNORM_FANCY).
template<class T >
Image< T >	maxNormalizeFancyLandmark (const Image< T > &src, const T mi, const T ma, const int nbiter=FANCYITER)
	Specialized implementation (corresponds to VCXNORM_FANCY).
template<class T >
Image< T >	maxNormalizeLandmark (const Image< T > &src, const T mi, const T ma)
	Specialized implementation (corresponds to VCXNORM_LANDMARK).
template<class T >
int	findPeaks (const Image< T > &src, const T mi, const T ma, double &sum)
	to find number of peaks in the image (used in landmark detection)
template<class T >
float	goodness_map (const Image< T > &src)
	to find the goodness of a map (used in landmark detection)
template<class T >
Image< T >	maxNormalizeStdev (const Image< T > &src)
	Specialized implementation (corresponds to VCXNORM_STDEV).
template<class T >
Image< T >	maxNormalizeStdev0 (const Image< T > &src)
	Specialized implementation (corresponds to VCXNORM_STDEV0).
Types of normalization (normtyp)

enum	MaxNormType { VCXNORM_NONE = 0, VCXNORM_MAXNORM = 1, VCXNORM_FANCY = 2, VCXNORM_FANCYFAST = 3, VCXNORM_FANCYONE = 4, VCXNORM_FANCYLANDMARK = 5, VCXNORM_LANDMARK = 6, VCXNORM_FANCYWEAK = 7, VCXNORM_IGNORE = 8, VCXNORM_SURPRISE = 9, VCXNORM_FANCYVWEAK = 10, VCXNORM_STDEV = 11, VCXNORM_STDEV0 = 12 }
const char *	maxNormTypeName (const MaxNormType m)
	Get a name in clear for a given type.
#define	NBMAXNORMTYPES 13
#define	VCXNORM_DEFAULT VCXNORM_FANCY
Additional parameters for fancynorm versions of the algorithm.

const int	FANCYITER = 5
	default number of iterations
const double	FANCYESIG = 2
	excitatory sigma as % of image size
const double	FANCYISIG = 25
	inhibitory sigma as % of image size
const double	FANCYCOEX = 0.5
	excitatory coefficient (strength)
const double	FANCYCOIN = 1.5
	inhibitory coefficient (strength)
const double	FANCYINHI = 2.0
	strength of global inhibition
const double	FANCYG = 2.1
	for sigmoid normalization
const double	FANCYH = 2.0
	for sigmoid normalization
const double	FANCYS = 1.0
	for sigmoid normalization
const int	LRLEVEL = 2
	for tuned long-range excitation:
Standard min/max bounds for maxNormalize().

const float	MAXNORMMAX = 10.0f
	upper bound for maxNormalize()
const float	MAXNORMMIN = 0.0f
	lower bound for maxNormalize()
const float	MAXNORMLANDMARK = 255.0f
	upper bound for maxNormalizeFancyLandmark()

Detailed Description

Define Documentation

#define NBMAXNORMTYPES 13

Have this value always equal to the largest of the VCXNORMs (used for range checking)

Definition at line 70 of file fancynorm.H.

Referenced by convertFromString(), and maxNormTypeName().

#define VCXNORM_DEFAULT VCXNORM_FANCY

Always points to the default maxnorm operation

Definition at line 88 of file fancynorm.H.

Enumeration Type Documentation

enum MaxNormType

Enumerator:

VCXNORM_NONE	no max-normalization, but may change range
VCXNORM_MAXNORM	non-iterative maxnorm
VCXNORM_FANCY	full implementation of fancy maxnorm
VCXNORM_FANCYFAST	fast implementation of fancy maxnorm
VCXNORM_FANCYONE	one-iteration only of fancy maxnorm
VCXNORM_FANCYLANDMARK	fancy maxnorm to find landmark
VCXNORM_LANDMARK	to find landmarks
VCXNORM_FANCYWEAK	Weak version of FancyOne.
VCXNORM_IGNORE	Do absolutely nothing, not even change range.
VCXNORM_SURPRISE	Like ignore but VCX will do something.
VCXNORM_FANCYVWEAK	Very weak version of FancyOne.
VCXNORM_STDEV	normalize map to have stdev=1, minval=0
VCXNORM_STDEV0	normalize map to have stdev=1, mean=0

Definition at line 51 of file fancynorm.H.

Function Documentation

template<class T >

int findPeaks	(	const Image< T > &	src,
		const T	mi,
		const T	ma,
		double &	sum
	)			`[inline]`

to find number of peaks in the image (used in landmark detection)

Definition at line 484 of file fancynorm.C.

References area(), findMax(), max(), maxNormalize(), segmentLandmark(), and VCXNORM_FANCY.

Referenced by goodness_map(), maxNormalizeLandmark(), SimulationViewerStats::save1(), and SimulationViewerNerdCam::writeStatusPage().

template<class T >

float goodness_map ( const Image< T > & src ) [inline]

to find the goodness of a map (used in landmark detection)

Definition at line 518 of file fancynorm.C.

References ASSERT, findMax(), findPeaks(), goodness_map(), Image< T >::initialized(), inplaceNormalize(), max(), MAXNORMLANDMARK, MAXNORMMIN, and segmentLandmark().

Referenced by ComplexChannel::combineOutputs(), SingleChannel::combineSubMaps(), goodness_map(), and maxNormalizeLandmark().

template<class T >

Image<T> maxNormalize	(	const Image< T > &	src,
		const T	mi,
		const T	ma,
		const MaxNormType	normtyp = `VCXNORM_DEFAULT`,
		int	nbiter = `FANCYITER`,
		const Image< float > *	lrexcit = `0`
	)			`[inline]`

Generic implementation (select algorithm with normtyp).

Normalize between mi and ma and multiply by (ma - mean). Versions with more arguments implement the core within-feature spatial competition for salience. See papers for details.

Definition at line 63 of file fancynorm.C.

References getMinMax(), maxNormalizeFancy(), maxNormalizeFancyFast(), maxNormalizeFancyLandmark(), maxNormalizeLandmark(), maxNormalizeNone(), maxNormalizeStd(), maxNormalizeStdev(), maxNormalizeStdev0(), VCXNORM_FANCY, VCXNORM_FANCYFAST, VCXNORM_FANCYLANDMARK, VCXNORM_FANCYONE, VCXNORM_FANCYVWEAK, VCXNORM_FANCYWEAK, VCXNORM_IGNORE, VCXNORM_LANDMARK, VCXNORM_MAXNORM, VCXNORM_NONE, VCXNORM_STDEV, VCXNORM_STDEV0, and VCXNORM_SURPRISE.

Referenced by DirectFeedChannel::applyMaxNorm(), ComplexChannel::combineOutputs(), MultiDirectFeedChannel::combineOutputs(), SoxChannel::combineOutputs(), HueChannel::combineSubMaps(), SingleChannel::combineSubMaps(), SaliencyMT::computeCMAP(), BeoSubSaliency::computeCMAP(), computeCMAP(), ObjDetChannel::doInput(), ContourChannel::doInput(), findPeaks(), SpectralResidualChannel::getOutput(), SingleChannel::getRawCSmap(), DirectionSpatioTemporalChannel::getRawCSmap(), main(), normalizMAP(), SingleChannel::postProcessMap(), RawVisualCortex::postProcessOutputMap(), VisualCortexSurprise::postProcessOutputMap(), BeobotVisualCortex::process(), BeobotVisualCortex::processEnd(), RegSaliency::runSaliency(), and BeobotBrainMT::threadCompute().

template<class T >

Image<T> maxNormalizeFancy	(	const Image< T > &	src,
		const T	mi,
		const T	ma,
		const int	nbiter = `FANCYITER`,
		const double	weakness = `1.0`,
		const Image< float > *	lrexcit = `0`
	)			`[inline]`

Specialized implementation (corresponds to VCXNORM_FANCY).

fancyNorm from Itti et al, JEI, 2001; FULL implementation.

Definition at line 265 of file fancynorm.C.

References ASSERT, CONV_BOUNDARY_CLEAN, CONV_BOUNDARY_ZERO, convolve(), downSize(), FANCYCOEX, FANCYCOIN, FANCYESIG, FANCYINHI, FANCYISIG, Image< T >::getHeight(), getMinMax(), Image< T >::getWidth(), Image< T >::initialized(), inplaceClamp(), inplaceNormalize(), inplaceRectify(), LRLEVEL, max(), min(), rescale(), sepFilter(), and sqrt().

Referenced by maxNormalize().

template<class T >

Image<T> maxNormalizeFancyFast	(	const Image< T > &	src,
		const T	mi,
		const T	ma,
		const int	nbiter = `FANCYITER`
	)			`[inline]`

Specialized implementation (corresponds to VCXNORM_FANCYFAST).

fancyNorm from Itti et al, JEI, 2001; FAST implementation.

Definition at line 206 of file fancynorm.C.

References ASSERT, buildPyrGaussian(), FANCYCOEX, FANCYCOIN, FANCYESIG, FANCYINHI, FANCYISIG, Image< T >::getHeight(), getMinMax(), Image< T >::getWidth(), Image< T >::initialized(), inplaceAttenuateBorders(), inplaceNormalize(), inplaceRectify(), log(), max(), rescale(), and sqrt().

Referenced by maxNormalize().

template<class T >

Image<T> maxNormalizeFancyLandmark	(	const Image< T > &	src,
		const T	mi,
		const T	ma,
		const int	nbiter = `FANCYITER`
	)			`[inline]`

Specialized implementation (corresponds to VCXNORM_FANCY).

fancyNorm from Itti et al, JEI, 2001; FULL implementation. adapted to find landmarks

Definition at line 331 of file fancynorm.C.

References ASSERT, CONV_BOUNDARY_CLEAN, FANCYCOEX, FANCYCOIN, FANCYINHI, Image< T >::getHeight(), getMinMax(), Image< T >::getWidth(), Image< T >::initialized(), inplaceNormalize(), inplaceRectify(), max(), min(), sepFilter(), and sqrt().

Referenced by maxNormalize().

template<class T >

Image<T> maxNormalizeLandmark	(	const Image< T > &	src,
		const T	mi,
		const T	ma
	)			`[inline]`

Specialized implementation (corresponds to VCXNORM_LANDMARK).

to find landmarks

Definition at line 382 of file fancynorm.C.

References ASSERT, findMax(), findPeaks(), Image< T >::getDims(), Image< T >::getVal(), goodness_map(), Image< T >::initialized(), inplaceNormalize(), max(), MAXNORMLANDMARK, MAXNORMMIN, Image< T >::resize(), and segmentLandmark().

Referenced by maxNormalize().

template<class T >

Image<T> maxNormalizeNone	(	const Image< T > &	src,
		const T	mi,
		const T	ma
	)			`[inline]`

Specialized implementation (corresponds to VCXNORM_NONE).

No max-normalization, just normalize between mi and ma.

Definition at line 130 of file fancynorm.C.

References inplaceNormalize(), and inplaceRectify().

Referenced by maxNormalize().

template<class T >

Image<T> maxNormalizeStd	(	const Image< T > &	src,
		const T	mi,
		const T	ma
	)			`[inline]`

Specialized implementation (corresponds to VCXNORM_MAXNORM).

maxNorm from Itti et al, IEEE-PAMI, 1998.

Definition at line 147 of file fancynorm.C.

References ASSERT, Image< T >::getHeight(), Image< T >::getVal(), Image< T >::getWidth(), Image< T >::initialized(), inplaceNormalize(), and inplaceRectify().

Referenced by maxNormalize().

template<class T >

Image<T> maxNormalizeStdev ( const Image< T > & src ) [inline]

Specialized implementation (corresponds to VCXNORM_STDEV).

The output image is the result of normalizing the input image to have stdev=1 and minval=0, by simply dividing by the stdev of the original image and then subtracting the minval of the resulting image. We set minval=0, rather than the more natural mean=0, because maxNormalize functions traditionally return images with only non-negative values, since negative values will likely eventually be subject to a rectification. The choice of minval=0 or mean=0 has no effect on the statistics of maps produced by summing several stdev-normalized maps; the resulting map itself will have the same stdev in either case, and the only difference will be a shift in its mean.

Definition at line 586 of file fancynorm.C.

References Image< T >::getDims(), getMinMax(), stdev(), and ZEROS.

Referenced by maxNormalize().

template<class T >

Image<T> maxNormalizeStdev0 ( const Image< T > & src ) [inline]

Specialized implementation (corresponds to VCXNORM_STDEV0).

The output image is the result of normalizing the input image to have mean=0 and stdev=1, by simply subtracting the mean of the original image and then dividing by the stdev of the original image. Beware that the resulting image will have negative values, which may be truncated by later rectification steps.

Definition at line 608 of file fancynorm.C.

References Image< T >::getDims(), mean(), stdev(), and ZEROS.

Referenced by maxNormalize().

const char* maxNormTypeName ( const MaxNormType m ) [inline]

Get a name in clear for a given type.

Definition at line 73 of file fancynorm.H.

References NBMAXNORMTYPES.

Referenced by ComplexChannel::combineOutputs(), IntegerComplexChannel::combineOutputsInt(), HueChannel::combineSubMaps(), SingleChannel::combineSubMaps(), convertFromString(), convertToString(), ObjDetChannel::doInput(), SingleChannel::postProcessMap(), RawVisualCortex::postProcessOutputMap(), VisualCortexSurprise::postProcessOutputMap(), and IntegerRawVisualCortex::postProcessOutputMap().