boofcv.alg.segmentation.slic

Class SegmentSlic<T extends boofcv.struct.image.ImageBase>

java.lang.Object

boofcv.alg.segmentation.slic.SegmentSlic<T>

Direct Known Subclasses:

SegmentSlic_F32, SegmentSlic_PlF32, SegmentSlic_PlU8, SegmentSlic_U8

public abstract class SegmentSlic<T extends boofcv.struct.image.ImageBase>
extends java.lang.Object

K-means based superpixel image segmentation, see [1]. The image is broken up into superpixels (clusters of connected pixels) in a grid like pattern. A connectivity rule of 4 or 8 is enforced across all the clusters. Clustering is done using k-means, where each point is composed on the 2D image coordinate and an intensity value in each color band, thus K = 2+numBands. Instead of computing the distance of each cluster's center from each point only points within a distance of S si considered. The difference in scale difference between pixels and image intensity is handled through a user configurable tuning parameter.

Deviations from paper:

• In the paper a LAB color space is always used. In this implementation a general purpose N-dimensional color space is used.
• To correctly support LAB or other color spaces a specialized implementation might be needed to ensure the intensity of a pixel is computed correctly.
• Small regions are merged into other regions based on how similar their color is. In the paper a small region is merged into the largest region it is connected to.

[1] Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Süsstrunk, SLIC Superpixels, EPFL Technical Report no. 149300, June 2010.

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	SegmentSlic.Cluster The mean in k-means.
static class	SegmentSlic.ClusterDistance Stores how far a cluster is from the specified pixel
static class	SegmentSlic.Pixel K-means clustering information for each pixel.

Field Summary

Fields

Modifier and Type	Field and Description
static int	BORDER
protected org.ddogleg.struct.FastQueue<SegmentSlic.Cluster>	clusters
protected boofcv.struct.ConnectRule	connectRule
protected int	gridInterval
protected boofcv.struct.image.ImageType<T>	imageType
protected T	input
protected org.ddogleg.struct.FastQueue<SegmentSlic.Pixel>	pixels
protected ClusterLabeledImage	segment

Constructor Summary

Constructors

Constructor and Description
SegmentSlic(int numberOfRegions, float m, int totalIterations, boofcv.struct.ConnectRule connectRule, boofcv.struct.image.ImageType<T> imageType)

Method Summary

Modifier and Type	Method and Description
abstract void	addColor(float[] color, int index, float weight) Performs a weighted add to the cluster's color at the specified pixel in the image
void	assignLabelsToPixels(boofcv.struct.image.GrayS32 pixelToRegions, org.ddogleg.struct.GrowQueue_I32 regionMemberCount, org.ddogleg.struct.FastQueue<float[]> regionColor) Selects which region each pixel belongs to based on which cluster it is the closest to
abstract float	colorDistance(float[] color, int index) Euclidean Squared distance away that the pixel is from the provided color
protected void	computeClusterDistance() Computes how far away each cluster is from each pixel.
org.ddogleg.struct.FastQueue<SegmentSlic.Cluster>	getClusters()
boofcv.struct.ConnectRule	getConnectRule()
boofcv.struct.image.ImageType<T>	getImageType()
abstract float	getIntensity(int x, int y) Intensity of the pixel at the specified location
org.ddogleg.struct.GrowQueue_I32	getRegionMemberCount()
protected float	gradient(int x, int y) Computes the gradient at the specified pixel
protected void	initalize(T input) prepares all data structures
protected void	initializeClusters() initialize all the clusters at regularly spaced intervals.
protected void	perturbCenter(SegmentSlic.Cluster c, int x, int y) Set the cluster's center to be the pixel in a 3x3 neighborhood with the smallest gradient
void	process(T input, boofcv.struct.image.GrayS32 output)
abstract void	setColor(float[] color, int x, int y) Sets the cluster's to the pixel color at that location
protected void	updateClusters() Update the value of each cluster using Maximization step.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Field Detail

BORDER

public static final int BORDER

See Also:

Constant Field Values

gridInterval

protected int gridInterval

input

protected T extends boofcv.struct.image.ImageBase input

segment

protected ClusterLabeledImage segment

clusters

protected org.ddogleg.struct.FastQueue<SegmentSlic.Cluster> clusters

pixels

protected org.ddogleg.struct.FastQueue<SegmentSlic.Pixel> pixels

imageType

protected boofcv.struct.image.ImageType<T extends boofcv.struct.image.ImageBase> imageType

connectRule

protected boofcv.struct.ConnectRule connectRule

Constructor Detail

SegmentSlic

public SegmentSlic(int numberOfRegions,
                   float m,
                   int totalIterations,
                   boofcv.struct.ConnectRule connectRule,
                   boofcv.struct.image.ImageType<T> imageType)

Method Detail

process

public void process(T input,
                    boofcv.struct.image.GrayS32 output)

initalize

protected void initalize(T input)

prepares all data structures

initializeClusters

protected void initializeClusters()

initialize all the clusters at regularly spaced intervals. Their locations are perturbed a bit to reduce the likelihood of a bad location. Initial color is set to the image color at the location

perturbCenter

protected void perturbCenter(SegmentSlic.Cluster c,
                             int x,
                             int y)

Set the cluster's center to be the pixel in a 3x3 neighborhood with the smallest gradient

gradient

protected float gradient(int x,
                         int y)

Computes the gradient at the specified pixel

setColor

public abstract void setColor(float[] color,
                              int x,
                              int y)

Sets the cluster's to the pixel color at that location

addColor

public abstract void addColor(float[] color,
                              int index,
                              float weight)

Performs a weighted add to the cluster's color at the specified pixel in the image

colorDistance

public abstract float colorDistance(float[] color,
                                    int index)

Euclidean Squared distance away that the pixel is from the provided color

getIntensity

public abstract float getIntensity(int x,
                                   int y)

Intensity of the pixel at the specified location

computeClusterDistance

protected void computeClusterDistance()

Computes how far away each cluster is from each pixel. Expectation step.

updateClusters

protected void updateClusters()

Update the value of each cluster using Maximization step.

assignLabelsToPixels

public void assignLabelsToPixels(boofcv.struct.image.GrayS32 pixelToRegions,
                                 org.ddogleg.struct.GrowQueue_I32 regionMemberCount,
                                 org.ddogleg.struct.FastQueue<float[]> regionColor)

Selects which region each pixel belongs to based on which cluster it is the closest to

getRegionMemberCount

public org.ddogleg.struct.GrowQueue_I32 getRegionMemberCount()

getClusters

public org.ddogleg.struct.FastQueue<SegmentSlic.Cluster> getClusters()

getImageType

public boofcv.struct.image.ImageType<T> getImageType()

getConnectRule

public boofcv.struct.ConnectRule getConnectRule()