org.openimaj.image.feature.dense.gradient

Class PHOG

java.lang.Object

org.openimaj.image.feature.dense.gradient.PHOG

All Implemented Interfaces:

FeatureVectorProvider<DoubleFV>, ImageAnalyser<FImage>

@Reference(type=Inproceedings,
           author={"Bosch, Anna","Zisserman, Andrew","Munoz, Xavier"},
           title="Representing shape with a spatial pyramid kernel",
           year="2007",
           booktitle="Proceedings of the 6th ACM international conference on Image and video retrieval",
           pages={"401","","408"},
           url="http://doi.acm.org/10.1145/1282280.1282340",
           publisher="ACM",
           series="CIVR \'07",
           customData={"isbn","978-1-59593-733-9","location","Amsterdam, The Netherlands","numpages","8","doi","10.1145/1282280.1282340","acmid","1282340","address","New York, NY, USA","keywords","object and video retrieval, shape features, spatial pyramid kernel"})
public class PHOG
extends Object
implements ImageAnalyser<FImage>, FeatureVectorProvider<DoubleFV>

This class is an implementation of an extractor for the PHOG (Pyramid Histograms of Orientation Gradients) feature described by Bosch et al. The PHOG feature is computed by creating a quadtree of orientation histograms over the entire image and appending the histograms for each cell of the quadtree into a single vector which is then l1 normalised (sum to unity).

In the original description, only orientations at edge pixels were counted; that restriction is optional in this implementation. If only edge pixels are used, then the feature describes the distribution of shape in the image. Conversely, if all pixels are used, the feature essentially describes the texture of the image.

As this class will typically be used to only construct a single feature from an image, it is built around a BinnedWindowedExtractor (or InterpolatedBinnedWindowedExtractor if interpolation is used). This will be much more efficient than a GradientOrientationHistogramExtractor in the single window case. If you do need to extract many PHOG-like features different rectangles of the same image, use a GradientOrientationHistogramExtractor coupled with a QuadtreeStrategy to achieve the desired effect.

Author:

Jonathon Hare (jsh2@ecs.soton.ac.uk)

Constructor Summary

Constructors

Constructor and Description
PHOG() Construct with the values used in the paper: 4 levels (corresponds to l=3 in the paper), 40 orientation bins (interpolated), signed gradients (called "shape360" in the original paper) and Canny edge detection.
PHOG(int nlevels, int nbins, boolean histogramInterpolation, FImageGradients.Mode orientationMode, ImageProcessor<FImage> edgeDetector) Construct with the given parameters.
PHOG(int nlevels, int nbins, FImageGradients.Mode orientationMode) Construct with the given values, using Canny edge detection and gradient histogram interpolation.

Method Summary

Modifier and Type	Method and Description
void	analyseImage(FImage image) Analyse an image.
Histogram	getFeatureVector() Extract the PHOG feature for the whole of the image last analysed with analyseImage(FImage).
Histogram	getFeatureVector(Rectangle rect) Extract the PHOG feature for the specified region of the image last analysed with analyseImage(FImage).

Methods inherited from class java.lang.Object

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

Constructor Detail

PHOG

public PHOG()

Construct with the values used in the paper: 4 levels (corresponds to l=3 in the paper), 40 orientation bins (interpolated), signed gradients (called "shape360" in the original paper) and Canny edge detection.

PHOG

public PHOG(int nlevels,
            int nbins,
            FImageGradients.Mode orientationMode)

Construct with the given values, using Canny edge detection and gradient histogram interpolation.

Parameters:

nlevels - number of pyramid levels (note this includes l0, so you might need 1 more)

nbins - number of bins

orientationMode - the orientation mode

PHOG

public PHOG(int nlevels,
            int nbins,
            boolean histogramInterpolation,
            FImageGradients.Mode orientationMode,
            ImageProcessor<FImage> edgeDetector)

Construct with the given parameters. The edgeDetector parameter can be null if you don't want to filter out non-edge pixels from the histograms.

Parameters:

nlevels - number of pyramid levels (note this includes l0, so you might need 1 more)

nbins - number of bins

histogramInterpolation - should the gradient orientations be interpolated?

orientationMode - the orientation mode

edgeDetector - the edge detector to use (may be null for gradient features)

Method Detail

analyseImage

public void analyseImage(FImage image)

Description copied from interface: ImageAnalyser

Analyse an image.

Specified by:

analyseImage in interface ImageAnalyser<FImage>

Parameters:

image - The image to process in place.

getFeatureVector

public Histogram getFeatureVector(Rectangle rect)

Extract the PHOG feature for the specified region of the image last analysed with analyseImage(FImage).

Parameters:

rect - the region

Returns:

the PHOG feature

getFeatureVector

public Histogram getFeatureVector()

Extract the PHOG feature for the whole of the image last analysed with analyseImage(FImage).

Specified by:

getFeatureVector in interface FeatureVectorProvider<DoubleFV>

Returns:

the PHOG feature

See Also:

FeatureVectorProvider.getFeatureVector()