Eigenregions For Image
Eigenregions For Image
Classification
Classification
IEEE Trans. on PAMI, 2004 IEEE Trans. on PAMI, 2004
Clément Fredembach, Clément Fredembach, Michael Schröder, Michael Schröder, Sabine Süsstrunk Sabine Süsstrunk
Outline
Outline
IntroductionIntroduction
Image segmentationImage segmentation EigenregionsEigenregions
Region classificationRegion classification Image classificationImage classification ConclusionConclusion
Introduction
Introduction
Image classification need features the Image classification need features the well express relevant image properties
well express relevant image properties
Eigenregions are geometrical features Eigenregions are geometrical features encompass area, location and shape
encompass area, location and shape
Eigenregions are obtained by Eigenregions are obtained by
analyzing segmented image regions
analyzing segmented image regions
with Principal Component Analysis
with Principal Component Analysis
(PCA)
Image segmentation(1)
Image segmentation(1)
Segment image into meaningful Segment image into meaningful regions
regions
Dominant Colors in Lab (DC Lab) Dominant Colors in Lab (DC Lab) method
method
Use k-means powered algorithm to Use k-means powered algorithm to segment images according to (CIE)
segment images according to (CIE)
Lab clusters
Lab clusters
Image segmentation(2)
Image segmentation(2)
Eigenregions(1)
Eigenregions(1)
Eigenregions are obtained by Eigenregions are obtained by
calculating the principal components of
calculating the principal components of
the region locations
the region locations
Segmented images have 64 * 48 Segmented images have 64 * 48 pixels, down-sample to 5 * 5
Eigenregions(2)
Eigenregions(2)
Eigenregions(3)
Eigenregions(3)
PCAPCA
– X: data matrix, N vectors of length 25X: data matrix, N vectors of length 25
– : mean over all observations of X: mean over all observations of X
– Y is defined as X with subtracted from each Y is defined as X with subtracted from each of its columns
of its columns
– C: covariance matrix, C = Y YC: covariance matrix, C = Y Y‧‧ TT
X
Eigenregions(4)
Eigenregions(4)
Region classification(1)
Region classification(1)
Based on database of 13500 real-Based on database of 13500 real-scene photographs
scene photographs
77776 regions obtained via 77776 regions obtained via segmentation
segmentation
Test the usefulness of eigenregions Test the usefulness of eigenregions with Automatic Color Correction (ACC)
with Automatic Color Correction (ACC)
Using 10 first eigenregions, color Using 10 first eigenregions, color features and texture features
Region classification(2)
Region classification(2)
Image classification
Image classification
Conclusion
Conclusion
Eigenregion is a geometrical feature coEigenregion is a geometrical feature co mbines region area, shape, and positio
mbines region area, shape, and positio
n information
n information
The largest variance in region geometrThe largest variance in region geometr y is due to the area and not to the sha
y is due to the area and not to the sha
pe or position
pe or position
PCA-based might not the most adaptePCA-based might not the most adapte d of region reconstruction
