textures and their usage in image...

TexturesImage Segmentation

SummaryTextures

Textures and their usage inImage Segmentation

Martin Petrıcek

April 14, 2011

Martin Petrıcek Textures and their usage in Image Segmentation


SummaryTextures

TexturesFeaturesIllumination invariance

Image Segmentation

Summary



SummaryTextures

Textures



Summary

FeaturesIllumination invariance

Texture features - colorspaces

I Several colospacesI RGB - acquisition colorspaceI HSV - perceptual colorspaceI L*A*B* - based on human vision



Summary


Texture features - CM

I Co-occurence matrixI Which colors occur in texture together?I QuantizationI RSCCM - Reduced Size Chromatic Co-occurence MatrixI Haralick Features



Summary


Texture features - Haralick Features

I Co-occurence matrix featuresI Angular second momentI ContrastI CorrelationI Sum of squares: varianceI Inverse difference momewntI Sum averageI Sum varianceI Sum entropyI EntropyI Difference varianceI Difference entropyI Information measures of correlationI Maximal correlation coefficient



Summary


Training

I Classify images into categoriesI Tedious

I Classify several image pairsI Must-link or cannot-link constraints (Same or different class)

I Selection of most relevant texture featuresI based on Laplacian score

I projections of same classes should be together

I based on constraint scoreI relevance with must-links and cannot-link constraints



Summary


Results

I Constraint score is better

100 %92 % 89 %



Summary


Illumination invariance

I Different light sources,different spectralcharacteristics

I Black body radiation,different temperatures

I Light sources with narrowspectrum

I Human eye can adapt

I Metameres



Summary


Illumination models

I Illumination modelsI DiagonalI Diagonal-OffsetI Affine

R ′ = R × αr + βrG ′ = G × αg + βgB ′ = B × αb + βb

R ′ = R × αrr + G × αrg + B × αrb + βrG ′ = R ×αgr +G ×αgg +B ×αgb +βgB ′ = R ×αbr +G ×αbg +B ×αbb + βb



Summary


Illumination invariance - solution

I Normalize colors beforecomputing features

I Assume diagonal model

I Normalize using FFT

I Use Gabor filtering or Locallinear transform on result



SummaryImage Segmentation

Image Segmentation




Training data

I We need to obtain training data

I Manually tracing examples istedious




Alternate approach

I Manually pick a bounding boxaround object

I Inside: object and some parts ofbackground

I Outside: only background

I Less precise

I Much faster, more images intraining set




Algorithm

I Split images into superpixels - patches of same texture

I Use patches to learn the system

I Learn using texture features from patches




Algorithm

I Fully segmented dataI Markov Random FieldsI Conditional Random FieldsI BoostingI SVM

I Partially segmented dataI Multiple Instance Learning

I 1. Identify positive in training dataI 2. Learn to distinguish positive and negative




Active learning

I Find out images with patches close to decision boundary

I Get user help with classifying these patches




Improvement - Background

I Use background information to improveaccuracy

I Cow is more likely to be on grass

I Car is more likely to be on road



Summary

LiteratureQuestions

Literature

I Rahat Khan, Damien Muselet and Alain Tremeau, Classicaltexture features and illumination color variations

I M. Kalakech, A. Porebski, P. Biela, D. Hamad and L.Macaire, Constraint score for semi-supervised selection ofcolor texture features

I Jaume Amores, David Geronimo, Antonio Lopez, MultipleInstance and Active Learning for weakly-supervisedobject-class segmentation

I R. Haralick, K. Shanmugan and I. Dinstein, Textural featuresfor image classification



Summary

LiteratureQuestions

Summary

Questions?


textures and their usage in image...

Documents