Report: Andreu Gonzalez
Model-based Texture Classification under Varying Illumination
Entire Report PDF
Abstract
This dissertation presents a
complete texture classification system to overcome the problem induced by
changes in the angle of illumination incident upon a 3D surface.
The system works on the basis
of a surface model, previously formed by means of a photometric stereo
technique. From this model, the system is able to render a 2D image of the
surface at any particular illuminant direction, thus providing a more
appropriate data for training the classifier.
Many laboratory experiments
are carried out in order to assess the accuracy of image prediction as an
individual component. The investigation considers a large diversity of textures,
including challenging situations such as rough, specular and anisotropic
surfaces. It is concluded that the predicted images, yet not being perfectly
accurate, are in all cases a much more reliable training data than a merely
single image from a single illuminant direction.
The technique is evaluated
using supervised statistical classification, which combines a bank of Gabor
filters for feature extraction with a linear Bayes classifier. The
classification performance is tested for different composite images, consisting
of a varying number of disjoint textures and configurations. It is shown that
our approach significantly reduces the misclassification rate, when compared
with a naive classification system. Furthermore, in some cases it even reaches
the level of accuracy that one would obtained with the proviso that training and
classification were performed under invariant illumination.
Table of Contents
1 Introduction
1.1
Motivation
1.2 Scope of the
research
1.3 Dissertation
organisation
1.4 Original work
2 Image acquisition and
illumination
2.1 Introduction
2.2 Characterisation of incident image
2.2.1 The reflectance function
2.2.2 Gradient space
2.2.3 The Lambertian image
2.3 The imaging process
2.3.1
Overview
2.3.2 Implications for texture
classification
2.4 Surface
description
2.4.1 First order
statistics
2.4.2 Histograms
3 The classification process
3.1 Introduction
3.2 Texture classification
3.2.1 Brief
description
3.2.2 How it works
3.2.3 Differences between classification and segmentation
3.3 Feature
extraction
3.3.1 Overview of feature
measures
3.3.2 Filtering for texture measure
3.3.3 Gabor functions
3.4 The discriminant
function
3.4.1 Theoretical
framework
3.4.2 A Bayes classifier
3.4.3 Practical
considerations
4 Surface recovery and rendering
4.1 Introduction
4.2 Surface recovery
4.2.1 Shape from
X
4.2.2 Photometric techniques
4.2.3 Related work in photometric
stereo
4.3 Rendering
5 Approach to the problem
5.1 Problematic
5.2 Model-based
solution
5.2.1 Recovery
stage
5.2.2 Training
stag
5.2.3 Classification
stage
5.3 Photometric
implementation
5.3.1 A simple photometric stereo
scheme
5.3.2 Collecting photometric
data
5.4 Feature
extraction
5.4.1 Multichannel
scheme
5.4.2 Filtering in the frequency
domain
5.4.3
Post-processing
5.5 Gabor filters
design
5.5.1 Filter
characterisation
5.5.2 Selection of
filters
6 Assessment of image prediction
6.1 Introduction
6.2 Limitations of the model
6.2.1 Non-Lambertian reflectance
6.2.2 Cast and self shadowing
6.3 Test textures
6.4 Accuracy of image
prediction
6.4.1 Influence of surface
roughness
6.4.2 Effect of varying
tilt
6.4.3 Effect of varying
slant
6.5
Discussion
7 Classification performance
7.1 Introduction
7.2 Experimental framework
7.2.1 Selection
criteria
7.2.2 Test
images
7.3 Accuracy of classification
7.3.1 Variation of tilt
angle
7.3.2 Variation of slant
angle
7.4 Dependence on image prediction
7.5 The effect of increasing number of textures
8 Summary, conclusions, and future work
8.1 Summary
8.2 Conclusions
8.3 Future
work
Appendix A: Test textures
Appendix B: Histogram description of surface gradient
Appendix C: Shell scripts
References
