Nonlinear recurrent mechanisms for the processing and representation of surface boundaries based on luminance and texture gradients [Elektronische Ressource] / von Axel Thielscher

Axel_2

Sujets

• Gesundheit

Informations

• Publié par: universitat_ulm
• Publié le: 01 janvier 2008
• Nombre de lectures: 31
• Langue: English
• Poids de l'ouvrage: 12 Mo

Extrait

Nonlinear Recurrent Mechanisms for the
Processing and Representation of Surface Boundaries
Based on Luminance and Texture Gradients



DISSERTATION


zur Erlangung des akademischen Grades eines


DOKTOR-INGENIEURS

(DR.-ING.)


der Fakultät für Ingenieurwissenschaften
und Informatik der Universität Ulm

von

Dr. biol. hum. Axel Thielscher
aus Kösching




Betreuer: Prof. Dr.-Ing. Jürgen Lindner
Prof. Dr. rer. nat. Heiko Neumann

Amtierender Dekan: Prof. Dr. rer. nat. Helmuth Partsch



Ulm, 30.04.08

Acknowledgements
First, I would like to thank Prof. Dr. J. Lindner for the interest he showed in the topic of this
PhD thesis (which is not in the mainstream of his research) and for kindly undertaking the
responsibilities as supervisor. Furthermore, I would like to thank Prof. Dr. M. Spitzer for
giving me the opportunity to realize this thesis when I was member of the Department of
Psychiatry (University of Ulm).
Prof. Dr. M.A. Cohen (CNS, Boston University) provided helpful comments on the stability
analysis in chapter 2.
My former colleagues PD Dr. G. Grön and M. Koelle offered substantial support and steady
motivation that enabled me to perform two fMRI-experiments to test some of the predictions
of the neural model.
I am grateful for Prof. Dr. H. Neumann's steady support during the last years, his motivating
and enlightening remarks and the frequent discussions which proved very helpful for my
work. I would like to thank Heiko for establishing a pleasant and at the same time productive
working atmosphere. In particular, I enjoyed the journal club meetings of his vision science
group, with its current and former members Dr. T. Hansen, Dr. M.S. Keil, Dr. P. Bayerl and
Dr. S. Corchs (among others).
Last but not least I would like to thank my wife Susanne for all her love and encouragement.

Table of Contents
Table of Contents
1 Introduction 5
1.1 Motivation and Overview 5
1.2 Architecture of the Human Visual System 8
1.3 How Texture- and Depth-Processing Relates to the Overall Task of 14
Object Recognition in the Ventral Visual Pathway
1.4 Shunting Cell Dynamics: Motivation and Functional Properties 20

2 A Model for the Recurrent Processing of Texture Boundaries 25
2.1 General Model Architecture: Physiological and Psychophysical 26
Foundations
2.2 Model Areas and Receptive Field Organization 31
2.3 Model Cell Dynamics: Overview and Stability Analysis 33
2.3.1 Boundedness 39
2.3.2 Test for the Applicability of the Cohen-Grossberg Theorem 40
2.3.3 Test for Global Stability Using Hirsch’s Theorem 42
2.3.4 Stability Analysis of a Hierarchy of Model Areas 48
2.4 General Model Behavior 51
2.5 Systematic Variation of Background Orientation Noise: Effect of 53
Feedback
2.5.1 Stimuli & Analysis of Model Activation Patterns 54
2.5.2 Results 55
2.5.3 Relation to Psychophysics 60
2.6 Functional Roles of Specific Model Connections 62
2.7 Functional Role of Model Area V2 64
2.7.1 Example: Lesioning Feedback from Model Area V4 to V2 65
2.7.2 Relation to Psychophysics 66
2.8 Variations of Texture Density 69
2.8.1 Simulation Results 70
2.8.2 Relation to Psychophysics 72
2.9 Modulation of V1 Responses by Feedback Activity 73
2.10 Segmentation of Real-world Images: An Example 75
1 2.11 Main Discussion 77
2.11.1 Key Model Features 77
2.11.2 Relation to Psychophysics 78
2.11.3 Relation to Physiological and Neuroimaging Data: The Functional Roles 81
of Areas V1 and V4
2.11.4 Other Models 88
2.11.5 Model Limitations and Putative Future Developments 91
2.12 Conclusion 92

3 Globally Consistent Depth Sorting of Overlapping 2D Surfaces 95
in a Model of Local Recurrent Interactions
3.1 Recurrent V1-V2 Contour Processing 98
3.2 Recurrent Depth Processing 102
3.2.1 A Recursive Scheme for Depth Sorting of Surface Contours Using Local 102
Relative Depth Cues
3.2.2 Outline of the Overall Model Architecture and Key Processing 105
Mechanisms
3.2.3 Dipole Dynamics: A Close-up 109
3.2.4 Interactions between Neighboring Dipoles 111
3.2.5 Interaction between Model Layers 114
3.2.6 The Impact of T-junctions on Dipole Activity 115
3.3 Simulation Results 117
3.3.1 Pre-Processing 118
3.3.2 Recurrent Depth Processing 120
3.3.3 Further Simulation Results 125
3.4 Discussion 130
3.4.1 General Model Framework: Surface Boundary Processing and 130
Depth Propagation
3.4.2 Depth Layers and Dipole Dynamics for Activity Propagation 132
3.4.3 General Model Framework 133
3.4.4 Other Models of Depth Processing 134
3.4.5 Model Properties, Limitations and Future Extensions 138

4 Summary 141

2 Table of Contents
A Model of Texture Border Processing: Supplements 149
A.1 Receptive Field Equations 149
A.2 Model Cell Dynamics: Reformulation of Equations for Stability Analysis 154
A.3 Boundedness 155
A.4 Stability Analysis of a Hierarchy of Model Areas: Supplements 157

B Model Determining the Depth of Surface Contours: Supplements 159
B.1 Recurrent V1-V2 Contour Processing 159
B.2 Feed-forward Scheme for the Detection of Corners 161
and T-Junctions
B.3 Interactions between Neighboring Dipoles: Supplements 163
B.4 Impact of the Model T-junction Detectors on Model Dipole Activity 166

Bibliography 169






3
4 1 Introduction
1 Introduction
1.1 Motivation & Overview
Recognition or identification of objects on the basis of two-dimensional images (such as
pictures taken by cameras or the retinal images of the human eyes) has to cope with several
non-trivial problems (Liter and Bülthoff, 1996): Subsequent images of the same object can
differ substantially from each other, depending on variations of viewpoint, distance,
illumination, etc. Furthermore, in real world situations, objects are often partially occluded by
others, in turn complicating a clear-cut recognition. The development of computational
frameworks and mechanisms enabling robust object recognition is in the focus of ongoing
and vivid research activities. However, it is a very challenging problem and no general
solution exists up to now. In contrast, humans and mammals achieve apparently effortless
and robust object recognition even in cluttered and noisy real-world environments. So one
way to improve the computational architectures and mechanisms used for image processing
might be to try to understand how the same tasks have been solved by Nature. Clearly, a
technical solution will never be a strict one-to-one implementation of biological vision.
However, the hope is to identify general principles, constraints, mechanisms and design
frameworks which can be used to guide the development of technical systems.
In the last few decades our knowledge about the structure and functionality of living beings
has dramatically increased. This increase has led to a distribution and parcellation of this
knowledge over various scientific disciplines and, at the moment, new scientific disciplines
emerge as an attempt to integrate research on the basis of life into a common framework.
Biological vision is a good example for that development. It is a traditional topic of Biology
(focusing on the physiology of the vision system) and Psychology (focusing on the overall
process of perception or Wahrnehmung). More lately, it has become one of the multi-
disciplinary research focuses of Neurosciences, a prominent member of the newly emerging
5 Life Sciences. Research on biological vision is demanding and requires the integration of
data obtained by several classical research disciplines such as Biology, Physics,
Psychology, Medicine and Mathematics into a common concept. In this work, I use the tools
developed in the field of Neuroinformatics to capture specific aspects of biological vision
(namely the processing of surface textures and depth information) into a coherent theoretical
framework, in turn allowing for a rigid testing of the underlying assumptions. The aim of my
work is two-fold: On the one hand the results presented here allow for a better understanding
of the neural mechanisms underlying biological object recognition, and on the other hand
they might help to improve future technical approaches to image processing.
Although our understanding of biological vision is still rather at the beginning, many important
key properties and principles have already been discovered. One central principle is the
initial processing of surface boundaries by the mammalian visual system as necessary pre-
requisite for object recognition. The visual system is faced with the problem that the same
surface can look very differently when changing illumination. Even different parts of a uniform
surface can have very different local brightness levels and colors due to reflections and
shadows originating from neighboring objects. However, when we look at such a surface we
perceive it as uniform and contiguous, i.e. our visual system ha