Giessen University

1
Robust contour extraction and junction detection
by a neural model utilizing recurrent long-range
interactions
Thorsten Hansen and Heiko Neumann

• Giessen University
• Dept. of Psychology

Ulm University Dept. of Neural Information
Processing
2
Overview of the Talk

• Motivation empirical evidence for recurrent
  long-range interactions
• 2. Approach and Model
• 3. Results Contour enhancement
• Corner detection
• 4. Conclusions

3
Sketch of V1 Architecture
long-range connections
McGuire et al. 1991
recurrent intercolumnar interactions
LGN
4
Specificity of Horizontal Long-Range Connections
in V1
Bosking et al. 1997
like connects to like plus
colinear arrangement
Long-range connections link neurons with
same orientation preference and collinear aligned
RFs.
5
Functional Implications of Lateral Long-Range
Interactions
Polat Sagi (1993)
Measurement of contrast detection thresholds for
foveal Gabor elements with and without flankers.
Colinear flanking Gabors (up to a distance of 10
wavelengths) facilitate contrast detection.
6
Key Mechanisms of the Proposed Model

• Excitatory long-range interactions between cells
  with collinear aligned RF (Bosking et al. 1997)
• Inhibitory short-range interactions
• Modulating feedback Initial bottom-up activity
• is necessary (Hirsch Gilbert 1991)

7
Model architecture
8
Recurrent Interaction
9
Results Contour Enhancement
Activity that fits into a more global context is
enhanced by top-down feedback.
10
Results Temporal Evolution
input image complex cells

long-range t1 t 2
t12
11
Quantitative Evaluation Contur Saliency
Saliency
Li, 1999
High saliency large values of (r,z)
12
Results Natural Images
input image complex cells
long-range
13
Simulation Physiological Data
Kapadia et al. 1995
response relative to single bar
bar flankers texture flankerstexture
enhancement for collinear bar suppression for
noisy textures
14
Properties of the Proposed Model
input image
complex cells
long-range
15
Definition of Corners and Junctions
Corners and junctions are points where two or
more lines join or intersect
16
Junctions for Object Recognition(Biederman 1987)
17
Junctions and Brightness Perception
Adelson (2000)
18
Junction Detection in Natural Images
Junctions often cannot be detected
locally (McDermott 2001)
13 pixel closeup
13 25 49
97 pixels
19
Neural Representation of Junctions
20
Read-out of Distributed Information
Orientation significance Length of the resulting
orientation vector in relation to the overall
activity
21
Corner and Junction Detection
Corner candidates high circular variance and
high overall activity
Corner points sufficiently large local maxima
of corner candidates
22
Results Localization Accuracy
generic junction configurations
23
Junction Detection on a Synthetic Image
Attneaves cat
complex cells
long-range
24
Junction Detection on Natural Images
Real world camera image
25
Junction Detection on Natural Images
cut-out of a plant image
Van Hateren van der Schaaf 1998
complex cells
long-range
input image
26
Evaluation using ROC Analysis
Comparison of the new scheme to standard methods
based on Gaussian curvature and
the structure tensor (black)
input image
27
Conclusions

• corners and junctions can be robustly
  represented
• by distributed activity within a cortical
  hypercolum
• recurrent colinear long-range interactions serve
• as a multi-purpose mechanism for
• contour enhancement
• noise suppression
• junction detection

Hansen Neumann (2004) Neural Computation 16(5).