Learning and Extracting Primal-Sketch Features in a Log-polar Image Representation

1
Learning and Extracting Primal-Sketch Features in
a Log-polar Image Representation

• Herman M. Gomes hmg_at_dsc.ufpb.br
• Departamento de Sistemas e Computação, COPIN,
  UFPB
• Campina Grande PB, Brasil
• Robert B. Fisher rbf_at_dai.ed.ac.uk
• Division of Informatics, Edinburgh University
• Edinburgh, UK

2
Contents

• Introduction
• Related Work
• Image Representation
• Proposed Approach
• Training
• Evaluation
• Results
• Concluding Remarks
• References

3
Introduction

• Traditional image feature extraction operators
• Cartesian domain (artefact of sensor
  architecture)
• Work independently of each other
• Designed by hand
• Primate visual system
• Mapping from retina to visual cortex is log-polar
• Learning
• Primal sketch Marr82
• Features like edges, bars, blobs, ends detected
  at a number of orientations and contrasts
• Grouping processes

4
Related Work

• Neural network learning of Edge features
  CTR95,PB92
• Limited to edges
• Comparable to Sobel or Canny performances
• Arbitrary features in the log-polar domain GF96
• Operators manually designed
• Poor sensitivity to the features contrast
• Limited to a fixed window size

Edge_at_0 ABS(fab-c-d-e)/3 Edge_at_60
ABS(abc-d-e-f)/3 ... Blob
MIN(x-a,x-b,x-c,x-d,x-e,x-f)
a
f
b
x
e
c
d
5
Image Representation

• Retina structure
• Uniform sampling, high density fovea
• Space-variant periphery

6
Image Representation

• Receptive field computation and reflectance
  estimation

7
Proposed Approach

• Training process

recep. field windows
Normalise Orientation
Compute Projection
Build Training Set
Train Neural Networks
Exemplars of Features
Training Set
feature class and contrast
8
Proposed Approach

• Testing process

position
orientation
Extract Recep. Fields
Normalise Orientation
Apply Neural Networks
Compute Feature Planes
Compute Projection
Feature class, position contrast and
orientation
Test Images
Planes
Edge
Bar
Blob
End
9
Proposed Approach

• Normalising the feature orientation

Edge
Bar
End
10
Training

• Principal Components from a set of synthetic
  features

11
Training

• Neural network architecture

PCA projected window
Principal components
Receptive field window
Neural network
Edge
...
Bar
N
Ñ
Blob
End
1x19
19x17
1x17
12
Evaluation

• Ground truth for untrained synthetic features

13
Evaluation

• Output of the Edge neural module

14
Evaluation

• Prediction errors of the symmetry operators

15
Results

• Testing on synthetic images

Edges
Bars
-Blobs
Blobs
Output
Retinal
Input
16
Results

• Testing on real images

Neural Outputs
Retinal Image
Logical Operators
Input Image
17
Concluding Remarks

• New learning-based approach to extracting primal
  sketch features
• Better results when compared to a previous
  approach
• More correctly classified features
• Good estimate for the features contrast
• Can be easily applied to different window sizes
  and new feature types
• Successfully being used as the core
  representation in the problem of learning
  structural relationships from sets of 2D
  image-based models GF2000.

18
References

• Marr82 D. Marr. Vision. W. H. Freeman and Co.,
  1982.
• CTR95 W. C. Chen, N. A. Thacker and P. I.
  Rockett. A neural network for probabilistic edge
  labelling trained with a step edge model. In
  Proc. of 5th Int. Conf. on Image Processing and
  its Applications, pages 618-621, Edinburgh, 1995.
• PB92 D. T. Pham and E. Bayro-Corrochano. Neural
  Networks for low-level image processing. In Proc.
  Of the IEE Int. Conf. on Artificial Neural
  Networks, pages 809-812, 1992.
• GF96 T. D. Grove and R. B. Fisher. Attention in
  Iconic Object Matching. In Proc. British Machine
  Vision Conf., Vol. 1, pages 293-302, Edinburgh,
  1996.
• GF2000 H. M. Gomes and R. B. Fisher.
  Structural Learning from Iconic Representations.
  Lecture Notes in Artificial Intelligence,
  1952399-408, 2000.