IMAGE PROCESSING WITH PULSE COUPLED NEURAL NETWORKS

1
IMAGE PROCESSING WITH PULSE COUPLED NEURAL
NETWORKS

• by
• Marius Schamschula, Rama Inguva, and John Johnson

2
Pulse Coupled Neural Networks

• Multi-layer 2D interconnection network
• No training, self-organizing
• Adjusted by threshold level and choice of
  interconnect kernel
• Used for image segmentation, scene analysis,
  iconification, foveation, intensity compression

3
Biological Foundations

• Based on studies of the cat visual cortex
• Pulse rate proportional to intensity
• Pulses from neighboring pixels interact via a
  series of local interconnections
• Pulses are grouped via a feedback mechanism

4
PCNN Scheme
5
PCNN Equations
6
PCNN Movie
7
Image Segmentation
Figures by Jason Kinser (http//www.ib3.gmu.edu/ja
ker/pcnn/mammo/breast.html)
8
Iconification
Figures by Jason Kinser (http//www.ib3.gmu.edu/ja
ker/pcnn/icons/icons.html)
9
Foveation
Figures by Jason Kinser (http//www.ib3.gmu.edu/ja
ker/pcnn/foveate/foveate.html)
10
Multichannel PCNN
11
PCNN Fusion

• Concurrently run several PCNNs
• Crosslink pulse patterns
• Inhibitive
• Additive
• Multiplicative
• Exponential

12
PCNN Fusion
13
PCNN Fusion Approach

• Balanced fusion of internal activity
• No feeding field

14
Multi-Sensor Fusion
North-Eastern Storm 3/4/1999 from GOES 8
Data(from http//rsd.gsfc.nasa.gov/)
Visible
IR2
IR3
IR4
IR5
15
Multi-Sensor Fusion
Fused output vs. m (100 iterations)
1.0
1.5
1.75
2.0
16
Multi-Sensor Fusion
Enhanced visible
Fused (m 2.0)over visible
17
PCNN Analysis

• What is the optimal pulse train?
• Can we clean up any noise in the original image?

18
PCNN Statistics

• Choose three images with 0, 25 and 50 random
  noise.
• Use linear threshold decay
• Exponential decay is non-repetitive

19
PCNN Statistics Results
0 Noise
25 Noise
50 Noise
20
Conclusion

• PCNNs are a unique tool
• PCNNs model real time-dependent neural activity
• Many properties are still unexplored
• Work is ongoing
• More information _at_
• http//www.caos.aamu.edu/PCNN/