Transcript ch12Boundarygabor

The Brain
from retina to extrastriate cortex
Neural processing responsible for
vision
• photoreceptors
• retina
– bipolar and horizontal cells
– ganglion cells (optic nerve)
•
•
•
•
optic nerves
optic chiasma (X)
lateral geniculate body
striate cortex
Lateral inhibition
• Edge detection and contrast enhancement
• Bipolar, Horizontal and Ganglion cells
Lateral inhibition
• If no activity in neighboring photoreceptors,
output = output of photoreceptor
Lateral inhibition
• if activity in neighboring photoreceptors,
– output is decreased, possibly absent
Lateral inhibition via addition
and negative weights
Optic nerve
• axons of the ganglion cells
– 1 million optic nerves
– 120 million photoreceptors
From light to vision
Lateral Geniculate Nucleus (LGN)
Geniculo-Striate Pathway
Striate
Cortex
Striate cortex
(primary visual centre)
• Neurons are edge detectors
fires when an edge of a particular
(LGN) orientation is present
Striate
Cortex
Striate cortex
(primary visual centre)
• Neurons are edge detectors
fires when an edge of a particular
(LGN) orientation is present
frequent output
Striate
Cortex
vertical bar
Striate cortex
(primary visual centre)
• Neurons are edge detectors
fires when an edge of a particular
(LGN) orientation is present
infrequent output
Striate
Cortex
diagonal bar
Edge detection
• each cell “tuned” to particular orientation
– vertical
– horizontal
– diagonal
• cats: only horizontal and vertical
• humans: 10 degree steps
• edges at particular orientations and
positions
Extrastriate cortex
(beyond the striate cortex)
V1
Extrastriate cortex
• Each area handles separate aspect of visual
analysis
–
–
–
–
“V1-V2 complex”: Map for edges
V3: Map for form and local movement
V4: Map for colour
V5: Map for global motion
• Each is a visual module
– connects to other areas
– operates largely independently
Douglas A. Lyon, Ph.D.
Chair, Computer Engineering Dept. Fairfield
University, CT, USA
[email protected], http://www.DocJava.com
Copyright 2002 © DocJava, Inc.
Background
• It is easy to find a bad edge!
• We know a good edge when we see it!
The Problem
• Given an expert + an image.
• The expert places markers on a good edge.
• Find a way to connect the markers.
Motivation
• Experts find/define good edges
• Knowledge is hard to encode.
Approach
•
•
•
•
Mark an important edge
Pixels=graph nodes
Search in pixel space using a heuristic
A* is faster than DP
Assumptions
• User is a domain expert
• Knowledge rep=heuristics+markers
Applications
• Photo interpretation
• Path planning (source+destination)
• Medical imaging
Photo Interpretation
Echocardiogram
•3D-multi-scale analysis
Path Plans, the idea
Path Planning-pre proc.
•hist+thresh
•Dil+Skel
Path Planning - Search
The Idea
• The mind selects from filter banks
• The mind tunes the filters
Gabor filter w/threshold
• The Strong edge is the wrong edge!
Canny Edge Detector
Mehrotra and Zhang
Sub bands for 3x3 Robinson
Sub Bands 7x7 Gabor
Gabor-biologically motivated
Log filters=prefilter+laplacian

2
x2  y2
1
2
2
e
2
2
1  x  y

1
4 
2
 
2
2
2
 
e


 f  f
 f (x, y)  2  2
x
y
2
2
2
x2 y2
2
2
Mexican Hat (LoG Kernel)
The Log kernel
Oriented Filters are steerable
Changing Scale at 0 Degrees
Changing Phase at 0 degrees
Summary
• Heuristics+markers= knowledge
• Low-level image processing still needed
• Global optimization is not appropriate for
all applications
• Sometimes we only want a single, good
edge