A General, Mechanistic Model of Spatial Pattern Detection
and David J. Heegerz
Department of Computer Science
zDepartment of Psychology
Stanford University
Stanford, California 94305
y
Patrick C. Teo
y
Abstract
Purpose: We propose a general, mechanistic, image-processing model of human spatial pattern
detection.
Methods: The model consists of three parts: a retinal component, a cortical component, and a
detection mechanism. The retinal stage involves luminance normalization (dividing by the local
mean luminance and then subtracting the new local average) followed by a static nonlinearity
similar to Graham and Hood (Vis. Res. 32:1373-1393, 1992). In the cortical stage, quadrature
phase linear lter outputs are squared and summed to yield local frequency and orientation
energy responses which are then divisively normalized (Heeger, Vis. Neurosci. 9:427-443, 1992).
Finally, detection is mediated by an ideal observer rule.
Results: The model accounts for a large body of data including: (1) luminance masking (Kortum & Geisler, Vis. Res., in press), (2) contrast sensitivity for dierent mean luminances and
grating areas (Rovamo et al, Vis. Res. 34:1301-1314, 1994; Van Nes & Bouman, J. Opt. Soc.
Am. 57, 1967), (3) spatial masking as a function of masker orientation (Foley, J. Opt. Soc. Am.
A. 11:1710-1719, 1994; Bradley & Ohzawa, Vis. Res. 26:991-997, 1986) and masker phase (Foley
& Boynton, Proc. SPIE 2054, 1994) and (4) subthreshold summation experiments (Graham,
1989).
Conclusions: Such a mechanistic, image-processing model of human spatial pattern detection
is useful because it predicts the results of any spatial pattern detection experiment. Moreover,
it can also be used as a measure of image integrity in designing and evaluating image processing
algorithms.
Slide set of talk presented at ARVO '95.
Teo and Heeger, \A General, Mechanistic Model of Spatial Pattern Detection," ARVO '95
A General Mechanistic Model of
Spatial Pattern Detection
2
Goals
Models of
Light Adaptation
(retinal)
Patrick C. Teo
Models of
Contrast Masking
(cortical)
?
Department of Computer Science
Stanford University
Can they be merged to predict both data sets?
David J. Heeger
Can they predict the results of other
spatial pattern detection experiments?
Department of Psychology
Stanford University
1
2
Retinal Component
Spatial Pattern Detection Model
+
Gain Control
+
Target+Mask
Image
?
d’
Corneal
Image
Gain Control
+
Gain Control
+
Mask Image
Retinal
Component
Cortical
Component
Luminance
Normalization
Optical
Blurring
Decision
Mechanism
3
Static
Non-linearity
4
Cortical Component
Decision Mechanism
Quadrature
Filter
Responses
Squaring
Normalized
Target+Mask
Images
Summing
σ
Divisive
Normalization
∑
∑
∑
|| 2
Squared Error
Norm
Pooled
Responses
Frequencydependent
Spatial Pooling
Spatial Pattern Detection Model
Luminance Masking
Cortical
Component
+
Decision
Mechanism
|| ||
Gain Control
+
2
∑
∑
∑
d’
4
Amplitude threshold (lg tds)
6
Amplitude threshold (lg tds)
5
Retinal
Component
d’
Normalized
Mask Images
Normalized
Response
Target+
Mask
Image
||
2 cpd, 1.91 lg tds
3
2
1
0
−1
0
1
2
3
4
Flashed background (lg tds)
4
2 cpd, 3.03 lg tds
3
2
1
0
−1
0
1
2
3
4
Flashed background (lg tds)
Mask
Image
Kortum & Geisler ’94
Model
7
8
Teo and Heeger, \A General, Mechanistic Model of Spatial Pattern Detection," ARVO '95
Luminance Masking
−1 0 1 2 3 4
4
3
2
1
0
4
3
2
1
0
−1 0 1 2 3 4
−1 0 1 2 3 4
−1 0 1 2 3 4
3.03
lg tds
−1 0 1 2 3 4
4
3
2
1
0
4
3
2
1
0
−1 0 1 2 3 4
1.91
lg tds
−1 0 1 2 3 4
4
3
2
1
0
−1 0 1 2 3 4
4
3
2
1
0
4
3
2
1
0
−1 0 1 2 3 4
−1 0 1 2 3 4
4
3
2
1
0
4
3
2
1
0
8 cpd
4
3
2
1
0
4
3
2
1
0
Orientation Masking
4 cpd
−15
Target threshold contrast (dB re 1)
4
3
2
1
0
2 cpd
Target threshold contrast (dB re 1)
Amplitude threshold (lg tds)
1 cpd
3
0 deg re vert
−20
−25
−30
−35
4.03
lg tds
−50
−15
90 deg re vert
−20
−25
−30
−35
−40
−30
−20
−10
Masker contrast (dB re 1)
−50
−40
−30
−20
−10
Masker contrast (dB re 1)
−1 0 1 2 3 4
−1 0 1 2 3 4
Flashed background (lg tds)
Foley & Boynton ’94
Model (scaled)
Kortum & Geisler ’94
Model
9
10
Orientation Masking
Target threshold contrast (dB re 1)
−15
−15
−20
−20
−25
−25
−30
−30
−35
Subthreshold Summation
11.25 deg
−35
−50
−40
−30
−20
−50
−10
45 deg
−15
−15
−20
−20
−25
−25
−30
−30
−35
−40
−30
−20
−10
90 deg
−40
−30
−20
−10
−50
Masker contrast (dB re 1)
6.0
0.8
2.0
0.6
1.0
0.4
0.2
−35
−50
∞
1
Relative contrast (1 cpd)
0 deg
−40
−30
−20
Model
3.8
−10
Masker contrast (dB re 1)
Foley & Boynton ’94
Model (scaled)
0
0
0.2
0.4
11
1
Two-Component Masking
(with additional fixed masker)
Target threshold contrast (dB re 1)
0.8
12
Orientation Masking
−15
0.6
Relative contrast (4 cpd)
No additional masker
Additional 90-deg masker
Threshold Contrast
Target
−20
Masker
Derrington &
Henning ’89
Model (raw)
−25
0.00275
0.0423
0.00630
(2.3 : 1)
0.103
(2.4 : 1)
0.0210
(7.6 : 1)
0.115
(2.7 : 1)
−30
−35
−50
−40
−30
−20
−10
Masker contrast (dB re 1)
Foley & Boynton ’94
Model (scaled)
13
14
Conclusions
Models of
Light Adaptation
(retinal)
Models of
Contrast Masking
(cortical)
A General
Mechanistic Model of
Spatial Pattern Detection
Can they be merged to
predict both data sets?
Can they predict the results
of other experiments?
15
Yes
Yes/No