pattern and object recognition - wofford college
TRANSCRIPT
Pattern recognition theories
How do we interpret lines and patterns as objects?
Why is object perception so difficult for computers?
Start simple: How do we recognize letters or other simple objects?
Template approach
Stimulus is compared to stored pattern
Examples? Bar code, bank check, scantron, etc.
Problems:
There are an infinite number of templates to
remember
Have to learn a template first
Any change in stimuli will not be recognized
Specialized receptors in visual cortex
Simple cells
e.g. Orientation specific
Complex cells
Combination of 2 simple features
Feature detectors
Stimulus
Cell’s
responses
Recognition by components
Biederman‟s RBC (recognition by component) theory
36 geons (3D)
basic building blocks
Emphasis on
intersections
Recognition with missing information possible
Pattern or object recognition
Bottom-up processing
Information from sensory receptors
Processing driven by stimulus
Data-driven
Top-down processing
Information from knowledge and expectations
Processing driven by higher level knowledge
Conceptually-driven
Examples
Tox-Doxn Pxocxssxng
To xllxstxatx, I cxn rxplxce xvexy txirx
lextex of x sextexce xitx an x, anx yox stxll
xan xanxge xo rxad xt – ix wixh sxme
xifxicxltx
The redundancy of stimuli provide more
features than required
Context and knowledge fills in the rest!
Beyond bottom-up processing
Depth perception
Depth cues: Relative size
Size constancy
Odor intensity
Controlled for sniff intensity
Perception of language
Speech segmentation
Word recognition
Flash stimulus
Word condition: FORK
Letter condition: K
Nonword condition: RFOK
Choose letter that was presented
K or M
Result:
Faster and more accurate when letter
part of original stimulus (word condition)
Word superiority effect
Treisman & Schmidt (1982)
Does prior knowledge change perception?
Method
Give Ss description of objects (“carrot, lake, tire”)
Flash display of #s/objects 200 ms; mask
Ask to report #s then objects
Results
Info significantly improves accuracy
Conclusion
“Top-down” knowledge changes perception
Able to “bind” features together more rapidly?
Palmer (1975)
Method Present scene
Ss ID flashed pics (a) or (b) or (c)
IV: type of picture
DV: accuracy
Results Appropriate pictures: 83%
Inappropriate pictures: 50%
Misleading pictures 40%
Conclusion Bottom-up perception interacts with prior knowledge
(top-down) to influence response
Gestalt principles of organization
Integrate info into meaningful whole
Heuristics: best-guess predictions
Laws of “perceptual organization”
Pragnanz: Good figure or simplicity
Similarity
Good continuation
Proximity
Common fate
Familiarity
Other heuristics
Occlusion heuristic
Light-from-above heuristic
Which gestalt law/heuristic?
Dalmation
http://michaelbach.de/ot/cog_dalmatian/index.html
“Biological motion”
http://michaelbach.de/ot/mot_biomot/index.html
CogLab: Apparent motion Data from Spring 09 (N = 8)
Expected result: For larger separations, the stimulus must "move" a
farther distance, which presumably requires a greater length of time.
Apparent motion/motion illusions
Pikler-Ternus display:
http://michaelbach.de/ot/mot_Ternus/index.html
“Rotating snake”
http://michaelbach.de/ot/mot_rotsnake/index.html
“Freezing rotation”
http://michaelbach.de/ot/mot_freezeRot/index.html
“Stepping feet”
http://michaelbach.de/ot/mot_feet_lin/index.html
Apparent motion factors
Color, shape, perceived depth, context
Optical illusions and visual
phenomenon
http://michaelbach.de/ot/index.html
Motion aftereffect: http://michaelbach.de/ot/mot_adapt/index.html
Lilac chaser: http://michaelbach.de/ot/col_lilacChaser/index.html
Optical illusions and visual phenomenon
Watercolor illusion: bright inside color spreads into enclosed area
Problems for computers
Stimulus on receptors is ambiguous
Inverse projection problem
Segmentation
Visual separation/overlap
Speech segmentation
Visual or verbal noise
Occlusions or obscured
Blurred or degraded
Changes in shadowing (lightness/darkness)
Human perception is different due to bottom-up AND top-down processing!
Chapter 3: Perception
Research questions What are the processes responsible for perception?
How do we recognize objects or words?
Why is perception difficult for computers?
Methods Name objects in pictures or read words (or letters)
With or without “noise”
With or without prior information (context)
Indicate what you see with an illusion figure
Results Requires combination of bottom-up and top-down processing
Use (gestalt) rules of perceptual organization
Experience-dependent plasticity: depends on experiences
Future directions