Dear Friends:
This is the Summary of the replies that I received regarding
my recent enquiry on wavelet receptive fields and the interplay
between neural morphology and function. Many thanks to
everybody who kindly contributed with invaluable information,
including:
Dr Savier M. Sauvan
Prof R. A. Young
Prof A. Pentland
Prof D. Heeger
Prof J. Nelson
Considering that external communication in our computing system
was severely disrupted during last week, it is possible that
some replies to my enquiry got lost. If this is the case, please
re-send your message.
There is little doubt that wavelets are important concepts
not only in computer vision, but also as means for modeling
biological visual systems. Perhaps it will bridge the gap
between the hypothesis that the brain processes Fourier
transform and the functional organization of the cortical
pathways according to scale-space retinotopical channels...
As far as the interplay between neural morphology and function
is concerned, to judge from the relatively less numerous
feedback from CVNET (see however Prof Nelson's reply), it
seems it has yet to come of age, particularly regaring
the cortical structure.
I would also like to add the following references that I
received recently from Prof. Young:
R. A. Young and R. M. Lesperance
A Physiological Model of Motion Analysis for Machine Vision
GM Publication GMR-7878, Jan 1993
R. A. Young
Oh Say can you See? The Phisiology of Vision
GM Publication GMR-7364, May 1991
R. A. Young
Bridging the gap between vision and commercial applications
SPIE Conference Proceedings, San Jose, Ca, vol 2411, 1995
R. A. Young
The gaussian derivative model for machine vision: visual cortex
simulation
GM Publication GMR-5323, Jul 1986
R. A. Young
The Gaussian derivative theory of spatial vision: analysis
of cortical cell receptive field line-weighting profiles
GM Publication GMR-4920, May 1985
R. A. Young
The Gaussian derivative model for spatial vision: I. Retinal
mechanisms
Spatial Vision, Vol2, No4, pp. 273-293, 1987
R. A. Young
Simulation of human retinal function with the Gaussian
derivative model
Proc IEEE omputer Society Conference on Computer Vision and
Pattern Recognition, June 22-26 1986, Miami, Florida
Best regards, Luciano
-----------------------------------------
REPLIES:
--------------------------------------------------------------------------
>
> Dear Dr. da Fontoura Costa,
>
> This paper might be useful to you:
>
> Field D.J. (1993). Scale-invariance and self-similar 'Wavelet' transforms:
> An analysis of natural scenes and mammalian visual systems. In: Wavelets,
> Fractals and Fourier Transforms: New Developments and new applications.
> Farge M., Hunt J., & Vassilicos (Eds.), pp. 151-193. Oxford University
Press.
>
> See you,
>
> Dr. Xavier M. Sauvan
> University Hospital Zurich
> Dept. of Neurology
> Frauenklinikstr. 26
> CH-8091 Zurich
-----------------------------------------------------------------------------
The Gaussian derivative functions are wavelets. They easily operate
at multiple scales by variation of the standard deviation of the
Gaussian envelope, and it is also a simple matter to translate them. Their
original application to vision in 1978 [Young, R. A., Orthogonal
basis functions for form vision derived from eigenvector analysis, ARVO, 1978]
was before wavelet functions were invented, and was based purely on analysis of
physiological data. It is interesting that Mallat and other major wavelet
investigators, who have been applying wavelet functions to the problem of
image representation for some time, have also converged on the
Gaussian derivative functions as being ideal wavelets for image representation
[Mallat, S. G., Multiresolution approach to wavelets in computer vision,
In J. M. Coombs et al., eds., Time-frequency methods and phase space, Proc.
of the Intl. Conf., Marseille France, 2nd ed., pp. 313-327 (1987).] Their
analysis
is based largely on the benefits and ease of mathematical representation
for image analysis. Koenderinck and colleagues [Solid Shape, MIT Press,
Cambridge, Mass. 1990] arrived at the same conclusion based largely on
mathematical considerations (usefulness of multiple-order derivative
estimators for shape analysis). Thus three independent lines of
investigation arrived at a common conclusion - the Gaussian deriviative
functions are very useful for vision.
If you are interested in further information I can send you papers.
- Dick
Richard A. Young Phone: 810-986-1471 (GM:8-226-1471)
Computer Science (AP50) FAX: 810-986-9356
GM Research & Devt. Center Internet: young@gmr.com
Warren, Mich. 48090-9055 GM_PROFS: RYOUNG--GMRCMSA
-----------------------------------------------------------------------------
I have published some works that briefly addresses this question, and
there is also wavelet software available for your experimentation.
The references for the papers are:
Pentland, A., (1993) Surface Interpolation Networks, {\sl Neural
Computation,} Vol. 5, No. 3, 356-442
Pentland, A., (1994) Interpolation using Wavelet Bases, {\sl
IEEE Trans. Pattern Analysis and Machine Vision,} Vol. 16, No. 4, pp.
410-414.
Unfortuately, many of the details of the comparison between human RF
and these wavelets is still waiting for publication in a forthcoming book.
Software implementing these wavelets is available by anonymous FTP
from whitechapel.media.mit.edu in the file ~ftp/pub/wavelet.reg.1.3.tar.Z
Enjoy!
Alex (Sandy) Pentland
----------------------------------------------------------------------------
From: IN%"heeger@white.stanford.edu" 13-JUN-1995 14:05:13.89
To: Luciano@IFQSC.SC.USP.BR
CC:
Subj: Gabors, wavelets, and cortical morphology
Received: from violet.Stanford.EDU by IFQSC.SC.USP.BR; Tue, 13 Jun 95 14:03 BRT
Received: by violet.Stanford.EDU (4.1/inc-1.0) id AA11684; Tue, 13 Jun 95
10:09:44 PDT
Date: Tue, 13 Jun 95 10:09:44 PDT
From: heeger@violet.Stanford.edu
Subject: Gabors, wavelets, and cortical morphology
To: Luciano@IFQSC.SC.USP.BR
Reply-to: heeger@white.stanford.edu
Message-id: <9506131709.AA11684@violet.Stanford.EDU>
X-Envelope-to: Luciano
Here's a long list of references for you. Hope this helps.
- DH
=======================================================================
Kevin Martin's group has done quite a lot of work trying to relate
cortical morphology and function. Here are some of their papers:
@article{Martin84,
author = "K A C Martin and D Whitteridge",
title = "Form, function and intracortical projections of spiny
neurones in the striate visual cortex of the cat",
journal = jphys,
year = 1984,
volume = 353,
pages = "463--504"}
@article{Martin88,
author = "K A C Martin",
title = "From single cells to simple circuits in the cerebral cortex",
journal = "Quarterly Journal of Experimental Physiology",
year = 1988,
volume = 73,
pages = "637--702"}
@article{Douglas87,
author = "R J Douglas and K A C Martin and D Whitteridge",
title = "Estimation of the amplitudes of theoretical unitary
excitatory post-synaptic potentials in neurones of the cat striate
cortex by a combination of biophysical and anatomical measurement",
journal = jphys,
year = 1987,
volume = 394,
pages = "110p"}
% intracellular division vs subtraction
@article{Douglas88,
author = "R J Douglas and K A C Martin and D Whitteridge",
title = "Selective responses of visual cortical cells do not depend on
shunting inhibition",
journal = "Nature",
year = 1988,
volume = 332,
pages = "642--644"}
@incollection{Douglas90,
author = "R J Douglas and K A Martin",
title = "Neocortex",
pages = "389-438",
booktitle = "The synaptic organization of the brain",
editor = "G M Shepherd",
year = 1990,
publisher = "Oxford UP",
address = "Oxford, UK"}
@article{Douglas91a,
author = "R J Douglas and K A C Martin and D Whitteridge",
title = "An intracellular analysis of the visual responses of neurones
in cat visual cortex",
year = 1991,
journal = jphys,
volume = 440,
pages = "659--696"}
@article{Berman91,
author = "N J Berman and R J Douglas and K A C Martin and D Whitteridge",
title = "Mechanisms of inhibition in cat visual cortex",
year = 1991,
journal = jphys,
volume = 440,
pages = "697--722"}
@article{Dehay91,
author = "C Dehay and R J Douglas and K A C Martin and C Nelson",
title = "Excitation by geniculocortical synapses is not vetoed at the
level of dendritic spines in cat visual cortex",
year = 1991,
journal = jphys,
volume = 440,
pages = "723--734"}
@article{Douglas91b,
author = "R J Douglas and K A C Martin and D Whitteridge",
title = "A functional microcircuit for cat visual cortex",
year = 1991,
journal = jphys,
volume = 440,
pages = "735--769"}
@article{Martin-jphys93,
author = "B Ahmed and J C Anderson and R J Douglas and K A C Martin and D
Whitteridge",
title = "A method of estimating net somatic input current from the action
potential discharge of neurones in the visual cortex of the anaesthetized cat",
year = 1993,
journal = jphys,
volume = 459,
pages = "134"}
================================================================================
I've done a lot work modeling V1 responses using wavelet transforms:
@article{Carandini-science94,
author = "M Carandini and D J Heeger",
title = "Summation and Division by Neurons in Primate Visual Cortex",
year = 1994,
journal = "Science",
volume = 264,
pages = "1333-1336"}
@article{Heeger-currdir94,
author = "D J Heeger",
title = "The Representation of Visual Stimuli in Primary Visual
Cortex",
year = 1994,
journal = "Current Directions in Psychological Science",
volume = 3,
pages = "159-163"}
@article{Heeger-jneurophys93,
author = "D J Heeger",
title = "Modeling simple cell direction selectivity with normalized,
half-squared, linear operators",
journal = jneuro,
year = "1993",
volume = "70",
pages = "1885-1898"}
@article{Heeger-visneuro92a,
author = "D J Heeger",
title = "Normalization of cell responses in cat striate cortex",
journal = "Visual Neuroscience",
year = "1992a",
volume = "9",
pages = "181--198"}
@article{Heeger-visneuro92b,
author = "D J Heeger",
title = "Half-squaring in responses of cat simple cells",
journal = "Visual Neuroscience",
year = "1992b",
volume = "9",
pages = "427--443"}
@incollection{Heeger-cmvp91,
author = "D J Heeger",
title = "Nonlinear model of neural responses in cat visual cortex",
pages = "119--133",
booktitle = "Computational Models of Visual Processing",
editor = "M Landy and J A Movshon",
year = 1991,
publisher = "MIT Press",
address = "Cambridge, MA"}
================================================================================
Other papers on V1 and wavelet transforms:
@article{Watson87a,
author = "A B Watson",
year = "1987a",
title = "The Cortex transform: rapid computation of simulated neural images",
journal = "Computer Vision Graphics and Image Processing",
volume = 39,
pages = "311--327"}
@article{Watson89,
author = "A B Watson and A J Ahumada",
year = 1989,
title = "A hexagonal orthogonal-oriented pyramid as a model of image
representation in visual cortex",
journal = "IEEE Transactions on Biomedical Engineering",
volume = 36,
pages = "97--106"}
-------------------------------------------------------------------------
June 13, 1995
Dear Dr. Fontoura Costa,
Here are three papers applying wavelet theory to early vision--
references with abstracts at the end of msg. Any use?
Gaudart-L Crebassa-J Petrakian-JP
Wavelet Transform in Human Visual Channels
APPLIED OPTICS, 1993, Vol 32, Iss 22, pp 4119-4127
Manjunath-BS Chellappa-R
A Unified Approach to Boundary Perception - Edges,
Textures, and Illusory Contours
IEEE TRANSACTIONS ON NEURAL NETWORKS, 1993,
Vol 4, Iss 1, pp 96-108
Casasent-DP Smokelin-JS
Neural-Net Design of Macro Gabor Wavelet Filters for
Distortion-Invariant Object Detection in Clutter
OPTICAL ENGINEERING, 1994, Vol 33, Iss 7, pp 2264-2271
There are two threads in visual science which symbolize my own
early hopes for a simple link between cell morphology to visual
function. They are the Octopus work led by J.Z. Young at the
Stazione Zoologica near Naples. This inspired Marc Collonier to
look at structure/function relationships in mammalian cortex
between dendritic or axonal arborization and RF structure. And
second, there is Jerry Lettvin's work in frog at MIT.
The link between ganglion cell classes and detector types drawn by
Lettvin, Maturana, McCulloch and Pitts (1959) was inferential but
largely correct (Pomeranz and Chung, 1970). However, this does
not necessarily mean that a particular dendritic arbor determines
the size of a particular RF ON-center. It may be that transient
ganglion cells look funny in some particular way, and when we see
that particular morphology, we know the cell has transient
functional properties.
The rigorous link between dendritic arborization and RF size came
from Heinz Wassle, Leo Peichel and Brian Boycott in Frankfurt;
e.g., Wassle, Peichl and Boycott, 1981 in Nature. (Other papers on
the tiling of the retina at end of msg.)
Turning to cortex, the classic claim of a link between dendritic
arborization and orientation processing is Collonier (1964). The
octopus has an enormous "oblique effect," namely superior
performance with horizontal and vertical cotours in behavioral
experiments, as well as some anatomic basis for this in its optic
lobe. Collonier spoke of converting circular LGN input to
elongated cortical RFs by the elongated stellate cell dendritic
arborization fields which he measured. The vertical axis was
favored by the dendrites. He also thought perceptual processing
should be superior along this favored axis, and pointed to the
decrease in acute-angle-expansion figural illusions when their
principal contours are vertical. Unfortunately for this argument,
the strength of many illusions as a function of the configuration's
overall orientation shows TWO minima, at horizontal as well as
vertical (a conventional "oblique effect").
In later work (Collonier and Sas, 1978 on **axonal** arborization) a
link was drawn between morphology and the preferred axis of
motion of V2 cells. (Note that the optimal motion axis is 90 deg to
the optimal orientation for a bar.)
Collonier's suggestions gave us all hope, but they have not been
substantiated (Martin and Whitteridge, 1984).
Why is tiling of the retina an open and shut case, while we still
struggle to link dendritic arborization with receptive field genesis
in cortex? Perhaps we need a technical advance. Changing from
the traditional transverse section to a retinal whole mount aided
the emergence of a simplified classification scheme for the cells.
Perhaps we need a conceptual advance. You can't see the tiling of
the retinal mosaic until you have the correct ganglion cell classes
and the realization that ON and OFF systems within one class tile
the retina separately. Perhaps we don't know the proper
dimensions of the cortex; in the retina, the spatial dimension is
paramount. If you miss a patch, the information is gone forever.
In its functinal architecture, the cortex expresses dimensions other
than spatial ones.
It has become harder to hope for simple links between cortical
morphology and RF function. Consider a V2 cell responding to a
global disparity. This feat probably requires processing
information from beyond the classic receptive field. Consider a V2
or V4 cell responding to a subjective contour induced by figures
outside its classic receptive field. Such feats of global information
pooling may require new concepts of spike transmission which
permit information to be exchanged among many neurons in
several areas during a single, temporally-extended (e.g., 50 ms)
transmission event. If the above examples don't do it for you,
consider face detection, or selective responses to "biological
motion" -- a walking Johansson (1973) figure (Oram and Perrett, J.
Cog Neurosci 1994, 6: 99). The convergence of 100 inputs from 10
areas onto a cell to confer these achievements is a far cry from a
ganglion cell's spreading dendritic umbrella catching X square
degrees of bipolar cell input. The game of linking function to the
topography of excitatory and inhibitory areas, and infering cortical
circuitry from RF topography ("simple cell connects to complex
cell") is over. The topography of ON and OFF areas does not look
like a face or a walking figure.
--jerry
Jerry Nelson, Ph.D.
Lab Neuropsychology
NIMH / NIH Bldg 49 1B80
Bethesda, MD 20892
jnelson@ln.nimh.nih.gov
tel 301/496-5625 ext 235 (lab), ext 274 (voice mail)
FAX 301/402-0046
tel+FAX pm & wkend: 703/448-4543
REFERENCES:
Collonier, M (1964): The tangential organization of the visual cortex.
Journal of Anatomy London 98: 327-344.
Collonier, M, Sas, E (1978): An anterograde degeneration study
of the tangential spread of aons in cortical areas 17 and
19 of the squirrel monkey (Saimiri sciureus).
J. Comp. Neurol. 179: 245-262.
<<"the differential fiber lengths in area 18 are related
to foveal acquisition or to a directional anisotropy of
visual movement perception along sector-zone axes.">>
Martin, KAC, Whitteridge, D (1984): The relationship of
receptive field properties to the dendritic shape of
neurones in the cat striate cortex.
J. Physiol. (Lond.) 356: 291-302.
<<There is no relationship betwen RFs and dendritic shape.
HRP-injected neurons were reconstructed in tangential
views.>>
Lettvin, JY, Maturana, HR, McCulloch, WS, Pitts, WR (1959):
What the frog's eye tells the frog's brain.
Proc. IRE 47: 1940-1959.
Pomeranz, B, Chung, SH (1970): Dendritic-tree anatomy codes
form-vision physiology in tadpole retina.
Science 170: 983-984.
<<Lettvin's evidence for the structure-funciton
relationship was weak, but correct.>>
-------------------------------------------------------------
TILING OF THE RETINA (complete retinal coverage by various
individual ganglion cell sub-types).
Wassle, H, Peichl, L, Boycott, BB (1981): Dendritic territories
of cat retinal ganglion cells.
Nature 292: 344-345.
Peichl, L (1989): Alpha and delta ganglion cells in the rat
retina.
J. Comp. Neurol. 286(1, 1 Aug): 120-139.
(->Max-Planck-Institut fur Hirnforschung, Neuroanatomische
Abteilung, Frankfurt, Federal Republic of Germany.)
<In the rat retina a distinctive class of large ganglion
cell was demonstrated by intracellular staining with
Lucifer Yellow and with reduced silver staining. They are
referred to as alpha cells because they resemble the alpha
cells of other mammalian retinae. A second class, called
delta cells, is also described. Both classes belong to the
type I group defined by Perry (Proc. R. Soc. Lond. [Biol.]
204:363-375, '79). The dendritic trees of both classes
stratify in either an inner or outer lamina of the inner
plexiform layer which presumably corresponds to an on/off
dichotomy in the response to light. Rat alpha cells
constitute 2-4% of all ganglion cells, and their density,
size, and detailed morphological appearance change with
retinal location. Inner and outer stratifying alpha cells
of the rat show significant differences compared to those
of other mammals. In central retina (at the large cell
density maximum) the densities and dendritic field sizes
of inner and outer alpha cells are approximately equal.
However, in peripheral retina outer alpha cells are up to
three times more numerous and have dendritic field areas
only one-third the size of those of the inner alpha cells.
The maximal density is about 110 alpha cells/mm2;
peripheral densities are about 30/mm2. The smallest
central dendritic field diameters are 220 microns.
Peripheral dendritic field diameters are 350-550 microns
for outer and 570-790 microns for inner alpha cells. Each
subpopulation is distributed in a regular mosaic, and the
territorial arrangement of the dendritic fields provides a
homogeneous coverage of the retina. The dendritic coverage
is three- to 3.6-fold for each subpopulation, irrespective
of their other quantitative differences.
Eccentricity-dependent receptive field sizes of the alpha
cells are predicted from the morphological data.>
Buhl, EH, Peichl, L (1986): Morphology of rabbit retinal
ganglion cells projecting to the medial terminal nucleus
of the accessory optic system.
J. Comp. Neurol. 253(2, 8 Nov): 163-174.
<Rabbit retinal ganglion cells were retrogradely labeled
following injection of rhodamine-labeled microspheres into
the medial terminal nucleus. The small fraction of
rhodamine-labeled neurons reached their peak concentration
within the visual streak and then decreased with
increasing eccentricity until none were encountered in the
far periphery. The same rabbits also received injections
of the fluorescent tracer Fast Blue into the superior
colliculus. No double-labeled neurons were observed, i.e.,
ganglion cells projecting to the medial terminal nucleus
(MTN) had no axon collaterals to the superior colliculus.
In fixed retinae rhodamine-labeled ganglion cells were
intracellularly injected with the fluorescent dye Lucifer
Yellow to reveal their full dendritic arborization.
MTN-projecting cells had medium-sized to large somata with
thin and frequently branched dendrites. The large
dendritic trees had a distinct morphology and were
predominantly unistratified in a narrow band that
presumably corresponded to the electrophysiologically
determined on-sublamina of the inner plexiform layer.
Dendritic field sizes were inversely related to ganglion
cell density, thus providing an eccentricity-independent,
constant dendritic coverage factor. Approximately five to
six dendritic fields from neighboring cells cover every
point of the retina. Published reports claim that the
physiological class of on-direction-selective ganglion
cells provides the sole retinal input to the MTN in the
rabbit. In this context morphological features of
MTN-projecting cells and their presumed functional
correlation with on-direction-selective ganglion cells are
discussed.>
Peichl, L, Buhl, EH, Boycott, BB (1987): Alpha ganglion cells
in the rabbit retina.
J. Comp. Neurol. 263(1, 1 Sep): 25-41.
(->Max-Planck-Institut fur Hirnforschung, Neuroanatomische
Abteilung, Frankfurt, Federal Republic of Germany.)
<In the rabbit retina a distinctive morphological class of
large ganglion cells was demonstrated by a combination of
intracellular staining with Lucifer Yellow and the
quantification of reduced silver-stained preparations. The
class is called alpha because of the qualitative and
quantitative resemblance to the alpha cells of the cat's
retina. Rabbit alpha cells change their size with location
on the retina. In the high ganglion cell density region of
the visual streak, their somata are about 15 micron in
diameter, and their dendritic fields have diameters as
small as 180-220 micron. The largest alpha cells in the
inferior periphery have soma diameters of 30 micron and
dendritic field diameters of 960 micron. There is a
considerable scatter of sizes at any retinal location.
Alpha cell density changes from about 55/mm2 in the streak
to about 3/mm2 in far periphery, and the cells make up
1-1.4% of the ganglion cell population. Dendritic trees
stratify in either an inner or an outer sublamina of the
inner plexiform layer, suggesting an on/off dichotomy in
the response to light. Each of the inner and outer
branching subtypes is distributed in a regular mosaic, and
the dendritic trees cover the retina completely and
economically. The possibility is discussed that the alpha
cells are the brisk transient/Y cells of physiology.>
Grunert, U, Greferath, U, Boycott, BB, Wassle, H (1993):
Parasol (P-Alpha) Ganglion-Cells of the Primate Fovea
-Immunocytochemical Staining with Antibodies Against
GABA(A)-Receptors.
Vision Res. 33, N1: 1-14.
(->Max Planck Inst Hirnforsch/Deutschordenstr 46/W 6000
Frankfurt 71//Germany; Umds/Div Anat & Cell Biol/London
Se1 9Rt//England)
63 refs
<Retinae of macaque monkeys were immuno-stained with
antibodies against GABA(A)-receptors. In peripheral retina
most ganglion cells were immunoreactive. In central
retina, around the fovea, staining in the ganglion cell
layer was selective and only 5-8% of all ganglion cells
were labelled: these had the largest cell bodies and their
dendrites occupied a broad stratum in the middle of the
inner plexiform layer. From comparison with Golgi-stained
ganglion cells it is concluded that the entire population
of parasol (Palpha)-cells at the fovea was labelled. The
mosaic and sampling properties of parasol cells were
determined by combining dendritic field measurements of
Golgi-stained cells with their density when
immuno-stained. There is convergence of 30-50 cones onto
each foveal parasol ganglion cell. The dendritic fields of
both ON- and OFF-parasol cells provide complete retinal
coverage. The Nyquist limits of their mosaics are 4 min of
arc.>
--------------------------------------------------------
3 PAPERS on WAVELETS IN EARLY VISION:
Gaudart-L Crebassa-J Petrakian-JP
Wavelet Transform in Human Visual Channels
Source: APPLIED OPTICS
1993, Vol 32, Iss 22, pp 4119-4127
Addresses:
UNIV-AIX-MARSEILLE-3, DEPT OPT PHYSIOL & OPTOMETRIE, AVE
ESCADRILLE NORMANDIE NIEMEN, F-13397 MARSEILLE 13, FRANCE
No. Cited Refs: 39
Abstract:
In this paper we report on an analysis of visual stimuli
models by a wavelet function. The human visual process is
compared with a wavelet transform. Wavelet functions have been
built from the Haar function. Two stimuli were analyzed by a
wavelet function: a sinusoidal luminance stimulus (spatial
frequency f) and a luminance-varying regular stimulus. The
theoretical results obtained from the wavelet transform are
compared with the physiological results of R. Blake [in
Frontiers in Visual Science, S. J. Cool and E. L. Smith, eds.
(Springer-Verlag, Berlin, 1978), pp. 209-219] and K. K. De
Valois [in Frontiers in Science, S. J. Cool and E. L. Smith,
eds. (Springer-Verlag, Berlin, 1978), pp. 277-285]. A
theoretical curve conforms to the shape of the contrast
sensitivity curves. Hence it can be concluded that the wavelet
transform is a new approach to human visual mechanisms.
Authors: Manjunath-BS Chellappa-R
Title: A Unified Approach to Boundary Perception - Edges,
Textures, and Illusory Contours
Source: IEEE TRANSACTIONS ON NEURAL NETWORKS
1993, Vol 4, Iss 1, pp 96-108
Addresses:
UNIV-CALIF-SANTA-BARBARA, CTR INFORMAT PROC RES, DEPT ELECT &
COMP ENGN, SANTA-BARBARA, CA 93106, USA
UNIV-MARYLAND, CTR AUTOMAT RES, DEPT ELECT ENGN, COLL-PK, MD
20742, USA
UNIV-MARYLAND, CTR AUTOMAT RES, DEPT COMP SCI, COLL-PK, MD
20742, USA
UNIV-MARYLAND, INST ADV COMP STUDIES, COLL-PK, MD 20742, USA
UNIV-SO-CALIF, INST SIGNAL & IMAGE PROC, LOS-ANGELES, CA
90089, USA
No. Cited Refs: 35
Abstract:
This paper presents a unified approach to boundary
perception. The model consists of a multistage system which
extracts and groups salient features in the image at different
spatial scales (or frequencies). In the first stage, a Gabor
wavelet decomposition provides a representation of the image
which is orientation selective and has optimal localization
properties in space and frequency. This decomposition is useful
in detecting significant features such as step and line edges at
different scales and orientations in the image. Following the
wavelet transformation, local competitive interactions are
introduced which help in reducing the effects of noise and
changes in illumination. Interscale interactions help in
localizing the line ends and corners, and play a crucial role in
boundary perception. The final stage groups similar features,
aiding in boundary completion. This approach is consistent with
some of the known neurophysiological observations regarding
biological visual information processing, as the different
stages can be identified with processing by simple, complex, and
hypercomplex cells in the visual cortex of mammals. Experimental
results are provided to indicate the performance of this model
in detecting boundaries (both real and illusory) in real and
synthetic images.
Authors: Casasent-DP Smokelin-JS
Title: Neural-Net Design of Macro Gabor Wavelet Filters for
Distortion-Invariant Object Detection in Clutter
Source: OPTICAL ENGINEERING
1994, Vol 33, Iss 7, pp 2264-2271
Addresses:
CARNEGIE-MELLON-UNIV, CTR EXCELLENCE OPT DATA PROC, DEPT ELECT
& COMP ENGN, PITTSBURGH, PA 15213, USA
No. Cited Refs: 14
Abstract:
We consider the detection of multiple classes of objects in
clutter with 3-D object distortions and contrast differences
present. We use a correlator because shift invariance is
necessary to locate and recognize one object whose position is
not known and to handle multiple objects in the same scene. The
detection filter used is a linear combination of the real part
of different Gabor filters, which we refer to as a macro Gabor
filter (MGF). A new analysis of the parameters for the initial
set of Gabor functions in the MGF is given, and a new neural
network algorithm to refine these initial filter parameters and
to determine the combination coefficients to produce the final
MGF detection filter is detailed. Initial detection results are
given. Use of this general neural network technique to design
correlation filters for other applications seems very attractive.
end