Cortical Maps of Resistive Anisotropic Networks (CORMORANT)
Funding Agency: CEC
Funding period: 1994-1997
URL:
Contact Person
Giacomo.M. BISIO
Department of Biophysical and Electronic Engineering
University of Genoa
Via Opera Pia, 11a
I-16145 Genova, ITALY
tel: (+39) 10 353 2756
fax: (+39) 10 353 2777
e-mail: bisio@dibe.unige.it
Participants
Coordinator:
DIBE-University of Genoa I
Partners:
IRST (Trento) I
Ruhr Universitat Bochum D
University of Bonn D
Abstract
The project will investigate how the
information processing capabilities of neural networks can be exploited in
the direct (VLSI) implementation of complex algorithmic tasks such
as those occurring in natural and artificial visual systems.
Research will address:
(i) the computational modeling of natural systems as formal neural networks;
(ii) the formalization of neural computation through an abstract
hierarchy of operators both for low- and high-level capabilities;
(iii) the architectural specification of networks of these operators,
specified as multilayer meshes of simple processing elements (lattice
networks);
(iv) the methodological issues related to the
programmability/design of these meshes specified at the circuit level.
Results
Analog approaches to the hardware implementation of machine vision systems
can be very effective as far as real-time performance, size, and power
consumption are concerned, but machine functionality, with regard both to
the variety of tasks to be performed, and to the conditions of application,
can be hindered by the limited flexibility and lack of programmability of
analog circuits.
A solution to this problem can be searched through a systematic approach to
the sensorial/perceptual computations, that goes from proper
model/algorithmic formulations, through adequate architectures
based on flexible primitives, to efficient VLSI circuital solutions.
The target machine vision problem considered for proving these statements
is
the estimation of depth
maps on the basis of phase-based disparity measurements. The proposed solutions are
validated focussing on chip functionality, tested at layout level and
through fabrication, and on the basis of specific computer
simulations in artificial and real environments.
Architectural and circuital solutions for microsystems to be used in
the emulation of
early vision tasks
The single-chip microsystem is characterised by three
levels of adaptability:
- a variety of algorithms (e.g., stereo depth
estimation, texture and motion analysis) can be specified in terms of
convolutions with spatial Gabor-like filters;
Gabor-like kernels can be related to the impulse response of
recurrent filters, specified as cooperative computational structures (i.e.,
lattice networks);
- the phase of these kernels can be varied through linear combinations
of
the recurrent filter responses at nearby nodes;
- the frequency and the
spatial extension of the Gabor-like kernel can be controlled by analog
tunable
circuits.
Analog VLSI Gabor convolution kenels have been prototyped (see www.prosoma.lu).
Validation of the lattice approach under VLSI constraints
- From the perspective of analog computation, an algorithm is
implementable if it can be reduced to local operations. The
phase-based approach to disparity estimation is local except for the
convolution operation. However, if the Gaussian envelope is replaced
by other kernels, convolution can be transformed into the solution of
a set of differential equations whose degree is related to the number
of connections per node necessary to implement the filter in a
discrete model.
- The first and simplest choice, i.e. the Cusp kernel that
presents a discontinuity in the first derivative, leads to a fourth
degree differential equation. A successive approximation of the
Gaussian kernel leads to the Quartic filter that improves the
smoothness of the disparity-field estimates but requires higher degree
differential equations. The modifications of the system's performance
due to the envelope's change are investigated in detail. The results
of the simulation experiments indicate that filters'
performances are comparable if they are modified in order to eliminate
sensitivity to the zero-frequency content of input signals.