ARTIF ICIAL

VISION

By
P.Sivasankar ,
ECE ,REC
Sivas1991@gmail.com
9600158854
Shyam paramasivam
ECE.REC
Shyam2291@gmail.com
9840350997

ABSTRACT
This paper focuses on two types of
neural networks; Space-variant
processing
Cellular
Neural
Networks, the other is SelfOrganized Fuzzy Inference Neural
Network in the view of achieving
artificial perception. Both the neural
networks can be considered as feedforward neural networks which
contain the ability of parallel
processing. Use of Cellular Neural
Networks is proposed as a tool for
artificial vision.
INTRODUCTION
In humans vision system, texture is
a basic cue for human beings to
recognize different kinds of objects
in the world. Moreover strict
constraints of size, weight, cost and
performance must be jointly
optimized in order to achieve proper
vision. For such reasons, the concept
of space-variant active vision
(SVAV) has been implemented
which utilises the properties of
Gabor filter.
In this paper, we will perform a
feasibility study, by considering
design of a foveated CNN for the
computation of optical flow, through
Gabor filtering.

OPTICALFLOW
COMPUTATION ON
LOG-POLAR GRIDS
Optical flow is the pattern of
apparent motions of objects,
surfaces, and edges in a visual scene
caused by the relative motion
between an observer and the scene.
In other words it is defined as the
vector field of apparent velocity of
objects in the image plane.

A more efficient and also

biologically plausible approach is
based on the estimation of energy
flow, with the assumption that
energy is locally conserved in both
space and time. Therefore, motion
can be viewed as orientation in
space-time, and velocity computed
by estimating location of energy
peaks in the spatio-temporal
frequency domain.
A very convenient tool to do this is
filtering with Gabor functions. A
Gabor filter is defined by an impulse
response that is a complex
exponential modulated by a
Gaussian function. A human vision
system contains 2D Gabor-type
Filtering, which performs quite
important functions in human visual
experiences. It is nothing but an
adaptive band-pass filtering method
that constructs a complete non

orthogonal basis set. In this paper,

we will consider space-time filters
obtained as product of two spatial
Gabor filters- and one time-filter
that is capable of extracting objects
moving with certain velocity.
Clearly, in this approach, huge
number
of
Gabor-filtering
processing has been required. This
will be a problem if we use
traditional digital signal processing.
Thus, an effective filtering approach
is required. Hence 2Gabor-type
filtering based Cellular Neural
Networks (CNN) are used. Based on
extremely high speed and high
performance nonlinear algorithm,
real-time Gabor type filtering
operation has been implemented.

an image buffer. The input image

has been loaded in this buffer and
distributed into next layer.

Self-Organized Fuzzy Inference

Neural Networks

Gabor-type Filtering CNN Array

Layer:

Self-Organized Fuzzy Inference

Neural Network, also known as
SOFIN, is a fuzzy neural network
which provides self-adaptive ability.
Here, SOFIN has been used to
simulate two major ability of
humans brain; one is inference, the
other is
memory. Following
describes the major function of
SOFIN.

This layer is used to collect the

information of texture which exists
in the input images. Clearly, second
layer looks like an array which is
constructed
by
many
CNN
processors. Each CNN processor
provides Gabor-type filtering with
different parameters.

Input Layer: This layer actually

does not contain any computing. In
fact, this layer can be considered as

Rectification Layer: This layer

provides two functions; one is
rectification operator, the other is to
collect the output from different
Gabor-type
Filtering
CNN.

Rectification operator is used to

compensate the effects of phase of
output of Gabor-type Filtering. In
other words, we only consider about
the degree of difference between
different responses of Gabor-type
filtering. Each node in this layer will
collect the output from different
Gabor-type Filtering CNN. Thus, the
output of this layer becomes to
multi-dimension values.

Output Term Layer: This layer can

be considered as a consequent
operator.
Output Linguistic Layer: This
layer is the defuzzification layer. In
training stage, the desired values
will be assigned in this layer and
testing stage, the testing results will
be generated by this layer.

Input Linguistic Layer: This layer

is the Layer 1 of SOFIN. This layer
contains no any computation but
uses to store the input variable. For
details please refer to previous
chapter.
Input Term Layer: This layer is the
Layer 2 of SOFIN. This layer
provides fuzzification function for
the input variables. In proposed
algorithm, previous node will
provide multi-dimension values.
Thus, this layer needs to fuzzify
each dimension in every node and
then, this layer needs to distribute
the value of each node and each
dimension to next layer. Here
Gaussian Probability Distribution
Function is used as the membership
function.
Rule Layer: This layer is used to
store the training results. Each
training stage maybe (or maybe not)
grow the connection and the scale of
this layer.

CONCLUSIONS
In this paper, it has been shown that
optical flow can be computed onchip together with foveated sensing
by cellular neural networks. Spacevariant templates have been design
for this purpose. A very promising

fact about feasibility of foveated

CNN ICs is that the number of
sensors, and therefore of cells,
necessary for real-world tasks is
very limited and already in the range
of realized prototypes. Current
studies are oriented at the design of
vision systems for autonomous
robots based on space-variant
CNNs, that will be employed for the
estimation of ego motion, obstacle
avoidance and object recognition.