Meas. Sci. Technol. 8 (1997) 13991405.

Printed in the UK

PII: S0957-0233(97)84508-0

The use of neural techniques in PIV

and PTV
I Grant and X Pan
Fluid Loading and Instrumentation Centre, Heriot-Watt University,
Edinburgh EH14 4AS, Scotland, UK
Received 29 May 1997, accepted for publication 6 October 1997
Abstract. The neural network method uses ideas developed from the physiological
modelling of the human brain in computational mechanics. The technique provides
mechanisms analogous to biological processes such as learning from experience,
generalizing from learning data to a wider set of stimuli and extraction of key
attributes from excessively noisy data sets. It has found frequent application in
optimization, image enhancement and pattern recognition, key problems in particle
image velocimetry (PIV). The development of the method and its principal
categories and features are described, with special emphasis on its application to
PIV and particle tracking velocimetry (PTV). The application of the neural network
method to important categories of the PIV image analysis procedure is described in
the present paper. These are image enhancement, fringe analysis, PTV and stereo
view reconciliation. The applications of common generic net types, feed-forwards
and recurrent, are discussed and illustrated by example. The key strength of the
neural technique, its ability to respond to changing circumstances by
self-modification or regulation of its processing parameters, is illustrated by
example and compared with conventional processing strategies adopted in PIV.

1. Introduction
1.1. Particle image velocimetry
Particle image velocimetry (PIV) is a technique which
allows the velocity of a fluid to be simultaneously measured
throughout an illuminated region, which is commonly
planar. Seeding, flow-following particles are introduced
into the flow and their motion is used to estimate the
kinematics of the local fluid. The velocity of the particles
is recorded using multiple exposure, quantitative, image
capture methods.
The extraction of fluid velocity information from
multiple-exposure images on film, holographic plate or
CCD sensor array is the key process in PIV data analysis.
The methodologies adopted to solve the paradigm have
been defined largely by the historical perspective of
the investigators, drawing on techniques and hardware
developed in parallel studies using image capture and
analysis in other regimes.
For low-density images, for which the individual
particle images are readily distinguishable, particle tracking
may be used.
When the seeding density is high,
image correlation methods are often adopted, allowing a
local average to be obtained. Early attempts at highdensity analysis also used image-transparency interrogation
by laser, producing a fringe pattern from which flow
parameters were derived. A full description of the method
has been given (Grant 1994, 1997).
Emphasis has recently been placed on obtaining
the third velocity component, normal to the illuminated
c 1997 IOP Publishing Ltd
0957-0233/97/121399+07$19.50

plane, in PIV studies. No universal scheme for three

velocity component measurements has so far been adopted.
The methods used all have considerable experimental
and processing difficulties. Most approaches may be
classified as stereographic (Chang et al 1984), holographic
(Thompson 1989) or multiple plane (light sheet) in
methodology (Utami and Ueno 1984) with some studies
combining the techniques (Grant et al 1991). Variants
of the stereogrammetry method making the method more
robust and simple to apply (Grant and Pan 1995) have been
reported.
1.2. The neural network
1.2.1. An introduction to the concept. The human
brain is thought to consistent of a three-dimensional
matrix of interconnected processing units, neurons, and
has the capacity for implementing simultaneous, non-linear,
processing strategies. Typically the brain consists of 1010
neurons, each of which is interconnected to 104 other
neurons. The input signals from other neurons are modified
by the interconnection efficiencies or weights. Each neuron
sums, or integrates, the net inputs and then acts as a
processing unit in that it either triggers, giving an electrical
output, or remains dormant. The synaptic connections
between the neurons are used to hold memories which
can be modified and updated both during the learning
(initialization) stage and during processing of data sets.
The distributed nature of the processing elements is used
in many neural networks.
1399

I Grant and X Pan

An important characteristic of neural networks which

lends them a degree of superiority over other systems is
their ability to learn by example, adapting their weights
in a manner determined by the processed data. A further
advantage of the neural network is its ability to tolerate
noise in an input pattern. If a network has been trained
sufficiently, it is capable of performing well even when
input patterns are noisy or incomplete.
1.2.2. The historical development of the neural network.
The first form of the neural network was devised by
McCulloch and Pitts (1943). They defined the adaptive
stimulusresponse neuron model. Two early rules for
training adaptive neuron elements were the perceptron rule
and the LMS algorithm (Widrow and Hoff 1960).
In the 1970s the adaptive resonance theory (ART)
was developed, a number of novel hypotheses about the
underlying principles governing biological neural systems
(Grossberg 1976). These ideas served as the bases for later
work involving three classes of ART architectures: ART1,
ART2 and ART3 (Carpenter and Grossberg 1983, 1987,
1990). These are self-organizing neural implementations of
pattern-clustering algorithms. Another important approach
to self-organizing systems was pioneered by Kohonen
(1982) with his work on feature maps.
In the early 1980s, the outer product rules and
equivalent approaches based on the early works of Hebb
for training a class of recurrent (signal feedback) networks,
now called the Hopfield model, were introduced (Hopfield
1982, 1984). More recently, Kosko (1987) extended
some of the ideas of Hopfield and Grossberg to develop
his adaptive bi-directional associative memory (BAM), a
network model employing differential as well as Hebbian
competitive learning laws.
1.2.3. The neuron model. A representation of the basic
features of a neuron is shown in figure 1. The inputs to
the neuron arrive along the dendrites, which are connected
to the outputs from other neurons by specialized junctions
called synapses. These junctions alter the effectiveness
with which the signal is transmitted. The axon serves
as the output channel of the neuron. It is also a nonlinear threshold device, producing a voltage pulse when the
activation level within the cell body rises above a certain
critical value.
The neuron model may be considered as a multipleinput, single-output operator (figure 2). The Xi are the input
signals from other neurons, the Wi are the interconnecting
weights and Y is the output signal.
In computational implementation each input is multiplied by a corresponding weight, analogous to a synaptic
inter-connective strength, and all of the weighted inputs
are then summed to determine the activation level of the
neuron, denoted NET. The NET signal is further processed
by an activation function f () to produce the output signal
of the neural computational element. The activation function f () may be, for instance, binary, sigmoid or threshold
linear (figure 3).
1400

Figure 1. The basic features of a biological neuron.

Figure 2. The artificial neuron.

Figure 3. The three representative non-linear activation

functions.

Figure 4. The feed-forwards network structure.

1.2.4. The network structure. The advantages of neural

computing are realized when the elements are connected
into networks. A group of neurons may be arranged in
a layer. Several layers may be interconnected. There are
two basic network communication structures used within
the layers, namely the feed-forwards network and the feedback network.

The use of neural techniques in PIV and PTV

The learning algorithm modifies the network weights to

produce outputs that are consistent, meaning that the
application of similar inputs will produce the same output
pattern. Therefore the network looks for regularities or
trends in the input signals. The learning process extracts
the statistical properties of the inputs and groups similar
input patterns into classes. Unsupervised learning is used,
for example, in the Kohonen self-organizing network and
in adaptive resonant theory (ART).
Figure 5. The feed-back network.

(i) In the feed-forwards network the output from a node,

or neuron, is passed on as input to other nodes in the
following layer without any feedback to nodes in the same
layer or the preceding layer (figure 4). The layers other
than the input and output layers are called hidden layers.
They act as filters. For example, the input signal could be
the pixel pattern for the letter A. The network could then
generate an output pattern, for example 010000001 (the
ASCII code for the input letter A). Even an imperfect or
partial letter could be recognized, depending on the prelearning examples and activation functions used.
(ii) Figure 5 shows a typical feed-back or recurrent
network. Each node receives the input signal from, and
sends an output signal to, every other node. The Hopfield
neural network is an example of this type of representation.
Neural networks developed from the two structures have
different properties and, therefore, different applications.
Some neural networks use both of the structures. The
characteristics of these structures are illustrated in the neural
network models described in the following sections.
1.2.5. Learning. The most important property of a
neural network is the ability to learn. The process of
learning consists of the neural network modifying its
weights in response to external inputs. The learning
rule determines the neuron inter-connective weights
modification algorithm. There are two types of learning:
supervised and unsupervised. Learning must take place
before the net becomes operational.
(i) In supervised learning a training set consists of both
inputs and target outputs. A training input is presented
to the net and the output of the network is calculated and
compared with the desired output. The difference is fed
back through the network in such a way that the weights are
modified to minimize the error between the desired output
and the current output. The same procedure is repeated
until the error has been reduced to an acceptably low level.
A typical supervised learning rule is back propagation.
Using this technique, a network can be made to memorize
information from the training set and perform decision
making based on these rules. However, a large training
set may be needed in order to produce satisfactory decision
making and the training phase can be very time consuming.
(ii) Unsupervised learning requires no target outputs
and consequently no comparisons with predetermined ideal
responses. Only the inputs are used to train the network.

1.2.6. Neural network models. Many neural network

models have been developed for specific applications. The
following are the most widely used models.
(i) The back propagation (perceptron) network is used
mainly to recognize input patterns with pre-defined and
learnt classes.
(ii) The Kohonen self-organizing network is capable of
constructing clusters of input patterns without knowledge of
their distribution. This property was used by Grant and Pan
(1993, 1995) in their multi-layer neural networks designed
for the analysis of PIV images.
(iii) The ART network is also used to classify input
patterns. Its advantage is that it can generate new classes
to cope with a continually varying environment without
destroying or damaging previously learned information.
(iv) The Hopfield neural network works differently
from the other network models.
The algorithm of
a Hopfield neural network is equivalent to an energy
optimization function derived for a specific case in a manner
defined by the mathematical model of the system.
2. Neural network applications in PIV
The analysis of PIV images typically involves pattern
recognition, classification and feature extraction. A neural
network used in pattern recognition will, in general, have
recognition rules established through the use of training
data sets or by self-learning. From this initialization the
network is able to proceed with independent identification
and decision making. The consequences of its operation
may be used in some cases to alter its rules or update its
memory. This adaptive approach is used in most neural net
models.
2.1. The neural net applied to tracking mode PIV
(PTV)
2.1.1. Introduction. Particle image tracking is a key
procedure in analysis of low-density PIV images. The
nearest neighbour and statistical analysis methods have
been used to group the particle images and quantify the
local displacement (Grant and Liu 1989). However, these
methods are unsatisfactory for flows having rapid direction
changes. The neural network method has been used for
the image tracking recognition to improve the efficiency of
analysis.
Work on the use of the neural network in tracking PIV
was reported by Teo et al (1991). A fuzzy ART (figure
6) was used to match the particle images from the second
1401

I Grant and X Pan

Figure 6. Adaptive resonance theory (ART).

frame to the images in the first frame. It consisted of two

layers of interconnected neurons. The bottom-up weights
determined the classes fuzzy maximum reference point and
top-down weights determined the classes fuzzy minimum
reference point. During processing, one frame was used
to create matching classes, then particle images in another
frame were matched to the most similar class in an optimal
way controlled by a vigilance parameter. The method was
reported to work well for particles that were well dispersed
(low image density) and for images that were taken a short
time apart. Cenedese et al (1992) studied a multi-layer
feed-forwards neural net and demonstrated the potential for
neural net methods applied to PTV.
The back-propagation network is a feed-forwards
network using a back-propagation learning algorithm. This
type of network has been used to distinguish all the image
pairs in a PIV image and provide different labels for each
pair (Yen Jia-Yush, Chen Ping-Hie and Chen Jian-Liang,
private communication). The velocities of all the identified
pairs are calculated and averaged to produce an average
velocity. The experimental results show that the results
from the proposed neural network match well those of the
auto-correlation process in the uniform flow region and a
78.1% success ratio in the stagnation flow region has been
claimed. This method is more suitable for uniform flow
than it is for turbulent conditions.
An analogous approach to the net optimization-function
approach has been adopted by Okamoto et al (1996), who
used a mechanical analogy involving spring interconnecting
constraints between particle images to define an energy
function which was minimized to obtain the best match
between particle images in two dimensions.
2.1.2. A case study: multi-layer, feed-forwards network
for time-coded, PTV image analysis. Grant and Pan
(1993, 1995) demonstrated the use of three- and fourlayer, feed-forwards neural net models (figure 4). They
were used successfully as competitive, adaptive filters in
feature-recognition tasks typical of PTV. The four-layer
net was able to distinguish directions in flow images
which carry a time signature. The models exhibited a
substantial improvement in performance compared with
1402

Figure 7. The input and output of the three-layer neural

network.

earlier statistical, non-neural, windowing methods (Grant

and Liu 1989, 1990).
Every layer consisted of the same number of nodes
distributed on a two-dimensional plane. The first layer
acted as an input buffer, whereas the following layers acted
as filters or selection modules. The filtering characteristics
of the layers were stored in the interconnecting weights
and were acquired during the processing stage, following a
self-learning algorithm approach inspired by the Kohonen,
self-organizing, feature map.
The analysis procedure commenced with a doubleexposure, low-image-density, PIV image being preprocessed to extract particle image centre pixels. The image
was then segmented into sub-images, each of which held
a representative particle image centre at its origin. The
sub-image dimensions were chosen to allow all possible
image partners, for the particle image at the origin, to fall
within the segment. A two-dimensional array was used to
represent the segment pixel by pixel (figure 7).
The segment was passed to the input layer of the
neural network ready for processing. The spatially adjacent
segments were consecutively processed. This was an
important consideration since it allowed the memory of
the flow displacements to be updated in a systematic and
meaningful fashion.
In the three-layer net the element of the array was set
to 1 if a particle image centre appeared at that pixel;
otherwise it was set to 0. A competition took place in
a winner-takes-all fashion on the output layer and gave the
best matched partner. If an element of the output array was
identified by the net as a partner particle image, it gave a
1 output while the others were 0. If all the elements in
the output layer were 0, no pair had been matched.
The four-layer net was applied to tagged PIV
images which had been obtained using the image coding
method whereby the image size was varied by altering
the brightness of the laser flash (Grant and Liu 1990).
Alternatively, if an electronically shuttered camera is used
with continuous illumination, the duration of the shutters
open time may be used in the coding (Grant and Wang
1995). The method is easily adapted for use in colourcoded images.

The use of neural techniques in PIV and PTV

Figure 8. (a) A typical turbulent flow simulation with a systematic change in direction. (b) The success ratio (SR) as a
function of the angular range. (c) The success ratio (SR) as a function of the image density.

The input of the four-layer net held the area information

of each particle image. The directionally sensitive layer was
inserted immediately after the input layer. Each node in this
layer had two inputs from the input layer. It compared the
sizes of the two input images and activated an output giving
the correct flow direction.
(i) The performances of the three- and four-layer
networks were compared with the conventional statistical
method. Local means and standard deviations of the
velocity were obtained using sub-image sampling on a
statistically significant numbers of particle image pairs
(Grant and Liu 1989). A significant improvement was
found using synthetic turbulence data with a systematic
change in direction. (figure 8(a)). Figure 8(b) and (c) are
the success ratios of matching as a function of the angular
change of direction and image density respectively.
(ii) In a first experimental application, a vortex
generator was used to produce a wing-tip type vortex A
Nd:YAG laser illuminated a plane normal to the mean flow.
The images, captured on 35 mm film, were processed using
the multi-layer, feed-forwards neural method (Grant and
Pan 1995). The image pairs were found to be correctly
matched on 91.7% of occasions (figure 9(a)). Conventional
statistical tracking does not monitor such flows efficiently
when large local changes in direction occur.
(iii) In a second experimental application, the flow
behind a circular cylinder was measured using laser
illumination in a plane containing the mean flow vector and
the normal to the cylinders principal axis. The illumination
was given a time signature (Grant and Liu 1990) and
the four-layer neural net was used (Grant and Pan 1995)
to measure the direction and magnitude of the particle

velocities. A success ratio of 95.9% was obtained (figure

9(b)).
2.2. Neural networks used in feature extraction for
PIV
2.2.1.
Introduction. Image enhancement techniques
are often applied to PIV images during a pre-processing
stage to aid extraction of particle image characteristics.
Owing to the noise often present in the PIV images,
normal filtering, stretching, binarization methods are not
necessarily adequate. Neural network methods have been
used widely to solve the image recognition and feature
extraction problems. Successful applications can be found
in object recognition, edge detection and image coding.
2.2.2. A case study: identifying particle image centres.
Carosone et al (1995) used the Kohonen neural network
(figure 10) in the recognition of partially overlapping
particle images. The Kohonen network worked as an
optimum classifier. It allowed single particle images to be
distinguished from overlapping particle images by shape
analysis. The input to the network was a vector set
consisting of geometrical features of a particle spot, such
as the first, second and third circularity measure, the aspect
ratio and the convexity. These parameters were invariant on
translation, rotation and scaling of an image. The network
classified the particle into two classes: single or overlap.
Different methods of calculating the image barycentre were
applied afterwards, according to the images class property.
Single-exposure and multiply exposed synthetic images
were tested. The neural network produced a sharp increase
in the number of identified barycentres in PTV images with
many overlaps.
1403

I Grant and X Pan

Figure 9. (a) The matched image pairs obtained from the tip vortex image using the neural net method. (b) The matched
image pairs obtained from the vortex shedding image using the neural net method.

Figure 10. A Kohonen network.

2.2.3. A case study: obtaining fringe characteristics.

Carosone and Cenedese (1996) applied a back-propagation
(BP) neural network to Youngs fringe pattern analysis,
which had the advantage of significantly accelerating the
analysis compared with the inverse FFT method. The
authors reported that a complexity analysis showed that
the BP neural network reduced the computational load by
a factor of 200 when run on the same computer. The
fringe image (256 256 pixels) was first compressed with
a Kohonen FSCL network to the size of a 256-word string.
It was sent to the BP network which employed 256 neurons
in the input layer, 96 neurons in the first hidden layer, 11
on the second hidden layer and two in the output layer. The
two outputs represented the fringe spacing and orientation.
The network was trained with large fringe images before
operation. The accuracy of the network was reported to be
96.3% orientation and 93.6% spacing recognition.
2.3. Using the neural network to solve the stereo vision
reconciliation problem
2.3.1. Introduction. Stereogrammetry is one of the
methods used in three-dimensional measurement. It utilizes
two or more images obtained from different view points to
estimate the three generalized co-ordinates of a point in
object space. A major difficulty in stereogrammetry, which
has lead to the most significant errors in general, is view
reconciliation, whereby sub-images on the two views are
1404

matched to identify common areas representing identical

regions of the object volume.
By using a neural network approach it is possible
to include constraints in the solution algorithm through
the definition of a cost function representing the system
constraints which is minimized using a distributed Hopfield
network (Hopfield 1982, 1984, Hopfield and Tank 1984).
The stereo view correspondence problem is developed as
an optimization problem with a cost, or energy, function
representing the system constraints. The approach taken
below (section 2.3.2) is to define a Lyapunov function
which represents the overall behaviour of the network and
which has a local minimum value when the network is in a
stable state, for which a best match of the views is obtained.
The importance of the Hopfield net (figure 5) is that,
while holding information about the system, it may take up
dynamically stable states. The optimization of the system
proceeds until the (system-defined) energy function finds
its way to one of these minima, where it becomes trapped.

2.3.2. A case study: matching stereogrammetry view

pairs obtained in PTV. Grant et al (1997) have applied
this approach to the reconciliation of stereo image pairs
obtained in PIV studies. Each of the stereo views is preprocessed to remove noise and to identify particle images.
Each particle image on the respective view is identified by
an index number. The stereo correspondence problem is
then represented by a two-dimensional array of neurons
having row and column numbers i and j respectively.
The row and column indices are used to represent the ith
neuron in the first stereo view and the j th neuron in the
second stereo view. A positive output from the neuron
(i, j ) indicates a possible match.
High success ratios have been reported for all image
densities. The performance of the network, applied to
simulated flow data for calibration purposes, was presented
as a success ratio curve plotted against the particle matching
density. This was defined as the averaged number of
candidates in each corresponding points interrogation area.
In most instances the success ratio was greater than 97%,
approaching 100% at lower image densities.

The use of neural techniques in PIV and PTV

3. Conclusion
The development and classification of the neural network
has been described with particular emphasis on areas
of possible application in particle image techniques.
In particular, image enhancement, feature extraction,
classification and optimization procedures have been
discussed and their particular usage in PIV described.
The unique self-learning capability of the neural
network has been shown to be of great use in PTV flow
diagnostics when the flow is changing significantly over the
recorded region. In particular, the ability of a network to
modify both the parameters defining its filtering procedures
and the rules chosen for evaluation makes the neural method
a powerful methodology suitable for the processing of
images obtained from complex flows.
The use of the method has been discussed by citing
its application in image tracking, fringe analysis, image
enhancement and stereo view reconciliation. The success
of its application in solving these diverse and complex
problems is an indication of its likely further development
and application in the field of PIV and PTV image
processing.
References
Carasone F and Cenedese A 1996 Youngs fringes analysis by
neural network Proc. 8th Conf. on Laser Anemometry,
Lisbon (Lodoan-IST) paper 21
Carosone F, Cenedese A and Querzoli G 1995 Recognition of
partially overlapped particle images using the Kohonen
neural network Exp. Fluids 19 22532
Cenedese A, Paglialunga A, Romano G P and Terlizzi M 1992
Neural net for trajectories recognition in a flow Int Symp on
Laser Anemometry, Lisbon paper 27.1
Carpenter G A and Grossberg S 1983 A massively parallel
architecture for a self-organising neural pattern recognition
machine Computer Vision Graphics Image Processing 37
54115
1987 Art2: self-organisation of stable category recognition
codes for analog output patterns Appl. Opt. 26 491930
1990 Art3 hierarchical search: chemical transmitters in
self-organising pattern recognition architectures Proc. Int.
Joint Conf. on Neural Networks, Washington, DC vol 2,
pp 303
Chang T P, Wilcox N A and Tatterson G B 1984 Application of
image processing to the analysis of three-dimensional flow
fields Opt. Eng. 23 2837 (reprinted in Grant (1994))
Grant I (ed) 1994 Selected Papers on PIV (Bellingham, WA: The
Society of Photo-Optical Instrumentation Engineers)
1997 Particle image velocimetry: a review Proc. Inst. Mech.
Eng. C 211 5576
Grant I, Fu S, Pan X and Wang X 1995 The application of an
in-lin, sterescopic, PIV system to 3 comonent velocity
measurements Exp. Fluids 19 21422

Grant I and Liu A 1989 Method for the efficient incoherent

analysis of particle image velocimetry images Appl. Opt. 28
17458 (reprinted Grant (1994))
1990 Directional ambiguity resolution in particle image
velocimetry by pulse tagging Exp. Fluids 10 716
(reprinted Grant (1994))
Grant I and Pan X 1993 The neural network method applied to
particle image velocimetry Proc. Optical Diagnostics in
Fluid and Thermal Flow, SPIE Technical Conf. 2005,
San Diego Paper 2005-43
1995 An investigation of the performance of multi layer
neural networks applied to the analysis of PIV images Exp.
Fluids 19 15966
Grant I, Pan X, Wang X and Romano F 1997 The neural
network method applied to stereo image correspondence
problem in 3 component PIV Appl. Opt. submitted
Grant I and Wang X 1995 Directionally-unambiguous, digital
particle image velocimetry studies using a image intensifier
camera Exp. Fluids 18 35862
Grant I, Zhao Y, Tan Y and Stewart J N 1991 Three component
flow mapping; experiences in stereoscopic PIV and
holographic velocimetry Proc. Fourth Int. Conf. on Laser
Anemometry, Advances and Applications (New York:
American Society of Mechanical Engineers) pp 36571
(reprinted in Grant (1994))
Grossberg S 1976 Adaptive pattern classification and universal
recording: Feedback, expectation, olfaction and illusions
Biol. Cybernetics 23 187202
Hopfield J 1982 Neural networks and physical systems with
emergent collective computational abilities Proc. Natl Acad.
Sci., USA 79 25548
1984 Neurons with graded response have collective
computational properties like those of two-state neurons
Proc. Natl Acad. Sci., USA 81 308892
Hopfield J and Tank D W 1984 Neural computation of
decisions in optimisation problems Biol. Cybernetics 52
14152
Kohonen Th 1982 Self-organised formation of topologically
correct feature maps Biolog. Cybernetics 43 5969
Kosko B 1987 Adaptive bidirectional associative memories
Appl. Opt. 26 494760
McCulloch W S and Pitts W A 1943 A logical calculus of the
ideas immanent in nervous activity Bull. Math. Biophys. 5
11533
Okamoto K, Koizumi M, Madarame H and Hassan Y 1996
Temporal mapping of the spring model for particle image
velocimetry Proc. Eighth Int. Symp. on Application of Laser
Techniques to Fluid Mechanics (Lisbon) paper 27.1
Thompson B J 1989 Holographic methods for particle size and
velocity measurementrecent advances Proc. SPIE 1136
30826 (reprinted Grant (1994))
Teo C L, Lim K B, Hong G S and Yeo M H T 1991 A neural
net approach in analysing photographs in PIV Proc. IEEE
Int. Conf. on Systems, Man and Cybernetics, Charlottesville,
VA vol 3, pp 15358
Utami T and Ueno T 1984 Visualization and picture processing
of turbulent flow Exp. Fluids 2 2532 (reprinted Grant
(1994))
Widrow B and Hoff M E 1960 Adaptive switching circuits 1960
IRE Western Electric Show and Convention Record vol 4,
pp 96104

1405