Image Processing using Pulse-Coupled Neural Networks

153 Pages·1998·9.163 MB·English

by  Thomas Lindblad PhD, Jason M. Kinser DSc (auth.)

Perspectives in Neural Computing
Springer-Verlag London Ltd.
Alsointhisseries:
J.G.Taylor
ThePromiseofNeuralNetworks
3-540-19773-7
MariaMarinaroandRobertoTagliaferri(Eds)
NeuralNets-WIRNVIETRI-96
3-540-76099-7
AdrianShepherd
Second-OrderMethodsforNeuralNetworks:FastandReliableTrainingMethodsfor
Multi-LayerPerceptrons
3-540-76100-4
JasonKingdon
IntelligentSystemsandFinancialForecasting
3-540-76098-9
DimitrisC.Dracopoulos
EvolutionaryLearningAlgorithmsforNeuralAdaptiveControl
3-540-76161-6
M.Karny,K.WarwickandV.Kfu-kova(Eds)
DealingwithComplexity:ANeuralNetworksApproach
3-540-76160-8
JohnA.Bullinaria,DavidW.GlasspoolandGeorgeHoughton(Eds)
4thNeuralComputationandPsychologyWorkshop,London,
9"11April1997:ConnectionistRepresentations
3-540-76208-6
MariaMarinaroandRobertoTagliaferri(Eds)
NeuralNets-WIRNVIETRI-97
3-540-76157-8
1.J.LandauandJ.G.Taylor(Eds)
ConceptsforNeuralNetworks:ASurvey
3-540-76163-2
Thomas Lindblad and Jason M. Kinser
Image Processing
using Pulse-Coupled
Neural Networks
Springer
ThomasLindblad,PhO
OepartmentofPhysics,Royallnstitute ofTechnology,24Frescativagen,
S-10405Stockholm,Sweden
]ason M.Kinser,OSc
The Institutefor Biosciences,BioinformaticsandBiotechnology,MSN4E3,George
Mason University,10900UniversityBoulevard,Manassas,VA20100, USA
SeriesEditor
].G.Taylor,BA,BSc,MA,PhO,FlnstP
Centrefor NeuralNetworks,OepartmentofMathematics,Kings College,
Strand,LondonWC2R2LS,UK
BritishLibraryCataloguinginPublicationData
Lindblad,Thomas
Imageprocessingusingpulse-coupledneuralnetworks.
(Perspectivesinneural computing)
J.Imageprocessing2.Neuralnetworks (Computerscience)
I.TitleII.Kinser,[asonM.
006.4
LibraryofCongressCataloging-in-PublicationData
Lindblad,Thomas
Imageprocessingusingpulse-coupledneural networksIThomas
Lindbladand[asonM.Kinser.
p. cm.--(Perspectivesinneuralcomputing)
Includesbibliographicalreferencesandindex.
l,Imageprocessing.2.Neuralnetworks(Computerscience)
I.Kinser,[asonM.,1962- . 11.Title. III.Series.
TA1637.L56 1998
621.36'7--dc21 98-13421
Apart fromanyfairdealingforthepurposesofresearch orprivate study,orcriticismorreview,
aspermitred underthe Copyright,DesignsandPatents Act1988,this publication may onlybe
reproduced,storedortransmitted,inanyformorbyanyrneans, withthe prior permissionin
writing ofthe publishers, or inthe caseofreprographiereproduction in accordancewith the
termsoflicencesissuedbytheCopyrightLieensingAgency. Enquiries concerningreproduction
outsidethoseterms shouldbesenttothepublishers.
ISBN978-3-540-76264-5 ISBN978-1-4471-3617-0(eBook)
DOI10.1007/978-1-4471-3617-0
©Springer-VerlagLondon 1998
OriginallypublishedbySpringer-VerlagLondonLimitedin1998.
The use ofregistered narnes,trademarks etc.in this publication does not imply,even in the
absence of a specific staternent, that such names are exempt from the relevant laws and
regulationsandthereforefreeforgeneraluse.
Thepublishermakes norepresentation,express orimplied,withregard totheaccuracy ofthe
informationcontainedinthisbookandcannotacceptanylegalresponsibilityorliabilityforany
errorsoromissionsthatmaybemade.
Typesetting:Camera-readybyauthor
34/3830-543210Printedonacid-freepaper
Preface
Image processing by electronic means has been a very active field for
decades. The goal has been, and still is, to have a machine perform the
same image functions which humans doquite easily. This goal is still far
from being reached. So we must learn more about the human
mechanisms and how to apply this knowledge to image processing
problems. Traditionally, the activities in the brain are assumed to take
place through the aggregate action of billions of simple processing
elements referred to as neurons and connected by complex systems of
synapses. Within the concepts ofartificial neural networks, the neurons
are generally simple devices performing summing, thresholding, etc.
However, weshownowthat the biologicalneurons are fairly complexand
perform much more sophisticated calculations than their artificial
counterparts. The neurons are alsovery specialised and it is thought that
there are several hundred types in the brain and messages travel from
oneneuron to anotheras pulses.
Recently, scientists have begun to understand the visual cortex of
small mammals. This understanding has led to the creation of new
algorithms that are achieving new levels of sophistication in electronic
image processing. With the advent of such biologically inspired
approaches, in particularwithrespectto neural networks, wehave taken
another step towards the aforementionedgoal.
In our presentation ofthe visual cortical models we will use the term
Pulse-Coupled Neural Network (PCNN). The PCNN is a neural network
algorithm that produces a series ofbinary pulse images when stimulated
with a grey scale orcolourimage. This network is differentfrom whatwe
generallymean by artificial neural networks in the sense thatit does not
train.
The goal forimage processing is to eventually reach a decision on the
content of that image. These decisions are generally far easier to
accomplish byexaminingthe pulse outputs ofthe PCNN rather than the
original image. Thus, the PCNN becomes a very useful pre-processing
tool. There exists, however, an argument that the PCNN is more than a
pre-processor. It is possible that the PCNN also has self-organising
abilities which make it possible to use the PCNN as an associative
memory. This is unusualforan algorithmthatdoesnot train.
vi Preface
Finally, it should be noted that the PCNN is quite feasible to
implement in specialised hardware. Traditional neural networks have
had a large fan-in and fan-out. In other words, each neuron was
connected to several other neurons. In electronics a different "wire" is
needed to make each connection and large networks are quite difficult to
build. The PCNN, on the other hand, has only local connections and in
mostcases these are always positive. This is quite plausible for electronic
implementation.
The PCNN is quite powerful and we are just beginning to explore the
possibilities.This text will review the theory and then explore its known
image processing applications: segmentation, edge extraction, texture
extraction, object identification, object isolation, motion processing,
foveation, noise suppression, and image fusion. This text will also
introduce arguments as to its ability to process logical arguments and its
use as a synergetic computer. Hardware realisation of the PCNN will
also be presented.
This text is intended for the individual who is familiar with image
processing terms and has a basic understanding of previous image
processing techniques. It does not require the reader to have an
extensive background in these areas. Furthermore, the PCNN is not
extremely complicated mathematically so it does not require extensive
mathematical skills. However, this text will use Fourier image
processing techniques and a working understanding of this field will be
helpfulin some areas.
The PCNN is fundamentally unique from many of the standard
techniques being used today. Many of these fields have the same basic
mathematical foundation and the PCNN deviates from this path.Itis an
excitingfield thatshows tremendous promise.
Acknowledgements
The work reported in this book includes research carried out by the
authors together with co-workers at various universities and research
establishments.Severalresearch councils,foundations and agencies have
supported the work and made the collaboration possible.Their support is
gratefully acknowledged.In particular we would like to acknowledge the
fruitful collaboration and discussions with the following scientists:
Kenneth Agehed, Randy Broussard, Age J. Eide, John Caulfield, Bruce
Denby, W. Friday, John L. Johnson, Clark S. Lindsey, Steven Rogers,
Thaddeus Roppel, Manuel Samuelides, Ake Steen, Geza Szekely, Mary
Lou Padgett, Ilya Rybak, Joakim and Karina Waldemark. 'Torrent'
thanks are due to Moyra Mann ofScience & Technology BookStore, UK.
Contents
1. Introduction andTheory 1
1.1 General Aspects 1
1.2The State ofTraditional Image Processing 2
1.2.1 Generalisationvs.Discrimination 2
1.2.2 The World ofInnerProducts 3
1.2.3The MammalianVisual System 4
1.2.4 Where DoWe GoFrom Here? 5
1.3Visual CortexTheories 5
1.3.1The Eckhorn Model 5
1.3.2The Rybak Model 7
1.3.3The Parodi Model 8
1.4The Visual Cortex and SimulationTheory 8
1.5Introduction to Applications 10
References 10
2. PCNNTheory 11
2.1The Original PCNN Model 11
2.2Time Signal Experiments 15
2.3PCNNAlterations 17
References 18
3. PCNN Image Processing 21
3.1 ImportantFeatures 21
3.2Image Fundamentals and the PCNN 23
3.3Blood Cell Identification 24
3.4AircraftIdentification 25
3.5Mammography 27
3.6 Aurora Borealis 29
References 30
viii Contents
4.The peNNKernel 33
4.1 1/RConnections 34
4.2Asymmetric Kernel 36
4.3On-Centre/Off-Surround Kernel 38
4.4 Discussion 38
References 38
5.Target Recognition 39
5.1 TraditionalTargetRecognition 40
5.2 Traditional Correlation FilterTargetRecognition 41
5.3Employingthe PCNN 44
5.4 Image Factorisationusingthe PCNN 46
References 47
6. Dealingwith Noise 49
6.1 Noise and the PCNN 49
6.2 Noise Reduction by a Signal Generator 51
6.3 FastLinking 54
6.4 Summary 56
References 57
7. Feedback 59
7.1 The Feedback Pulse-Coupled Neural Network 59
7.2 Sample Problems 61
7.3 Summary 63
References 63
8. ObjectIsolation 65
8.1The Fractional Power Filter 66
8.2 ObjectIsolation System 68
8.3An Example 68
8.4 Dynamic ObjectIsolation 75
8.5Shadowed Objects 78
References 79
Contents ix
9. Foveation 81
9.1The FoveationAlgorithm 82
9.2TargetRecognition bya PCNN Based FoveationModel 85
9.3Summary 89
References 89
10. ImageFusion 91
10.1The Multi-SpectralPCNN 92
10.2Pulse-CoupledImageFusionDesign 94
10.3An Example 96
10.4ExampleFusing aWavelet FilteredImages 98
10.5Summary 99
References 100
11.SoftwareandHardwareRealisation 101
11.1Software 101
11.2ParallelImplementation 120
11.3ASimplified Implementation 122
11.4OpticalImplementation 123
11.5Implementationin VLSI 125
11.6Implementationin FPGA 126
References 130
12.Summary,Applications andFutureResearch 131
12.1Object Isolation 133
12.2Dynamic ObjectIsolation 133
12.3ColourImageProcessing 134
12.4Histogram Considerations 135
12.5Maze Solution 136
12.6Foveation 137
12.7ObjectIcons 139
12.8SyntacticalInformationProcessing 140
12.9FusedPCNN 141
12.10HardwareImplementations 141
12.11FutureResearch 144
References 145
Appendix: Whereto FindPCNNComputerCode 149
Index 151
Chapter 1
Introduction and Theory
1.1 General Aspects
Humans have an outstanding ability to recognise, classify and
discriminate objects with extreme ease. For example, if a person was in a
large classroom and was asked to find the light switch it would not take
more than a second or two. Even if the light switch was located in a
different place than the human expected or it was shaped differently
than the human expected it would not be difficult to find the switch.
Another exampleis thatofrecognisingdogs.Ahuman needs to see only a
few examples and then he is able to recognise dogs even from species
that he has not seen before. This recognition ability also holds true for
animals, to a greater or lesser extent. A spider has no problem
recognising a fly. Even a baby spider can do that. At this level we are
talking about a few hundred to a thousand processing elements or
neurons. Nevertheless the biological systems seem to do their job very
well.
Computers, on the other hand, have a very difficult time with these
tasks.Machines need a large amount ofmemory and significant speed to
even come close to the processing time of a human. Furthermore, the
software for such simple general tasks does not exist. There are special
problems where the machine can perform specific functions well, but the
machines donot performgeneralimage processing and recognition tasks.
In the early days ofelectronic image processing,many thought that a
single algorithm could be found to perform recognition. The most popular
of these is Fourier processing. It, as well as many of its successors, has
fallen short of emulating human vision. It has become obvious that the
human uses many elegantly structured processes to achieve its image
processinggoals, and we are beginningto understand only a fewofthese.
One of the processes occurs in the visual cortex, which is the part of
the brain that receives information from the eye. At this point in the
T. Lindblad et al., Image Processing using Pulse-Coupled Neural Networks
© Springer-Verlag London 1998