A SURVEY ON INDUSTRIAL VISION SYSTEMS,

1
APPLICATIONS AND TOOLS
Elias N. Malamas,

Euripides G.M. Petrakis2, Michalis Zervakis

Department of Electronic and Computer Engineering

Technical University of Crete
Chania Crete Greece
[email protected], [email protected], [email protected]

Laurent Petit, Jean-Didier Legat

Microelectronics Laboratory
Universite Catholique de Louven
Louven-La-Neuve Belgium
[email protected], [email protected]

ABSTRACT
The state of the art in machine vision inspection and a critical overview of real-world applications
are presented in this paper. Two independent ways to classify applications are proposed, one
according to the inspected features of the industrial product or process and the other according to
the inspection independent characteristics of the inspected product or process. The most
contemporary software and hardware tools for developing industrial vision systems are reviewed.
Finally, under the light of recent advances in image sensors, software and hardware technology,
important issues and directions for designing and developing industrial vision systems are
identified and discussed.
Keywords: Machine vision, automated visual inspection, image processing, image analysis.

1 INTRODUCTION
Machine vision provides innovative solutions in the direction of industrial automation [1]. A
plethora of industrial activities have benefited from the application of machine vision technology
on manufacturing processes. These activities include, among others, delicate electronics
component manufacturing [2], quality textile production [3], metal product finishing [4], glass
manufacturing [5], machine parts [6], printing products [7] and granite quality inspection [8],
integrated circuits manufacturing [9] and many others. Machine vision technology improves
1

This work was supported by project HIPER (BE97-5084) under programme BRIGHT-EURAM of the European
Union (EU).
2
Corresponding author.

productivity and quality management and provides a competitive advantage to industries that
employ this technology.

1.1 OVERVIEW ON INDUSTRIAL VISION SYSTEMS

Traditionally, visual inspection and quality control are performed by human experts [10].
Although humans can do the job better than machines in many cases, they are slower than the
machines and get tired quickly. Moreover, human experts are difficult to find or maintain in an
industry, require training and their skills may take time to develop. There are also cases were
inspection tends to be tedious or difficult, even for the best-trained experts. In certain
applications, precise information must be quickly or repetitively extracted and used (e.g., target
tracking and robot guidance). In some environments (e.g., underwater inspection, nuclear
industry, chemical industry etc.) inspection may be difficult or dangerous. Computer vision may
effectively replace human inspection in such demanding cases [11].
Main Processor or
Computer with
Image Processing
Software

Network Interface

Image Processing
Hardware

Illumination

Camera(s)

Manufacturing Process
Control Systems
Robot(s), PLC(s), etc.

Figure 1: A typical industrial vision system.

Figure 1 illustrates the structure of a typical industrial vision system. First, a computer is
employed for processing the acquired images. This is achieved by applying special purpose
image processing analysis and classification software. Images are usually acquired by one or
more cameras placed at the scene under inspection. The positions of the cameras are usually
fixed. In most cases, industrial automation systems are designed to inspect only known objects at
fixed positions. The scene is appropriately illuminated and arranged in order to facilitate the
reception of the image features necessary for processing and classification. These features are
2

also known in advance. When the process is highly time-constrained or computationally intensive
and exceeds the processing capabilities of the main processor, application specific hardware (e.g.,
DSPs, ASICs, or FPGAs) is employed to alleviate the problem of processing speed. The results of
this processing can be used to:

Control a manufacturing process (e.g., for guiding robot arms placing components on
printed circuits, painting surfaces etc.).

Propagated to other external devices (e.g., through a network or other type of interface
like FireWire) for further processing (e.g., classification).

Characterize defects of faulty items and take actions for reporting and correcting these
faults and replacing or removing defective parts from the production line.
The requirements for the design and development of a successful machine vision system

vary depending on the application domain and are related to the tasks to be accomplished,
environment, speed etc. For example, in machine vision inspection applications, the system must
be able to differentiate between acceptable and unacceptable variations or defects in products,
while in other applications, the system must enable users to solve guidance and alignment tasks
or, measurement and assembly verification tasks.
There exists no industrial vision system capable of handling all tasks in every application
field. Only once the requirements of a particular application domain are specified, then
appropriate decisions for the design and development of the application can be taken. The first
problem to solve in automating a machine vision task is to understand what kind of information
the machine vision system is to retrieve and how this is translated into measurements or features
extracted from images. For example, it is important to specify in advance what defective means
in terms of measurements and rules and implement these tasks in software or hardware. Then, a
decision has to be made on the kind of measurements to be acquired (e.g., position or intensity
measurements) and on the exact location for obtaining the measurements.
For the system to be reliable, it must reduce escape rates (i.e., non-accepted cases
reported as accepted) and false alarms (i.e., accepted cases reported as non-accepted) as much
as possible. It is a responsibility of the processing and classification units to maintain system
reliability, but the effectiveness of classification depends also on the quality of the acquired
images. An industrial vision system must also be robust. Thus, it should adapt itself automatically
and achieve consistently high performance despite irregularities in illumination, marking or
3

background conditions and, accommodate uncertainties in angles, positions etc. Robust

performance is difficult to achieve. High recognition and classification rates are obtained only
under certain conditions of good lighting and low noise. Finally, an industrial vision system must
be fast and cost efficient.
In this survey, we emphasize the important attributes of an industrial machine vision
inspection system such as, flexibility, efficiency in performance, speed and cost, reliability and
robustness. In order to design a system that maintains these attributes it is important to clearly
define its required outputs and the available inputs. Typically, an industrial inspection system
computes information from raw images according to the following sequence of steps:
1. Image acquisition: Images containing the required information are acquired in digital
form through cameras, digitisers etc.
2. Image processing:

Once images have been acquired, they are filtered to remove

background noise or unwanted reflections from the illumination system. Image restoration
may also be applied to improve image quality by correcting geometric distortions
introduced by the acquisition system (e.g., the camera).
3. Feature extraction: A set of known features, characteristic for the application domain, is
computed, probably with some consideration for non-overlapping or uncorrelated features
[12], so that better classification can be achieved. Examples of such features include size,
position, contour measurement via edge detection and linking, as well as and texture
measurements on regions. Such features can be computed and analyzed by statistical or
other computing techniques (e.g. neural networks or fuzzy systems). The set of computed
features forms the description of the input image.
4. Decision-making: Combining the feature variables into a smaller set of new feature
variables reduces the number of features. While the number of initial features may be
large, the underlying dimensionality of the data, or the intrinsic dimensionality, may be
quite small. The first step in decision making attempts to reduce the dimensionality of the
feature space to the intrinsic dimensionality of the problem. The reduced feature set is
processed further as to reach a decision. This decision, as well as the types of features and
measurements (the image descriptions) computed, depends on the application. For
example, in the case of visual inspection during production the system decides if the
produced parts meet some quality standards by matching a computed description with

some known model of the image (region or object) to be recognized. The decision (e.g.,
model matching) may involve processing with thresholds, statistical or soft classification.
At the last level of decision-making and model matching mentioned above, there are two
types of image (region or object) models that can be used namely, declarative and procedural.
Declarative models consist of constraints on the properties of pixels, objects or regions and on
their relationships. Procedural models are implicitly defined in terms of processes that recognize
the images. Both types of models can be fuzzy or probabilistic, involving probabilistic constraints
and probabilistic control of syntactic rules respectively. A special category of models is based on
neural networks.
Model-based approaches often require that descriptions (e.g., features) of the image at
different levels of specificity or detail be matched with one of possible many models of different
classes of images. This task can become very difficult and computationally intensive if the
models are complex and a large number of models must be considered. In a top-down approach
to model matching, a model might guide the generation of appropriate image descriptions rather
than first generating the description and then attempting to match it with a model. Another
alternative would be to combine top-down and bottom-up processes. The above control strategies
are simplified when one is dealing with two-dimensional images taken under controlled
conditions of good lighting and low noise, as it is usually the case in industrial vision
applications. Image descriptions and class models are easier to construct in this case and complex
model matching can be avoided. Model-based approaches to visual inspection tasks [13] have
been applied in a variety of application fields and many of them are reviewed in the following
sections.

1.2 DEVELOPMENT APPROACHES AND ENVIRONMENTS

The development of a machine vision system begins with understanding the applications
requirements and constraints and proceeds with selecting appropriate machine vision software
and hardware (if necessary) to solve the task at hand. Older machine vision systems were built
around low-level software, requiring full programming control. They were based on simple frame
grabbers providing low-level interface capabilities with other system components. They were also
characterized by low-level user interfaces, low-level image analysis capabilities and difficulties in
system integration and maintenance. Eventually, machine vision inspection systems became more
modular, providing more abstract capabilities for system development and maintenance and
reaching higher level of robustness.

Todays applications need environments that are developed in short time and are adjusted
to modifications of the manufacturing process. In addition, the system must be simple to operate
and maintain. The key here is to select an appropriate development environment providing
Graphical User Interfaces (GUIs) or other programming tools (see Section 3 of this survey).
Through GUIs and visual programming tools, even non-vision experts but authorized users like
e.g., manufacturing engineers, are allowed to interact with the application and specify sequences
of operations from pull-down menus offering access to large pools of tested algorithms.
Programming is easier in this case, since the algorithms are selected based on knowledge of what
they do and not on how they do it. The use of GUIs shifts the effort of application development to
the manufacturing engineer from the programmer expert, as in the earlier days of machine vision
systems. This feature not only results in faster and cheaper application developments, but also
allows addressing several applications with a single piece of re-configurable software (i.e., the
application development tool).
Industrial vision systems must be fast enough to meet the speed requirements of their
application environment. Speed depends on the task to be accomplished and may range from
milliseconds to seconds or minutes. As the demands of processing increase, special purpose
hardware is required to meet high-speed requirements. A cost saving feature of industrial vision
systems is their ability to meet the speed requirements of an application without the need of
special purpose hardware. PCs and workstations are nowadays fast enough so that this can be
achieved in many application domains, especially in those with less demanding run time
requirements [14, 15].
Advances in hardware technology in conjunction with the development of standard
processing platforms have made the production and maintenance of industrial automation
systems feasible at relatively low cost. Pentium PCs with Windows NT (Windows 2000, XP) or
UNIX based systems like Linux are considered the main alternatives with Windows being
preferred to achieve labor saving application development with maximum portability based on
ready-to-use software (e.g., commercially available software). Linux is becoming eventually a
standard especially in cases where customized or cost saving solutions are preferred. Linux is
sometimes offered as open-source freeware and appears to be the ideal solution in the case of
dedicated applications where independency on vendor specific software has to be achieved.
However the limited availability of application development tools (e.g., interfacing software) is a
serious drawback of Linux.

1.3 APPLICATIONS OF INDUSTRIAL VISION SYSTEMS

Interesting surveys specializing in a single application field include among others Ref. [16] for
automatic PCB inspection, Ref. [17] for wood quality inspection, and Ref. [18] for automatic fruit
harvesting. Other important general reviews that cover all the fields of visual inspection have
been published in Ref. [13], whereas model-based approaches to visual inspection are considered
in [19] and [20] and more recently in [21, 22] and [23]. In Ref. [21], a classification of automated
visual inspection applications is presented based on the type of images to be processed. Binary,
gray-scale, color, and range image systems are considered, each one showing certain
characteristics in the context of the particular application field being used. In Ref. [22] and [23]
on the other hand, machine vision systems are classified according to the qualitative
characteristics of the objects or processes under inspection. Three classes are presented, namely
dimensional verification, surface detection, and inspection of completeness.

1.4 CONTRIBUTIONS AND STRUCTURE OF THE SURVEY

In this survey, we present an overview of machine vision applications in the industrial
environment. Two independent ways of classifying industrial vision applications are proposed.
First, industrial vision applications are classified according to the inspected features of the
industrial product of process in four categories, namely: (a) Dimensional quality, (b) Structural
quality, (c) Surface quality and (d) Operational quality. Industrial vision applications are also
classified in terms of flexibility according to the so-called Degrees of Freedom (DoFs) that
form inspection independent features. This classification enables the evaluation of tools intended
towards similar industrial vision applications. A variety of software and hardware solutions
available for the development of applications are presented. Finally, the future trends of machine
vision technology are also discussed.
The rest of this survey is organized as follows: In Section 2, a review of the recent
literature in industrial vision along with our proposed classifications of industrial vision
applications is presented. Issues related to the development of industrial vision systems are
discussed in Section 3. A variety of software and hardware tools that can be used to assist the
development of machine vision inspection systems in the industry are also presented in this
section. Future trends in the field are presented and discussed in Section 4, followed by
concluding remarks in Section 5.

2 CLASSIFICATION OF INDUSTRIAL VISION APPLICATIONS

In modern industrial-vision-system research and development, most applications are related to at
least one of the following four types of inspection:
1. Inspection of dimensional quality,
2. Inspection of surface quality,
3. Inspection of correct assembling (structural quality) and
4. Inspection of accurate or correct operation (operational quality).
A formalization of the above categorization is attempted in the following, by probing
further onto the characteristics of products being inspected. Table 1 gathers some of the most
ordinary inspected features of products [1].
Dimensions, shape, positioning, orientation, alignment, roundness,
corners
Assembly
Holes, slots, rivets, screws, clamps

Dimensional
Structural

Foreign objects

Dust, bur, swarm

Surface

Pits, scratches, cracks, wear, finish, roughness, texture,

seams-folds-laps, continuity

Operational

Incompatibility of operation to standards and specifications

Table 1: Potential features of inspected products.

Notice that, despite the inherent differences in the nature of the four categories of
inspection, they are all reduced to the action of confirmation of quality standards satisfaction,
which is, in most cases, a binary (yes/no) decision. Figure 2 illustrates this relationship.
T he inspection
Problem :
Q u ality V erification

Q u ality of
D im ensional
C haracteristics

Q u ality of
Su rface
C haracteristics

A ssem bly Stru ctu ral

Q u ality

O perational
Q u ality

Figure 2: Major categories of industrial vision applications.

Industrial vision applications may also be classified based on features whose
measurements do not affect the inspection process (may take any value) allowing the system to

be independent on these types of features. The set of such features defines the so-called Degrees
of Freedom (DoFs) of the inspection process. Some of the most common DoFs met in the
industrial world are shown in Figure 3 and concern shape, geometrical dimensions, intensity,
texture, pose, etc. The DoFs of objects are strongly related to the variances of their characteristics
and are considered to be a measure of the flexibility of the vision system.
Object Degrees of Freedom (DOF):
Flexibility Characterization

Pose

Size

Shape

Colour

Texture

Illumination

Figure 3: Major DoFs of industrial vision systems.

The less the DoFs the stronger the dependency of the inspection system on the application
for which it is originally designed. Therefore, systems with low DoFs are less likely to be
expandable. High levels of variability, on the other hand, are characteristic of more general or
expandable systems. To allow many DOFs, the system must employ sophisticated image
classification approaches based on carefully selected models and algorithms, as to minimize its
dependency on the inspected item and its potential deformations. Moreover, the more the DoFs of
a system the greater its potential for expandability. For example, the system can be enhanced to
detect new types of defects if additional image processing and analysis functions are introduced
to the system and applied independently from the old ones to capture more image features (e.g.,
capture surface in addition to dimensional characteristics). The above considerations concerning
the proposed classification based on DoFs reveals a known trade-off in the design of inspection
systems between flexibility, complexity and cost which is not obvious in other classifications.
Table 2 illustrates the relationship between DOFs and quality inspection systems
developed for applications discussed in the survey. Most applications focus on the inspection of a
single characteristic (e.g., size). The remaining characteristics (e.g., finish, texture, etc.) can be
considered as DOFs for these applications indicating the flexibility of the vision system.
However, not all of them are equally important. In the rest of this survey, only the important
characteristics that relate to an application are discussed.

QUALITY
INSPECTION

S
U
R
F
A
C
E

D
I
M
E
N
S
I
O
N
A
L

STRUCTURAL

OPERATIONAL

REFERENCE

APPLICATION
FIELD

DEGREES OF FREEDOM
(DOFS)

[24]
[25]
[26]
[27]
[28]
[9]
[29]
[8]
[30]
[31]
[32]
[3]
[17]
[33, 34]
[35]
[36]
[37]
[38]
[12]
[39]
[7]
[40]
[15]
[41]
[42, 43]
[44]
[45]
[18, 46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
[56]
[57]

Mini resistor painting

Aluminium sheet casting
Railroad line inspection
Oil seals
Chicken meat defects
Wafer surface inspection
Surface approximation
Granite surface inspectopn
Directional texture
Surface roughness
Surface defects
Textile seam defects
Internal wood defects
Wood veneer surface
Surface corrosion
Machined parts inspection
Solder joints inspection
Solder joints inspection
Solder joints inspection
External screw threads
Banknotes inspection
Image segmentation
Object classification
Object classification
Circular parts
Packaging
Line segment measurement
Fruit harvesting
Packaging
Packaging

Resistor orientation
Sheet width
Illumination/Rail-foot head position,
Illumination
Illumination/Skin
Distortion/Scale/Orientation/Position
Illumination
Texture
Illumination/Rotation of direction
Orientation
Pose
Translation/Rotation
Wood density
Scale/Intensity
Shape
Scale/Translation/Orientation
Orientation
Orientation/Position
Position/Orientation/Size
Thread position
Position
Shape/Texture
Scale/Orientation
Shape
Peripheral occlusion
Position/Orientation
Orientation/Scale
Maturity/Illumination/Occlusion
Shape/Size
Illumination/Shape

Automotive industry
Object assembly
Railroad parts inspection
Railroad parts inspection
Automotive industry
Automotive industry
PCB inspection
Laser butt joint welding
Wrist watch quality

Size/Shape/Pose
Orientation (limited)
Illumination/Shape
Illumination/Shape/Texture
Illumination/Position
Illumination/Position/Shape/Size
Illumination
Welding path Shape/Gap size/Beam position
Hands Shape/Size/Orientation/Distortion

Table 2: Classification of industrial vision systems.

2.1 DIMENSIONAL QUALITY

Checking whether the dimensions of an object are within specified tolerances or the objects have
the correct shape, are ordinary tasks for industrial vision systems. Such tasks involve inspection
of geometrical characteristics of objects in two or three dimensions and are related to the first of

10

the four types of inspection defined in the previous section, namely the inspection of dimensional
quality.
Various industries are involved in the development of vision systems for automated
measurement of dimensional quality. In packaging industry, the tasks vary from measurements of
the fill level in bottles, to sell by date inspection and to airbag canister inspection (e.g., online
gauging systems that measure height, concentricity and diameter of airbag canisters). In Ref.
[48], a vision-guided system for the automated supply of packaging machines with paper and foil
material is presented. The system enables the manipulator to locate the proper bobbin, depalletize
it and transfer it to the requesting machine. A similar application is addressed in [44] were, a
vision system is used to determine the correct position of pallets and recognize the arrangement
pattern of sacks on the pallets. The system enables a robot mechanism to grasp the sacks and pass
them along a rotating cutting disk.
A popular and demanding real-time application is the inspection and classification of
solder joints on Printed Circuit Boards (PCBs). A typical inspection system for this application
consists of a camera with appropriate illumination placed on top of the PCB conveyor system
[37]. Processing PCB images consists of two major stages: First a pre-processing is performed in
order to remove noise and make the tracking of solder joints on the image of the PCB easy. Then,
the solder joints are classified according to the types of defects. The usual classification is
concerned with the quantity of the solder paste placed on a joint. Four classes are defined, namely
good, excess solder, insufficient and no solder. Simulation results on geometric models of joints
have shown that efficient classification can be achieved only by an optimal feature selection, so
that the classes do not overlap [12]. Current research has shown that histogram-based techniques
[38] perform better than two and three-dimensional feature-based techniques [37], both in terms
of system and computational complexity. The major problem is that two-dimensional features
alone are insufficient for correct classification and an extra classifier is required to separate
overlapping classes. In Ref. [58], it is shown that a combination of histogram and 2D, 3D featurebased techniques can overcome the performance of other techniques relying only on topological
features. Many PCB inspection systems rely on neural networks for the design of classifiers that
can deal with both distribution (histogram) and topological features of defects.
An approach to the problem of cutting two-dimensional stock sheets is reported in [47]. A
machine vision system is employed to acquire images of irregularly shaped sheets. Then, a
genetic algorithm is applied to generate part layouts that satisfy the manufacturing constraints

11

(e.g., minimization of trim loses). This method is particularly useful for the leather and apparel
industries, were irregular parts are commonly used.
An automatic visual system for the location of spherical fruits on trees, under natural
conditions, is presented in [18]. The system utilizes a laser range finder that provides range and
attenuation data of the inspected fruit surface and shape analysis algorithms are employed to
process the acquired reflectance and range images to locate the fruit and, finally, to determine the
position of the fruit on the tree. Experimental results report 74% detection over green (low degree
of ripeness) fruits and 100% detection of visible fruits. In [46] the above system is embedded in
the AGRIBOT integrated robotic system, aimed at the automatic harvesting of fruits.
The problem of measuring line segments, a primary machine vision problem, is addressed
in [45]. A heuristic algorithm for line segment measurement is proposed and used to assess the
efficiency of a machine vision system in accurately measuring properties of line segments, such
as length, angle and straightness. A similar application concerns the detection of circular parts
with peripheral defects or irregularities [43]. A two-stage Hough transform is applied for the
detection of circular machine parts.
A model-based computer vision system for the estimation of poses of objects in industrial
environments, at near real-time rates, is presented in [15]. A demanding real time application is
the detection of high quality printed products [7]. This application deals with products with high
degree of resemblance, where minor differences among them makes the application very difficult
to cope with, considering its real-time nature. An original algorithm based on morphological
operations facilitates the detection of flaws at near-pixel resolution. The system is applied for the
inspection of banknotes, which is clearly a very delicate application, considering the requirements
in the validity of the produced printings.
An interesting application in this category deals with the inspection of screw threads for
compliance with manufacturing standards [39]. Edge detection algorithms (based on linear
interpolation to the sub-pixel resolution) are applied to detect regions of interest. Each such
region is matched with multiple models of threads, since the dimensions and positions of the
inspected threads are allowed to vary. The system has been tested on the production line and has
been shown to perform better than other competitive methods, such as manual measurement.
Active Shape Models as the basis of a generic object search technique are employed in [49]. The
approach is based on the identification of characteristic or landmark points (i.e., points that
exist in all aspects of the object) in images and on the recording of statistics concerning the
12

relationships between the positions of the landmark points in a set of training examples. The
effectiveness of the approach is demonstrated on inspection of automotive brake assemblies.

2.2 SURFACE QUALITY

Inspecting objects for scratches, cracks, wear, or checking surfaces for proper finish, roughness
and texture, are typical tasks of surface quality inspection. Significant labour savings are
achieved in textile, wood and metal industries employing vision systems for fault detection and
quality verification.
In [3], the quality of textile seams is assessed using feature classification (based on selforganizing neural networks). The system also enables the expedition of seam quality judgement,
compared to human inspection, by locating seams on images of low contrast and then inspecting
the waviness of the seam specimens (this information is in fact three-dimensional).
CATALOG [17] is a system for internal wood log defects detection, based on Computer
Axial Tomography (CAT or CT). Sequences of CT image slices are acquired and each one is
segmented into two-dimensional regions. Each segmented image slice is analyzed and is
characterized as defect-free or defect-like. The correlation of defect-like regions across a CT
sequence enables the three-dimensional reconstruction of the log defects. In [33], the use of a
decision tree in combination with a modular neural network topology is shown to be more
efficient than a single large neural network alone for the classification of wood veneer. The
design of this topology is based on normalized inter-class variation of features for separating
between classes. An improved version of this topology [34], based on intra-class variation of
features, allows for the reduction of the complexity of the neural network topology and results in
improved classification accuracy. A review of research activities for locating defects on wood
surfaces is presented in [17].
Machine vision can also be used for the inspection and visualization of defects on ground
or machined components (e.g., cracks, pitting and changes in material quality). Segmentation
techniques for the detection of characteristic surface faults (e.g., indentations, protrusions) are
proposed in [32]. Similar techniques are applied for the detection of scratches during machine
polishing of natural stone. The assessment of surface roughness of machined parts is addressed in
[31]. Fourier transform is applied first for the extraction of roughness measures. Then, neural
networks are employed for the classification of surfaces based on roughness. The inspection of
defects on objects with directionally textured surfaces (e.g., natural wood, machined surfaces and
textile fabrics) is addressed in [30]. A global image restoration scheme based on Fourier
13

transform is applied. High frequency Fourier components corresponding to line patterns are
discriminated from low-frequency ones corresponding to defective regions. An alternative
approach for the inspection of randomly textured color images is presented in [8]. This method
considers both color and texture image properties and introduces a color similarity measure that
allows the application of the watershed transform. The problem of recovering depth information
for surface approximation of objects is examined in [29]. This is achieved using stereo image
pairs, a Scanning Electron Microscope (SEM) and involves computation of disparity estimates
utilizing a feature-based stereo algorithm.
The use of a finite-window robust sequential estimator for the detection and analysis of
corrosion in range images of gas pipelines is presented in [35]. Deviations from the robust surface
fit (which correspond to statistical outliers) represent potential areas of corrosion. The algorithm
estimates surface parameters over a finite sliding window. The technique is shown to be robust in
that it estimates the pipeline surface range function in the presence of noise, surface deviations
and changes in the underlying model. Despite the fact that the method exhibits real-time
execution capability, it fails to interpret correctly the combinations of high magnitude and high
frequency ripples with large patches of corrosion.
Surface inspection is also applied to the aluminium strip casting process. Infrared (IR)
temperature measurements (providing a measure of the distribution of surface temperature) are
used to evaluate the quality of aluminium sheets. A two level process for the inspection of
aluminium sheets is addressed in [25]. First, the system inspects both sides of an aluminium sheet
and captures images of potential defective areas. These images are then classified according to
defect type and stored for review by experts. In [9], machine vision is applied for the inspection
of wafer surfaces in Integrated Circuits (IC) production. A fuzzy membership function is used to
cope with the wide range of shape variations of the dimple defects.
Potential applications of surface quality inspection also include detection of damages on
railroad tracks [26], where on-board detection and classification of defects is performed in real
time. Exhaustive (100%) quality inspection of painting of metal film mini resistors is addressed
in [24], were detection of low quality products is achieved by the acquisition of a line pattern
image of a correctly painted resistor, which is compared with each acquired line pattern image.
Inspection of machined parts (e.g., circular oil seals) is reported in [8], where both surface and
dimensional qualities are verified. The centre of each circular seal is computed and the intensities
of its circumferential pixels are inspected.

14

In food industry, the inspection of the quality of goods is of primary interest. In [28], an
intelligent system for the detection of defects on chicken meat before packaging is presented. The
system relies on the analysis of chromatic content of chicken images, the extraction of potential
defective areas by morphological processing and their classification according to a predefined list
of defects.

2.3 STRUCTURAL QUALITY

Checking for missing components (e.g., screws, rivets, etc.) on assembled parts or checking for
the presence of foreign or extra objects (e.g., leaves, little sticks) are typical tasks of this class of
quality inspection.
In semiconductor and electronic industries, the tasks vary from connector presence,
capacitor polarity checking, Integrated Circuit (IC) pin gauging, IC identification, IC alignment
and positioning, to information gathering tasks such as automatic defect classification on
electronic circuit boards etc. For example, a connector inspection system is designed in [16],
which is fast and capable of detecting bent pins on connectors with 20-1000 pins.
The work in [55] deals with inspection of structural quality of PCB components. The
inspected objects (electronic components) are assumed to have little variations in size or shape
but significant variation in grey-level appearance. Statistical models of the intensity across the
objects structure in a set of training examples are built. Subsequently, a multi-resolution search
technique is used to locate the model that matches a region of an input image. A fit measure with
predictable statistical properties is also used to determine that this region is a valid instance of the
model. The method demonstrates failure rates that are acceptable for use in a real environment
(i.e., 1 in 1000 samples).
Template matching methods for the detection of anomalies in a car assembly line in real
time is proposed in [53]. Templates corresponding to four image regions of a car are selected by a
human supervisor and are analysed by the system. This work is part of an integrated system for
automatic inspection of a complete car assembly line. A second part of the same system aimed at
the inspection of the condition of vehicle doors is presented in [54]. In order to detect whether a
vehicle door is open or closed, a line-fitting algorithm is applied.
The detection of components of railroad lines is addressed in [26] and [52]. Filtering
techniques for detecting rail clamps and neural networks for detecting screws are employed in
[51]. The detection of wooden ties of rail lines in real time is presented is addressed in [52]. The
15

system enables the detection of tie boundaries on rail line images. An adaptive edge detector
based on a modified Marr-Hildreth operator is employed to cope with the steep transitions in the
image resulting from wood grain. A stochastic model-based inspection algorithm (based on
Bayesian estimation) for the detection of assembly errors on rigid objects in presented in [50].
The image models describe the appearance of a complex three-dimensional object in a twodimensional monochrome image. This method is applied for verifying correct assembly in a gear
assembly line and a VHS cassette production line.

2.4 OPERATIONAL QUALITY

Inspection of operational quality is related to the verification of correct or accurate operation of
the inspected products according to the manufacturing standards.
The inspection of laser butt joint welding is addressed in [56]. A camera captures the
welding seam track and determines the proper welding path and gap size. A noise-eliminating
step is applied first. Then, the welding path and gap are calculated on segmented welding images.
Segmentation is based on Laplacian filters. The information computed above enables the control
of the laser for the automatic welding of butt joints. Quality inspection of wristwatches is
addressed in [57]. All inspected watches are first synchronized with a reference clock. Images of
watch hands are acquired by a camera and are classified as hour, minute, second, and 24-hour
hands. The difficulty of this task stems from the overlapping of hands, as well as from the
existence of a curved magnifying glass over the date window of the watch, which corrupts the
clarity of captured images of hands. To compensate for such problems, the time that a watch
shows is detected and compared with the time of the reference clock using neural network
classifiers.

3 DEVELOPMENT OF INDUSTRIAL VISION SYSTEMS

Todays machine vision systems can be regarded as consisting from standard platform
components. The migration to standard PC-based platforms also standardized networking, backup
and storage technologies. Powerful Graphical User Interface (GUI) environments running on PCs
coupled (if necessary) with image processing accelerators provide the core technologies
necessary for building powerful, user-friendly machine vision environments at moderate cost.
System development involves integration of software and hardware tools into a complete
application. Todays machine vision systems are offering far easier integration of various
components originating from various software and hardware vendors. Even conventional

16

programming environments such as C and C++ allow for software components to be embedded
into a single system.
With the advent of new hardware for sensors, grabbers and computers, machine vision for
industrial inspection tackles even more sophisticated problems. High complexity algorithms can
nowadays be implemented for real time vision and new sensors (e.g., CMOS sensors) offering
high dynamic range allow for more reliable, flexible and faster image acquisition than traditional
CCD sensors, even under poor lighting conditions. At the same time, image-processing software
has become user friendly and powerful utilizing software libraries implementing some of the
most popular image processing and analysis algorithms. Most of these environments support
both, visual programming in combination with flexible GUI interfaces and traditional
programming. Both programming practices can be combined to facilitate application
development. Visual programming can be employed to accelerate applications prototyping
whereas the final application can be implemented and optimized using standard programming
methods and languages.
The current trend in industrial vision is to use commercial products instead of customized.
This reduces the effort and risk in developing new products and allows for immediate
exploitation of new hardware. When higher performance is needed, specialized DSP processors
can be used. The selection of the appropriate software tools is of crucial importance. A software
tool must have the following desirable features:

Multi-processing level support: The type of processing in an industrial vision system

varies from low level (e.g., filtering, thresholding), to medium level (e.g., segmentation,
feature computation) and high level (e.g., object recognition, image classification etc.). An
image software package must support all levels of functionality. Otherwise, different
software tools must be adopted and integrated into the same system.

Ease of manipulation:

Graphical user interfaces, visual programming and code

generation are typical features facilitating application development. Image functions must
be categorized by type and scope so that even a non-expert may choose the appropriate
function based mostly on what it does rather than on how it is done.

Dynamic range and frame-rate support: New types of sensors (e.g., CMOS sensors)
offer high dynamic range and faster image acquisition (e.g., 16 bits per pixel instead of 8

17

bits per pixel). Image software must support the processing of such high dynamic range
images at variable frame rates.

Expandability: The software package must be able to accommodate new or better

algorithms substituting old ones. In addition, the software package must easily adjustable
to new requirements of the running application without major programming effort.

Dedicated hardware support: The software package must be able to work in

collaboration with hardware (e.g., DSPs, ASICs, or FPGAs) to alleviate the problem of
processing speed in the case of computationally intensive applications.

In the following, a survey of popular software and hardware products for image processing
and industrial inspection is presented. The tools of each category are discussed separately for ease
of presentation. This list is by no means complete. However, it presents either the most
commonly used or best-suited tools for industrial vision applications.

3.1 SOFTWARE TOOLS

This review includes image processing and analysis tools, as well as, tools based on neural
networks, fuzzy logic and genetic algorithms.
3.1.1 IMAGE PROCESSING AND ANALYSIS TOOLS
Image processing is usually performed within rectangles, circles or along lines and arcs. Image
processing operators include filtering (e.g., smoothing, sharpening), edge detection, thresholding,
morphological operations etc. Such operations can be used to improve image quality (e.g.,
remove noise, improve contrast) and to enhance or separate certain image features (e.g., regions,
edges) from the background. Image processing operations transform an input image to another
image having the desired characteristics.
Image analysis transforms images to measurements. In particular, image analysis is
related to the extraction and measurement of certain image features (e.g., lines, and corners) and
transforms these image features to numbers, vectors, character strings etc. For example, lines,
regions, characters, holes, rips, tears can be gauged or counted. Image analysis involves feature
extraction operations (e.g., Hough transform for line and curve detection) in conjunction with
operations that measure average light intensity, texture, and shape characteristics such as Fourier
descriptors, moments, edge thinning, edge connectivity and linking etc. The ultimate goal of
image analysis is geared towards pattern recognition i.e., the extraction of features that can be
used by classifiers to recognize or classify objects.
18

An image processing environment to be suitable for industrial inspection, must (at least)
contain algorithms for edge and line detection, image enhancement, illumination correction,
geometry transforms, Region of Interest (RoI) selection, object recognition, feature selection and
classification. Table 3 provides a review of some of the most popular image processing tools
offering the desired functionality. These tools offer adequate features and performance for several
applications involved in the industrial sector. In terms of combined software and hardware, there
are four alternatives: (I) IM-PCI with IPL of Sherloc32/MVTools, (II) MaxPCI with PC Image
Flow or WiT and (III) Matrox Genesis with Matrox Imaging Library and (IV) Philips Trimedia
VLIW processor board with Rhapsody.

Software Package

Library

Visual
Programming

Command
Line

Dedicated H/W
Available

Source
Code

Khoros

Yes

Yes

No

No

Yes

SCIL-Image

Yes

Yes

No

No

Yes

LeadTools

Yes

Yes

No

IPL Lib

Yes

Yes

No

Yes

No

Sherlock32 / MVTools

Yes

Yes

No

Yes

Yes

Image-Pro plus

Yes

Yes

No

No

No

OPTIMAS

Yes

Yes

No

No

No

WiT

Yes

Yes

Optional

Yes

No

PC Image Flow

Yes

Yes

Datacube

Yes

No

Intel Image Processing Lib.

Yes

No

MMX

HALCON

Yes

Yes

No

No

No

VISION97

Yes

Yes (frame grabber)

No

No

AdOculos

Yes

Yes

No

MIL

Yes

Yes

Matrox

Yes

No

Rhapsody

Yes

No

No

Yes

No

No

No

Table 3: Image processing and analysis software tools.

3.1.2 NEURAL NETWORKS (NNS)
Neural Networks (NNs) are being successfully applied across a wide range of application
domains in business, medicine, geology and physics to solve problems of prediction,
classification and control. Neural networks are composed of a number of similar elementary
processing units (neurons) connected together into a network [59]. Neurons are arranged in layers
with the input data initializing the processing at the input layer. The processed data of each layer
passes through the network towards the output layer. Neural networks adapt the weights of their
19

neurons during a training period based on examples, often with a known desired solution
(supervised training). After sufficient training, the neural network is able to relate the problem
data to the appropriate solution spaces, i.e. generate input/output relations, thus offering a viable
solution to a new problem through examples [60]. They are capable of handling a variety of
image classification tasks in industrial vision environments, ranging from simple gauging to
advanced classification problems, such as fault detection, optical character recognition, operation
prediction, engine monitoring and control etc. They can be used either as standalone techniques
(e.g., wood [33], seam [3], surface roughness [31] inspection) or in conjunction with other
methods (e.g., solder joint inspection) [37]. Neural networks have been applied in all classes of
quality inspection introduced in Section 2, namely dimensional quality [36-38], surface quality
[3, 31, 34], structural quality [51] and operational quality [57]. They are applicable in almost
every situation where a relationship between input and output parameters exists, even in cases
where this relationship is very complex and cannot be expressed or handled by mathematical or
other modelling means.
Table 4 summarizes the features of the most commonly used neural network tools.
Beyond general purpose and stand-alone tools, there exist library tools, such as the SPRLIB and
the ANNLIB (developed by the Delft University Technology at Netherlands) emphasizing on
image classification and pattern recognition applications. Almost all tools provide a plethora of
neural architectures, covering the most popular, as well as some less known. Some of them
provide the ability for user defined topologies as well.

20

Package
Name

New
Algorithms

Braimaker
Neuro
Solutions

Back-propagation
Offers user
defined
neural
topologies
and
components

G2
NeurOnLine
SPRLIB
ANNLIB

Types of NN

Can build
exotic
network
architectures
using the
same data
types

Recurrent
back propagation,
back propagation
through time

Industrial
Applications

Package
User
or
Interface
Library
Software Graphical
package
Summaries of Software Graphical
applications
package
included

Back propagation, Detailed

RBF, Rho, auto
petrochemical
associative
application
included
Back-propagation,
pseudo-Newton,
LevenbergMarquadt,
conjugate gradient
descent, BFGS,
Kohonen maps

ILIB
Neural
Connection
DataEngine,
v.i,ADL

Trajan 3.0

Allows
building of
hybrid
networks

Code
Generation
or DLL
C
C++ /DLL

Software Graphical
package object
oriented
C/C++
Libraries

C/C++
Libraries
Software Graphical
Package

Multilayer
Perceptron, RBF,
Kohonen,
Bayesian
Multilayer
Perceptron,
Kohonen, Feature
map, Fuzzy,
Kohonen network
Offers all of the
above
architectures and
training
algorithms

Software Graphical
Package

C++ / DLL

Software Graphical
Package

DLL

Table 4: Neural network tools.

3.1.3 FUZZY SYSTEMS (FSS) AND NEURO-FUZZY SYSTEMS (NFSS)

A variety of industrial vision applications of diverse nature have been benefited by the use of
Fuzzy Systems (FSs) and Neuro-Fuzzy Systems (NFSs). Examples of such applications are cork
quality inspection [61], identification of mechanical component dimensional tolerances [62], IC
product quality control [63] etc. Fuzzy sets are based on decisions on linguistic variables, which
get linguistic values described by fuzzy sets, called membership functions [64]. Their basic
processing elements are fuzzy sets instead of numerical values. A fuzzy set can be considered as
an extension of a classical (crisp) set in the form that a crisp set permits only full membership or
no membership, whereas a fuzzy set permits partial membership with a certain degree. Thus, a

21

fuzzy set, say A, in a domain U is characterized by its membership function A that takes values
in the real interval 01. For each x U , A ( x ) expresses the degree of membership of this
value to the set A, where 1 denotes full membership and 0 denotes no membership at all. For
example a dimensional tolerance in length can define fuzzy sets small, medium and large
with respect to its values, as illustrated in Figure 4:

small

medium

large

length (nm)
0

60

90

120

lm ax

Figure 4: Example of membership function definition for dimensional tolerance.

Classical decision making systems usually try to avoid vague, imprecise or uncertain
information. Fuzzy systems on the other hand, deliberately make use of this information through
the membership functions that describe the degree to which a measurement belongs to certain sets
or classes. The membership function is the essential component of a fuzzy set. Thus, the
operations intersection, union and complement of fuzzy sets are defined via the membership
functions of the sets involved. Input values are matched with the preconditions of if-then rules
on fuzzy sets, describing the systems behavior. This kind of structured knowledge provides the
system with a rule based processing (or rule based reasoning) mechanism. This mechanism
supports independent rules (i.e., changes in one rule do not effect the result of other rules). FSs
and NNs differ mainly on the way they map inputs to outputs, the way they store information or
make inference steps. Table 5 lists the most popular software and hardware tools based on FSs as
well as on merged FSs and NNs methodologies.
Neuro-Fuzzy Systems (NFS) form a special category of systems that emerged from the
integration of Fuzzy Systems and Neural Networks [65]. There are two major subcategories
characterizing this integration namely: Neural Network Fuzzy Systems (NNFSs) incorporating
FSs represented on a Neural Network topology and Fuzzy Neural Networks (FNNs) which
include Neural structures with a number of fuzzyfied parts [62]. NNFSs aim at providing fuzzy
systems with automatic tuning methods typical to neural networks but, without altering their
22

functionality. Neural Networks learning ability can be used to build membership functions and
rules to encode systems behavior. The learning techniques employed are mainly based on
multilayer feed forward networks with the back-propagation algorithm. Fuzzy systems offer their
users better in-sight of neural black box structures, by encoding structured information in the
form of rules and by offering tools for exploring this knowledge.
FNNs retain the basic properties and architectures of NNs and simply fuzzify some of
their elements. The system obtains increased flexibility in storing, recalling and associating
information. Not only binary but also continues values can be given as inputs, which highly
increases system robustness. Usually, fuzzified components result in higher training speed of
NNs. Fuzzy Logic also makes neural models understandable and this increases users flexibility.
There exist a large variety of Neuro-Fuzzy topologies in the literature, such as the Fuzzy
ARTMAP in [66], the MCFC for speech recognition in [67], the ASAFES2 network in [68] etc.
The most popular topology though is the ANFIS [69], which follows the general structure of
Neural-Fuzzy systems [70] and moreover has an extra layer to perform normalization of rules
firing strengths.
Recently, Genetic Algorithms (GAs) [71] have also been employed to deal with certain
tasks ns certain application domains. For example, an application of genetic algorithms in
industrial vision is pattern detection, which defers from pattern matching in the sense that the
item sought is not known in advance. The use of genetic algorithms has also been reported for the
classification of objects [41], the detection of circular objects [42], image segmentation [40] as
well as for the minimization of trim loses in cutting two-dimensional stock sheets [47] etc. Sugal
2.1 (TRAJAN Software Ltd.) and the Generator (NEW LIGHT INDUSTRIES Ltd.) are the two
packages that can simulate genetic algorithms for various purposes and provide complete control
over these algorithms. The main difference between them is that, Generator is designed to interact
with Excel while, Sugal is autonomous.

23

Software
Package
Mathematica
FIDE
TILShell/
FuzzyCLIPS
FCM

MatLab

DataEngine, V.i
lib

FuzzyTECH+N
euroFuzzy
Module
NeuFrame v.4

Requirements
Windows/Unix,
Mathematica S/W
DOS/Windows,
4MB RAM, 4MB
HD, >386
Windows, CLIPS
(for Fuzzy
CLIPS)
Windows

Windows, Unix,
Linux, HP,
Solaris, Matlab
S/W
Windows,
Labview(for the
V.i library),
Pentium, 64MB
RAM, 50MB HD
Windows
Windows

Code
Generation

Features*

Hardware Support

No

FS only, Variety of built-in

functions
FS only, High and low
level debugging, waveform
generator, composer
FS only, Programming
language, simulation

No

ANSI C/ Java/
MatLab/
Assembly
C code
C Code,
Assembly

FS only, Neurally
optimized description,
high complexity of fuzzy
systems

C code

FS and NFS, ANFIS,

Graphical Editors

DLL

FS and NFS, Graphical

programming language,
Signal and Image
processing algorithms

C code,
Assembly, DLL

FS and NFS, Fuzzy Design

Wizard, Interface to other
S/W
FS and NFS, Variety of
Neural Topologies

C, C++, Java,
MatLab

MC6805, MC68HC05,
MC68HC11, MC68HC16,
MC68HC33x
VY86C570 dedicated fuzzy
processor
NEC 17K/75X/ 78K0/78K3
and V
Panasonic MN1500,
68HC11, 8051, Inmos
Transputer T805
No

No

Motorola 68HC12 MCU

No

*: FS=Fuzzy System, NFS=Neuro-Fuzzy System

Table 5: Fuzzy and neuro-fuzzy software systems.

3.2 HARDWARE TOOLS

Software implementations are often insufficient to meet the real time requirements of many
industrial vision applications. The ever-increasing computational demands of such applications
call for hardware tools implementing image processing algorithms. In the following ASICs,
DSPs, FPGAs and general-purpose processors are considered as possible alternatives in dealing
with the problem of processing speed. The choice among them has to be made taking into
account issues such as, size of chip, power dissipation and performance. However, issues such as
flexibility of usage, programming environment are now becoming of great importance for the
application developers. Table 6 summarizes the characteristics of some commonly used image
processing boards.

24

Board
TriMedia
(Philips)
IM-PCI
(Imaging
Technology)
MaxPCI
(Datacube)
PCI/C6600
(Texas
Instruments)
Genesis
(Matrox)
Mpact 2
(Chromatic)
Mamba
(Coreco)
TPE3 (AG
Electronics)
VigraVision
(Visicom)

Video Display

166 MHz

Image
Acquisition
Yes

Yes

Image
Processing
Library
Yes

ASIC

40 MHz

Optional

Optional

Yes

ASIC

40 MHz

Yes

Optional

Yes

DSP C6201

200 MHz

Optional

Optional

Yes

DSP C80

60 MHz

Yes

Yes

Yes

ASIC

125 MHz

Yes

Yes

Yes

Pentium II

466 MHz

No

Optional

Yes

PowerPC 7400

400 MHz

Optional

Optional

Yes

FPGA Xilinx
Virtex

Up to 300 MHz

Yes

Yes

Yes

Processing chip

Clock Speed

DSP TM-1300

Table 6: Image processing boards.

3.2.1 APPLICATION SPECIFIC INTEGRATED CIRCUITS (ASICS)

There are several ways to perform hardware image processing. The first one is to build a circuit
dedicated to the application using an ASIC. As algorithms become more complex, the future of
ASIC design will use more and more Intellectual Property (IP) blocks available on the market
either as a hardware black box units (i.e., layout cells) or software packages (in a hardware
description language such as VHDL or Verilog). Development time decreases this way because
these cells have already been validated on various technologies. Then, the design of ASICs
consists of assembling IP blocks and putting some glue logic between them for their interface.
The main disadvantage of the ASIC approach is that the circuit is usually limited to work
for one application (e.g., the system developed in [29] for stereo vision). To overcome this
limitation, programmability can be achieved using processor cores that are included in the ASIC.
A few companies (e.g., ARM, I.C.COM, Clarkspur, DSP Group, Argonaut) propose RISC or
DSP cores, mainly dedicated to portable applications (i.e., with low power consumption, but also
with performances below what is achieved with packaged chips). One of the most attractive
products comes from Argonaut with the ARC processor core. This is a 32-bit general purpose
RISC processor offering a lot of flexibility. The core size is very small (i.e., lower than 16 kgates,
about 2 mm2 area in a 0.35-m technology) and the processor has high clock speed (i.e., greater
25

than 100 Mhz). A software tool allows for building a configuration adapted to the specifications
of the application. For example, the synthesis can be targeted for area or for speed, the instruction
cache size and the register file size can be adapted accordingly etc. VHDL code is then produced
and can be synthesized in any technology.
Reconfigurability is another way of dealing with the limited applicability of ASICs.
Reconfigurable multiprocessor networks compromise the trade-off between the need for low
network diameter and the limited number of interconnection links among processors. In [72], an
ASIC architecture of this kind is presented. It enables the implementation of a variety of image
processing algorithms for low and intermediate level computer vision, such as FFT, edge
detection, template matching, Hough transform etc. The use of general purpose processing
elements implementing several image processing tasks on the same architecture, is another way
of dealing with the limitations of ASICs. Several approaches have been reported in the literature,
including [73], with on-chip photo detectors for early visual processing [74], and [75] for highlevel image processing.
3.2.2 DIGITAL SIGNAL PROCESSORS (DSPS)
An alternative is to use a chip instead of a processor core. Many different processor architectures
have being proposed. Each has its own advantages and disadvantages. The use of DSP boards for
the fast execution of image processing algorithms has been extensively used in industrial vision
applications with hard real-time constraints. Some popular DSP architectures are the TriMedia
Mediaprocessor by Philips Semiconductors [76], IM-PCI by Imaging Technology, MaxPCI by
Datacube, Texas Instruments (TI) TMS320Cxx Family, the Genesis Vision processor by Matrox
based on the TIs TMS320C80 DSP and on the PCI platform, the Mpact media processor by
Chromatic Research Inc. [77] etc.
3.2.3 GENERAL PURPOSE PROCESSORS
General-purpose processors develop faster than DSPs. They run at much higher frequencies than
their predecessors, approaching this way the performance of DSPs. Intel with the MMX and the
SIMD Instruction on the Pentium, AMD with the Athlon architecture and Motorola with the
Altivec on the PowerPC, propose new coprocessor architectures dedicated to intensive
computations on large sets of data, fitting well for image processing applications. The main
advantage of the general-purpose processors is of course the programming environment, allowing
the user to develop applications without having any knowledge of the inside architecture. On the
other hand, the power consumption and the size of these chips are often prohibitive, especially for
26

embedded applications. A review of the RISC architectures available today capable of

performing image processing tasks can be found in [14].
3.2.4 FIELD PROGRAMMABLE GATE ARRAYS (FPGAS)
FPGAs are now competitive to ASICs both, in terms of capacity (i.e., number of equivalent gates
contained in one chip) and performance. This allows to quickly having prototype of the circuit
that has to be designed and able to operate in real conditions. The main advantage compared to
ASICs is that FPGAs can be reprogrammed. Complex FPGAs allow to design reconfigurable
systems that can efficiently implement real-time image processing algorithms. FPGA-based PCI
boards are an attractive alternative to DSP systems. Recent FPGAs include processor core such as
the ARM or PowerPC cores with RAM and peripherals. This feature allows the designer to
develop a 1-chip reconfigurable hardware-software solution and highly facilitates redesign and
testability of large circuitry with the same components. Thus, the FPGA chip can be reused in
case of changes in the specifications or simply when errors have been introduced in the design
process.
3.2.5 NEURAL, FUZZY, AND NEURO-FUZZY HARDWARE SYSTEMS
There exist also platforms that are capable of implementing fuzzy, neural, or hybrid systems.
Most of them are based on general-purpose micro-controllers, which are fast enough to execute
assembly programs that describe fuzzy or neural systems. On the other hand, there are dedicated
processors, such as the SGS-Thomson WARP family of fuzzy controllers, for the acceleration of
fuzzy-oriented applications. In between the two methodologies, there exist processors, such as the
Motorola 68HC12, which offers a 4 instruction set of fuzzy operations, in addition to its generalpurpose architecture. All of the presented systems are supported by software packages, which
enable either the high-level or the low-level description of fuzzy or neural systems and the
generation of speed-optimized code for each platform. Table 7 summarizes the characteristics of
fuzzy-oriented hardware products.

27

Processing Chip

Clock Speed

Description

Software Support

CNAPS server

Board

CNAPS

20MHz

NN, Kohonen, BP, other

NeuroChip PCI

SAND

50MHz

Siemens SYNAPSE

MA-16

50MHz

ZISC036

16MHz

8-bit NM6403

50MHz

Ni1000

NNP

8, 16bit MCS 96

0-20MHz

NN, feedforward, RBF,

Kohonen, other
NN, MLP, RBF,
Kohonen, other
NN, on-chip learning,
RBF, Kohonen, other
NN, forward
propagation, Sobel
transform, other
NN, on-chip learning,
variety of topologies
NN, on-chip learning,
variety of topologies
FS, General Purpose,
Fast on-chip peripherals

BrainMaker,
CNAPS-C, BuildNet,
CodeNet
NeuroLution

16 bit 68HC12

8MHz

8 bit WARP 1.1, 2.0

40MHz

16bit ADSP 21xx

8 bit 80C5x5
8-bit COP8

33MHz
12 MHz
10 MHz

16-bit L-Neuro 1.0,

2.0, 2.3

60 MHz

IBM ZISC
RC Module
NeuroMatrix
Nestor Ni1000
AAC NNP
Intel 8XC196Kx

Motorola 68HC12
MCU
SGS-Thomson
WARP
FDG EZ-LAB
Rigel R-535J
National
NeuFuz/COP8
Philips L-Neuro

FS, General Purpose, 4

fuzzy instructions
implemented
FS,Fuzzy dedicated
Processor
FS,General Purpose
FS,General Purpose
NFS,General Purpose
NFS, Neuro-Fuzzy
Processor

SYNAPSE S/W
S/W Development
Tools
C++ compiler,
assembler, debugger
NestorACCESS
NNP S/W
Any C code
generating S/W
(FuzzyTECH,
TilShell, FCM)
FuzzyTECH, FCM
(HC-11ver.)
FuzzyStudio v3.0
FID
FLASH
NeuFuz 4
Several

Table 7: Fuzzy-oriented hardware systems.

4. FUTURE DIRECTIONS
The major trade-off in industrial vision is that, for a vision system to be viable, it should satisfy
increasingly demanding performance criteria against budget constraints on behalf of the end user
[78]. There are three technological trends that enable the alleviation of this gap: (a) Rapid
development in semiconductor technology along with development of multipurpose mainstream
Operating Systems, (b) Improvements in Human-Computer Interfaces, and (c) Advances in solid
state imaging sensors.

4.1 SOFTWARE AND HARDWARE TECHNOLOGY

The trend to fast and cheap multipurpose solutions (software or hardware) in image processing
[79] is especially important to the industrial environment, because there is an increasing demand
for cost-effective and high quality products that can only be fulfilled only by very sensitive,
automated quality monitoring. Moreover, the industry is a volatile and competitive environment,
28

thus industrial applications must be flexible enough to adjust to new production methods. These
requirements call for multipurpose hardware and mainstream Operating Systems (OS), so that
fast image processing can be achieved with moderate cost. The industrial OS standard in
contemporary vision systems is UNIX, since it is the most popular platform for image processing.
However, Windows NT seems to be a challenge, while a wide range of inexpensive statisticalanalysis and standard image manipulation applications is available for this platform. A problem
that arises in heavy-duty real-time applications is that a single system cant handle the data
streams of several hundred Mbytes per second. Systems integrators have addressed the challenges
of constantly accelerating process speeds, increasing optical solutions, and more complex
classification criteria in the industrial area with specialized hardware, frame grabbers, DSPs, or
even arrays of RISC systems working in parallel.
Referring to semiconductor technology, the advances in the field have allowed for the
integration of image processing and feature extraction algorithms onto silicon. More specifically,
they enable hardware implementation of algorithms that are essential to industrial vision (e.g.
contour extraction, color inspection, morphology, etc.), but are time consuming if implemented
on software run from a host computer. The trends of this integration are in the form of either an
ASIC system running as a specialized processor and supervised by the host, or a co-processor
embedded in the architecture of the host (e.g. the MMX technology imported in PC-compatible
systems). It is clear that the latter is the optimum solution in terms of cost for most applications,
except some demanding real-time cases where a network of ASIC-boards is used to cope with
processing demands [7, 25].

4.2 HUMAN-COMPUTER INTERFACES

Another major demand of state-of-the-art systems in industrial inspection is fast prototyping,
which requires adjusting an established inspection system to variations in the manufacturing
process with the lowest cost and the minimum time possible. This calls for flexible and
comprehensive image processing libraries that include special modules for all aspects of
industrial inspection, along with multipurpose hardware. This is the reason that model-based
solutions have not been adopted widely by the industry, since they require the development of a
model for the images to be processed, which is either very complex or very tedious, especially for
applications concerning the inspection of highly irregularly-shaped objects (e.g., plastic parts
industry). A solution to this problem to a certain extend, has been given by the use of CAD-based
vision systems [21].

29

Interfacing the industrial vision system with the human operator is very important for the
end user. This interface should enable any operator to efficiently adjust the parameters of the
system and handle some low-level problems occurring during inspection, without having to call
the software supplier. Windows-style tools (such as the ones described in Section 3) for PC-based
applications are the common case for contemporary systems. Along with the simplification of
software interfaces, the use of hardware interfaces especially for on-line intervention, such as the
mouse or the light pen, have also contributed to the acceptance of machine vision in the industrial
field [80].

4.3 IMAGE SENSORS

The third technological trend that boosts the performance of industrial vision systems is the
advances in imaging sensors. Solid-state technology has allowed the elimination of thermionic
technology from the capturing of images, which was inappropriate for such applications due to
slow frame rates, increased device volume, increased noise [81] etc. The introduction of solidstate technology in image capturing has led to some breakthroughs in industrial vision, since they
offer a number of advantages as opposed to the predecessor technology. Some of these
advantages are smaller device sizes, robustness against EM noise, higher resolutions,
asynchronous triggering (capturing the image the time it is needed), stop-motion techniques
(capturing fast-moving objects) [81], on-chip signal processing [82-84], robustness against
changes of lighting conditions [85, 86] etc. The most important technologies used in integrated
imaging sensors are the Charge-Coupled Device (CCD), Charge-Injection Device (CID) and
Complementary MOS (CMOS) [87].
Table 8 summarizes the most important characteristics of the three families of sensors.
We have chosen typical devices at the 300K pixel resolution level, since in practice the majority
of industrial vision systems perform image capturing at this level of resolution [27, 37, 47].
Notice at this point that there is a trade-off between speed and resolution, since higher resolution
supplies the image processing system with larger image-data streams, thus image processing
algorithms are slowed down. A popular solution to this trade-off is the use of high resolution
when detailed analysis is required and objects under inspection remain stationary [28] or a
network of processing boards is available [7], as opposed to the use of low resolution or
consideration of only small Regions of Interest (ROIs) when detail is not important [9, 44, 47] or
the inspection controls high speed actions [26].

30

Technology@
Model
Resolution
(Array Size)
ROI* Processing
Capability
On-Chip Processing
Consumption
Pixel Size
Dynamic Range&
Data Bits
Data Rate
Full Frame Rate

CCD@
DALSA CA-D8-0512W1

CID@
CIDTEC RACID810/8112

CMOS@
VLSI Vision VV55003

512x512

512x512

648x484

No

Yes

Yes

No
6.750W
10m
54dB
8
25MHz
77fps+ (max)

Yes
11/22mW
20m
57.5dB
8
10-20MHz
30-70fps+

Yes
< 125mW
7.5m
57dB
10
20-30MHz
30fps+

Table 8: Typical characteristics of the three solid-state imaging sensor families.

(*ROI: Region of Interest, &Dynamic Range=20log(ratio), +fps: frames per second, 1: reference
[88], 2: reference [89], 3: reference [90])
The future trends of imaging sensor technology are very prosperous. Image resolution can
reach 16Mpixels (4096x4096) for area scan devices and can serve the most demanding
applications [81]. Frame rates can be as high as desired and can reach up to 60Kfps with the
capabilities of region-of-interest (ROI) processing on behalf of CID and CMOS cameras. Such
high speeds are of course possible only for the acquisition of ROI windows of very small extent,
or for quite low resolution frames. Data bit resolution is increasing and the technology moves
from the 8-bit era to the 16-bit era. Color cameras are also advancing their capabilities mostly
using CCD sensor technology. Commercially available cameras use single CCD sensor of size
, 1/3, or even chip for standard PAL/NTSC video quality. Digital cameras are also
becoming available at affordable prices, using 3 CCD chips for 24 bit true color acquisition
separating the image into Red, Green and Blue channels. Using advanced CCD imager with 680K
pixel capacity, they achieve high DV picture quality. Digital color cameras capture still images in
standard JPEG format, whereas interconnection, communication and video transmission is often
performed through the IEEE 1394 interface, also known as FireWire, which tends to become a
standard. The integration of IEEE 1394 interface into the computers operating system is
achieved via available device drivers. Camera systems achieving 640x480 frame acquisition with
24bit true color are becoming of widespread use, while megapixel cameras are gaining their share
in todays market.
Several criteria are used to evaluate image sensors, the most important being the
following [91-93]: a) Responsivity, which is a measure of signal level per unit of optical energy.
CMOS sensors are slightly better than CCD in this category, due to the fact that gain elements are
easier to place on their chip. b) Dynamic Range defined as the ratio of a pixels saturation level to
31

its signal threshold. CCD sensors are better because they have less on-chip circuitry, which
reduces the noise and increases the sensitivity of the sensor. c) Uniformity, indicating the
consistency of response for different pixels under identical illumination conditions. Circuitry
variations affect the uniformity of pixels on an image sensor. CMOS sensors are more sensitive to
these variations because of the more additional circuitry on sensor. Newer CMOS devices have
added feedback to the amplifiers to compensate these variations, but this only works well under
illuminated conditions. CCD has better uniformity because the lack of any amplification in the
sensor itself. d) Speed of operation, with CMOS sensors operating faster because most of the
circuitry is on board. Thus, the signals communicate less distance and dont have to be piped to
other chips on the printed circuit board. CCD imagers still operate adequately fast for most
applications, but anticipated demanding applications will consider CMOS sensors instead. e)
Reliability, in which respect CMOS sensors are superior to CCDs because of the high level of
integration contained on the chip. More integration means less external connections that are
susceptible to corrosion and other problems associated with solder joints in harsh environments.
Overall, CCDs offer superior image performance and flexibility at the expense of system size.
CMOS imagers offer more integration, lower power dissipation, and smaller system size at the
expense of image quality and flexibility. For next-generation applications, CMOS evolves in
order to get around the lowquality problem. Improvements are incorporated by the use of
microlenses, which are small lenses manufactured directly above the pixel to focus the light
towards the active portion, and the minimization of the space circuitry in the CMOS pixel.
On-chip A/D conversion and signal processing have been enabled from the advances in
semiconductor technology, thus eliminating the need for separated chips. The trend is to move
from the imaging sensor to the image processing sensor, with on-chip capabilities for image
processing algorithms such as low-pass filtering [94], velocity measuring [83], edge detection,
smoothing [73] etc. Although CCD is a mature technology that is commonly used in industrial
vision applications, the potential of the alternative technologies (CID and CMOS) is very high,
considering their on-chip intelligent and autonomous post-processing.
Many applications in industrial vision require stand-alone operation, which means that
there is a need for intelligent cameras providing fast processing capabilities inside the camera.
The major challenge in this direction is to maintain an easy-to-program feature by providing the
end-user with commercial image processing libraries. The integration of a general-purpose
processor or a DSP inside the camera offers such features, since compiling tools are widely
available for these kinds of architectures. Unfortunately, even when using instruction-level

32

parallelism like in VLIW processors, the performance achieved is often not sufficient to handle
the amount of data generated by high resolution and dynamic-range cameras. One
complementary approach is to add a coprocessor exploiting data-level parallelism, i.e. capable of
performing the same operation on several different data. The combination of the two approaches
(i.e. intelligent camera and coprocessor) should result in very powerful systems able to offer
flexible programming facilities together with increased performance for most of the algorithms
included in image processing libraries.

5 CONCLUSIONS
The state of the art in machine vision inspection research and technology is presented. The
cardinal factors affecting the development of automated inspection systems are discussed and
related to the literature in the field. Contemporary applications of machine vision in the industry
are also reviewed and classified according to (a) Their measured parameters (i.e., dimensions,
surface, assembly and operation) and (b) The systems Degrees of Freedom. Tools and
techniques either dedicated to specific application requirements or targeted towards a wider
variety of applications with similar requirements are presented. This review covers a wide range
of software and hardware products including integrated image processing software packages,
image processing libraries, neural network, fuzzy, neuro-fuzzy tools, genetic algorithms as well
as hardware tools. The paper concludes with a critical perspective and a summary of future
directions in the field, as they are determined by the increasing performance criteria imposed to
the industrial vision technology and the restricted end-user budget. Trends on technological fields
affecting industrial vision, including semiconductor technology, human-computer interfacing and
imaging sensors, are also emphasized.

REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]

E. R. Davies, Automated Visual Inspection, in Machine Vision, vol. Chapter 19, 2nd
Edition ed: Academic Press, 1998, pp. 471-502.
J. L. C. Sanz and D. Petkovic, Machine Vision Algorithm for Automated Inspection of
Thin-Film Disk Heads, IEEE Trans. on PAMI, vol. 10, pp. 830-848, 1988.
C. Bahlmann, G. Heidemann, and H. Ritter, Artificial Neural Networks for Automated
Quality Control of Textile Seams, Pattern Recognition, vol. 32, pp. 1049-1060, 1999.
J. W. Tucker, Inside Beverage Can Inspection: An Application from Start to Finish,
presented at Proc. of the Vision '89 Conference, 1989.
A. R. Novini, Fundamentals of Machine Vision Inspection in Metal Container Glass
Manufacturing, presented at Vision '90 Conference, 1990.
J. Ker and K. Kengskool, An Efficient Method for Inspecting Machine Parts by a
Fixtureless Machine Vision System, presented at Vision '90 Conference, 1990.

33

[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]

T. Torres, J. M. Sebastian, R. Aracil, L. M. Jimenez, and O. Reinoso, Automated RealTime Visual Inspection System for High-Resolution Superimposed Printings, Image and
Vision Computing, vol. 16, pp. 947-958, 1998.
L. Shafarenko, M. Petrou, and J. Kittler, Automatic Watershed Segmentation of
Randomly Textured Color Images, IEEE Trans. on Image Processing, vol. 6, pp. 15301543, 1997.
H. Li and J. C. Lin, Using Fuzzy Logic to Detect Dimple Defects of Polisted Wafer
Surfaces, IEEE Transactions on Industry Applications, vol. 30, pp. 1530-1543, 1994.
A. Mital, M. Govindaraju, and B. Subramani, A Comparison between Manual and
Hybrid Methods in Parts Inspection, Integrated Manufacturing Systems, vol. 9, pp. 344349, 1998.
P. Kopardekar, A. Mital, and S. Anand, Manual, Hybrid and Automated Inspection
Literature and Current Research, Integrated Manufacturing Systems, vol. 4, pp. 18-29,
1993.
O. Oyeleye and E. A. Lehtihet, A Classification Algorithm and Optimal Feature
Selection Methodology for Automated Solder Joint Inspection, Journal of
Manufacturing Systems, vol. 17, pp. 251-262, 1998.
C. Taylor, J. Graham, and D. Cooper, System Architectures for nteractive KnowledgeBased Image Interpretation, Philosophical Transactions of the Royal Society of London
(A) Mathematical and Physical Sciences, vol. 324, pp. 457-465, 1988.
P. Baglietto, M. Maresca, M. Migliardi, and N. Zingirian, Image Processing on HighPerformance RISC Systems, Proceedings of the IEEE, vol. 84, pp. 917-930, 1996.
M. Magee and S. Seida, An Industrial Model-Based Computer Vision System, Journal
of Manufacturing Systems, vol. 14, pp. 169-186, 1995.
M. Moganti, F. Ercal, C. H. Dagli, and S. Tsunekawa, Automatic PCB Inspection
Algorithms: A Survey, Computer Vision and Image Understanding, vol. 63, pp. 287313, 1996.
S. M. Bhandarkar, T. D. Faust, and M. Tang, CATALOG: A System for Detection and
Rendering of Internal Log Defects Using Computer Tomography, Machine Vision and
Applications, vol. 11, pp. 171-190, 1999.
A. R. Jimenez, A. K. Jain, R. Ceres, and J. L. Pons, Automatic Fruit Recognition: A
Survey and New Results Using Range/Attenuation Images, Pattern Recognition, vol. 32,
pp. 1719-1736, 1999.
R. T. Chin and C. A. Harlow, Automated Visual Inspection: A Survey, IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 6, pp. 557-573, 1982.
R. T. Chin, Automated Visual Inspection: 1981 to 1987, Computer Vision Graphics
Image Processing, vol. 41, pp. 346-381, 1988.
T. S. Newman and A. K. Jain, A Survey of Automated Visual Inspection, Computer
Vision and Image Understanding, vol. 61, pp. 231-262, 1995.
B. C. Eduardo, Review of Automated Visual Inspection 1983 to 1993-Part I:
Conventional Approaches, SPIE-Intelligent Robots and Computer Vision XII, vol. 2055,
pp. 128-158, 1993.
B. C. Eduardo, Review of Automated Visual Inspection 1983-1993-Part II: Approaches
to Intelligent Systems, SPIE-Intelligent Robots and Computer Vision XII, vol. 2055, pp.
159-172, 1993.
J. M. R. Sanches and M. S. Piedade, An Automatic Visual Inspection System of the
Quality of Painting of Metal Film Mini Resistors, presented at International Conference
on Signal Processing Applications and Technology (ICSPAT'97), 1997.
W. Miller, J. Jaskot, B. McCoy, and E. Schiller, A Distributed System for 100%
Inspection of Aluminum Sheet Products, presented at International Conference on
Signal Processing Applications and Technology (ICSPAT'96), 1996.
34

[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]

I. Khandogin, A. Kummert, and D. Maiwald, Automatic Damage Detection for Railroad

Tracks by the Analysis of Video Images, presented at International Conference on
Signal Processing Applications and Technology (ICSPAT'97), 1997.
B. C. Jiang and S. J. Jiang, Machine Vision Based Inspection of Oil Seals, Journal of
Manufacturing Systems, vol. 17, pp. 159-166, 1998.
M. Barni, V. Cappellini, and A. Mecocci, Colour-Based Detection of Defects on Chicken
Meat, Image and Vision Computing, vol. 15, pp. 549-556, 1997.
A. J. Lacey, N. A. Thacker, S. Crossley, and R. B. Yates, A Multi-Stage Approach to the
Dense Estimation of Disparity from Stereo SEM Images, Image and Vision Computing,
vol. 16, pp. 373-383, 1998.
D. M. Tsai and C. Y. Hsieh, Automated Surface Inspection for Directional Textures,
Image and Vision Computing, vol. 18, pp. 49-62, 1999.
D. M. Tsai, J. J. Chen, and J. F. Chen, A Vision System for Surface Roughness
Assessment Using Neural Networks, Int. J. Adv. Manuf. Technology, vol. 14, pp. 412422, 1998.
M. L. Smith, G. Smith, and T. Hill, Gradient Space Analysis of Surface Defects Using a
Photometric Stereo Derived Bump Map, Image and Vision Computing, vol. 17, pp. 321332, 1999.
P. R. Park and M. S. Packianather, A Decision Tree of Neural Networks for Classifying
Images of Wood Veneer, International Journal of Advanced Manufacturing Technology,
vol. 14, pp. 280285, 1998.
M. S. Packianathe and P. R. Park, Neural Networks for Classifying Images of Wood
Veneer, Part II,, International Journal of Advanced Manufacturing Technology, vol. 16,
pp. 424-433, 2000.
K. L. Boyer and T. Ozguner, Robust Online Detection of Pipeline Corrosion from Range
Data, Machine Vision and Applications, vol. 12, pp. 291-304, 2001.
T. H. Hou and M. D. Pern, A Computer Vision-Based Shape-Classification System
Using Image Projection and a Neural Network, International Journal of Advanced
Manufacturing Technology, vol. 15, pp. 843-850, 1999.
T. H. Kim, T. H. Cho, Y. S. Moon, and S. H. Park, Visual Inspection System for the
Classification of Solder Joints, Pattern Recognition, vol. 32, pp. 565-575, 1999.
S. Janannathan, Automatic Inspection of Wave Soldered Joints Using Neural Networks,
Journal of Manufacturing Systems, vol. 16, pp. 389-398, 1997.
R. J. Hunsicker, J. Patten, A. Ledford, C. Ferman, and M. Allen, Automatic Vision
Inspection and Measurement System for External Screw Threads, Journal of
Manufacturing Systems, vol. 13, pp. 370-383, 1994.
P. Andrey, Selexionist Relaxation: Genetic Algorithms Applied to Image Segmentation,
Image and Vision Computing, vol. 17, pp. 175-187, 1999.
R. Cucchiara, Genetic Algorithms for Clustering in Machine Vision, Machine Vision
and Applications, vol. 11, pp. 1-6, 1998.
J. Y. Goulermas and P. Liatsis, Genetically Fine-Tuning the Hough Transform Feature
Space, for the Detection of Circular Objects, Image and Vision Computing, vol. 16, pp.
615-625, 1998.
D. M. Tsai, A Machine Vision Approach for Detecting and Inspecting Circular Parts,
Int. J. Adv. Manuf. Technology, vol. 15, pp. 217-221, 1999.
M. Kavoussanos and A. Pouliezzos, An Intelligent Robotic System for Depalletizing and
Emptying Polyethylene Sacks, International Journal of Advanced Manufacturing
Technology, vol. 14, pp. 358-362, 1998.
Y. R. Shiau and B. C. Jiang, Study of a Measurement Algorithm and the Measurement
Loss in Machine Metrology, Journal of Manufacturing Systems, vol. 8, pp. 22-34, 1999.

35

[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
[55]
[56]
[57]
[58]
[59]
[60]
[61]

[62]
[63]
[64]

A. R. Jimenez, R. Ceres, and J. L. Pons, A Vision System Based on a Laser RangeFinder Applied to Robotic Fruit Harvesting, Machine Vision and Applications, vol. 11,
pp. 321-329, 2000.
S. Anard, C. McCord, and R. Sharma, An Integrated Machine Vision Based System for
Solving the Nonconvex Cutting Stock Problem Using Genetic Algorithms, Journal of
Manufacturing Systems, vol. 18, pp. 396-414, 1999.
H. K. Tonshoff, C. Soehner, and G. Isensee, Vision-Guided Tripod Material Transport
System for the Packaging Industry, Robotics and Computer-Integrated Manufacturing,
vol. 13, pp. 1-7, 1997.
J. J. Hunter, J. Graham, and C. J. Taylor, User Programmable Visual Inspection, Image
and Vision Computing, vol. 13, pp. 623-628, 1995.
D. Tretter, C. A. Bouman, K. W. Khawaja, and A. A. Maciejewski, A Multiscale
Stochastic Image Model for Automated Inspection, IEEE Trans. on Image Processing,
vol. 4, pp. 1641-1654, 1995.
I. Khandogin, A. Kummert, and D. Maiwald, DSP Algorithms for the Automatic
Inspection of Fixing Devices of Railroad Lines, presented at International Conference on
Signal Processing Applications and Technology (ICSPAT'98), 1998.
J. Velten, A. Kummert, and D. Maiwald, Real Time Railroad Tie Inspection
Implemented On DSP And FPGA Boards, presented at International Conference on
Signal Processing Applications and Technology (ICSPAT '99), 1999.
K. H. Kim, Y. W. Kim, and S. W. Suh, Automatic Visual Inspection System to Detect
Wrongly Attached Components, presented at International Conferernce on Signal
Processing Applications and Technology (ICSPAT'98), 1998.
Y. K. Chung, K. H. Kim, and H. B. Kim, Extracting of Line Features for Automatic
Inspection of Vehicle Doors, presented at International Conference on Signal Processing
Applications and Technology (ICSPAT), 1997.
T. F. Cootes, G. J. Page, C. B. Jackson, and C. J. Taylor, Statistical Grey-Level Models
for Object Location and Identification, Image and Video Computing, vol. 14, pp. 553540, 1996.
J. Y. Jeng, T. F. Mau, and S. M. Leu, Gap Inspection and Alignment Using a Vision
Technique for Laser Butt Joint Welding, International Journal of Advanced
Manufacturing Technology, vol. 16, pp. 212-216, 2000.
M. Moreira, E. Fiesler, and G. Pante, Image Classification for the Quality Control of
Watches, Journal of Intelligent and Fuzzy Systems, vol. 7, pp. 151-158, 1999.
S. Jagannathan, D. Seebaluck, and J. D. Jenness, Intelligent Inspection of Wave Soldered
Joints, Journal of Manufacturing Systems, vol. 11, pp. 137-143, 1992.
S. Haykin, Neural Networks: Prentice-Hall, 1999.
N. K. Bose and P. Liang, Neural Network Fundamentals with Graphs, Algorithms, and
Applications: McGraw-Hill, 1996.
J. Chang, G. Han, J. M. Valverde, N. C. Griswold, J. F. Duque-Carrillo, and S. E., Cork
Quality Classification System using a Unified Image Processing and Fuzzy-Neural
Network Methodology, IEEE Transactions on Neural Networks, vol. 8, pp. 964-973,
1997.
T. Sarkodie-Gyan, C. W. Lam, D. Hong, and A. W. Campbell, An Efficient Object
Recognition Scheme for a Prototype Component Inspection, Mechatronics, vol. 7, pp.
185-197, 1997.
Y. H. Chen, Computer Vision for General Purpose Visual Inspection: a Fuzzy Logic
Approach, Optics and Lasers in Engineering, vol. 22, pp. 181-192, 1995.
J. C. Bezdek, J. Keller, R. Krisnapuram, and N. R. Pal, Fuzzy Models and Algorithms for
Pattern Recognition and Image Processing: Kluwer, 1999.

36

[65]
[66]
[67]
[68]
[69]
[70]
[71]
[72]
[73]
[74]
[75]
[76]
[77]
[78]
[79]
[80]
[81]
[82]
[83]
[84]
[85]

S. K. Pal and S. Mitra, Neuro-Fuzzy Pattern Recognition: Methods in Soft Computing:

Willey, 1999.
G. A. Carpenter, S. Crossberg, N. Markuzon, J. H. Reynolds, and D. B. Rosen, Fuzzy
ARTMAP: A Neural Network Architecture for Incremental Supervised Learning of
Analog Multidimentional Maps, IEEE Transactions on Neural Networks, vol. 3, 1992.
U. H. Kuo, J. P. Hsu, and C. I. Kao, MCFC: A Fuzzy Neural Network Model for Speech
Recognition, Journal of Intelligent and Fuzzy Systems, vol. 4, pp. 257-268, 1996.
K. C. Zikidis and A. V. Vasilakos, ASAFES2: A Novel Neuro-Fuzzy Architecture for
Fuzzy Computing Based on Functional Reasoning, Fuzzy Sets and Systems, vol. 83, pp.
63-84, 1996.
J. S. Jang, ANFIS: Adaptive-Network-based Fuzzy Inference Systems, IEEE
Transactions on Systems, Man and Cybernetics, vol. 23, pp. 63-84, 1993.
C. T. Lin and C. S. G. Lee, Neural Fuzzy Systems, A Neuro-Fuzzy Synergism to Intelligent
Systems. NJ: Prentice Hall, 1996.
D. E. Golberg, Genetic Algorithms in Search, Optimization and Machine Learning:
Addion-Wesley, 1999.
S. M. Bhandarkar and H. R. Arabnia, Parallel Computer Vision on a Reconfigurable
Multiprocessor Network, IEEE Transactions on Parallel and Distributed Systems, vol. 8,
pp. 292-309, 1997.
T. Komuro, I. Ishii, and M. Ishikawa, General-Purpose Vision Chip Architecture for
Real-Time Machine Vision, Advanced Robotics, vol. 12, pp. 619-627, 1999.
D. W. Hammerstrom and D. P. Lulich, Image Processing Using One-Dimensional
Processor Arrays, Proceedings of IEEE, vol. 84, pp. 1005-1018, 1996.
C. L. Wang, P. B. Bhat, and K. Prasanna, High-Performance Computing for Vision,
Proceedings of the IEEE, vol. 84, pp. 931-946, 1996.
S. Fathnam and G. Slavenburg, Processing the New World of Interactive Media: The
Trimedia VLIW CPU Architecture, IEEE Signal Processing Magazine, vol. 15, pp. 108117, 1998.
S. Purcell, The Impact of Mpact 2: The Mpact 2 VLIW Media Processor Improves
Multimedia Performance in PCs, IEEE Signal Processing Magazine, vol. 15, pp. 102107, 1998.
J. McKrory and M. Daniels, The Impact of New Technology in Machine Vision,
Sensor Review, vol. 15, pp. 8-11, 1995.
E. J. Delp, Image-Processing Software and Hardware, IEEE Signal Processing
Magazine, vol. 15, pp. 8-11, 1997.
D. L. Dechow, Fundamentals of Machine Vision Systems Integration, presented at The
Vision Show, 1997.
D. Braggins, Recent Developments in Image Capture for Machine Vision, Sensor
Review, vol. 15, pp. 11-14, 1995.
M. Loinaz and B. Ackland, Video Cameras: CMOS Technology Provides On-Chip
Processing, Sensor Review, vol. 19, pp. 19-26, 1999.
M. Schanz, W. Brockherde, R. Hauschild, B. J. Hosticka, and A. Teuner, CMOS
Photosensor Arrays With On-Chip Signal Processing, presented at Proc. Of European
Solid State Circuits Conference (ESSCIRC'97), Southampton U.K., 1997.
E. R. Fossum, CMOS Image Sensors: Electronic Camera-On-A-Chip, IEEE
Transactions on Electron Devices, vol. 44, pp. 1689-1698, 1997.
K. Yamada, T. Nakano, and S. Yamamoto, Robot Vision Robust to Changes of Lighting
Conditions Using a Wide-Dynamic Range Visual Sensor, Electrical Engineering in
Japan, vol. 120, pp. 34-40, 1997.

37

[86]
[87]
[88]
[89]
[90]
[91]
[92]
[93]
[94]

K. Yamada, T. Nakano, and T. Yamamoto, A Vision Sensor Having an Expanded

Dynamic Range for Autonomous Vehicles, IEEE Transactions on Vehicular
Technology, vol. 47, pp. 332-341, 1998.
D. Braggins, Creating the Right Image, Sensor Review, vol. 16, pp. 16-21, 1996.
MotionVision Cameras-CA-D8: 1O/P, 25MHz, : Dalsa Specification Datasheet, 2000.
RACIC810/811 Random Access Charge Injection Device Imaging Sensors, : CIDTEC
Specification Datasheet, 2000.
VV5500/6500 Multi Format Digital Output Sensor, : VLSI Vision Inc. Specification
Datasheet, 2000.
D. Litwiller, CCD vs. CMOS: Facts and Fiction: Laurin Publishing, 2001.
C. Seibold, Comparison of CMOS and CCD Image Sensor Technologies, : ECE 442
Senior Design, 2002.
CMOS vs. CCD and the Future of Imaging, , 2001.
R. Hauschild, M. Hillebrand, B. J. Hosticka, J. Huppertz, T. Kneip, and M. Schwarz,
ACMOS Image Sensor with Local Brightness Adaptation and High Intrascene Dynamic
Range, presented at European Solid State Circuits Conference (ESSCIRC'97), The
Hague, The Netherlands, 1999.

38