Content uploaded by Osslan Vergara
Author content
All content in this area was uploaded by Osslan Vergara on Sep 13, 2016
Content may be subject to copyright.
Jorge LuisGarcía-Alcaraz
Aidé AracelyMaldonado-Macías
GuillermoCortes-Robles Editors
Lean
Manufacturing in
the Developing
World
Methodology, Case Studies and Trends
from Latin America
Editors
Jorge Luis García-Alcaraz
Departamento de Ingeniería Industrial y
Manufactura
Instituto de Ingeniería y Tecnología
Universidad Autónoma de Ciudad Juárez
Chihuahua
Mexico
Aidé Aracely Maldonado-Macías
Autonomous University of Ciudad Juarez
Chihuahua
Mexico
Guillermo Cortes-Robles
Institute of Technology of Orizaba
Av. Instituto Tecnologico
Orizaba
Mexico
ISBN 978-3-319-04950-2 ISBN 978-3-319-04951-9 (eBook)
DOI 10.1007/978-3-319-04951-9
Springer Cham Heidelberg New York Dordrecht London
Library of Congress Control Number: 2014934662
Springer International Publishing Switzerland 2014
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or
information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed. Exempted from this legal reservation are brief
excerpts in connection with reviews or scholarly analysis or material supplied specifically for the
purpose of being entered and executed on a computer system, for exclusive use by the purchaser of the
work. Duplication of this publication or parts thereof is permitted only under the provisions of
the Copyright Law of the Publisher’s location, in its current version, and permission for use must
always be obtained from Springer. Permissions for use may be obtained through RightsLink at the
Copyright Clearance Center. Violations are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of
publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for
any errors or omissions that may be made. The publisher makes no warranty, express or implied, with
respect to the material contained herein.
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)
Contents
Part I Introduction
1 Lean Manufacturing in Production Process
in the Automotive Industry............................. 3
Jesús Salinas-Coronado, Julián Israel Aguilar-Duque,
Diego Alfredo Tlapa-Mendoza and Guillermo Amaya-Parra
Part II Main Lean Manufacturing Tools and Techniques
2 Troubleshooting a Lean Environment ..................... 29
Moisés Tapia-Esquivias, Manuel Darío Hernández-Ripalda,
José Antonio Vázquez-López and Alicia Luna-González
3 Statistical Process Control ............................. 47
Manuel Iván Rodríguez-Borbón
and Manuel Arnoldo Rodríguez-Medina
4 Statistical Process Control: A Vital Tool
for Quality Assurance................................. 65
Jorge Meza-Jiménez, Miguel Escamilla-López
and Ricardo Llamas-Cabello
5 Process Improvement: The Six Sigma Approach ............. 87
Diego Tlapa-Mendoza, Jorge Limón-Romero,
Yolanda Báez-López and Jesús Salinas-Coronado
6 Creating the Lean-Sigma Synergy ........................ 117
Francisco J. Estrada-Orantes and Noé G. Alba-Baena
xiii
7 Automatic Product Quality Inspection Using Computer
Vision Systems ...................................... 135
Osslan Osiris Vergara-Villegas, Vianey Guadalupe Cruz-Sánchez,
Humberto de Jesús Ochoa-Domínguez,
Manuel de Jesús Nandayapa-Alfaro and Ángel Flores-Abad
8 Critical Success Factors for Kaizen Implementation .......... 157
Denisse Rivera-Mojica and Lizeth Rivera-Mojica
9 Critical Success Factors Related to the Implementation
of TPM in Ciudad Juarez Industry ....................... 179
José Torres
10 Critical Success Factors for the Implementation of JIT ........ 207
Lizeth Rivera-Mojica and Denisse Gabriela Rivera-Mojica
11 Supplier Selection in a Manufacturing Environment .......... 233
Rodrigo Villanueva-Ponce, Jaime Romero-González
and Giner Alor-Hernández
12 Megaplanning: Strategic Planning, Results Oriented
to Improve Organizational Performance ................... 253
Ernesto Alonso Lagarda-Leyva, Javier Portugal-Vásquez,
Lilia Guadalupe Valenzuela-Preciado,
Tania Mendoza-Gurrola and Roger Kaufman
Part III Human Factors and Ergonomics in Lean Manufacturing
13 Human Factors and Ergonomics for Lean
Manufacturing Applications ............................ 281
Arnulfo Aurelio Naranjo-Flores and Ernesto Ramírez-Cárdenas
14 Low Back Pain Risk Factors: An Epidemiologic Review ....... 301
Lilia Roselia Prado-León
15 Lean-Six Sigma Framework for Ergonomic Compatibility
Evaluation of Advanced Manufacturing Technology .......... 319
Aide Maldonado-Macías, Jorge Luis García-Alcaraz,
Rosa María Reyes-Martínez and Juan Luis Hernández-Arellano
xiv Contents
Chapter 7
Automatic Product Quality Inspection
Using Computer Vision Systems
Osslan Osiris Vergara-Villegas, Vianey Guadalupe Cruz-Sánchez,
Humberto de Jesús Ochoa-Domínguez,
Manuel de Jesús Nandayapa-Alfaro and Ángel Flores-Abad
Abstract The ability to visually detect the quality of a product is one of the most
important issues for the manufacturing industry because of the demand made by
the consumers is increasing. This process is typically carried out by human
experts; unfortunately experts frequently make mistakes because this process
could be tedious and tiring even for the most trained operators. A lot of solutions
have been proposed to solve this problem, such as the use of lean manufacturing
and computer vision systems. This chapter, presents a detailed explanation about
the stages involved to create a system to automatically verify the quality of an
object using computer vision and digital image processing techniques. First, a
revision of state of art researches is presented. This work also focuses on a dis-
cussion of the issues involved with computer vision applications. Afterwards, a
detailed explanation about the design of two study cases to inspect fabric and apple
defects, and its correspondent results are presented. Finally, a point of view about
the trends in automatic quality inspection systems using computer vision is
offered.
Keywords Quality control Computer vision Automation
7.1 Introduction
Visual inspection is the result of a processing carried out by a part of the brain of
the luminous information that arrives to the eyes, and is one of the main data
sources of the real world (Sannen and Van Brussel 2012). The information
O. O. Vergara-Villegas (&)V. G. Cruz-Sánchez H. de Jesús Ochoa-Domínguez
M. de Jesús Nandayapa-Alfaro Á. Flores-Abad
Universidad Autónoma de Ciudad Juárez, Instituto de Ingeniería y Tecnología, Ciudad
Juárez, Chihuahua, Mexico
e-mail: overgara@uacj.mx
J. L. García-Alcaraz et al. (eds.), Lean Manufacturing in the Developing World,
DOI: 10.1007/978-3-319-04951-9_7, Springer International Publishing Switzerland 2014
135
perceived with the sense of sight is processed in distinct ways based on the specific
characteristics needed for the future tasks to execute. As a result of an image
analysis process, the representation of an object is obtained. Immediately a deci-
sion is taken to define what to do with the visual information, which typically
implies the recognition of the object(s) detected inside a scene and the reactions
realized by a body part.
Every day, human beings recognize objects inside a particular scene observed by
means of the vision system. This process is done unconsciously (with a minimum
effort) even with a lack of the complete knowledge or description of the object to be
recognized. Recognition is a term used to describe the ability of human beings to
identify the objects around based on previous knowledge (Lee et al. 2010).
The task of visual inspection to recognize objects and to evaluate its quality
constitute one of the most important processes in several industries such as the
manufacturing industry and alimentary industry, in which, due to the customer
demands, it is mandatory to assure the quality of a product (Satorres et al. 2012;
Razmjooy et al. 2012; Peng et al. 2008; Sun et al. 2010). The issue of inspecting
objects in order to detect defects such color, scratches, cracks or checking surfaces
for a proper finish is related to visual quality inspection.
Typically, the quality inspection is made by human experts. However, the
experts frequently make mistakes, because the process could be tedious and tiring
even for well-trained operators. The problem increases usually due to, the work-
days of the inspectors are very long (more than 6 h).
This leads to several industries to look for alternatives to avoid the mistakes
made by human inspectors. One of the alternatives adopted by many industries, to
remain competitive, is the promoting of lean manufacturing, in which the practices
can work synergistically to create a streamlined, high quality system that produces
finished products at the pace of customer demand with little or no waste (Sullivan
et al. 2002; Abdulmalek and Rajgopal 2007). Unfortunately, the existent evidence
suggests that several organizational factors may enable or inhibit the implemen-
tation of lean practices among manufacturing plants (Shah and Ward 2003).
Another alternative is to provide to a computer with the ability of inspect and
recognize objects automatically. The use of a computer together with other
mechanisms, such as cameras, sensors, the knowledge provided by the human
expert, and complex algorithms, allows having a capable tool for automatic
inspection of the product quality. Then, automation becomes a necessary task for
inspection and recognition of objects in order to guarantee the quality of a product.
In this chapter, the issue of the automatic inspection of the quality of an object
using computer vision (CV) is addressed. The rest of the chapter is organized as
follows.
In Sect. 7.2, a revision of several current designs which use computer vision to
inspect the quality of an object is presented. A brief explanation of the issues
involved to create a computer vision system using digital image processing
techniques is shown in Sect. 7.3.InSect. 7.4, two study cases are discussed, the
first case addressing the problem of fabric color defects detection and the second
case addressing the problem of golden apples defect detection. A point of view
136 O. O. Vergara-Villegas et al.
about the trends in the design of computer vision systems is presented in Sect. 7.5.
Finally, Sect. 7.6, presents the conclusions obtained with this work.
7.2 Literature Review
In the literature there exist several researches which have addressed the problem of
automatic inspection of an object quality using a CV system. Machine vision has
been a widely used technology in the industry for the past three decades. It has
been an excellent tool for many industrial inspection tasks such as: brake disks,
printed circuit board (PCB), float glass, electric contacts, tiles, chickens, fruits,
vegetables, fabric, gears, chip alignment, led, to name a few.
The work proposed by (Lerones et al. 2005) defines a solution for the automatic
raw foundry brake disk dimensional characterization and visual inspection for the
automotive industry. To solve the problem, three CV techniques were used: (a) a
calibrated 3D structured-light, for dimensional characterization and inspection, (b)
a 3D uncalibrated structured-light, for local fault detection, and (c) a common 2D-
vision technique for further local fault recognition. The industrial results show that
the described system is appropriate for brake disk dimensional characterization as
well as for the detection of hard masses, featheredges, pores, hole jump obstruc-
tion, ventilation slot obstruction and veining, providing more efficiency on the
production line and bettering working conditions for human operators.
A region-oriented segmentation algorithm for detecting the most common peel
defects of citrus fruits is shown in (Blasco et al. 2007). The histogram of an input
image is computed to obtain the peak in the lower values corresponding to the
background, since it was a uniform black color that contrasted against the orange
color of the fruits. Then, the main stage was performed to detect a region of
interest consisting of the sound peel, the steam and the defects. This is made by a
region growing algorithm followed by a region merging. The algorithm is robust
against different varieties and species of citrus fruit and does not need previous
manual training or adjustments to adapt the system to work with different batches
of fruit or changes in the lighting conditions.
In the chapter presented in (Peng et al. 2008), an online defects inspection
method of float glass based on machine vision is presented. The method inspects
defects through detecting the change of image gray levels caused by the difference
in optic character between glass and defects. Initially, the noise is reduced by
means of image filtration based on gradient direction. To remove the back ground
stripes, a downward segmentation is implemented. The possible defect core and its
distortion are segmented with fixed threshold method and the OTSU algorithm
with gray range restrict. Finally, fake defects are eliminated through the method
based on defect texture detection. The success of this inspection method provides a
reference for defects detection on other materials, such as armor plate.
A CV for the automation of online inspection to differentiate freshly slaugh-
tered wholesome chickens from systemically diseased chickens is presented in
7 Automatic Product Quality Inspection Using Computer Vision Systems 137
(Yang et al. 2009). The system consisted of a camera used with an imaging
spectrograph and controlled by a computer to obtain line-scan images quickly on a
chicken processing line of a commercial poultry plant. The system scanned
chicken carcasses to locate the region of interest (ROI) of each chicken to extract
useful spectra from the ROI as inputs to the differentiation method, and to
determine the condition of each carcass as being wholesome or systemically
diseased. The high accuracy obtained with the evaluation results showed that the
machine vision system can be applied successfully to automatic online inspection
for chicken processing. Table 7.1, shows a summary of 11 CV systems contrasting
its main features and showing the accuracy recognition rates.
7.3 Computer Vision and Digital Image Processing
CV is concerned with modeling and replicating the human vision using computer
software and hardware. It is also the main theory for building artificial systems to
extract information from images. The main challenge is to combine knowledge in
computer science, electrical engineering, mathematics, physiology, biology, and
cognitive science in order to understand and simulate the operation of the human
vision system. Specifically, CV is a combination of concepts, techniques and ideas
from Digital Image Processing (DIP), Pattern Recognition (PR), Artificial Intel-
ligence (AI) and Computer Graphics (CG).
CV describes the automatic deduction of the structure and the properties of a
(possible dynamic) three-dimensional world from either a single or multiple two-
dimensional images of the world. The computer vision great trick is to extract
descriptions of the world from pictures or sequences of pictures (Nalwa 1993).
As a consequence, computer vision systems need the digital image processing
techniques to enhance the quality of the acquired images for future use or inter-
pretation. Then, DIP is concerned with taking one array of pixels as input and
producing another array of pixels as output, which, in some way, represents an
improvement of the original array. DIP is the science of modifying digital images
by means of a computer (Forsyth and Ponce 2012).
Frequently, computer vision and digital image processing are erroneously used
as a same term. The main difference is in the goals, not in methods. For example, if
the goal is to enhance an image for later use it, then this may be called digital
image processing, on the other hand, if the goal is to emulate human vision like
object recognition, defect detection or automatic driving, then it is closer to
computer vision.
There are no clear-cut boundaries in the continuum from DIP at one end to CV
at the other. However, one useful paradigm is to consider three types of com-
puterized processes: low-level, mid-level and high-level. Low-level processes
involve primitive operations for image preprocessing such as denoising and de-
blurring, this level is characterized by the fact that both inputs and outputs are
images. Mid-level processes involve tasks such as segmentation and description of
138 O. O. Vergara-Villegas et al.
individual objects. Finally, the high-level involves making sense of an ensemble of
recognized objects and performing the cognitive functions normally associated
with vision (Gonzalez and Woods 2009).
As afore mentioned, two broad categories are defined; methods whose inputs
and outputs are images and methods whose inputs may be images, but whose
outputs are attributes extracted from those images. According to (Gonzalez and
Woods 2009) the fundamental stages comprising a CV system for digital image
processing are: (a) image acquisition, (b) preprocessing, (c) segmentation, (d)
feature extraction (representation and description), (e) Recognition (interpretation
and classification), and (f) knowledge base.
In Fig. 7.1 a graphical representation of each stage and the mode of interaction
of each with the other are shown. Additionally, in the following subsections a brief
explanation of each stage is presented.
Table 7.1 A summary of 11 computer vision systems to detect defects in objects
Author
(year)
Country Object Technique Defect Accuracy
(%)
Lerones et al.
(2005)
Spain Brake disk 3D structured
light
Circularity, defective
grinding, hard
masses, feather
edges
80
Blasco et al.
(2007)
Spain Citrus Region oriented
segmentation
Peel 95
Peng et al.
(2008)
China Float glass Texture reverse
subtraction
Bubbles, lards,
optical distortion
98
Yang et al.
(2009)
USA Chickens Thresholding Wholesome, diseased 89
Sun et al.
(2010)
China Electric
contacts
Blob analysis
and PSO
Deckle edge, back
cracks, side
cracks,
eccentricity
96.7
Mery et al.
(2010)
Chile Corn tortillas SVM Size, hedonic scale,
production level
96
Gadelmawla
(2011)
Saudi
Arabia
Spur gears Tolerances Outer diameter,
diameter pitch,
circular pitch
99
Perng et al.
(2011)
China LED Thresholding Missing components,
incorrect
orientations
95
Razmjooy
et al.
(2012)
Iran Potato SVM and ANN Size, color 96.86
Moz
ˇina et al.
(2013)
Slovenia Pharmaceutical
tablets
PCA Debossing shallow 84
Lin and Fang
(2013)
China Tile Corner
convergence
and
clustering
Alignment 90.5
7 Automatic Product Quality Inspection Using Computer Vision Systems 139
7.3.1 Image Acquisition
Before any image processing can start, an image must be captured by a camera and
converted to a manageable entity. Thus, in order to acquire a digital image, an
image sensor and the ability to digitize the signal produced by that sensor are
needed (Wandell et al. 2002). The sensor can be a television camera, a line scan
camera, video, scanner, etc. If the output of the sensor is not digital, then an analog
to digital converter is necessary to digitize the image. The digital image is obtained
as a result of sampling and quantization of an analog image or created already in
digital form.
Typically, a digital image is represented as a bi-dimensional matrix of real
numbers. The convention f(x,y) is used to refer an image with size M9N, where
xdenote the row number, and ythe column number. The value of the bi-dimen-
sional function f(x, y) at any pixel of coordinates (x0, y0), denoted by f(x0, y0), is
called the intensity or gray level of the image at that pixel.
7.3.2 Preprocessing
After a digital image has been acquired, several preprocessing methods can be
applied in order to enhance the data of the image prior to the computational
processing. In this stage the image is processed and converted into a suitable form
Form ColorSize
Texture Area
Knowledge Base
Criteria
Image acquisition Segmentation
Defects
Form
Color
Color Texture Moments Mean
Feature
selection
Rotation
Scale
Translation
Feature extraction and selection
Preprocessing
RGB YIQ
Quality criteria
definition
Recognition
Compare criteria
Judge
Fig. 7.1 Fundamental steps of a computer vision for digital image processing
140 O. O. Vergara-Villegas et al.
for further analysis (Choi et al. 2011). Most of the computer vision applications
require taking care in designing the processing stage in order to achieve acceptable
results. Preprocessing operations are also called filtration.
Examples of such operations include smoothing, exposure correction and color
balancing, noise reduction (denoising), increasing sharpness, image deblurring,
image plane separation, normalization, etc. The image obtained after this stage is
the input to the segmentation step.
7.3.3 Segmentation
Autonomous segmentation is one of the most difficult stages in DIP. Segmentation
is the process of partitioning a digital image into multiple segments (sets of pixels,
also known as superpixels). The goal is to simplify and to change the represen-
tation of an image into something that is more meaningful and easier to analyze
(Padma and Sukanesh 2013). A coarse segmentation process delays the outcome of
a satisfactory solution in a DIP problem. On the other hand, a weak segmentation
process in most of the cases will lead to errors.
As a result of the segmentation process, the raw pixel data that constitute the
boundaries among the regions or information about which pixel belongs to which
region is obtained. The image segmentation is the operation that marks the tran-
sition between low-level to mid-level image processing. Among the most com-
monly used segmentation methods are: thresholding, contour detection, region
based, morphological watersheds and region growing.
7.3.4 Feature Extraction
Once an image has been segmented, the resulting individual regions can be
described. Feature extraction, also called image representation and description, is
the operation performed to extract and highlight features with some quantitative
information which is essential to distinguish one class of objects from another. It is
a critical step in most computer vision solutions because it marks the transition
from pictorial to non pictorial data representation.
In order to store the characteristics extracted from an object an n91array
called feature vector is built. The feature vector is a compact representation of an
image and its content can be symbolic, numerical or both. The main challenge in
this step is that the features extracted must be invariant to changes in rotation,
scale, translation and contrast. Obtaining the invariants ensures that the computer
vision system will be able to recognize objects even when they appear with dif-
ferent contrast, size, position and angle inside the image (Mullen et al. 2013).
The important features extracted include points, straight lines, regions with
similar properties, color, textures, shapes, and a combination. The boundary
7 Automatic Product Quality Inspection Using Computer Vision Systems 141
descriptors such as statistical moments or Fourier descriptors are often used to
tackle the problem of invariance.
7.3.5 Recognition
This stage constitutes the high-level of image processing. Recognition is the
process to assign a label to an object, based on the information provided by its
descriptors. Moreover, it involves assigning meaning to a set of recognized objects
(Xiao and Wang 2013). The recognition algorithms analyze the numerical prop-
erties of various image features and classify data into categories. This stage is
composed of two phases: training and testing.
All the classification algorithms are based on the assumption that the image in
question depicts one or more features and that each of these features belongs to
one of several distinct and exclusive classes. The classes may be stated a priori by
an analyst specifying the number of desired categories (supervised classification)
or automatically clustered (unsupervised classification) into sets of prototypes.
Quite often, the algorithms used for recognition include artificial neural net-
works (ANN), support vector machines (SVM), distance and similarity measures
(k-Nearest Neighbors, Bayesian, Euclidean, Manhattan), and matching.
7.3.6 Knowledge Base
The a priori knowledge about a specific image processing problem is coded in the
form of a knowledge database. The database may be as simple as detailing regions
of an image where the information of interest is known to be located, thus limiting
the search that has to be conducted for seeking information; or can be quite
complex such as inter related list of all major possible defects in a material
inspection problem (Gonzalez and Woods 2009). The knowledge base guides the
operation of each processing module and also controls the interaction among all
the modules.
7.4 Study Cases
As described in Sect. 7.2, several computer vision systems have been created in
different industries. Even, when almost all the systems are build based on the steps
explained in Sect. 7.3, not always all the steps are used. Sometimes, other more
specific steps are included. The purpose of the current section is to present two
study cases, one is to offer the specific details on how to build a computer vision
142 O. O. Vergara-Villegas et al.
system using digital image processing techniques to inspect the color quality
fabric, and the other to inspect defects in apples.
7.4.1 Automatic Inspection of Fabric Quality
In the textile industry, defects detection is of crucial importance in the quality
control of manufacturing (Bissi et al. 2013). The common inspection process is
made by human experts, but since this task is dull and repetitive many errors are
frequently committed. It has been estimated that the price of fabrics is reduced
from 45 to 65 % due to the presence of defects (Kumar 2008).
The fabric is affected by yarn quality, loom defects and stamping process. The
textile industry has the necessity to verify the quality of a stamping fabric process.
The following subsections explain the design and implementation of a computer
vision system to inspect color degradation in fabric.
7.4.1.1 Definition of the Quality Criteria
In the textile industry a common error that causes defects in the stamping process
of a fabric roll is that some of the colors being used goes finishing. For a human
inspector, this color degradation can be detected until the color difference between
the original and the degraded fabric is very noticeable. For this example, the
quality of a fabric is defined by the conservation of colors at the stamping process.
The input to the system is a criterion about the quantity of color variations
accepted to consider a fabric without defects (good quality). At this stage a range
between 1 and 10 % of color degradation is accepted for a good quality fabric,
while out of that range the fabric is considered of bad quality.
7.4.1.2 Acquisition of Fabric Images
The stamping process was simulated using artificial fabric images, due to real
stamping process imaging could not be achieved. Three databases of images were
selected, each one containing what is called a texton that is the main element that
compose a texture. The public databases used were: ‘‘Artificial Color Texture’’,
‘‘Forrest textures’’, and ‘‘Texture Gallery’’. A subset of ten textons was selected
from each database.
7.4.1.3 Preprocessing of Fabric Images
To automatically simulate the creation of a fabric roll, each texton was repeated
one after another 60 times. This operation was made randomly with rotations of 90
7 Automatic Product Quality Inspection Using Computer Vision Systems 143
and 180clockwise, scaling of doubled and halved of the original size, scaled and
rotated together, and with addition of 10 and 40 % of salt and pepper noise with
zero mean and a variance of 400. A random degradation in red, green and blue
(RGB) colors of 1, 2, 5, 10, 20, 40 and 50 % was applied.
At the end of this process 1400 images were created, 700 images of good
quality fabric and 700 images of bad quality fabric. In Fig. 7.2, several textons
with different changes are depicted. In Fig. 7.3 an example of good quality fabric
and bad quality fabric is shown.
7.4.1.4 Feature Extraction from Fabric Image Rolls
At this stage, a feature vector to store the statistical features associated with each
fabric roll was computed. The process of the texture feature extraction was carried
out in three RGB (Red, Green, Blue) planes and in the intensity plane of the HSI
(Hue, Saturation, Intensity) model. In summary, the features extracted to handle
image invariance were: 10 normal moments, 10 central moments, 7 Hu moments,
6 Sidharta Maitra moments, mean, variance and standard deviation in red, green
and blue, and mean for the intensity of the HSI model. At the end of this step, a
feature vector with 43 characteristics was obtained for each fabric roll.
(d) (e) (f)
(a) (b) (c)
Fig. 7.2 Examples of
textons. aOriginal texton:
125_exhaustive_p1,
b125_exhaustive_p1 rotated
90clockwise,
c125_exhaustive_p1
doubled, dOriginal texton:
S_S_cloth2, eS_S_cloth2
rotated counterclockwise, and
fS_S_cloth2 rotated
clockwise and halved
Fig. 7.3 Examples of fabric
rolls. aGood quality fabric,
and bBad quality fabric
144 O. O. Vergara-Villegas et al.
7.4.1.5 Feature Selection
This stage consists on the selection of the more discriminate features to distinguish
one fabric from another. In order to perform this stage, the so called testor theory
was used. The algorithm used was the well known BT algorithm to obtain typical
testors.
The BT algorithm defines a testor for two classes T
0
and T
1
as: ‘‘A set
t=(i1,…,is) of columns of the table T(and the respective features xi1,…, xis), is
called testor for (T
0
,T
1
)=T, if after eliminating from Tall of the columns except
those belonging to t, does not exist any row in T
0
equal to T
1
’’.
A testor is called irreducible (typical) if upon eliminating any columns it stops
being testor for (T
0
,T
1
), where sis called the length of the testor. The BT algo-
rithm for typical testors is based in the idea of generate boolean vectors of
a=(0,….., 0, 1) until arrive to a=(1,1,….,1, 1). For each case, it is verified if
the set of columns that correspond to the coordinates of the n-tuple generated is a
typical testor.
At this point, it is needed to have the learning matrix which contains the
descriptions of objects in terms of a set of features; the matrix of differences, which
it is obtained from the learning matrix comparing the values of object features
pertaining to different classes; and the basic matrix, that is formed exclusively by
basic rows (incomparable rows).
After computing the BT algorithm, a measurement of the importance of a single
feature was calculated. After the typical testors were obtained, an irreducible
combination of fewer features was obtained, each feature was very important to
assure the difference from one class to another (bad vs good quality). Therefore, if
a feature appears in many testors, it cannot be eliminated because it helps to
preserve the class separation.
After this phase, the feature vector was reduced by a 76.7 %. Only 10 features
were selected from the original set of 43 features to be used in the final recognition
step. The features selected were: 2 central moments, 3 Maitra moments, red and
blue color mean, color green variance, illumination mean and blue standard
deviation.
7.4.1.6 Recognition of Fabric Images
In the recognition step, the voting algorithm (Alvot), which is based on partial
precedence or partial analogies, was used. The premise of the algorithm is the fol-
lowing: if an object may looks like another, but not totally, and that the parts that look
alike can offer information about possible regularities. Thus, a final decision is taken.
The model of voting algorithms is described in six steps: (a) obtain the system of
support sets, (b) computation of the similarity function, (c) evaluation by row for a
given support set, (d) evaluation by class for a given support set, (e) evaluation by
class for all the system of support sets, and (f) obtaining the solution rule.
7 Automatic Product Quality Inspection Using Computer Vision Systems 145
After the computation of all the stages involved in voting algorithm, the dis-
crepancies among the object inspected and its correspondent model were deter-
mined and a decision about the membership to good or bad quality class of the
inspected object is emitted.
7.4.1.7 Experimentation and Results
Five different types of tests were carried out: Case (a)Validation of the system
learning ability: Consists on validating if the system performs the training in a
correct way. In these test, the system does not fail when an image of the training
set is used as input. Case (b)Rotation changes: The goal is to verify if the system
can handle rotation effects at 90 and 1808with regard to the original fabric images.
Case (c)Scale changes: The goal is to verify if the system can handle scaling
effects of halved and doubled the size of the original fabric images. Case
(d)Rotation and scale changes: This test was made to verify the case b and c
together. Case (e)Noise insertion: The objective is to evaluate if the system can
handle images contaminated with salt and pepper noise, with a probability of 10 %
for good quality and 40 % for bad quality.
The results obtained in each case are depicted in Table 7.2.
As can be seen in Table 7.2, the techniques utilized were effective and an
average recognition of 71.2 % was obtained. The main contribution is that the
system can handle invariance to rotation, scale and noise in a separated and in a
combined way.
7.4.2 Automatic Inspection of Apple Quality
The creation of computer vision applications for agricultural industry has
increased considerably in recent years (Cubero et al. 2011). The agricultural
industry has the necessity to verify the quality of fruits and vegetables during the
storage process. Typical target applications of inspection systems include grading
and quality estimation from external parameters or internal features. Much of the
sorting and grading processes are still not automatic. Manual inspection of fruit is
tedious and can cause eye fatigue. Frequently, the inspection process is subject to
errors due to different judgment emitted by different persons.
This study case, addresses the problem of design and implement a computer
vision system for the automatic inspection of the quality fruit, particularly the
golden apples. Apples are very susceptible to damage and the presence of bruises
on the apple skin affects not only the appearance of the apple, which is an
important indicator of quality, but also accelerates its deterioration (Santos and
Rodrigues 2012). Two main problems were detected for the implementation of a
computer vision system for apple grading: (1) how to acquire the whole surface
146 O. O. Vergara-Villegas et al.
image of an apple by cameras at an on-line speed, and (2) how to quickly identify
the stem, calyx and the presence of defects (Li et al. 2002).
The solution presented for apple classification tackles both problems. Addi-
tionally, the system not use only the typical numerical information obtained from a
computer vision system but also a symbolic knowledge obtained from a human
expert is added to enhance the ability of the system to classify the apples. The
approach presented is called a neuro-symbolic hybrid system (NSHS).
7.4.2.1 Definition of the Quality Criteria
A hybrid system offers the possibility to integrate two or more knowledge rep-
resentations of a particular domain in a single system. One of the main goals is to
obtain the complementary knowledge that allows improving the efficiency of the
global system. A specific example of a hybrid system is the so called NSHS, which
is mainly based on the symbolic representation of an object obtained from a human
expert in the form of rules and a computer vision system to obtain the numerical
information.
The quality criteria to evaluate a golden apple were obtained from a human
expert in apple classification. For the case of golden apples, a category is assigned
depending on the value of the external attributes. There exist four categories:
category extra, category I, category II, and category III. In this example, only the
category extra is evaluated, by that an apple can belong to one of two classes: good
or bad quality. In Fig. 7.4, an example of good and bad quality apples are depicted.
Additionally, in Table 7.3 a resume of the external attributes of a golden apple
with its associated variable name, as well as the value and type are shown.
7.4.2.2 Image Acquisition
For the problem of category extra golden apple inspection, 148 images were
obtained by means of a digital camera. The golden apples were acquired from crop
fields at Puebla city in Mexico. For the complete set of images an operation of
rotating 90 and 180clockwise, doubled and halved the scaling and noise addition,
similar to the study case of fabric was performed. At the end of this process the set
Table 7.2 Results obtained for fabric roll color inspection
Case # images % of success % of failure
a 200 100 0
b 360 69.7 30.3
c 360 64.15 35.85
d 240 72.31 27.69
e 240 50.2 49.8
Total 1400 556.36 143.64
Average 71.2 28.8
7 Automatic Product Quality Inspection Using Computer Vision Systems 147
of 148 apples were divided into two categories: bad (74) and good (74) quality. In
Fig. 7.5, an example of different apples after changing in rotation and scale are
depicted.
7.4.2.3 Image Preprocessing
The stage consists of the image conversion from RGB color model to YIQ
(Luminance, Inphase, Quadrature) color model. The main reason to perform such
conversion is to facilitate the image feature extraction. The YIQ model was
computed to separate the color from the luminance component because of the
ability of the human visual system to tolerate more changes in reflectance than to
changes in shade or saturation. The main advantage of the model is that reflectance
(Y) and the color information (I and Q) can be processed separately. Reflectance is
proportional to the amount of light perceived by human eye.
7.4.2.4 Apple Feature Extraction
The characteristics of each image were extracted based on the information defined
by the human expert in the form of rules and by means of image processing in the
Fig. 7.4 Examples of category extra golden apples. aGood quality apple, and bBad quality
apples
Table 7.3 Quality criteria determined to measure the quality of an apple
Attribute Acronym Type Value
Lengthened defect LD Range 0–6
Spotted defect SD Range 0–2.5
Various defects VD Range 0–5
Stem S Binary True/false
Red color RC Range 0–255
Green color GC Range 0–255
Blue color BC Range 0–255
148 O. O. Vergara-Villegas et al.
form of numerical data. These two types of knowledge information were combined
in order to obtain the overall representation of an apple.
The number of rules defined by the experts were four, an example of a single
rule is the following: ‘‘IF an apple has the corresponding color, and has the stem,
and has lengthened defects that do not exceed 2 cm, and has several defects that
do not exceed 1 cm
2
, and has spotted defects that do not exceed cm
2
, THEN the
apple belongs to category extra with good quality’’.
In order to obtain the numerical values, the same process of feature extraction
for the problem of fabric was performed.
At the end of this stage, the rules were compiled with a knowledge based
artificial neural network (KBANN), in order to obtain a representation of the
information that further can be combined with the numerical results extracted from
the computer vision system. The combination was made using the method called
Neusim, which is based on the cascade correlation Fahlman algorithm.
7.4.2.5 Apple Classification
The output of the apple feature extraction stage was a joined representation of the
symbolic and numerical knowledge. In order to further classify a golden apple, a
refinement of that knowledge must be made. That refinement is performed by
running again the Neusim method, but now not to join knowledge but use it as a
classifier.
The main advantage of use Neusim algorithm is that one can see the number of
hidden units added during the learning process, this is very useful to monitor the
complete process of incremental learning. The output of this stage is the decision
about the quality of an apple in one of two classes, bad or good.
Fig. 7.5 Golden apples of category extra. aOriginal apple image, bApple rotated 180
clockwise, and cApple halved the original size
7 Automatic Product Quality Inspection Using Computer Vision Systems 149
7.4.2.6 Experiments and Results
From the total set of 128 golden apple images, 74 were used for the stage of
training and 74 for the stage of recognition. In order to carry out the experiments,
three different approaches were selected: (a) a connectionist approach which uses
only the data obtained from the computer vision system, (b) a symbolic approach
which uses only the data obtained from the compiled rules, and (c) a NSHS, which
is a combination of connectionist and symbolic approach.
For the case of the tests using the connectionist approach, three scenarios were
defined: (a) using the numerical data obtained from the overall 148 images
(100 %), (b) using the data obtained only from 111 images (75 %), and (c) using
only data from 74 images (50 %). Three rules were used to obtain the results
referring to the test case using the symbolic approach. The first rule called 7
involves the following seven attributes: LD, SD, VD, S, RC, GC, and BC. The
second rule called R5 considers the following five attributes: RC, GC, BC, S, and
LD. Finally, the third rule called R4 includes the following four attributes: LD, SD,
VD, and S. For the case of NSHS approach, a combination of connectionist and
symbolic was made. Three rules R7, R5, and R4 were combined with 100, 75 and
50 % of the total number of examples. The overall results obtained are shown in
Table 7.4.
One of the typical problem causing failures in computer vision systems is the
lack of the complete description of an object. This can be observed by verifying
the results obtained in Table 7.4 with symbolic and connectionist approaches. This
disadvantage can be withdrawn by using a method to complete the information
with the data defined by the knowledge of a human expert. The systems which
allow this combination types are called NSHS, as can be seen with results shown
in Table 7.4; these systems are very efficient for complementing the necessary
knowledge for an automatic object inspection by means of a computer vision
system.
For example, in the pure symbolic approach, the rule R4 was not enough to
classify the apples correctly, but when it is integrated with the group of numeric
examples (100, 75, 50 %), a substantial improvement is obtained, because the
knowledge that does not contain the rule, is complemented with the numerical base
of examples.
7.5 The Future of Computer Vision Systems
Even when the last few years have shown many computer vision systems which
have been proven to be very efficient for the particular task for which they were
created, there are still many challenges to be solved. The researchers divide these
challenges into two categories, issues referring to hardware design techniques and
issues concerned to software algorithms (Andreopoulos and Tsotsos 2013).
150 O. O. Vergara-Villegas et al.
7.5.1 Issues Concerning Hardware
This challenge concerns the design and build of specific hardware to solve the
typical problems of computer vision, due to that there is a significant gap in terms
of the input size and the computational resources needed to reliably process those
inputs. The solution needs to take into account variables such as real time
acquisition huge storage space, distributed information, depth information, parallel
execution, and portability. Following several of the hardware trends are explained.
Designing of sensors such Microsoft Kinect: With the invention of the low-cost
Microsoft Kinect sensor, high-resolution depth and visual RGB sensing has
become available for widespread uses. The complementary nature of the depth and
visual information provided by the Kinect sensor opens up new opportunities to
solve fundamental problems in computer vision and to design and build other
powerful sensors based on this technology. The main topics that will be solved
include preprocessing, object tracking and recognition, human activity analysis,
hand gesture analysis, and indoors 3-D mapping.
Distributed smart cameras: This implies the design and implementation of real
time distributed embedded systems that perform computer vision using multiple
cameras. This approach has emerged thanks to a confluence of simultaneous
advances in disciplines such: computer vision, image sensors, embedded com-
puting, and sensor networks. Processing images in a network of distributed smart
cameras introduces several complications such as the tracking process. The dis-
tributed smart cameras will represent key components for future embedded
computer vision systems and that smart cameras will become an enabling tech-
nology for many new applications.
Table 7.4 The results obtained for the three approaches proposed for tests
Approach Compiled rules % of examples used Accuracy (%)
Connectionist – 100 95.14
– 75 91.21
– 50 90.54
Symbolic R7 – 93
R5 – 90.12
R4 – 14.19
NSHS R7 100 96.62
R7 75 95.27
R7 50 90.54
R5 100 95.27
R5 75 95.94
R5 50 96.62
R4 100 91.22
R4 75 93.24
R4 50 94.59
7 Automatic Product Quality Inspection Using Computer Vision Systems 151
Multicore processors: The recent emergence of multi-core processors enables a
new trend in the usage of computers. Computer vision applications, typically
requires heavy computation and lots of bandwidth, which makes difficult to run the
applications in real time. The use of multi-core processors can potentially serve the
needs of such workloads. In addition, more advanced algorithms can be developed
utilizing the new computation paradigm. The main advantage offered will be the
execution of parallel algorithms, for example by taking multiple cameras inputs of
a scene, extracting useful features, and performing statistical inference almost at
the same time. In resume, the fact of parallelizing the workload will speed up the
performance of the computer vision systems.
Graphics processing units (GPU): Computer vision tasks are computationally
intensive and repetitive, and they often exceed the capabilities of the CPU, leaving
little time for higher level tasks. However, many computer vision operations map
efficiently onto the modern GPU, whose programmability allow a wide variety of
computer vision algorithms to be implemented. The GPU provides a streaming,
data parallel arithmetic architecture which carries out a similar set of calculations
on an array of image data. The single instruction multiple data (SIMD) capability
of the GPU makes it suitable for running tasks, which often involve similar cal-
culations operating on an entire image.
Mobile devices: Since mobile devices are gaining increasingly powerful pro-
cessors, great quality cameras and high speed network access, they will be suitable to
implement computer vision algorithms. The ubiquitous platform offered by a mobile
device can be used for a wide range of applications such as: smart surveillance,
assembly line troubleshooting aid, controlling specific devices with gestures or
screen touches, and entertainment. Additionally, the sensors often included in mobile
devices such as: GPS, gyroscopes and accelerometers can contribute to enhance the
computer vision tasks. The main disadvantage is the life time battery.
7.5.2 Issues Concerning Software
One of the main problems to solve is to find a way to generalize the knowledge as
well as the human vision system, enhance the way for the representation of objects,
and the ability for learning and inferencing the models used. Following several of
the software trends are explained.
Multispectral and hyper spectral image analysis: The data of the images is
obtained at specific frequencies across the electromagnetic spectrum. The spectral
imaging allows the extraction of additional information which human eye fails to
capture with its receptors for red, green and blue. After the representation in the
electromagnetic spectrum, the images are combined to form a three-dimensional
hyper spectral data cube for processing and analysis. The main difference between
multispectral and hyper spectral images is the number of bands obtained. The main
areas for the application of this type of imaging are agriculture, mineralogy,
physics, and surveillance.
152 O. O. Vergara-Villegas et al.
RGB-D images for representation and recognition: This type of images are
formed with the classical red, green and blue information and adding information
about the scene depth, which is obtained by a technique called structured light
implemented with an infra red laser emitter. This type of images is obtained using
a kinect sensor and the data containing visual and geometric information of the
scene. This image offers the advantage to obtain the depth information using only
one device instead of using the classical pair of images. The features obtained with
RGB-D images will be very useful to improve the process of shape estimation, 3D
mapping and localization, path planning, navigation, pose recognition and people
tracking.
Real time display of obscured objects: This kind of algorithms could help car
drivers and airplane pilots to see through fog, and submarines explore under the
sea. The algorithms could provide safety features for future intelligent transpor-
tation systems. This means that computer vision software modules can distinguish
specific objects more accurately from the rest of a scene, even in complete
darkness.
Deep and feature learning: One of the main challenges for computer vision and
machine learning is the problem of learning representations of the perceptual
world. The learning and perceptual methods allows automatically learn a good
representation of the input unlabeled data, offering a time reduction against typical
learning algorithms which spend a lot of time to obtain the input feature repre-
sentation. Since these algorithms mostly learn from unlabeled data, they have the
potential to learn from vastly increased amounts of data (since unlabeled data is
cheap), and therefore also achieving a vastly improved performance. The repre-
sentation of multilevel hierarchies obtained are useful not only for low level
features such as edge or blob detectors, but also are useful for high level concepts
such as face recognition.
Mid-level patch discovery: The technique allows discovering a set of dis-
criminative patches that can serve as a fully unsupervised mid-level visual rep-
resentation. The desired patches need to satisfy two requirements (1) to be
representative, they need to occur frequently enough in the visual world, and (2) to
be discriminative, that is, they need to be different enough from the rest of the
visual world. The patches could correspond to parts, objects, visual phrases, etc.
but they are not restricted to be any one of them. The patches are simple to
compute, and offers very good discriminability, broad coverage, better purity, and
improved performance compared to visual world features. These approaches can
be used mainly on scene classification, beating bag-of-words, spatial pyramids,
object bank, and scene deformable-parts models.
Augmented reality: Allows enhancing the senses of a user by manipulating
virtual objects superimposed on top of the real world scenes. In other words, AR
bridges the gap between the real and the virtual in a seamless way. Specifically,
because of the improved ease-of-use of augmented reality interfaces, these systems
may serve as new platforms to gather data, as imagers may be pointed by users to
survey and annotate objects of interest to be stored in different kind of systems
7 Automatic Product Quality Inspection Using Computer Vision Systems 153
such as educational learning, games, training skills and geographical information
systems (GIS).
Gesture control: Human–computer interfaces (HCI) needs to evolve from
mouse-keyboard based interaction, multi-touch screens and other exotic approa-
ches such as using special gloves or other devices, until reach the use of gestures to
translate human actions into control applications. There were and still are a series
of attempts to produce computer control scripts via gesture-based interfaces, and
the research literature is abundant in chapters on this subject. The main challenge
is that the computer-based control via gestures has to be robust and in real-time.
Any lag in the result can lead to users abandoning it.
7.6 Conclusions
In this chapter the design and implementation of two computer vision systems to
verify the quality of fabric and apples was presented. The main challenges to build
the computer vision systems were explained exhaustively. A brief explanation and
contrasting of several computer vision systems already presented in the literature
were discussed. Additionally, a highlighting was made about its ability or accuracy
to detect defects of an object. Finally, a brief explanation about the future of
computer vision systems for quality inspection was presented.
Despite the numerous studies developed in the computer vision area, there is
still not a standardized method which could be proposed for the assessment of the
quality of different types of objects. The particular characteristics of an object
require a computer vision system to be customized; this implies an exhaustive
research process and not only the purchase of expensive equipment. Besides, in
order to obtain a better performance of the system, the acquisition of the knowl-
edge from the human experts and the techniques to represent it in terms of
numerical information is mandatory.
References
Abdulmalek, F., & Rajgopal, J. (2007). Analyzing the benefits of lean manufacturing and value
stream mapping via simulation: A process sector case study. International Journal of
Production Economics, 107(1), 223–236.
Andreopoulos, A., & Tsotsos, J. (2013). 50 years of object recognition: Directions forward.
Computer Vision and Image Understanding, 117(8), 827–891.
Bissi, L., Baruffa, G., Placidi, P., Ricci, E., Scorzoni, A., & Valigi, P. (2013). Automated defect
detection in uniform and structured fabrics using Gabor filters and PCA. Journal of Visual
Communication and Image Representation, 24(7), 838–845.
Blasco, J., Aleixos, N., & Moltó, E. (2007). Computer vision detection of peel defects in citrus by
means of a region oriented segmentation algorithm. Journal of Food Engineering, 81(3),
535–543.
154 O. O. Vergara-Villegas et al.
Choi, J., Ro, Y., & Plataniotis, J. (2011). A comparative study of preprocessing mismatch effects
in color image based face recognition. Pattern Recognition, 44(2), 412–430.
Cubero, S., Aleixos, N., Moltó, E., Gómez-Sanchis, J., & Blasco, J. (2011). Advances in machine
vision applications for automatic inspection and quality evaluation of fruits and vegetables.
Food and Bioprocess Technology, 4(4), 487–504.
Forsyth, D., & Ponce, J. (2012). Computer vision: A modern approach (2
nd
ed.). New Jersey,
USA: Pearson.
Gadelmawla, E. (2011). Computer vision algorithms for measurement and inspection of spur
gears. Measurement, 44(9), 1669–1678.
Gonzalez, R., & Woods, R. (2009). Digital image processing (3rd ed.). New Jersey, USA:
Prentice Hall.
Kumar, A. (2008). Computer-vision-based fabric defect detection: A survey. IEEE Transactions
on Industrial Electronics, 55(1), 348–363.
Lee, S., Kim, K., Kim, J., Kim, M., & Yoo, H. (2010). Familiarity based unified visual attention
model for fast and robust object recognition. Pattern Recognition, 43(3), 116–1128.
Lerones, P., Fernández, J., García-Bermejo, J., & Zalama, E. (2005). Total quality control for
automotive raw foundry brake disks. International Journal of Advanced Manufacturing and
Technology, 27(3–4), 359–371.
Li, Q., Wang, M., & Gu, W. (2002). Computer vision based system for apple surface defect
detection. Computers and Electronics in Agriculture, 36(2–3), 215–236.
Lin, K., & Fang, J. (2013). Applications of computer vision on tile alignment inspection.
Automation in construction, 35(November), 562–567.
Mery, D., Chanona-Pérez, J., Soto, A., Aguilera, J., Cipriano, A., & Veléz-Rivera, N., et al.,
(2010). Quality classification of corn tortillas using computer vision, Journal of food
engineering, 101(4): 357–364.
Moz
ˇina, M., Tomaz
ˇevic, D., Pernuš, F., & Likar, B. (2013). Automated visual inspection of
imprint quality of pharmaceuticals tablets. Machine Vision and Applications, 24(1), 63–73.
Mullen, R., Monekosso, D., & Remagnino, P. (2013). Ant algorithms for image feature
extraction. Experts Systems with Applications, 40(11), 4315–4332.
Nalwa, V. (1993). A guided tour of computer vision (1
st
ed.). Boston, Massachusetts, USA:
Addison-Wesley Longman Publishing Co.
Padma, A., & Sukanesh, R. (2013). Wavelet statistical texture features-based segmentation and
classification of brain computed tomography images. IET Image Processing, 7(1), 25–32.
Peng, X., Chen, Y., Yu, W., Zhou, Z., & Sun, G. (2008). An online defects inspection method for
float glass fabrication based on machine vision. International Journal of Advanced
Manufacturing Technology, 39(11–12), 1180–1189.
Perng, D., Liu, H., & Chang, C. (2011). Automated SMD LED inspection using machine vision.
The International Journal of Advanced Manufacturing Technology, 57(9–12), 1065–1077.
Razmjooy, N., Somayeh, B., & Soleymani, F. (2012). A real-time mathematical computer
method for potato inspection using machine vision. Computers and Mathematics with
Applications, 63(1), 268–279.
Sannen, S., & Van Brussel, H. (2012). A multilevel information fusion approach for visual
quality inspection. Information Fusion, 13(1), 48–59.
Santos, J., & Rodrigues, F. (2012). Applications of computer vision techniques in the agriculture
and food industry: A review. European Food Research and Technology, 5(6), 989–1000.
Satorres, S., Gómez, J., Gámez, J., & Sánchez, A. (2012). A machine vision system for defect
characterization on transparent parts with non-plane surfaces. Machine Vision and Applica-
tions, 23(1), 1–13.
Shah, R., & Ward, P. (2003). Lean manufacturing: context, practice bundles, and performance.
Journal of Operations Management, 21(2), 129–149.
Sullivan, W., McDonald, T., & Van Aken, E. (2002). Equipment replacement decisions and lean
manufacturing. Robotics and Computer Integrated Manufacturing, 18(3–4), 255–265.
Sun, T., Tseng, C., & Chen, M. (2010). Electric contacts inspection using machine vision. Image
and Vision Computing, 28(6), 890–901.
7 Automatic Product Quality Inspection Using Computer Vision Systems 155
Wandell, B., Gamal, A., & Girod, B. (2002). Common principles of image acquisition systems
and biological vision. Proceedings of the IEEE, 90(1), 5–17.
Xiao, B., & Wang, G. (2013). Generic radial orthogonal moment invariants for invariant image
recognition. Journal of Visual Communication and Image Representation, 24(7), 1002–1008.
Yang, C., Chao, K., & Kim, M. (2009). Machine vision system for online inspection of freshly
slaughtered chickens. Sensors and Instruments for Food Quality and Safety, 3(1), 70–80.
156 O. O. Vergara-Villegas et al.
