IMAGE-BASED OBJECT DETECTION AND IDENTIFICATION
SALIM MOHAMMED HUMAID ALWAILI
A thesis submitted in fulfillment of the
requirements for the award of the degree of
Master of Science (Remote Sensing)
Faculty of Geoinformation and Science Engineering
Universiti Teknologi Malaysia
May 2007
PSZ 19:16 (Pind. 1/97)
UNIVERSITI TEKNOLOGI MALAYSIA
BORANG PENGESAHAN STATUS TESISυ
JUDUL:
IMAGE-BASED OBJECT DETECTION AND IDENTIFICATION
____________________________________________________________________
____________________________________________________________________
SESI PENGAJIAN:
Saya
II(2006/2007)
SALIM MOHAMMED HUMAID ALWAILI
(HURUF BESAR)
mengaku membenarkan tesis (PSM/Sarjana/Doktor Falsafah)* ini disimpan di Perpustakaan
Universiti Teknologi Malaysia dengan syarat-syarat kegunaan seperti berikut:
1. Tesis adalah hakmilik Universiti Teknologi Malaysia.
2. Perpustakaan Universiti Teknologi Malaysia dibenarkan membuat salinan untuk tujuan pengajian
sahaja.
3. Perpustakaan dibenarkan membuat salinan tesis ini sebagai bahan pertukaran antara institusi
pengajian tinggi.
4. ** Sila tandakan (4)
9
SULIT
(Mengandungi maklumat yang berdarjah keselamatan atau
kepentingan Malaysia seperti yang termaktub di alam AKTA
RAHSIA RASMI 1972)
TERHAD
(Mengandungi maklumat TERHAD yang telah ditentukan oleh
organisasi/badan di mana penyelidikan dijalankan)
TIDAK TERHAD
Disahkan oleh
--------------------------------------------------------
----------------------------------------------------------
(TANDATANGAN PENULIS)
(TANDATANGAN PENYELIA)
Alamat Tetap:
PO Box 113, PC 113
NATIONAL SURVEY AUTHORITY
SULTANATE OF OMAN
Tarikh:
15TH OF MAY 2007
CATATAN:
EN. SAMSUDIN BIN AHMAD
Nama Penyelia
Tarikh:
15TH OF MAY 2007
*
**
Potong yang tidak berkenaan.
Jika tesis ini SULIT atau TERHAD, sila lampirkan surat daripada pihak berkuasa/organisasi
berkenaan dengan menyatakan sekali sebab dan tempoh tesis ini perlu
υ
Tesis dimaksudkan sebagai tesis bagi Ijazah Doktor Falsafah dan Serjana secara
penyelidikan, atau disertasi bagi pengajian secara kerja kursus dan penyelidikan, atau
Laporan Projek Sarjana Muda (PSM).
“I hereby declare that I have read this thesis and in my opinion
this thesis is sufficient in terms of scope and quality for the
award of the degree of Masters of Science (Remote Sensing)”
Signature
: ..........................................................
Name of Supervisor I : EN. SAMSUDIN BIN AHMAD
Date
: 15TH OF MAY 2007
ii
I declare that this thesis entitled “Image-Based Object Detection and
Identification” is the result of my own research except as cited in the
references. The thesis has not been accepted for any degree and is not
concurrently submitted in candidature of any other degree.
Signature
:............................................................
Name
: Salim Mohammed Humaid Alwaili
Date
:15th of May 2007
iii
To the spirit of my beloved father,
To my mother,
To my wife Ziyana,
And to my kids Bader, Ahmed, and Mohammed
iv
ACKNOWLEDGEMENT
I wish to express my sincere appreciation to the Head of Remote Sensing
section Professor Dr. Mazlan Bin Hashim, for his encouragement, guidance,
patience, and support. I will always be indebted to him for her persistence in
ensuring that this thesis is up to the academic world standards. I am also very
thankful to my supervisor En. Samsudin Bin Ahmad for his guidance.
I am also thankful to the staff in the Faculty of Geoinformation and Science
Engineering for their assistance during the past three semesters.
Last, but not least, I am extremely grateful and indebted to my sponsors, the
National Survey Authority. In particular, I would like to thank the Head of the NSA
Brigadier Mohammed Said Al-Kharousi and the Director of Support Services.
v
ABSTRACT
Object detection is a process of identifying and locating objects in image
scene. It is a process of combining digital image processing and computer vision.
There are two main approaches for object detection, namely contour-based and
region-based detection for both approaches. Segmentation plays an important role in
object-detection. Detection and extraction of the spatial and spectral responses of
different targets of interest could only be carried out with different approaches. In
this study, object-based detection is carried out to identify and recognize typical
military targets (Plane1, Plane2, Tank) located on natural background using spectral
and spatial approaches. This is useful in military intelligence or strategic planning
where it assists in identifying the aforementioned targets. In the spectral-based
object detection, the spectroradiometer readings of the targets are individually
observed on different backgrounds (white platform used as calibration plane, sand
and camouflage). These spectroradiometer observations are reduced and examined
for forming the unique spectral signatures of the targets. Spatial-characteristics of
the targets are also examined using object-based detection where multi-resolution
segmentation were employed for this purpose. Results of the study obtained that all
the three targets can be extracted with more than 20% differentiation, while the
spectral characteristics of the targets are also good indications for object detection
with overall accuracy of RMSE = ± 0.6.
vi
ABSTRAK
Pengesanan objek daripada imej adalah satu proses di mana objek sasaran
dikenalpasti dan ditentukan. Proses ini menggabungkan pemprosesan imej digital
dengan ‘ computer vision’. Terdapat dua pendekatam utama dalam proses
pengesanan objek iaitu Berdasarkan Kontur dan Berdasarkan Bahagian. Teknik
Segmentasi merupakan teknik utama bagi kedua-dua pendekatan ini. Pengesanan
penentuan elemen spatial dan spektral objek sasaran adalah fokus utama dalam
kajian ini, di mana sasaran utama yang digunakan mempunyai kepentingan
ketenteraan, iaitu Kapal Terbang Pejuang 1, Kapal Terbang Pejuang 2 dan Kereta
Kebal. Bagi Pendekatan Spektral, cerapan alat spektroradiometer dijalankan bagi
semua objek sasaran dengan pelbagai variasi latarbelakang. Terdiri daripada
permukaan putih yang digunakan sebagai rujukan kalibrasi, latarbelakang pasir dan
bahan penyamaran. Cerapan spektroradiometer ini kemudiannya dilaraskan sebelum
pencarian ciri-ciri ‘spectral signatures’ sasaran dibuat. Penentuan ciri-ciri spatial
objek dijalankan dengan menggunakan ‘multi-resolution segmentation’. Hasil
daripada kajian ini mendapati kaedah spektral boleh digunakan untuk mengenalpasti
sasaran yang diuji dengan ketepatan RMSE = ± 0.6.
vii
TABLE OF CONTENTS
CHAPTER
1
2
TITLE
PAGE
TITLE
i
DECLEARATION
ii
DEDICATION
iii
AKNOWLEDGMENT
iv
ABSTRACT
v
ABSTRAK
vi
TABLE OF CONTENTS
vii
LISTS OF TABLES
ix
LIST OF FIGURES
x
LIST OF APPENDIX
xi
INTRODUCTION
1
1.1
The Importance of Object Detection
1
1.2
Problem Statement
4
1.3
Objectives
4
1.4
Scope of the Study
5
1.5
Significance of the Research
6
1.6
Chapter Organization
6
OBJECT DETECTION APPROACHES AND
APPLICATIONS
7
2.1
Introduction
7
2.2
Approaches to Object Detection and Recognition
7
2.3
Image Segmentation
11
2.3.1 Contour Segmentation
11
2.3.2 Region Segmentation
12
2.3.3 Object Oriented Approach
14
Reflectance and Object Detection
14
2.4
viii
3
2.5
Automatic Object Detection
15
2.6
Summary
17
METHODOLOGY
18
3.1
Introduction
18
3.2
Materials
20
3.2.1 Target Sets
20
3.2.2 The Data Capture
20
3.3
4
5
3.2.2.1 Digital Camera Nikon Coolpix 7900
21
3.2.2.2 Spectro-radiometer
21
3.2.2.3 Images of Target Sets
21
3.2.2.4 Reflectance Measurement
24
Data Pre-processing
25
3.3.1 Radiometer Correction
25
3.3.2 Geometric Rectification
27
3.3.3 Image Segmentation
28
3.3.4 Multiresolution Segmentation
29
3.3.4.1 Classification of Segmented Image
30
3.3.4.2 Classification Based Segmentation
33
3.3.4.3 Feature Space Optimization
34
RESULTS AND ANALYSIS
35
4.1
Introduction
35
4.2
Object Detection Based on Reflectance
36
4.3
Spectral Assessment
41
4.4
Segmentation Results
43
4.5
Assessment of Object Detection Based on Spatial Entities
46
4.6
Discussion
48
CONCLUSIONS AND RECOMMENDATIONS
50
5.1
Conclusions
50
5.2
Recommendations
51
ix
LIST OF TABLES
TABLE No.
3.1
TITLE
Descriptive statistics and the amount shift applied to the
PAGE
26
red and blue bands to normalize radiometric effects
3.2
The result of rectification process
27
3.3
Signatures of the class of interest of the target set
32
4.1
Separation of spectral reflectance curve at RGB bands of
37
the targets for the white board set
4.2
Separation of spectral reflectance curve at RGB bands of
39
the targets for the sand background set
4.3
Separation of spectral reflectance curve at RGB bands of
41
the targets for the camouflage set
4.4
Assessment for the spectral reflectance curves for the 3
42
target sets for the RGB bands
4.5
Overall assessment for the spectral reflectance curves for
42
the 3 target sets for the RGB bands
4.6
Accuracy assessment for the segmentation process
45
4.7
Results of feature space optimization
45
4.8
A matrix for feature space optimization
46
4.9
Results of spatial assessment
48
x
TABLE OF FIGURES
FIGURE No.
TITLE
PAGE
2.1
Block diagram of object detection
10
2.2
Region-based segmentation diagram
13
2.3
Automatic Target Detection Process
16
3.1
Flow chart for the methodology carried out in the study
19
3.2
The target set
22
3.3
Configuration of spectro-radiometer with respect with
25
respect target platform for calibration purpose
3.4
The histograms for the RGB bands of the calibration
26
board
3.5
The image segmentation process in the eCognition
29
system adopted in the study
3.6
Image classification
32
4.1
Reflectance curve for the white board target set
36
4.2
Reflectance curve for the sand target set
38
4.3
Reflectance curve for the camouflage target set
40
4.4
Results of the object detection using image segmentation
43
4.5
Assessment of spatial entities based on generated
47
template
xi
LIST OF APPENDIXES
APPENDIX No.
APPENDIX I
TITLE
Description of Generic Shape Features
PAGE
57
CHAPTER 1
INTRODUCTION
1.1
The Importance of Object Detection
Object detection is a process of identifying and locating objects in a scene. It
is a process of combining digital image processing and computer vision (Gonzalez et
al, 2001). The main goal of an object detection process is to detect and identify
objects in a scene.
Selected parameters, which are commonly used for object
detection such as the interpretation parameters aimed to recognizes and identify
objects, will be described in the following sections. The commonly used parameters
in object detection for remote sensing images are elaborated in the following section.
Image-based object detection has been widely used in the remote sensing
field.
Skilled interpreter and computer-assisted interpretation are two essential
factors in digital-recognition objects. In the manual-interpretation approach, seven
interpretation elements are usually used to interpret images; namely size, shape,
shadow, tone, color, texture and pattern. The size of an object is defined by the most
commonly parameters such as length, area, perimeter, area and volume. The shape
of an object is determined by its geometric characteristics such as linear, curvilinear,
circular, elliptical, radial, square, etc. The shadow is a silhouette caused by sideilluminated objects, often. The image tone of an object refers to the gray scale of the
image, typically ranging from bright to total dark. It is used in extracting the scene
of the height or relief of the object of interest. The color tone is also used in the
interpretation of color images. Texture is a group of small patterns often used in
extracting groups of objects. Pattern is usually used in the spatial arrangement of
2
objects on the ground (systematic, unsystematic, etc.). Apart from the above seven
interpretation elements, other elements are used in manual interpretation such as the
associated relationships as the combination of the elements, geographic
characteristics and configuration of the objects.
A well-trained interpreter can
interpret the image using the above elements easily as a dichotomous key (Jensen
2000).
The use of digital interpretation approach is widely used in various
applications.
In this approach, the computer systems (including appropriate
software) are used to extract object features using basic interpretation elements. The
most common elements used for digital recognition are tone, color and texture. Tone
and color of individuals in the scene are the two dominant interpretation elements
used in the computer-based object detection based on statistical pattern recognition
techniques (Hardin and Thomson, 1992). At present, both the manual and computerbased interpretation are used hand-in-hand, where each of these approaches
compliment each other as fully computer-based interpretation is rather limited due to
each of the dedicated systems being built for specific purposes.
The image-based object detection has formed wide applications, which
includes military and biomedical imaging. In military applications, recognizing and
detecting of targets within specific areas has been the main focus. Such targets are
crucial inputs into intelligence planning and monitoring.
In the biomedical
applications, object-detection has been carried out for assisting medical practitioners
in diagnosis of diseases/sickness (Yuan-Hui Yu and Chin-Chen Chang, 2006). These
include also recognition of fingerprints and eye pupil as biometric identification,
apart from the study of human perception to vision and recognition ability.
Image-based object detection and recognition is important, particularly for
intelligence where it is widely used in the planning of military operations. It is also
very important for assisting pilots, tanker operators, and shooters to detect and
identify targets of interest. Using this technique the operator could identify the target
of interest faster and more precisely. Computer-assisted object detection could be
used for automatic detection of any object or target of interest. This is referred to as
3
automatic target recognition (ATR), which generally refers to the self-directed or
aided target detection and recognition by computer processing of images from a
variety of sensors such as forward looking infrared (FLIR), synthetic aperture radar
(SAR), inverse synthetic aperture radar (ISAR), laser radar (LADAR), millimeter
wave (MMW) radar, multispectral/ hyperspectral sensors, low-light television
(LLTV), video, etc. ATR is an extremely important approach that is useful for
targeting and surveillance missions of defense weapon systems operating from a
variety of platforms (BHANU et al, 1997).
Automatic target recognition aims to reduce the workload of human operators
such as image analysts, pilots, and tankers, who are tasked with a large scope of
activities ranging from assessing the battlefield/battlespace situation over large areas
and volumes to targeting individual targets on land, sea, or air. It is also used to
reacquire targets when used on unmanned lethal weapon systems such as missiles.
The effectiveness and efficiency of locating and detecting targets is greatly increased
with this application (Clark et al, 1997).
In the context of medical imaging, object detection is widely used in
detecting symptoms peculiar to disease or related to malfunction or strange organs.
It uses the concept of detecting foreign bodies inside the human body. Yfantis et al.
(2000) proposed an algorithm for detecting cancer using ultrasound image. The
algorithm is based on calculating a number of attributes in the section of interest in
the image and finding the distance in the multivariate space of the vector of attributes
in the section from 2 predetermined centroids (cancerous and non-cancerous). If the
distance from the non-cancerous centroid is smaller than the distance from the
cancerous centroid the region is classified as non-cancerous, otherwise it is
cancerous. Gender classification could be also done with the help of object detection
using face detection and principal component analysis technique (Zehang et al,
2002).
In this study, artificial steel objects on a sand background will be detected
using the object-oriented approach and spectral reflectance characteristics.
The
image sets in this study are captured using digital optical camera in a controlled
4
laboratory environment where a specrtoradiometer device is used for acquiring
reflectance measurements from the object targets.
1.2
Problem Statement
Remotely sensed images are formed as the resultant of the interaction of
electromagnetic radiation (EMR) with the earth surface. The magnitudes of these
interactions are dependent on the wavelength of the EMR and the characteristics of
the earth's surface. Extraction of any feature from these images can therefore be
made based on the spectral and spatial entities of the EMR interactions. Two typical
algorithms established for object detection in remote sensing were based on both
these the spatial and spectral interactions. In this study, the region-based algorithm is
used for detecting and identifying spatial entities (shape and area), while the spectral
reflectance characteristics of the objects are used for recognizing the target of
interest.
1.3
Objectives
The objectives of this study are:
(1) To detect and extract selected targets from remotely sensed images
based on the spectral and spatial characteristics;
(2) To perform accuracy assessment of the objects extracted in objective
(1).
5
1.4
Scope of the Study
The scopes of the study are as follows:
(i) The images used are captured in the laboratory using a digital camera,
of a set of known miniature military targets with sand background.
The targets comprised two steel model air fighters (Plane1, Plane2)
and one steel model tank.
(ii) The images of the targets are captured within the visible spectrum of
EMR at different sets:
(a) The white board, captured as a control;
(b) Targets on the white board (capturing reflectance for
individual target);
(c) Sand reflectance;
(d) Camouflage reflectance;
(e) Targets on sand (capturing reflectance for individual target);
and
(f) Reflectance for the targets covered with camouflage.
(iii)The corresponding spectral responses of the targets with the
background were also observed using spectroradiometer in the range
of 400-700nm, equivalent to the wavelength of the camera used to
capture the digital image.
(iv) The spectral responses of the targets are reduced and analyzed for the
absolute responses within the given wavelength range for determining
the best differentiation among the targets.
(v) The detection and extraction of objects using spectral entities is
performed based on region growing and region merging algorithm
where it has to satisfy the condition (Hsiao et al, 2005).
6
1.5
Significance of the Research
Detection and extracting the spatial and spectral responses of different targets
are usually done through different approaches. This research helps users to identify
and recognize different kinds of targets (Plane1, Plane2, Tank), which are sited on
natural background using spectral and spatial approaches. It is useful for military
purposes and could be utilized in different applications.
1.6
Chapter Organization
This report consists of five chapters.
Chapter 1 is an introduction and
background of the project. Chapter 2 reviews some previous similar works for animage-based object detection/recognition. This includes segmentation approaches,
the importance of reflectance in object detection and automatic target detection.
The methodology of the study is described in Chapter 3, where the materials
and method used in this study is elaborated. The results of the study are presented,
analyzed and discussed in Chapter 4. Finally, Chapter 5 concludes the findings of
this study, apart from highlighting the future recommendations for improving this
study.
CHAPTER 2
OBJECT DETECTION APPROACHES AND APPLICATIONS
2.1
Introduction
This chapter reviews some previous similar works on object detection. The
review focuses on object-based approaches; namely contour-based, region-based,
boundary-based and feature extraction and morphology operations.
The image
segmentation approaches are also reviewed, which cover contour segmentation and
region segmentation. The relation between reflectance and object detection is also
discussed in this chapter.
2.2
Approaches to Object Detection and Recognition
Image-based object detection and recognition have been widely used for the
identification of artificial targets in the natural environment. There are at least three
main approaches for such detection, which are (i) contour-based (Caron et al, 2002),
(ii) region-based approach (Ternovskiy et al., 2003; Rizon et al., 2006), (iii)
boundary-based approach (Chen and Gao, 2005; Lau and Ozawa, 2004; Zouagui et
al., 2003). There are some derived approaches such as the one that depends on
feature extraction and morphological operation (Wang et al., 2005) and others
depending on the statistical repartition (Vincent et al., 2000), which are used for
different applications.
8
Contour-based object detection can be achieved using two common
approaches: (i) high pass filtering, and (ii) matching technique. The first approach
includes Canny filters and adaptive filtering, while the matching approach is based
on matching the object contour with the geometrical models. Both these approaches
work only when the shape of the object to be detected is known. It needs a contour
template, which defines the object of interest (Caron et al., 2002).
The main
advantages of using the contour-based approach are: (a) it preserves most of the
shape information about the object, which is necessary for identifying and
recognizing objects; and (b) 3-D models can be made from 2-D models.
The
contour-based approach can be used to extract shapes and morphology for features
such as the shape of any desired object. Most of the contour detection algorithms
extract broken or incomplete boundaries due to poor imaging conditions.
The second approach to object detection is based on region-based approach,
which describes the regions within the image. It is used as shape-based detection.
There are three steps in detecting objects using the region-based approach, namely:
(i) Fitting method: uses a fixed number of rectangles to fit the shape of the
object which represents the shape as vector of numeric parameters
(ii) Eigenmodes of finite element model: the eigenmodes method uses the finite
element model to represent the shape. Shape similarity computed as the
modal deformation energy needed to align 2 objects (Sebastian and Kimia,
2001)
(iii)Statistical calculation for the grey level: can be used by calculating mean and
median of the grey level, the minimum and maximum gray level values and
the number of pixels with values above and below the mean.
The advantage of using region-based approach is the simplicity of identifying
the topology and the texture of the objects (Chen and Gao, 2005). An example of
region-based detection application is image classification using maximum likelihood
analysis.
Lastly, the boundary-based approach for object detection has been reported
by Lau and Ozawa (2004), Chen and Gao (2005) and Zouagui et al., (2003). This
9
approach assumes that the pixel values change at the object's boundaries where they
are then converted to those values to a closed region boundaries. The assumption
that the adjacent pixels in the same region have similar visual features is also a basic
primitive in the region-based approach. Boundary-based approach is the best for
determining edge information, and this can be achieved in the following procedures:
(i) by local filtering using Canny edge detector;
(ii) Using snake model where an active contour model is used to minimize the
energy; and
(iii) Using a balloon model that is used to solve and generalized the problems
encountered by the main method.
It is a tool used for deducing the region in order to use it for region-based approach.
The advantages of using boundary-based approach are: (i) the ease of getting
information about the interior characteristics of the objects, and (ii) it does not
require order boundary points to preserve the shape of objects.
Apart from the three above-mentioned approaches, object detection can also
be carried out using feature extraction and morphological operations (Wang et al.,
2005). The main aim of feature extraction is to discriminate different regions by
combining features within the image.
Feature extraction is carried out in the
sequence of procedures, such as: (i) initial segmentation, (ii) morphology processing;
and (iii) object detection.
The initial segmentation stage is essential to segment the image to regions
before extracting features within images. This is performed using texture properties
of the image. The image texture is described using either statistical or texture
measurements based on the co-occurrence matrix. The result of this stage is a binary
image where it is be used in the morphology processing stage (César et al., 2007)
The morphology processing starts with the binary image that is obtained from
the first stage. It consists of the following processes: (1) object selection; (2) label
extraction; and (3) boundary decision. The goal of object selection is to remove the
noise from the image using dilation and erosion operations. Label extraction is
meant to identify the presence of homogenous regions. As a result, labeled image
10
can be obtained which can identify the objects numbers of the original image. The
purpose of boundary decision is to find the contour for every object, allowing for
easily locating the object in the original image. As such, object detection operation
can be performed. An example of this approach is detecting artificial objects in a
natural background (Wang et al., 2005). Figure 2.1 below illustrates object-detection
using the morphological approach.
Original image
Initial segmentation
Feature extraction
Fuzzy c-mean clustering
Mathematical morphology processing
Selecting and labeling objects
Finding location of objects in labeled image
Locating the objects areas in original image
Object detection
Object detection
Source: Wang et al, 2005
Figure 2.1: Block diagram of object detection
11
2.3
Image Segmentation
Image segmentation is defined as the process of assigning pixels to regions
having common properties. In other words, subdividing an image into constituent
regions in order to locate and extract outlines of certain objects in the image. It is the
first step in the object detection/recognition process. Meanwhile, the level to which
the segmentation takes place depends on the application to be used for and the
interesting objects to be detected (Zouagui et al., 2003; Lau and Ozawa, 2004).
Image segmentation algorithms are generally based on two basic properties of
the intensity values: (i) discontinuity and (ii) similarity. Using the first property, the
approach is to subdivide an image based on rapid changes in the intensity values
such as the edges of the image. The similarity property of the intensity is based on
partitioning an image into regions that are similar according to the predefined
criteria. Examples of the second category are region growing, region splitting and
region splitting and merging. Image segmentation can be achieved using: (a) contour
segmentation; (b) region segmentation; and (c) object-oriented approach (Gonzalez
et al., 2002).
2.3.1
Contour Segmentation
The contour-based segmentation approach is called the valley-following
method or the snake method. It starts with edge detection followed by a linking
process that seeks to make use of curvilinear continuity (Malik et al., 2001;
Zagorchev et al., 2007). The information that could be taken from the boundaries of
region segmentation, could define contours. Contours can be discrete or continuous.
Discrete contours are defined by a sequence of points whereas continuous contours
are defined by a parametric curve such as B-Splines and non-uniform rational BSplines. In the discrete representation of the contour segments, the adjacent contour
points are connected to create a continuous boundary. As a result, the contour
defined by a parametric curve represents a continuous boundary. This segmentation
approach is widely used for medical analysis. Fear et al. (2000) use this approach for
12
cancer detection and cancer contour creation. Liang et al. (2006) also use this
approach to analyze and interpret X-rays and MRI images.
2.3.2
Region Segmentation
Region segmentation is commonly used to subdivide an image into its
constituent regions or objects. The subdivision of the image into appropriate level
depends on the problem or the dimension of the object to be detected.
The
subdivision stops once the objects of the interest have been isolated. It consists of
processes in a sequence, such as: (i) region growing; (ii) region merging; and (iii)
region splitting.
Region growing is a procedure that groups pixels or sub regions into larger
regions based on the predefined criteria. It has advantages such as the boundaries of
the growing regions are perfectly thin and connected and the algorithm used is very
stable (Gonzalez et al., 2002).
Region merging segmentation simply merges regions that have similar
characteristics. In other words, if there are two adjacent regions having similar
properties, these two regions are merged. However, region splitting is a process that
aims to divide regions that are heterogeneous. Further explanation on the other
region segmentation is given in Gonzalez et al. (2002). Figure 2.2 illustrates region
segmentation methods. Figure 2.2 (a) represents region growing method, while (b)
illustrates region splitting method. Figures (c) and (d) show one of the splitting
techniques. It has a representation called quadtree (which means a tree in which
nodes have exactly four descendants).
13
(a) Region merging method
(b) Region splitting method
R
R1
R2
R3
R41
(c) Quadtree
R43
R44
source Gonzalez et al 2002
R1
R3
(d) Partitioning image
R42
R4
R2
R41 R42
R43 R44
source Gonzalez et al 2002
Figure 2.2: (a) Region merging method, (b) Region splitting method , (c) Qadtree;
and (d) Partitioning image
14
2.3.3
Object Oriented Approach
The object-oriented approach depends mainly on image segments, and not on
single pixels. The expected results of the object-oriented segmentation are real world
objects with the proper shape and classification, which cannot be achieved by a
pixel-based approach. This approach has the advantages of extracting additional
information that can be derived based on image object, apart from giving meaningful
statistics and textures. In the object-oriented approach, it is possible to take context
and semantic information into account. The input and the output could be vectorbased. The final output could be carried out to analyze aerial images and to update to
update GIS information (Benz et al, 2004). Quartel et al. (2006) use this approach to
extract beach morphology using video images.
Xu et al. (2004) also use this
approach but for land cover mapping using Quick Bird images using eCognition
software.
2.4
Reflectance and object detection
Reflectance is an important property of an object by which one could
differentiate object type and materials. There are some approaches that have been
used to find relations between reflectance of objects and recognition. Different
objects having different subjects result in different reflectance magnitudes.
Accordingly, the characterization of earth surface features based on their reflectance
is a central concept in remote sensing. Two approaches described below concerning
the importance of reflectance in object detection are: (i) reflectance ratio
computation, (ii) the shape and reflectance properties.
Any object appearance depends on its shape, the optical properties of its
surface (reflectance), the surrounding distribution of the light, the viewing position,
roughness and material. A method for computing the ratio of the reflectance of a
region to that of its background is one of the approaches to use reflectance for object
detection within an image. This is in turn resulted in a physical region with each
scene that is invariant to the shape of the object as well as the intensity and direction
15
of illumination.
From this photometric invariant (reflectance ratio), valuable
information could be used for recognition process. To do such a process the image is
first segmented followed by computing the reflectance ratio for each individual
segment with respect to their background.
Shape and reflectance are two important properties for an object. Russell et
al. (2007) use such properties to detect objects. The shape itself is not enough to
recognize any objects as a result reflectance play important role in detecting and
recognizing an object. Russell et al. (2007) use two different experiments in order to
detect and recognize object, one using gray scale image and the other one using color
image. From the two experiments used, good differentiation between the faces of the
male and the female has been achieved.
2.5
Automatic Object Detection
Automatic object detection (ATR) is the process of detecting and identifying
objects in a scene. It has been widely used in the extracting appropriate information
from a remote sensing data. There are many application of ATR such as image
analysis and interpretation. It is used to assist operators in image analysis to reduce
the workload. Advancement in computing contributed significantly toward ATR
usage. Two main approaches commonly used in ATR are (i) image-based and (ii)
model-based approaches (Bhanu et al, 1997). The image-based approach requires
combining different views using small reference images. The model-based approach
depends on boundary segmentation and planar surface extraction to describe the
scene. Target detection is done using trained neural networks or generic algorithms
(Vasile and Marino, 2005).
Automatic target detection could also be achieved by the matching technique
(Huttenlocher et al., 1997). This can be accomplished by matching target models
with images as sets of oriented edge pixels. The matching method, however, is very
tedious particularly for producing a meaningful match of high significance. It has a
disadvantage of artifacts appearing if the image used a considerable inherent clutters.
16
In order to reduce such problem, the orientation and the location of pixels have to be
carefully considered.
Typical data-processing for any ATR approach uses the following sequence
of steps: (i) detection of objects, (ii) discrimination of objects, (iii) classification of
objects, (iv) recognition of objects; and (v) identification of objects. In the detection
stage, the target is detected based on the presence of its signature captured in the
image. While in the discrimination stage any artifacts or noise are removed. In the
third stage, the targets are classified according to their signatures and how they are
differentiated from each other and to what class they are classified. This stage need
predefinition of the targets by image analyst. The fourth stage is recognition, which
is meant to recognize targets by their class. The final stage is the identification stage,
which is the ability to precisely define the targets (Paget et al., 2001). All the
procedures for all ATR are shown Figure 2.3 below.
Objects Detection
Objects Discrimination
Objects Classification
Objects Recognition
Objects Identification
Source : Paget et al, (2001).
Figure 2.3: Automatic Target Detection stages
17
2.6
Summary
The approaches to image-based object detection and applications were
reviewed. In this chapter, the object detection/recognition process has two main
approaches, which are (i) contour-based, (ii) region-based. There are also hybrid
approaches
such
as
boundary-based
approaches,
feature
extraction
and
morphological operations, which are derived from the above two main ones. The
contour-based approach employs high pass filters and matching technique, whereas
the region-based detection is achieved by the fitting method, eignmode of finite
element model and statistical calculation of gray level.
Object detection/recognition cannot be done without image segmentation,
which is one of the crucial procedures in image analysis and interpretation.
Accordingly, there are diverse segmentation approaches that can be used for object
detection ranging from contour-based segmentation and region-based segmentation.
Object detection, on the hand, can also be based on characteristics of the
object of interest. Detection of objects can be done by computing reflectance ratio
and by obtaining shape and reflectance properties of the object to be detected.
Automatic target detection (ATR) was also reviewed in this chapter.
CHAPTER 3
METHODOLOGY
3.1
Introduction
This chapter highlights the methodology carried out for the study. There are
two main parts of the methodology, namely: (1) materials used; and (2) the method
employed in the study. Detailed description of all materials used is given in the
following sections along with a step-by-step explanation on the methods used in the
object detection in this study. The flowchart of the methodology employed for the
entire study is shown in
Figure 3.1.
19
Image capture
Spatial-based
•
Digital camera (RGB)
•
Lab-based spectroradiometer
object detection
object detection
Image processing
• Geometric correction
•
Spectral-based
Spectral reflectance
Radiometric correction
Reduction of
spectral observation
Image segmentation
• Region-growing approach
•
Object oriented
Analyze of SRC's
for target detection
Template
Evaluation
• shape
Evaluation
Results on spatialbased object detection
Results on spectral -
End
based object detection
End
Figure 3.1: Flowchart of the methodology carried out in the study
20
3.2
Materials
There are four types of materials used in this study, which are: (a) Plane 1, (b)
Plane 2, (c) Tanker; and (d) Sand used as natural background.
3.2.1
Target Set
Specially-made miniaturized target sets are configured for this study. Figure
3.2 (a) illustrates the target set with selected targets, three steel miniaturized models,
namely Planes (Palne1, Plane2) and tank are placed on a rectangular board as
background. The dimension of the board are measured and demarcated, to identify
the entire test points and the position of the object/targets. Plane 1 represents F-15
fighter with the dimension 36cm main length and wings of 12cm length, whereas
Plane2 represents F-14 fighter with the length of 15cm wings of 6cm each side. The
tanker is Sherman Tank M4A3 model with a length of 17cm and width of 8cm.
3.2.2
The Data Capture
Two types of data capture are carried out in this study. Firstly, to capture
digital image using digital camera positioned on a goniometer to produce an image of
the target set. This image is used to examine the extraction of the known targets
based on optical-based object detection algorithm.
Configuration of the digital
image capture is illustrated in Figure 3.2 (a), where the digital camera is used instead
by spectro-radiometer.
The second data capture is carried out using spectro-radiometer of the target
set on the wavelength range of 400 -700 nm for the RGB of the visible wavelength.
The data being captured for the reflectance is stored as txt file and are later analyzed
using statistical analysis software. The technical descriptions of the digital camera
and the specroradiometer are given in the following sections.
21
3.2.2.1 Digital Camera Nikon Coolpix 7900
The digital camera used in the study is Nikon model Coolpix 7900. It has
data capture capability of 7 megapixel (7MP resolution) with 3x optical Zoom
offering focal length of 38-114mm. Three types of image sizes can be acquired,
which include:
(i) 7 million pixels at 3072 x 2304 scene,
(ii) 5 million pixels at 2592 x 1944scene; and
(iii) 3 million pixels at 2048 x 1536 scene.
3.2.2.2 Spectro-radiometer
The spectro-radiometer is used for capturing the reflectance of the individual
objects of the target sets with various background settings. The spectro-radiometer
employed is ASD FieldSpec Pro JR, which is manufactured by Analytical Spectral
Devices Inc. It has the ability to acquire the data for full spectrum range of 3502500nm with rapid data collected at 1/10th of the second per spectrum. It is a full
workstation with field laptop and Gun to measure reflectance. This instrument is
designed to collect solar reflectance, radiance and irradiance measurements. It can
be used for different applications such as remote sensing, geology, ecology, forestry,
plant physiology, and many others (ASD manual).
3.2.2.3 Images of Target Set
Figure 3.2 (a) represents the configuration of the specto-radiometer in the
laboratory set up to measure the reflectance of the target. The goniometer is used for
placement of spectro-radiometer with target to appropriate illumination geometry for
the light source and objects of interest. Figure 3.2 (b) shows the white board used
for calibration, fully grided for assisting geometric rectification. It is divided into 5
divisions in x-axes and 10 divisions in y-axes, respectively. Each division represents
22
10cm intervals. Figure 3.2 (c) shows camouflage used to cover the targets. On the
mean time, Figure 3.2 (d) demonstrates the setting of the object on the white board.
However, Figure 3.2 (e) shows the setting of the targets on the natural environment
(sand). The axes origin (x,y) is set to the left bottom corner the background board.
The orientation of the targets and the position of the control points used for
rectification are shown in Figure 3.2 (f).
(a)
L
R
G
x
O
y
R: Radiometer O: Object set L: source of light, G: goniometer
(b)
4
3
2
1
2
3
4
5
6
7
8
9
Figure 3.2: the target set: (a) Configuration of target set placed on goniometer,
(b) the white board background used for calibration…./continue
23
(c)
(d)
X
Y
(5,10)
(e)
Plane1
Plane
Tank
X
(0,0)
Y
Figure 3.2: the target set: (c) the camouflage on the top of targets,
(d) the targets sited on the white board background, (e) the targets sited on the sand
background…./continue
24
(f)
4
3
7
Plane1
5
Tank
Plane
1
6
2
X
Y
Figure 3.2: the target set: (f) the targets sited on the sand background with the
control points used for rectification.
3.2.2.4 Reflectance Measurement
The reflectance measurement of the target set is made using spectroradiometer in the controlled environment. The captured reflectance sets are: (i) the
white board which is used for calibration, (ii) targets (Plane1, Plane2,Tank) on the
white board, (iii) camouflage only, (iv) sand only, (v) targets (Plane1, Plane2,Tank)
on the sand; and (vi) targets (Plane1, Plane2,Tank) covered with camouflage. The
output of this process is spectral reflectance curves for each set mentioned and for
each target in different environment. This is done to establish spectral reflectance
curves in order to discriminate between the reflectance behaviors for different target
sets (refer to Figure 3.2 (c) - (f)).
Calibration for reflectance measurement is
performed by capturing the reflectance of objects of interest from different angles
and taking the average as the ideal one for calibration. Figure 3.3 is a diagram
representing calibration method been taken. Reflectance is captured from different
angles (α, α1, α2) with constant height of the sensor (radiometer) from the
background, i.e. 2 m. Then a graphical representation for the 3 reflectance readings
are averaged as the calibrated spectral reflectance.
25
α, α1, α2 are the angles for
Radiometer
observation set for the spectral
reflectance of background
α2
α
2m
α1
+
Demarcated grid for
geometric reference
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
Target background
Figure 3.3: configuration of spectro-radiometer with respect to target platform for
calibration purpose
3.3
Data Pre-processing
Image processing started with a data preprocessing stage where the image is
radiometrically corrected and geometrically rectified. The preprocessed data is then
subjected to segmentation approach adopted in the study is elaborated further in
Section 3.3.3.
3.3.1
Radiometric Correction
The radiometric correction is aimed to reduce radiometric errors or distortion.
These are due to: (i) sensor sensitivity (the digital camera), (ii) the source of
illumination, (iii) light fallout on object due to scattering, and (iv) absorption of
radiation inside the laboratory. The histogram of the three bands of the digital data is
shown in Figure 3.4 below, which shows the radiometric value of white background.
The shift in the peaks of the red, green and blue bands is actually the amount of the
"offset" values needed to radiometracally the data. A simple approach to normalize
26
these three bands for radiometric shift is carried out, such that the green band is set as
the reference while the red and the blue are shifted by the amount of differentiation
the center peak values with the green peak respectively. Table 3.1 below tabulates
the descriptive statistics and the amount of shift applied to the red and blue bands to
normalizing of radiometric effects. Once the "offset" for the red and blue bands is
applied, the resultant image of white platform is totally "white" with one
"superimposed" histogram for the RGB bands.
Figure 3.4: The histograms for the RGB bands of the calibration board
Table 3.1: descriptive statistics and the amount of shifted applied to the red and blue
bands to normalizing of radiometric effects
Band
Min
Max
Mean
Stdev
Amount of radiometric
shift
to
normalized
radiometric effects
1
0
248
105.9
24.525952 105.9 -99.9 = 6
20.601978 106.5 -99.9 = 6.6
2
0
255
106.5
3
0
255
99.9
22.359181
Zero
green)
shift
(reference
27
3.3.2
Geometric Rectification
The Geometric rectification aims to assign absolute coordinates of the
object to the digital image so accurate measurements of the image can be made. The
image to map approach is used. The control points used are acquired from the target
set for this study, which is the white board of (50cm x 100cm). This rectification
process has two steps: (i) transformation of image into the map coordinate system,
(ii) the resampling of the image brightness into the transformation coordinate system.
The 1st degree polynomial equation (Equation 3.1) is used for transformation
due to small size of the image used. The resampling method used the nearest
neighbor strategy. It does not need a lot of computation and it preserves the original
image from loosing original information.
Table 3.2 shows the result of the
rectification process. RMS for X is 0.14 pixels and RMS for Y is 0.01 pixels, with
overall RMSXY of =
(0.14) 2 + (0.01) 2 =0.14 pixels.
X = a1 + a 2 x + a3 y
Y = b1 + b2 x + b3 y
}
Equation 3.1
where,
X, Y= are the absolute coordinates obtained from the template measurements,
x,y = image coordinates of pixels, and
a1 ,a2 , a3, b1 ,b2 , b3 are coefficients of transformation.
Table 3.2: The result of rectification process
ID
X
Y
Xuncorrected
Yuncorrected
ΔX
ΔY
1
2
1
441.97
1000.97
0.01
0
2
7
1
1983.98
1105.98
-0.02
0
3
9
4
194.06
2583.56
-0.02
0
4
1
4
116.50
25.50
0.02
0
28
5
6
3
1663.97
449.10
0.10
-.01
6
4
1
1059.03
1043.03
0.01
0
7
4
3
1047.02
407.02
-0.10
.01
RMSx
0.14
3.3.3
RMSy
0.01
RMSXY
0.14
Image Segmentation
Image segmentation is the process of partitioning an image into segments or
regions in order to analyze and interpret it. In this study, it is aimed to detect the
shape of selected targets. The image segmentation of the eCognition system is used
for this purpose and such process is achieved through the following procedures:
(a) importing of the image;
(b) multiresolution segmentation;
(c) classification of the segmented image;
(d) classification-based segmentation; and
(e) feature space optimization.
Figure 3.5 represents the flowchart segmentation process used in the study.
29
Load/import image/s and
define working units
Multiresolution Segmentation
Classification process
Classification Based Segmentation
Feature Space Optimization
Figure 3.5: The image segmentation process in the eCognition system adopted in the
study
3.3.4
Multiresolution Segmentation
Multiresolution segmentation is a method of generating image objects. It
produces homogenous segments in any chosen image resolution. It is the bottom up
region growing-based approach, where small pixels are merged into bigger ones.
Multiresolution segmentation has the following parameters:
(i) Image weight: used to define the layer weight for the selected image layers to
define image importance, the higher the number the higher layer importance.
(ii) Level: to determine which level the segmentation will be saved, and
determines the hierarchy of the file.
30
(iii) Scale parameter: used to determine the heterogeneity of the output image.
Increasing the value creates larger image object. The scale parameter refers to
exclusively to the weights of the image layers (weight* scale parameter).
Chosen the best scale parameters depend on the image and objects used. In
this study, the scale parameter is 100.
(iv) Compactness and smoothness: used to define how shape homogeneity is
described. As the smoothness increases, compactness decreases. The total
scale is 1. In this project, Compactness is 0.2 and the smoothness is 0.8.
(v) Segmentation mode: there are 3 segmentation modes (a) normal, (b) sub-
object line analysis; (c) spectral difference segmentation. Normal mode is the
default mode where the scale parameters can be used to create sub-objects for
calculation of line features. Whereas the spectral difference mode is to join
neighboring objects according to color differences.
The parameters to be used rely on the image used and object to be segmented.
To get to the result, classification for the multiresolution segmentation must be done.
3.3.4.1 Classification of Segmented Image
The classification process is used to assign objects to certain classes
according to the class's description. The class description for the main classes in this
study (Plane1, Plane2, Tank and the background) uses the manual interpretation
parameters to interpret and identify the objects. eCognition software used two main
approaches: (i) knowledge-based, and (ii) fuzzy approach. The knowledge approach
depends on knowledge of the interpreters about the object to be classified, whereas
the fuzzy approach depends on the class description using fuzzy sets of object
features defined by membership function.
To achieve good classification, the
approach to be used should combine both approaches.
Classification can be performed either using supervised or unsupervised
approaches. For the supervised approach, training sets for each class are generated
before the classification takes place. While the unsupervised classification starts
31
with grouping of spectral classes using clustering method, labeling of spectral classes
formed are needed for the final output. In the eCognition, system the classification
uses the following sequence of steps:
(a) Open class hierarchy: in order to start supervised classification and collecting
training samples.
(b) Insert class: for every feature has its own class to distinguish between
different classes.
(c) Select samples: In the case of target set (plane 1) the camouflage color of
dark green is one class whereas the grey color of targets is another class, the
rockets have different class etc.
(d) Start classification using nearest neighbor which is used by the software.
(e) Apply standard nearest neighbor to all classes
(f) Classify: where the system take care to apply the classification according to
training sets.
Figure 3.6 (a) below shows the selection of training sets for classification.
Samples as demarcated by the polygon for seven classes the signature of these
classes is given in Table 3.3. The classes are marked according to the signature of
the segmented item as it appears in Figure 3.6 (b) (class 2 (a), (b), etc.), where the
results of classification are represented. As it appears, editing for the resulted classes
is necessary to get the object of interest with the appointed class.
32
1 Background
2 Tank
3 Parts from Plane1
4 Plane2
5 Parts from Plane1
6 Parts from Plane1
7 Parts from the background
(a)
(b)
Figure 3.6: (a) selected training sample, (b) the resultant classification
Table 3.3: signatures of the class of interest of the target set
Class
Mean
Std Dev
R
G
B
R
G
B
1(background)
112.20
109.85
102.42
14.19
12.93
13.36
2 (a)
35.04
54.39
51.92
15.75
14.58
16.24
2 (b)
57.38
60.52
55.65
16.58
16.21
15.98
2 (c)
60.02
68.81
57.99
19.23
21.14
22.44
2 (d)
35.57
54.63
58.30
15.58
11.35
10.83
3 (a)
105.27
128.42
120.62
21.78
22.63
27.90
3 (b)
117.02
143.33
138.44
20.26
20.84
26.38
3(c)
108.60
129.81
117.18
18.71
20.30
21.83
3 (d)
134.28
142.73
150
16.53
24.58
29.45
33
4 (a)
105.26
128.42
120.62
21.78
22.63
27.90
5 (a)
62.16
85.15
76.08
28.27
29.53
36.64
5 (b)
123.08
147.08
154.50
21.55
25.02
33.55
5 (c)
90.55
115.97
128.97
46.97
51.07
52.26
6 (a)
84.25
102.63
102.55
32.02
35.54
46.47
7 (a)
53.20
57.14
55.27
20.27
20.04
21.47
7 (b)
93.97
93.39
87.97
13.21
13.09
12.95
7 (c)
52.81
55.34
58.42
13.21
12.97
12.69
7 (d)
142.96
140.06
134.25
10.64
10.87
10.70
7 (e)
153.28
171.63
187.13
21.69
25.30
28.50
7 (f)
53.20
57.14
55.48
20.27
20.04
21.47
7 (g)
92.29
90.90
86.10
11.61
11.20
10.35
In order to perform the next process of segmentation, a structure group has to
be defined. It is aimed to merge classes together to perform classification-based
segmentation. This stage has two main steps: (i) definition of structure group (ii)
object generation. The definition of structure groups aims to create image objects
based on the classification. In this project, all classes of the same objects are joined
in this stage. In the case of Plane 1, a single class formed by combining other classes
that identify Plane 1.
3.3.4.2 Classification Based Segmentation
Classification based segmentation is another approach for object oriented
segmentation which further segment the classified image.
It differs from
multiresolution segmentation in that the earlier method depends on the homogeneity
whereas the latter segmentation uses knowledge base to generate object. This is
done by editing structure groups, which was discussed earlier.
34
3.3.4.3 Feature Space Optimization
Feature space optimization is a process to find a well-suited feature
combination to separate classes in conjunction with a nearest neighbor classifier. It
takes selected classes to be separated and results in a best separation with certain
features. In order to do feature space optimization the criteria have to be initially
selected to separate the objects. The features must be predefined before the generic
shape features of the object of interest are identified. The following steps show
feature space optimization processes in the eCognition:
(1) Select the classes to calculate the optimal feature space: the objects of interest
(2) Select the initial feature space to be reduced, the features to be optimized
with (generic shape features)
(3) Select the image object level: to what level shall be used (level2).
(4) Enter the maximum dimension according to the features selected: how many
features to be optimized e.g. 11 features.
(5) Start the optimization
(6) The result of this process shown in the optimized feature space which defines
the best separation distance (the distance between closest samples of the
chosen class) and the dimension used: what is best feature for optimizing
between objects of interest
(7) Separation reports of the optimized targets.
In this study, the generic shape features used are based on Baat and Benz
(2001) as the followings: (a) area, (b) length/width ratio, (c) length, (d) width, (e)
border length, (f) shape index, (g) density, (h) asymmetry,(i) compactness, (j) elliptic
fit, and (k) rectangular fit. Full description about those parameters is given in the
Appendix 1.
CHAPTER 4
RESULTS AND ANALYSIS
4.1
Introduction
Object detection from the captured images was performed using the spectral
and spatial approaches. There are three main image sets used for reflectance capture,
which include the following:
(a) white board set consists of: (i) white board (calibration), (ii) Plane 1 on the
white board, (iii) Plane 2 on the white board, (iv) Tank on the white board
(b) Sand set, consists of: (i) Sand only, (ii) Plane 1 on the sand, (iii) Plane 2 on
the sand, (iv) Tank on the sand
(c) Camouflage set consists of: (i) Camouflage, (ii) Plane 1 covered with
camouflage, (iii) Plane 2 covered with camouflage, (iv) Tank covered with
camouflage.
Image segmentation is carried out adopting region-based and object-oriented
approaches. The image set used was captured with the sand background. For other
image sets such as the camouflage, it is difficult to detect objects because of them
being covered. Hence, this image set is not worth preserving for further analysis.
The results of image-based object detection using reflectance and segmentation
processes are presented in this Chapter.
36
4.2
Object Detection Based on Reflectance
The following graphs show the results of the reflectance capture for the
different target sets. The first set includes the results of: (a) calibration of the white
board; and (b) Comparison of spectral reflectance curve for the objects on the white
board background is shown in Figure 4.1.
Calibration
G
B
1.4
R
Reflectance
1.2
1
R1
R2
R3
0.8
0.6
0.4
0.2
0
400
450
500
550
600
650
700
750
800
Wavelength (nm)
(a) R1: the reflectance at α, R2: reflectance at α1, R3: reflectance at α 2
Comparison for the whiteboard targets
1.2
R
G
B
Reflectance
1
Calibration
0.8
Plane1w
0.6
Plane2w
0.4
Tankw
0.2
0
400
450
500
550
600
650
700
750
800
Wavelength (nm)
(b)
Figure 4.1: reflectance curve for the target sets: (a) reflectance of the white
background used for calibration of RGB bands; and (b) comparison of object inserted
on the white background.
37
The spectral reflectance represented for all the target sets are on the visible
band of the spectrum, equivalent to the image capture using Nikon Coolpix 7900
digital camera operating in red, green and blue bands. Blue band corresponds to
400-500nm, while Green band has the wavelength of 500-600nm and Red band
equivalent to 600-750nm wavelength. Figure 4.1 (a) shows the calibration where
there are three RGB curves, which correspond to the positions captured in order to
determine the amount of shift to normalize each band respectively.
Speckles
occurred in all the three RGB bands at wavelength of less than 450nm. Minimal and
acceptable spectral curves occur after the 450nm range, with smoother curve in
wavelength larger than 600nm.
Figure 4.1 (b) shows all the three targets placed on the white background are
well separated in the wavelength larger than 500nm with optimum separation in the
larger wavelengths. This separation is minimal and non-uniform in the wavelengths
less than 500 nm. These fluctuations are mainly due to the existence of noises as
mentioned above. Separation between Plane 1 and Plane 2 is 10% and separation of
these two plane targets to the tank is more than 20% due to their differing materials,
as shown in Table 4.1.
Table 4.1: separation of spectral reflectance curve at (a) Blue band, (b) Green band,
(c) Red band for the white board target set
λ = 500
Plane1
Plane2
Tank
λ = 600
Plane1
Plane2
Tank
Plane1
φ
10%
20%
Plane1
φ
10%
20%
Plane2
10%
φ
30%
Plane2
10%
φ
30%
Tank
20%
30%
φ
Tank
20%
30%
φ
(a)
(b)
λ = 700
Plane1
Plane2
Tank
Plane1
φ
10%
20%
Plane2
10%
φ
30%
Tank
20%
30%
φ
(c)
38
B
0.6
G
Sand
R
Reflectance
0.5
0.4
0.3
0.2
0.1
0
400
450
500
550
600
650
700
750
800
Wavelength (nm)
(a)
Comparison between the reflectance curves for
targets
B
G
R
Reflectance
0.8
Sand
0.6
Plane1S
0.4
Plane2s
TankS
0.2
0
400
450
500
550
600
650
700
750
800
Wavelength (nm)
(e)
Figure 4.2: (a) Spectral reflectance Sand, (b) Comparison of spectral reflectance
curve for the objects on the sand.
Figure 4.2 (a) shows the reflectance response of the sand background with
RGB bands. Speckles occurred mostly at wavelengths less than 500nm, the blue
range. The reflectance smoothly increases with the green and red band ranges
starting from the wavelength 500nm onwards.
Figure 4.2 (b) shows all the three targets placed on the sand background are
well separated in wavelengths larger than 500nm with optimum separation in the
larger wavelengths. Fluctuations also appear in this target set, which is mainly
because of noise existence.
This separation is minimal and non-uniform in
wavelengths less than 500 nm. Separation between Plane 1 and Tank is 30%, while
39
Plane 2 has the same reflectance curve as the sand background. Table 4.2 shows that
Plane 2 and the tank have a separation of 40% at the 500nm wavelength, 15% at
600nm wavelength and 0% at 700nm. It also shows that the separation between
Plane 2 and Plane1 is 10% at 500nm, 20% at 600nm and 30% at 700nm. Table 4.2
also shows that there is no separation between Plane 2 and the background, which is
due to the smaller size of Plane 2 and the effects of sand reflectance as the
background with respects to other materials.
Table 4.2: separation of spectral reflectance curve at (a) Blue band, (b) Green band,
(c) Red band for the sand background set
λ = 500
Plane1
Plane2
Tank
λ = 600
Plane1
Plane2
Tank
Plane1
φ
10%
30%
Plane1
φ
10%
20%
Plane2
10%
φ
40%
Plane2
10%
φ
15%
Tank
30%
40%
φ
Tank
20%
15%
φ
(a)
(b)
λ = 700
Plane1
Plane2
Tank
Plane1
φ
10%
30%
Plane2
30%
φ
0%
Tank
20%
0%
φ
(c)
40
Camouflage
G
B
0.5
R
Reflectance
0.4
0.3
0.2
0.1
0
400
450
500
550
600
650
700
750
800
w a velength (nm)
(a)
Comparison of the Camouflaged curves
0.5
B
G
R
Reflectance
0.4
Camouflage
0.3
Plane1C
0.2
Plane2C
TankC
0.1
0
400
450
500
550
600
650
700
750
800
Wavelength (nm)
(b)
Figure 4.3: (a) spectral reflectance curve for Camouflage material, (b) Comparison
of spectral reflectance curve of the objects covered with camouflage with camouflage
Figure 4.3 (a) shows the reflectance response of the camouflage material with
RGB bands. Speckles occurred mostly at wavelengths less than 500nm - the blue
range - while speckles also continued with other bands. These noises appear more
clearly in this target set due to camouflage material used.
Figure 4.3 (b) shows all the three targets covered by camouflage on the sand
background. They are well separated in wavelengths larger than 500nm with best
separation in the larger wavelengths. The fluctuation also appears in this target set,
which is mainly because of noise existence and the covering camouflage of material.
Separation between Plane 1 and Plane 2 is 10% and separation of these two planes
41
targets to the tank is more than 10%, as shown in Table 4.3. This closer separation
between the targets in this set is because of the effect of the camouflage on the
reflectance at RGB bands.
Table 4.3: separation of spectral reflectance curve at (a) Blue band, (b) Green band,
(c) Red band for the camouflage set
λ = 500
Plane1
Plane2
Tank
λ = 600
Plane1
Plane2
Tank
Plane1
φ
10%
10%
Plane1
φ
10%
10%
Plane2
10%
φ
10%
Plane2
10%
φ
10%
Tank
10%
10%
φ
Tank
10%
10%
φ
(a)
(b)
λ = 700
Plane1
Plane2
Tank
Plane1
φ
10%
10%
Plane2
10%
φ
10%
Tank
10%
10%
φ
(c)
4.3
Spectral Assessment
Apart from the segmentation analysis mentioned in the next section, the
absolute reflectance curves for the targets are also analyzed. Table 4.4 tabulates the
reflectance capture for the three targets used. Table 4.5 represents the assessment for
all RGB bands. Among those readings, Plane 1 covered with camouflage has the
lowest reflectance where it has a mean reflectance of 0.0276 and the highest
reflectance is for the calibration (white board with mean reflectance of 1.0047).
From the standard deviation results for the reflectance curves, it appears all the
curves are acceptable.
42
Table 4.4: Assessment for the spectral reflectance curves for the three target sets for
RGB bands
R
Target
Range
G
B
Mean
Std Dev
Range
Mean
Std Dev
Range
Mean
Std Dev
White board set
Plane1
0.35
0.1860
0.0629
0.07
0.2175
0.0109
0.03
0.210
0.0049
Plane2
0.41
0.3006
0.0770
0.06
0.3469
0.0131
.03
0.323
0.0064
Tank
0.29
0.0626
0.0468
0.07
0.0449
0.0116
0.03
0.035
0.0048
Plane1
0.35
0.1860
0.0629
0.07
0.2175
0.0109
0.03
0.210
0.0050
Plane2
0.35
.01254
0.0604
0.25
0.2610
0.0710
0.15
0.455
0.0409
Tank
0.64
0.4863
0.0722
0.03
0.4889
0.0086
0.04
0.484
0.0067
Sand set
Camouflage set
Plane1
0.14
0.0171
0.0270
0.04
0.0185
0.0104
0.05
0.027
0.0081
Plane2
0.35
0.1090
0.0540
0.09
.1418
0.0197
0.13
0.198
0.0297
Tank
0.19
0.0492
0.0430
0.07
.0508
0.0133
0.06
0.070
0.0114
Table 4.5: Overall assessment for the spectral reflectance curves for the three target
sets for the RGB bands
Object
Minimum
Maximum
Mean
Std. Deviation
Calibration
0.98
1.05
1.0047
0.00932
Plane1w
0.17
0.25
0.2133
0.00796
Plane2w
0.31
0.38
0.3342
0.01467
Tankw
0.01
0.08
0.0380
0.00915
Sand
0.12
0.55
0.4048
0.12356
Plane1s
0.17
0.25
0.2133
0.00796
Plane2s
0.14
0.55
0.4043
0.11680
Tanks
0.46
0.51
0.4870
0.00780
Camouflage
0.11
0.38
0.2361
0.06867
Plane1c
0.00
0.06
0.0276
0.01265
Plane2c
0.08
0.33
0.1990
0.06292
Tankc
0.00
0.11
0.0694
0.02165
43
4.4
Segmentation Results
There are four main outputs from the segmentation process, namely: (a)
multiresolution segmentation, (b) classification of segments, (c) segmentation of
defined segments; and (d) final classification of segmented image. Figure 4.4 shows
the image of the outputs.
The most component of the segmentation results is the final output of the
process, as shown in Figure 4.4 (d). The ultimate shape of the tested targets against
the background is shown in Figure 4.4 (d), which is the shape expected from the
segmentation process.
(a)
(b)
Figure 4.4: results of the object detection using image segmentation (a) the result of
multiresolution segmentation, (b) the resulted image of the classification
44
(c)
Legend
Plane1
Plane2
Tank
(d)
Background
Figure 4.4: results of the object detection using image segmentation (a) the result of
multiresolution segmentation, (c) the resulted image of the classification based
segmentation, (d) the image after classification (cont’d)
Table 4.6 shows the details of final classification of the targets. It is clearly
shown that all the three targets tested are segmented as one object, respectively. This
indicates effectiveness of the adopted process. In Table 4.6, objects refer to the
number of segments or regions detected within each of the tested targets. Plane 1,
Plan 2 and the Tanker are identified as unique object, which is expected. The
signature statistics of these three targets are also uniquely depicted by zero standard
deviation. In case of the background, variations of 118 segments are normal as sand
is used. In natural environment, these variations could be greater because of the
45
atmospheric effects.
The main signature features of the targets with the sand
(simulating the natural settings), the Tanker, the Plane 2 differs to sand by 10% and
30% respectively with variation of ± 20% from sand background. The Tanker is not
likely to be successfully detected and identified. Adding the worst case of 10%
obscuring by the atmospheric effects, the Tanker is confirmed as not being a good
candidate for this detection approach. Taking the same natural setting and possible
effects, Plane 2 is marginally detected, but with the weakest magnitude.
Table 4.6: Accuracy assessment for the segmentation process
Class
Objects Mean
StdDev Minimum
Maximum
Plane1
1
0
1
1
1
Background 118
0.6979 0.213
0.004926354 1
Plane2
1
0.2356 0
0.2356
0.2356
Tanker
1
0.7964 0
0.7964
0.7964
Feature space optimization is used to discriminate between different features/
objects.
The generic shape features have been used after classification-based
segmentation. The parameters have been mentioned earlier in Chapter 3. This
process is done using the sets presented in Table 4.7.
Table 4.7: Results of feature space optimization
Optimization Set
Detected and identified
All objects together (Plane1,
11 (area, length, width, elliptic fit, rectangular fit,
Plane2, Tank, Background)
border length, asymmetry, density, shape index,
length/width, compactness)
Plane1 and the background
3 (area, border length, width)
Plane2 and the background
10 (area, length, width, elliptic fit, rectangular fit,
border length, asymmetry, density, shape index,
length/width)
Tank and the background
1 (area)
Plane1 and plane2
3 (area, border length and width)
Plane 1 and the tank
3 (area, border length and width).
46
Plane2 and the tank
5 (area, rectangular fit, elliptic fit compactness and
asymmetry)
Plane1, plane2 and the tank
5 (area, rectangular fit, elliptic fit compactness,
asymmetry )
As it appears from the above table, different parameters are used for
discriminating between the shapes of the targets. These parameters are the best ones
for differentiation purposes. To discriminate between the three (Plane 1, Plane 2,
Tank) targets and the background, 11 features can be used to differentiate between
them. Among the 11 features, the area feature is the only feature, which is used in all
sets. On the other hand, feature space optimization using generic shape feature could
be represented by the matrix shown in Table 4.8.
Table 4.8: A matrix for Feature Space Optimization
Object
Plane1 Plane2 Tank Background
Plane1
0
3
3
3
Plane2
3
0
5
10
Tank
3
5
0
1
10
1
0
Background 3
4.5
Assessment of Object Detection Based on Spatial Entities
The spatial assessments in this study focus on the analysis of shape of the
tested targets (Plane1, Plane2, Tank).
This is carried out by comparing the
dimensions of the targets using a template (created in ground truth) with the
segmented shape derived. Figure 4.5 shows the template used and the parameters for
the assessment. Figure 4.5 (a) illustrates the template used to compare Plane1 and
Plane2, which are identified by the main axis, minor axis1 for the right wing and
minor axis2 for left wing. Tank is identified by major axis and a minor axis. From
the comparison shown in Table 4.9 below, it appears that almost all objects fit the
47
template. There is a minute difference for the axis of Plane 1, which is because of
the segmentation process and measurement analysis.
d
j
a
b
l
k
c
e
m
Plane1
f
g
h
i
n
Tanker
Plane 2
(a)
Shape determinants: Plane1 Major axis is dbe, minor axis are
ab, bc, Plane2 major axis is jgi, minor axis fg, gh, Tanker
major axis is kl, and minor axis is mn.
Major axis
Major axis
Major axis
(b)
Figure 4.5: Assessment of spatial entities based on generated template (a) the linear
graphs typical to objects tested, (b) template overlaid the original image.
48
Major axis
Major axis
Major axis
(c)
Figure 4.5: Assessment of spatial entities based on generated template (c) the
template overlaid on ultimate output of the segmentation approach (continued)
Table 4.9: Results of spatial assessments
Object
Plane1
Georectified
ΔMajor
Δminor
Δminor
axis
axis1
axis2
RMS
Major
Minor
Minor
Major
Minor
Minor
axis
axis1
axis2
axis
axis1
axis2
36
12 (ab)
12
36.6
11.7(a'b')
11.7(b'c')
0.6
0.3
0.3
1.161895
(bc)
(b'd'e')
6 (fg)
6 (gh)
15(j'g'i')
6 (f'g')
6 (g'h')
0
0
0
0
8 (mn)
-
16.9 (k'l')
7.8 (m'n')
0.1
0.2
(bde)
Plane2
Segmented image
15
(jgi)
Tank
17
0.223607
(kl)
4.6
Discussion
Spectral and spatial object detection approaches were used in this study. It
concentrated mainly on region-merging and object oriented based segmentations
(Gonzalez et al., 2002 ; Xu et al., 2004). As it appears in Section 4.4, the image is
segmented to certain regions then these regions are classified using the supervised
classification approach. Those segments are then merged to form objects using
49
classification-based segmentation, which uses the object-oriented approach to detect
the shape of the three targets. Accordingly, the spectral approach was done using
reflectance-based approach (Russell et al., 2007), where the objects could be detected
from the spectral reflectance responses at the RGB band of the spectrum. For the
three target sets - the white background, sand background and the camouflage sets the separations between the three targets are clear for all visible bands.
CHAPTER 5
CONCLUSIONS AND RECOMMENDATION
5.1
Conclusions
Following are the conclusions drawn from the study carried out:
(a) The selected targets labeled using the object-based detection and
identification using both spectral and spatial entities show good outputs. In
the case of spectral approach, the segmentation of the three selected targets is
very distinct in the visible range with optimum separation of 10% between
Plane 1 and Plane 2 and 20% between the two planes and the Tanker. Even
in the natural background such as sand, Plane 1 and the Tanker targets are
recognizable at reasonably good separation of 30% at all visible band tested,
while Plane 2 and other targets are separated by an average separation of
10%. However, the three targets are separated by 10% for all spectral RGB
bands.
(b) Spectral and spatial assessments have been carried out for the three targets
derived from the segmentation process using the object-based approach. In
the case of the targets sited on the white background, all targets have minimal
standard deviations of less than 0.007 for all the RGB band of the spectrum.
Even the natural background set of sand spectrally have good signature of
less than 0.05 standard deviations for all the visible bands. The camouflage
set shows good results for all targets from their standard deviation values of
less than 0.03.
The output objects have very good correlation with the
original target set. The RMS values for Plane 1 is 1.1, Plane 2 is zero and
51
Tanker is 0.2. This is an indication of proper application of the spatial
approach in this study.
5.2
Recommendations
The recommendation and the outlooks for this study consist of:
(i) Martial and environment.
This recommendation is to use material and environment that simulate the real
world. In the real world, the aircrafts are sited on airports where the background
is made from asphalt. The environment used for this project is applicable for the
tank case. It could be more beneficial for spectral capture using real material, i.e.
using the models from the same materials used in reality. This could be done by
contacting the Plane or the military equipment factories.
If the reflectance
measurements have to be done indoor, it is recommended to measure reflectance
in a proper way, such as measuring in a dark room where there is no interfaces
from other sources of light such as from the window or doors. The floor should
be of a similar color and material to the background used in the study, in order to
have good results without interference from other subjects.
If the objects
materials are from materials used in the real life, the reflectance should be better.
Furthermore, the settings of the experiment should closely mimic the real world
environment.
(ii) Satellite image
Satellite images will help significantly if used in this research. They will give
experience of using satellite image whether they are RADAR images or highresolution optical images in object detection.
(iii)Thermal Camera
Another recommendation is to use a thermal camera to measure the temperature
of the objects. Using the output of such a process could be combines with the
results of the object-oriented approach to detect the shape of the object. This will
incorporate the thermal characteristics of the objects into the analysis.
52
LIST OF REFERENCES
Abe, Y. and Hagiwara, M.,(1997), Hierarchical Object Recognition from a 2D
Image using a Genetic Algorithm, IEEE, International Conference on Computational
Cybernetics and Simulations, 3: (2549 – 2554)
Baatz, M., and Benz, U., (2001), Definciens Imaging, eCognition object oriented
image analysis manual
Benz, U. C., Hofmann, P., Willhauck, G., Lingenfelder, I., and Heynen, M., (2004),
Multi-resolution object-oriented fuzzy analysis of remote sensing data for GIS-ready
information, ISPRS Journal of Photogrammetry & Remote Sensing, 58: (239– 258)
Bhanu, B. and Peng J., ( 2000) Adaptive Integrated Image Segmentation and Object
Recognition, IEEE, Image Processing, 6 (1): 427-441
Bhanu, B., Dudgeon, D.E., Zelnio, E.G., Rosenfeld, A., Casasent, D., Reed, I.S.
(1997), Guest Editorial Introduction to Special Target Detection and Recognition,
IEEE, Image Processing, 6 (1): 1-6
Canny, J.F., (1986), A computational approach to edge detection, IEEE Transaction
Pattern Analysis, 8 (6): 679–698
Caron, Y., Vincent, N., Markis, P.,(2002), Artificial object detection in natural
environments, RFAI Team Publication, First Annual Meeting on Health, Science and
Technology, Ecole Doctorale SST, Tours (France), 1 : (1-10)
53
César, A.B., Castañón, Jane, S, F., Sandra, F., Arthur, G., Luciano, d.F.,
Costab,(2007), Biological shape characterization for automatic image recognition
anddiagnosis of protozoan parasites of the genus Eimeria, Science Direct, Pattern
Recognition,40 :( 1899 – 1910)
Chen, H., and Gao, Q., (2005), Efficient Image Region and Shape Detection by
Perceptual Contour Grouping, IEEE,International Conference on Mechatronics &
Automation, 2: (793- 798)
Clark, F.O. and Daniel, P. H., (1997), Automatic Target Recognition by Matching
Oriented Edge Pixels, IEEE Transactions on Image Processing, 6 (1):103
Cohen, L.D., (1991), On active contour models and balloons, Computer Vision,
Graphics, Image Process, Image Understanding 53 (2): 211–218
Conzalez, R. C., (2002), Digital Image Processing, Richard E. Woods, 2nd
edition, New Jersey
Cubron, C. A. and Peddle, D.R, (2006) A low cost filed & laboratory gonimeter
system for estimating hyperspectral bidirectional reflectance, Can. J Remote
Sensing , 32 (3): 244-253
Fear, E.C., Stuchly, M.A., (2000), Microwave detection of breast cancer, IEEE,
Transactionson Microwave Theory and Techniques, 48(11): 1854 - 1863
Gevers, T., and Smeulders, A.W. M., (1998), Image Indexing using Composite Color
and Shape Invariant Features, IEEE, 16th International Conference on Computer
Vision, PP: 576-581, Bombay, India.
Gool, L. V., Moons, T., Ungureanu, D., (1996), Geometric/Photometric Invariants
for Planar Intensity Patterns, 4th European Conference Computer Vision, 1: (642651)
54
Hardin, P.J., and Thomson, C.N., (1992), Fast Nearest Neighbor Classification
Method For Multispectral Imagery, Professional Geographer, 44(2):191-201
Hsiao, Y.T., Chuang, C.L., Jiann J.A., Chien, C.C., (2005), A contour based image
segmentation algorithm using morphological edge detection, IEEE International
Conference on Systems, Man and Cybernetics, 3: (2962 – 2967)
Huttenlocher, D. P., and Olson, C. F., (1997), Automatic Target Detection by
Matching Oriented Edge Pixels, IEEE Transaction, Image Processing, 6 (1) :1103113
Jensen, J.R, (1996) , Introductory Digital Image Processing, A Remote Sensing
Perspective, Prentice Hall, 2nd edition, New Jersey
Jensen, J.R, (2000), Remote sensing of the Environment, an Earth Resource
Perspective, Prentice Hall, New Jersey
Kass, M., Witkin A., Terzopoulos, D., (1988), Snake: active contour models, Int. J.
Computer Vision, 1 (4): 312–331
Lau, P. Y. and Ozawa, S. (2004), A region-based approach combining marker-
controlled active contour model and morphological operator for Image
Segmentation, Process 26th Annual International Conference of the IEEE EMBS San
Francisco, CA, USA, 1 (3) :1652-1655
Lazzerini, B., and Marcelloni, F., (2001), A Fuzzy Approach to 2-D Shape
Recognition, IEEE Transaction of fuzzy systems, 9 (1): 5-16
Liang, J., McInerney, T., Terzopoulos, D., (2006) , United Snakes , Medical Image
Analysis, 10 : (215–233)
Malik, J., Belongie, S., Leung, T., and Shi, J., (2001),Contour and Texture Analysis
for Image Segmentation, International Journal of Computer Vision, 43(1): 7–27,
Kluwer Academic Publishers
55
Nayar, S.K. and
Bolle, R.M. (1996), Reflectance based object recognition,
International Journal of Computer Vision, 17 (2) :219-240
Paget, R., Homer, J., and Crisp, D., (2001), (Automatic) target detection in synthetic
aperture radar imagery via terrain recognition, IEEE, International Conference on
Image Processing, 3: (54-57)
Quartel, S., Addink, E. A., (2006), Object-oriented extraction of beach morphology
from video images, International Journal of Applied Earth Observation and
Geoinformation, 8 : (256–269)
Ralesc, A. L., and Shananhan, J. G., (1999), Pereptual Organisation for inferring
Object Boundaries in an Image, Pattern Recognition, 32 : ( 1923-1933)
Rizon, M., Yazid,H., Saad, P., Md Shakaff, A. Y., and Saad , A. R., (2006), Object
Detection using Geometric Invariant Moment, American Journal of Applied Sciences
2 (6): 1876-1878
Russell, R., Biederman, I., Nederhouser, M., Sinha, P., (2007), The utility of surface
reflectance for the recognition of upright and inverted face, Vision Research, 47:
(157–165)
Sarkar, S. and Boyer, K.,(1993), Perceptual Organization in Computer Vision: A
review and a proposal for classificatory structure, IEEE Transaction on Systems,
Man and Cybernetics, 23 (2) : 382~398
Sebastian, T. B., and
Kimia, B. B., (2001), Curves vs skeletons in object
recognition, International Conference on Image Processing, Thessaloniki, Greece,
IEEE, 3 : ( 22-25)
Ternovskiy, I., Nakazawa, D., Campbell, S., and Suri, R., E., (2003), Biologically
Inspired Algorithms For Object Recognition, IEEE International Conference on
Integration of Knowledge Intensive Multi-Agent Systems, PP: 364-367
56
Vasile, A. N. and Marino, R. M.
(2005), Pose-Independent Automatic Target
Detection and Recognition Using 3D Laser Radar Imagery, Lincoln Laboratory
Journal, 15 (1): 61-78
Vincent, N., Makris, P., Brodier, J., (2000), Compressed Image Quality and Zipf’s
Law, 5th International Conference on Signal Processing , Beijing (China), 2 : (10771084
Wang, L., Li, X., and Fang K. , (2005),
Object Detection Based on Feature
Extraction and Morphological Operations, IEEE, International Conference on
Neural Networks and Brain, 2: ( 1001 – 1003)
Xu, W., Wu, B., Huang, J., Zhang, ,Y., Tian, Y., (2004), A Segmentation and
Classification Approach of Land Cover Mapping Using Quick Bird Image,
International Geoscience and Remote Sensing Symposium, IEEE, 5 : ( 3368-3370)
Yfantis, E. A., Lazarakis, T., Bebis, G., and Gallitano, G. M., (2000), An Algorithm
for Cancer Recognition and Ultrasound Images, International Journal on Artificial
Intelligence Tools, 9 (2) : 265-276
Yuan-Hui, Y., and Chin-Chen, C.,(2006), A new edge detection approach based on
image context analysis Image and Vision Computing, Image and Vision Computing,
24 (10) : 1090-1102
Zagorchev, L., Goshtasby A., and Satter, M., (2007), R-snakes, Image and Vision
Computing , 25 (6) : 945-959
Zehang, S., Bebis, G., Xiaojing, Y., Yuan, X., and Louis, S.J., (2002), Genetic
Feature Subset Selection for Gender Classification: A Comparison Study, 6th IEEE
Workshop on Applications of Computer Vision, PP : 165 – 170
Zouagui, T., Benoit-Cattin, H., and Odet, C..(2003), Image segmentation functional
model, Pattern Recognition, 37: (1785 – 1795)
57
APPENDIX I: DISCRIPTION OF GENERIC SHAPE FEATURES
(i) Area: In nongeoreferenced data the area of a single pixel is 1. Consequently,
the area of an image object is the number of pixels forming it. If the image
data is georeferenced, the area of an image object is the true area covered by
one pixel times the number of pixels forming the image object.
(ii) Length/width: There are two ways to compute the length/width ratio of an
image object.
(a) The ratio length/width is identical to the ratio of the eigenvalues of the
covariance matrix with the larger eigenvalue being the numerator of
the fraction.
γ =
eig1 ( S )
l
=
, eig1 ( S ) > eig 2 ( S )
w eig 2 ( S )
(2)
(b) The ratio length/width can also be approximated using the bounding
box. Feature value range: (0;1)
γ =
l a 2 + ((1 − f ).b) 2
=
w
A
(3)
(iii)Length: The length can also be computed using the length-to-width ratio
derived from a bounding box approximation. It is approximated as follows:
l=
A.γ
(4)
58
Another possibility which works better for curved image objects is to
calculate the length of an image object based on its sub-objects. Feature
value range: (0 ; depending on shape of image object)
(iv) Width: Also the width of an image object is approximated using the length-to
width ratio. In eCognition the width is approximated as follows:
w=
A
γ
(5)
Again, for curved image objects the use of sub-objects for the calculation is
the superior method. Feature value range: (0 ; depending on shape of image
object)
(v) Border length: The border length e of an image object is defined as the sum
of edges of the image object that are shared with other image objects or are
situated on the edge of the entire scene. In nongeoreferenced data the length
of a pixel edge is 1. Feature value range: (4 ; depending on shape of image
object)
(vi) Shape index: Mathematically the shape index is the border length e of the
image object divided by four times the square root of its area A. Use the
shape index s to describe the smoothness of the image object borders. The
59
more fractal an image object appears, the higher its shape index. Feature
value range: (1 ; depending on shape of image object)
s=
(vii)
e
(7)
4. A
Density: The density d can be expressed by the area covered by the
image object divided by its radius. eCognition uses the following
implementation, where n is the number of pixels forming the image object
and the radius is approximated using the covariance matrix:
d=
n
(8)
1 + var( X ) + Var (Y )
Use the density to describe the compactness of an image object. The ideal
compact form on a pixel raster is the square. The more the form of an image
object is like a square, the higher its density.
Feature value range: (0;
depending on shape of image object)
(viii)
Asymmetry: The lengthier an image object, the more asymmetric it is.
For an image object, an ellipse is approximated which can be expressed by
the ratio of the lengths of minor and major axes of this ellipse. The feature
value increases with the asymmetry. Feature value range: (0 ; 1)
K = 1−
n
m
(10)
(ix) Compactness: In eCognition the compactness c, used as a feature, is
calculated by the product of the length m and the width n of the
60
corresponding Object and divided by the number of its inner pixels a.
Feature value range: (0; inf )
c=
n.m
a
(11)
(x) Elliptic Fit: As a first step in the calculation of the elliptic fit is the creation of
an ellipse with the same area as the considered object. In the calculation of
the ellipse also the proportion of the length to the width of the Object is
regarded. After this step the area of the object outside the ellipse is compared
with the area inside the ellipse that is not filled out with the object. While 0
means no fit, 1 stands for a complete fitting object. Feature value range: (0;
1)
(xi) Rectangular fit: A first step in the calculation of the rectangular fit is the
creation of a rectangle with the same area as the considered object. In the
calculation of the rectangle also the proportion of the length to the width of
the object in regarded. After this step the area of the object outside the
rectangle is compared with the area inside the rectangle, which is not filled
out with the object. While 0 means no fit, 1 stands for a complete fitting
object. Feature value range: (0; 1)