WEB APPLICATION FOR CLASSIFICATION OF OAK SPECIES LEAVES DIPLOMA THESIS Author: Rareș -Dacian PETRIȘOR Supervisors : Assoc . Prof. Eng. Rodica… [625806]

FACULTATEA DE AUTOMATIC Ă ȘI CALCULATOARE
201 8

WEB APPLICATION FOR CLASSIFICATION
OF OAK SPECIES LEAVES

DIPLOMA THESIS

Author: Rareș -Dacian PETRIȘOR
Supervisors : Assoc . Prof. Eng. Rodica HOLONEC, PhD
Prof. Eng. Honoriu VĂLEAN, PhD

FACULTATEA DE AUTOMATIC Ă ȘI CALCULATOARE

DECAN
Prof.dr.ing. Liviu MICLEA Vizat,

DIRECTOR DEPARTAMENT AUTOMATICĂ
Prof.dr.ing. Honoriu VĂLEAN

Autor : Rareș -Dacian PETRIȘOR

Web Application for Classification of Oak Species Leaves

1. Enunțul temei: O scurtă d escriere a temei proiectului de diplomă

2. Conținutul proiectului: (enumerarea părților componente) Pagina de prezentare ,
Declarație privind autenticitatea proiectului, S inteza proiectului , Cuprins, Titlul
capitolului 1, Titlul capitolului 2,… Titlul capitol ului n, Bibliografie, Anexe.

3. Locul documentației: Universitatea Tehnică din Cluj -Napoca , Institutul Național
de Cercetare -Dezvoltare în Silvicultură “ Marin Drăcea ” Cluj -Napoca

4. Consultanți: prof. dr. ing. Ioan Tăut

5. Data emiterii temei:

6. Data predării:

Semnătura autorului

Semnătura c onducător ului științific

FACULTATEA DE AUTOMATIC Ă ȘI CALCULATOARE

Declarație pe proprie răspundere privind
autenticitatea proiectului de diplomă

Subsemnatul(a) Rareș -Dacian PETRIȘOR ,
legitimat(ă) cu CI seria CJ nr. 039575 , CNP [anonimizat] , autorul lucrării:
Web Application for Classification of Oak Species Leaves
elaborată în vederea susținerii examenului de finalizare a studiilor de licență la
Facultatea de Automatică și Calculatoare , specializarea Automatică și Informat ică
Aplicată (în limba engleză) , din cadrul Universității Tehnice din Cluj -Napoca, sesiunea
Iulie 2018 a anului universitar 2017 -201 8, declar pe proprie răspundere, că această
lucrare este rezultatul propriei activități intelectuale, pe baza cercetăr ilor mele și pe baza
informațiilor obținute din surse care au fost citate, în textul lucrării, și în bibliografie.
Declar, că această lucrare nu conține porțiuni plagiate, iar sursele bibliografice au
fost folosite cu respectarea legislației române și a co nvențiilor internaționale privind
drepturile de autor.
Declar, de asemenea, că această lucrare nu a mai fost prezentată în fața unei alte
comisii de examen de licență.
In cazul constatării ulterioare a unor declarații false, voi suporta sancțiunile
adminis trative, respectiv, anularea examenului de licență .

Data Rareș -Dacian PETRIȘOR

(semnătura)

FACULTATEA DE AUTOMATIC Ă ȘI CALCULATOARE

SINTEZA
proiectului de diplomă cu titlul:
Web Application for Classification of Oak Species Leaves

Autor: Rareș -Dacian PETRIȘOR
Conducător științific: conf. dr. ing. Rodica HOLONEC
prof. dr. ing. Honoriu VĂLEAN
1. Cerințele temei:

2. Soluții alese:

3. Rezultate obținute:

4. Testări și verificări:

5. Contribuții personale:

6. Surse de documentare:

Semnătura autorului

Semnătura conducătorului științific

1 Contents
1 INTRODUCTION ………………………….. ………………………….. ………………………….. ………………………… 2
1.1 GENERAL CONTEXT ………………………….. ………………………….. ………………………….. …………………….. 2
1.2 OBJECTIVES ………………………….. ………………………….. ………………………….. ………………………….. …. 3
1.3 SPECIFICATIONS ………………………….. ………………………….. ………………………….. …………………………. 3
2 BIBLIOGRAPHIC STUDY ………………………….. ………………………….. ………………………….. ……………… 5
3 DESIGN AND IMPLEMENT ATION ………………………….. ………………………….. ………………………….. .. 13
3.1 RASPBERRY PI IMAGE ACQUISITION SYSTEM ………………………….. ………………………….. ………………….. 14
3.1.1 Technology ………………………….. ………………………….. ………………………….. ……………………… 14
3.1.2 Hardware layout and setup ………………………….. ………………………….. ………………………….. . 15
3.1.3 Image acquisition and transmission script ………………………….. ………………………….. ………. 17
3.2 IMAGE PROCESSING AND CLASSIFICATION ………………………….. ………………………….. ……………………… 19
3.3 WEB APPLICATION ………………………….. ………………………….. ………………………….. ……………………. 20
4 CONCLUSION ………………………….. ………………………….. ………………………….. …………………………. 21
4.1 OBTAINED RESULTS ………………………….. ………………………….. ………………………….. …………………… 21
4.2 FURTHER DEVELOPMENT ………………………….. ………………………….. ………………………….. …………….. 21
5 REGULI DE FORMATARE ………………………….. ………………………….. ………………………….. …………… 22
5.1 FORMATAREA PAGINII ………………………….. ………………………….. ………………………….. ……………….. 22
5.2 TITLURI ȘI STILURI ………………………….. ………………………….. ………………………….. …………………….. 22
5.3 FIGURI , TABELE ȘI ECUAȚII ………………………….. ………………………….. ………………………….. …………… 23
5.3.1 Figuri ………………………….. ………………………….. ………………………….. ………………………….. …. 23
5.4 TABELE ………………………….. ………………………….. ………………………….. ………………………….. …….. 23
5.5 ECUAȚII ………………………….. ………………………….. ………………………….. ………………………….. ……. 23
5.6 REFERINȚE BIBLIOGRAFI CE ………………………….. ………………………….. ………………………….. …………… 24
6 REFERENCES ………………………….. ………………………….. ………………………….. ………………………….. . 25

Introduction
2 1 Introdu ction
1.1 General context
The main purpose of this diploma project is the development of an application for
the classification of leaf images. The leaves belong to species of Quercus genus.
Oak sp ecies in Romania have a significant ecological and economical importance,
covering approximatively 19% of the Forest Fund. These species are encountered in pure
or mixed stands, at altitudes of up to 1000 m. However, as any tree species, they are
vulnerabl e to diseases and pathogen attack. The pathogens can be biotic (e.g. bacteria,
fungi, insects) or abiotic (e.g. late frosts, severe draughts, strong winds).
Due to the morphological similarities that exist between leaves, it is important to
accurately dist inguish between species of the Quercus genus, in order to apply adequate
treatments to the affected trees. Currently, in Romania, this classification is done mainly
manually (in the field or in a laboratory) , both by Forest Districts and by the National
Institute for Research and Development in Forestry “Marin Drăcea ”. This project offers a
tool that increases the automation of this classification operation if it is done in the
laboratory .
The project offers a Raspberry Pi based image acquisition system, c lassifiers for oak
species leaves and a web application that serves as a user -friendly interface between the
two components mentioned before and also access o a relational database in order to
store the obtained results. The classifiers were developed to d istinguish between here
species of oak, that are most commonly encountered in Romania: Quercus robur (common
oak), Quercus petraea (sessile oak) and Quercus frainetto (Hungarian oak).
/* will be modified accordingly, when changes are made to chapter titles
This diploma projects contains the following chapters: Introduction , Bibliographic
Study , Analysis, Design, Implementation and Conclusion . In Introduction an overview
about the project is given . Its purpose, objectives and specifications are presented in
detail. The Bibliographic Study chapter contains an analysis of existing research in the
field of classifier development for leaf image classification and image processing
algorithm, together with theoretical aspects. Algorithm details, classifier types an d
obtained results are discussed. In Analysis, Design, Implementation a detailed
description about each component of the developed application is given. These details
include descriptions of the technologies used, algorithms, use cases, diagrams. The
Conclusion chapter provides information about the obtained results (classification
accuracy), a comparison with results obtained in extant literature and further
development directions for this application.
*/

Introduction
3 1.2 Objectives
This project’s objectives are listed b elow:
• Development of an image acquisition system for laboratory use. This allows the user to
take leaf images and send then to the web application for analysis;
• Development of classifiers for leaf images. This includes three sub -objectives:
o Creating an ima ge database from leaf samples in order to provide training and
validation data sets for the classifiers. This involves collecting leaf samples from
the field and the construction of a laboratory stand for standardized image
acquisition;
o Development of an i mage processing algorithm for the extraction of leaf
parameters provided to the classifiers;
o Classifier development itself, by using data obtained from the previous sub –
objective.
• Development of a web application that allows image acquisition using the dev eloped
system, classification and storage of results.

1.3 Specifications
For each objective presented before, a set of specifications was given. More details
are presented below.
The image acquisition system has to respect the following hardware and software
specifications:
• Hardware specifications:
o Board: Raspberry Pi 3 Model B+;
o Camera: standard webcam with USB connector, with a minimum image
resolution of 720 x 576 px.
• Software specifications:
o Communication protocol with the web application: SCP (Secure Copy
Protocol);
o Data transmission: through Wi -Fi, using a static IP;
o The image acquisition and data transmission operations are to be done by a bash
script.
In the leaf image classifier development , several operations were involved. For
each operation, a serie s of specifications were imposed:
• Image database:
o Dataset size: approximatively 50 leaf samples for each species (150 leaves
in total), both for the training dataset and for the validation dataset;
o Image acquisition: to be done in a standardized manner, by creating a stand
in which the camera is positioned directly above the leaf.
• Image processing algorithm:
o To include the following operations done on the image: conversion to
grayscale, binarization, cropping, rotation, normalization and leaf data
extractio n;
o Implementation: using MATLAB’s Image Processing toolbox.

Introduction
4 • Classifier development:
o Type: Neural Network (NN);
o Requirements: to classify images belonging to three species of the Quercus
genus;
o Implementation: using MATLAB’s Neural Network Toolbox and Machi ne
Learning Toolbox.
The web application represents the connection between the previously mentioned
items, and also provides an easy to use interface for users that do not have the expertise
to work directly with those 2 components. The following specifica tions were imposed:
• Functional requirements:
o Image upload using the Raspberry Pi based system and from local storage;
o Image classification by calling the developed classifiers ;
o Classification result manual editing;
o Deletion of an uploaded image.
• Architectu re:
o MVVM (Model -View -View Model);
o Three -layer architecture (Data Access Layer, Business Layer, Application
Layer);
• Database:
o Relational database;
o Technology: Microsoft SQL Server.
• Technologies:
o ASP.NET (C# ; for backend );
o Entity Framework (for Object Relati onal Mapping);
o Razor View Engine, HTML, CSS, JavaScript, jQuery (for the frontend part) .
• Other constraints:
o Image acquisition using the Raspberry Pi system should be done by running
a batch script from the backend;
o Image classification should be done by ca lling a MATLAB function using a
COM object.

Bibliographic Study
5 2 Bibliographic Study
The Quercus genus contains over 200 species, mostly trees. They occupy large areas
in temperate and subtropical regions of the Northern Hemisphere. The number of
European oak species is lo w compared to East Asian and, especially, North American
species. Morphologically, they are characterized by alternate leaves, that are caducous,
marcescent or persistent, lobate. The flowers are unisexual monoecious. The male flowers
have the perigone div ided in 4 -7 lobes and 4 -12 stamina, grouped in log, thin, pendent
aments. The female flowers have a slightly dented perigone and are solitaire or groped in
a spike inflorescence containing 2 -8 flowers. The fruit is an achene (the acorn) that is
cylindrical , ellipsoidal or hemispherical, contained in a cup with numerous scales
(imbricated, free or concrescent) and has an annual or biannual maturation and a
hypogeal germination. Oak species have a light behavior. In Romania, they occupy 19% of
the Forest Fund [1].
For this research , three species were chosen to develop classifiers: Quercus robur
(common oak), Quercus petraea (sessile oak) and Quecus frainetto (Hungarian oak). Since
the species belong to the same genus, morphologica l similarities are high. However, some
differences do exist. A detailed description about each species and its leaf is given in the
following paragraphs.
Common oak occupies approximatively 130000 ha. It is encountered in pure or
mixed stands in plain and low hills areas, such as the Western plain, central Transylvania
or north -eastern Moldavia . The leaves are 6 -20 cm long and 3 -10 cm wide, and have
variable shapes, obovate or rhombic -obovate, having the maximum width in the fore
third. They have rounded ti p, narrowed and auriculate base, are sessile or have a short
petiole and have 4 -8 pairs of unequal, asymmetric lobes [1].
Sessile oak is encountered at altitudes that can be higher than the common oak’s
area, of up to 1000 m i n the Southern Carpathians. Its leaves are 8 -16 cm, rhombic –
obovate, variable, having the maximum width between middle and tip, a usually cuneate
base, gradually narrowing towards the tip. They have a 1 -2.5 cm petiole and have 5 -8 (10)
pairs of rounded lob es, gradually getting smaller towards the tip. The largest pair is the
third or the fourth from the base [1].
Hungarian oak is encountered in hilly areas of Banat and Transylvania and also in
Muntenia and Oltenia. It is encoun tered at altitudes of up to 600 m. Hungarian oak leaves
are large, up to 20 cm, obovate -elliptic. The maximum width is in the half containing the
tip. They are sessile or have a short petiole, have deep lobes, the lobes are lobulated and
are deep [1].
In this project, an image database was created using samples collected from the field.
It was decided to create a new database. Other papers, such as [2] or [3], use an existing
database (Plant scan database and Flavia , respectively).
Leaf recognition has been treated extensively in recent literature and many
methodologies have been proposed for tree species. Various parameters and classifiers
were employed. The classi fiers include k -nearest neighborhood classifiers, Neural

Bibliographic Study
6 Network, Support Vector Machine. The parameters taken into consideration were shape
features, color histograms, Moment -Invariants, Fourier Descriptors, geometric and digital
features. Accuracies of u p to 96% were obtained.
Reference [4] used a k -nearest neighborhood classifier that was implemented and
tested on 640 leaves belonging to 32 plant species (20 images per species) . Image
acquisition was performed in daylight, us ing a smartphone camera. The image resolution
was 1980 x 1024 px. The image preprocessing algorithm consisted of the following steps:
rotation (such that the tip of the leaf is oriented vertically towards the upper part of the
image), conversion to graysca le (because information regarding color is not required by
the proposed approach), thresholding (using Otsu’s method, having as a result the binary
image), opening operations (a number of three successive sets of opening operations
consisting of erosions a nd dilations for hole removal that occur in the processed image
due to noise), inverse threshold (in order to obtain a black background), edge extraction
(obtaining the leaf edge using Suzuki’s algorithm) and edge filtering (elimination of small
contours t hat possess a relatively small length compared to the largest contour). From the
processed image, a series of features were extracted. A brief description of each
feature/feature class is given. The convex hull (obtained by using the boundary points of
the leaf) parameters were the number of vertices, area and perimeter . Morphological
features included the leaf length, width, perimeter and area. The first two parameters
were determined by the finding of the minimum and maximum coordinates along the X
and Y axes. Another feature type employed by this study were the distance maps, namely
the vertical and horizontal maps and the centroid radial map. In the first case, lines were
drawn on the segmented image. Each line consisted of minimum and maximum
coordinate values, which intersected the leaf on the corresponding axis. The distance of
each line was estimated and stored. In the second case, the Euclidean distances between
16 points taken on the bounding box and the leaf centroid (the point of intersection of t he
diagonals of the bounding box) were calculated. The color histogram was also taken into
consideration as a feature. It was computed for a cropped part of the image (a central part,
determined using the length and width of the leaf bounding box). The cro pping was
performed in detriment of the usage of the entire image because the white background
around the leaf would have affected the color histogram. The accuracy obtained was
83.5%. this was further enhanced to 87.3% by using data from a color histogram . The
classification was done on a server and it used a mobile application to take and classify
images.
Three plant types (Pittosporum tobira , Betula pendula and Cercis siliquastrum ) were
analyzed by [2] used two approaches: o ne based on Moment -Invariant model and one
based on Centroid -Radii Model. Both of them implied using shape modelling techniques.
In the first case, four normalized central moments were studied in different combinations
(individually, in joint 2 -D and 3 -D feature spaces) , in order to produce optimum results.
The second case employed an edge detector to extract the leaf shape. A feature vector
containing 36 radii at a 10 -degree angular separation was constructed. A combination of
these models was tested to ve rify if a better accuracy could be obtained. A dataset of 180
images (3 classes of 60 images each) was used to train Neural Network classifiers. The
classification accuracy ranged between 90 -100%.

Bibliographic Study
7 Artificial Neural Networks were also used by [5] to classify plants. 534 leaves
belonging to 20 species were collected. It is to be noted that only healthy, representative
leaves that did not present any defects were collected since the focus was put on having
high quality samples. No resolution or orientation constraints were imposed upon the leaf
images. 400 leaf samples were used to train the classifiers. The 134 test samples yielded
a 92% recognition accuracy. An image processing algorithm was applied to the images,
having as purpos e the accurate extraction of leaves from the background. It consisted of
four steps (conversion to grayscale, thresholding, binarization and filtering). A different
method than the one used by other research papers was employed for the determination
of the grayscale image. It was obtained by the processing of the blue band component of
the original image. Image processing techniques were used to extract leaf shape features:
aspect ratio (ratio between the maximum width and the major axis length) , width rat io
(the ratio between the width at the half of the major axis and the maximum width) , apex
angle (the angle at the apex between leaf edges and the width at three quarters of the
major axis) , apex ratio (the ratio between the angle made at the apex by the w idth at 90%
of the major axis and the angle at 75% of the major axis) , base angle (the angle between
leaf edges with the width at one quarter of the major axis) , centroid deviation ratio,
moment ratio (the ratio between the deviation on Y axis and the devi ation on X axis from
he leaf centroid) and circularity (the ratio between 4 x π x leaf area and the leaf perimeter
squared) . These features were given as inputs to the neural network. The classification
accuracy was obtained even without taking into consideration types of leaf margins, vein
and removal of the petiole. A software sol ution was developed in which the user provides
the leaf image as an input and it requires inly two clicks to return a classification result.
Reference [6] developed a leaf recognition system of woody species in Central
Europe b ased on images. A new dataset containing 151 species , with at least 50 leaves per
species was used , from the Middle European Woody Plants 2012 dataset. Among the
species included in the dataset were: Quercus frainetto, Alnus glutinosa, Kerria japonica,
Cornus mas, Elaganus angustifolia, Aesculus hippocastanum, Betula pendula, Acer
campestre, Carpinus betulus, Syringa vulgaris, Ulmus levis . This dataset contained native or
cultivated tree and shrub species. Image acquisition was performed using a scanner wit h
a resolution of 300 DPI. The images had a white background and were stored in a PNG
format. This paper took a more complex approach related to the dataset, since it was
botanically supervised, it provided a unique approach for compound leaves and it
cont ained a suitable quantity of high quality leaf samples. For image preprocessing, the
following operations were involved: conversion to grayscale, binarization (using Otsu’s
method, with a manual correction and contour extraction (using a four -neighborhood
method and keeping only the longest boundary, while others, such as holes, were
considered as noise). Features based on leaf boundary and texture were investigated . Two
types of features were tested: image moments and Fourier descriptors. The classificatio n
was done using a nearest neighborhood classifier. It was based on Fourier descriptors
normalized to translation, rotation, scaling and the starting point of the boundary.
Moreover, the size of the leaf was used as a separate feature, provided it was know n.
Access to the classifiers was provided through a web application.

Bibliographic Study
8 Probabilistic Neural Networks and image and data processing techniques were used
by [7] for the implementation of a general purpose automated leaf recognition solution
for plant classification. The image processing algorithm consisted of the following steps:
conversion to grayscale, binarization, smoothing and filtering using a Laplacian filter. 12
digital morphological features were computed, based on the foll owing basic geometric
features: diameter, physiological length, physiological width , leaf area and leaf perimeter.
An input vector of 5 principal variables was given to the PNN. These variables are based
on 12 leaf features that were extracted and orthogon alized using Principal Component
Analysis. Training was done on 1800 leaves belonging to 32 plant species. An accuracy
greater than 90% was obtained. This solution provided an accurate artificial intelligence
approach for plant classification that has as a dvantages a fast execution time and an easy
implementation.
Cotton, orange and lemon leaves were classified using Support Vector Machine and
Neural Network classifiers by [8]. The purpose was to detect diseases. Leaf samples we re
collected, and image acquisition was done using a digital camera, under controlled
illumination conditions. The classification was based on color feature extraction from the
RGB color model. The RGB color pixel indexes were extracted from an identified Region
of Interest. The classifiers were implemented using MATLAB toolboxes. An accuracy of
69% was obtained.
Reference [9] used image processing techniques and Gray Level Cooccurrence
Matrix and Principal Component Analysis me thods to classify leaf images. GLCM and PCA
were used to extract leaf texture features. These algorithms were trained on 390 leaves
belonging to 13 plant species with 65 new or deformed leaf images. Accuracies of 78%
and 98% were obtained by the GLCM metho d and PCA method, respectively.
The classification and authentication of medicinal plant materials and herbs widely
used for Indian herbal medicine preparation was the target for [10] . It was important to
determine the quality and authenticity of these leaves. Texture features were used for
classification, using PCA method. That method was used to prevent misclassification of
look -alike leaves. For the detection of the leaf edge, a Sobel filter was used. Shape,
venation, textur e and color were taken into consideration as leaf features. The following
leaf parameters w ere used: aspect ratio, compactness, centroid, eccentricity, dispersion.
An accuracy of 89.2% was obtained.
A different approach was proposed by [11] . Most of the automatic identification
methods are based on features of leaf shape, venation and texture. Leaf tooth is commonly
used in traditional plant identification. However, it is ignored in most of the cases. In this
paper, an identific ation method based on sparse representation of leaf tooth features was
proposed. Using this method, the image corners were identified first. The abnormal image
corner was identified using the PauTa criteria. After that, the top and bottom leaf tooth
edges are detected, to effectively correspond to the extracted image corners. A feature
vector based on four leaf tooth features (leaf num, rate, sharpness and obliqueness) is
created. A sparse representation -based classifier was used to identify plant samples.
Testing revealed a 76% accuracy rate. The image preprocessing algorithm involved the
following steps: grayscale conversion, binarization through adaptive image thresholding

Bibliographic Study
9 on the grayscale image, application of a Roberts cross operator to obtain an edge i mage,
application of a dilation operator in morphology operations to remove holes in the image
and the thin operator in mathematical morphology to make leaf edge one pixel thick. It
was implemented using MATLAB Image Processing Toolbox.
Reference [12] used a statistical model classifier to compute the matching store of
the template and query leaf. The method proposes the extraction of leaf contours (the
lines between the centroid and each contour point of the image). A length hi stogram was
created for the distribution of distances in the leaf contour. The method achieved a 97.2%
accuracy. The study shows that the features are invariant to scale and rotations since the
features are approaching positive correlation in terms of coef ficient correlation. It also
proves that the method offers better performances than Zernike moments and curvature
scale space.
Three techniques were presented by [3] for classifying leaf shapes. These
techniques were Sup port Vector Machine – Binary Decision Tree, Probabilistic Neural
Network and Fourier Moment technique. The three techniques were applied on an
existing image database (the Flavia database), which contained 1600 leaves belonging to
32 classes, each class co ntaining 50 -60 leaf samples of the same type. It was observed that
SVM -BDT performs better than the PNN with Principal Component Analysis and the
Fourier Moment technique. The advantages of the SVM -BDT architecture over the other
two methods were: efficien t computation of the decision tree architecture, high
classification accuracy of the SVM, improvement in recognition speed when dealing with
a large number of classes. The image preprocessing algorithm included the conversion to
grayscale, binarization, sm oothing and application of a Laplacian filter to extract the leaf
contour. Five basic geometric features of the leaf were extracted: diameter, length, width,
area and perimeter. From these features, twelve digital morphological features were
computed: smoo th factor, aspect ratio, form factor, rectangularity, narrow factor,
perimeter ratio of diameter, perimeter ratio of length and width and vein features. The
SVM -BDT had the highest performance (96%), the PNN had a 91% accuracy, while the
Fourier Moment tec hnique had an accuracy of 62%. The SVM -BDT yielded a better
performance than the other two methods because of a high generalization performance
without the need to add a prior knowledge. The PNN also had a high performance since
the PCA reduces the dimensi on of the input vector, hence improving the speed of the PNN.
Fourier Method implies a more extensive computational work for the calculation of the
moments. Because of that, a lower performance was obtained.
A plant recognition system that used color and s hape information was used by [13] .
It obtained an accuracy of 93.3% using an Artificial Neural Network and an accuracy of
85.9% using a k -Nearest Neighborhood classifier on the Flavia dataset. The method
proposed by [14] was base d on a combination of leaf shape feature and texture. A Gabor
filter was used to model the leaf texture. The shape was obtained using a set of curvelet
transform coefficient combined with invariant moments. The efficiency was tested using
a neuro -fuzzy controller and a feedforward backpropagation multi -layered perceptron.
The highest accuracy obtained was of 87.1% on 930 images belonging to 31 species.

Bibliographic Study
10 Reference [15] used vein patterns of scanned le af images to classify three
leguminous species (soybean, red and white beans) . The main difference of this approach
was that it did not take into consideration features such as shape, size, texture or color, as
in the case of most of the existing studies. Image segmentation was implemented using an
adaptive threshold and the unconstrained hit or miss transform. Image acquisition was
performed with a standard scanner. An accuracy of 84.1% was obtained using the
Penalized Discriminant Analysis method. This a ccuracy was increased to 88.4% if the
leaves were cleared using a chemical process. The main drawback was the increase of
time and cost.
The concept of bag of words was employed by [16] (borrowed from text
classification) to ex tract a bag of contours from leaf shapes. It was tested on the Swedish
leaf dataset with an accuracy of 96.6%. the classifier was of nearest neighborhood type. It
also implemented the concepts of Local Constrained Linear Coding and Spatial Pyramid
Matching for shape representation.
Medicinal plant leaf recognition was included in an automated system by [17] . Five
species were considered: Desmodium gyrans , Butea monosperma , Malpighia glabra ,
Helicteres isora and Gymnema sylvestre . The features used for leaf recognition were Grey
Tone Spacial Dependency Matrices (GTSDM) , grey textures and Local Binary Pattern
(LBT) operators. The grey level features (textures), also known as first order statistical
central moments, included the mean , variance, skewness and standard deviation. The first
order features do not take into account the relative layout between gray values. The
GTSDM features are second order features. Usually, they are more sensitive and intuitive
than in the case of first o rder features. This category of features is based on the GTSD
matrix, in which texture is related to the statistical distribution of gray tones. The matrix
is a tabulation of probability of gray levels x occurring in an image at a distance d and an
angle a from grey level y in two points (p and q). These probabilities give a cooccurrence
matrix. Six different classifiers were analyzed by this method. The LBP based textures are
operators that are known for their invariance to local grayscale variations and m onotonic
photometric changes and for a high descriptive power. The image is assumed to be
constituted of micropatterns of the input image (i.e., spots, lines, flat areas, edges). The
value is calculated on the basis of a circular neighborhood of p pixels o f gray levels q and
radius r around the central pixel of gray level value v. The LBP histogram was considered
as a feature. Image acquisition was done using a camera. Images were taken in a JPE
format, having a size of 3420 x 4320 x 3. The working dataset was formed by dividing the
images into a size of 50 x 50 x 3. A dataset of 250 images from 5 different plant species
was used. An accuracy of 94.7% was obtained without using any image preprocessing
algorithms.
Reference [18] used 64 features derived from shape only with a 71.5% accuracy. A
new method was proposed, the Distributed Hierarchical Graph Neuron (DHGN) for
pattern recognition and a k -Nearest Neighborhood for pattern classification. The DHGN
approach implements the re cognition based on neuron -value adjacency comparison
approach. The strength of the firing of each neuron is based upon the matching function
between the input element and the stored elements within the neuron (known as biases).
The bias entries represent p air entries (value, position) that correspond to pattern

Bibliographic Study
11 elements. The entries were obtained from synaptic responses between adjacent neurons
(i.e., input values of these neurons). In the case of Graph Neuron implementation, the
neuron firing mechanism is different from other intelligent techniques based on neurons.
Its synaptic plasticity is independent of weight adjustment mechanism based on input
strength. Hence, the value matching function between adjacent neurons is employed in
the output computation o f each neuron. The database contained 1600 images from 100
plant species. It contains both common species and exotic species, such as species from
the Eucalyptus genus ( Eucalyptus urnigera and Eucalyptus neglecta ). The training dataset
consisted of 70% of the images, while the rest was used for testing purposes. The
classifiers were implemented in Python.
The bisection of leaves was proposed as a method for leaf classification by [19] . A
series of preprocessing techniques were ap plied to leaf images. 7 low cost morphological
features, together with 3 extra features using leaf half images were extracted. In the case
of leaves that had similar morphological regions, the image was split into two regions. For
these leaves, the structu ral features of one half are similar, while the features of the other
half are different. By taking into consideration this, the leaf was oriented based on the
major axis. Two parts where obtained by performing a vertical slice through the leaf
centroid. T he features considered were area, extent and eccentricity. These
characteristics were extracted from each half and their proportion to each other were
used as features. These features were given as input to a Probabilistic Neural Network
classifier that yi elded a 92.5% accuracy. The training dataset consisted of 1120 samples
of 32 different species from the Flavia database, while testing was performed on 160
images.
Reference [20] implemented a locally modified linear discriminan t embedding
algorithm. It was applied on the ICL plant leaf database. The method was tested on 750
leaf images belonging to 50 different species. An accuracy of 93.5% was obtained.
A method based on vein, shape, color and texture features of the leaf was used by
[21] . This paper proposed an approach based on Zernike moments and other type of
features, such as geometric parameters, color moments and the Gray Level Co -Occurrence
Matrix. For leaf classification, two methods were c onsidered: one based on distance
measure, and one based on Probabilistic Neural Networks. The geometric features
included the following items: aspect ratio (the ratio between the minor axis length and
the major axis length), circularity (ratio between the leaf area and the perimeter squared),
irregularity (the ratio between the radius of the maximum circle enclosing the region and
the minimum circle that can be contained in the region), vein features (defined as ratios
between pixel numbers that are part of the vein and the area of the leaf), solidity (ratio
between the leaf area and the area of the convex hull), convexity (ratio between the
perimeter of the convex hull and the perimeter of the leaf). The color moments were
extracted from the Red, Green, Blu e components of the leaf image by employing statistical
parameter computations (mean, standard deviation, kurtosis, skewness). From the Gray
Level Co -Occurrence Matrix, the following features were extracted: energy, contrast,
homogeneity, entropy and corre lation. The Zernike Moments were based on Zernike
polynomials. These polynomials are orthogonal to the unit circle. 54 different features
were extracted and given as input to a Probabilistic Neural Network classifier. Flavia and

Bibliographic Study
12 Foliage dataset were used f or testing purposes. The accuracy on both dataset s was around
95%. The usage of the Principal Component Analysis as a feature selection method
provided a slight accuracy improvement (of approximatively 2%).
Probabilistic Neural Networks were also used by [22] for the classification of broad
flat leaves. The shape was extracted using some user inputs. The user selected a base point
and a number of reference points on the leaf bla de. A rotation was performed, such that
the base of the leaf was on the left part of the image, and the lea f was horizontally aligned.
Based on those, a binary image was extracted. Out of this image, several features were
extracted (area, perimeter , eccentricity , minor axis length, major axis length, equiv alent
diameter, convex area, extent ). Another set of features was based on a more complex
algorithm. Slicing by the major axis, on a parallel direction with the minor axis as
performed. After that, a normalization was done on the feature points, consideri ng the
ratio between the slice length and the leaf major axis length. The training dataset for the
Probabilistic Neural Network consisted of 1200 samples from 30 plant species. A ten -fold
cross validation technique was used by the proposed framework. A 91. 4% accuracy was
obtained.

Design and Implementation
13 3 Design and Implementation
This project has as main objective the design and implementation of an image
classification application. It contained three main components:
• Raspberry Pi image acquisition system;
• MATLAB image processi ng and classification functions;
• Web application.
A block diagram overview of the application is presented in Figure 3.1. the web
application serves both as a connection between the other components and as a User
Interface, through which the user interacts with the aforementioned components. Image
acquisition is done using a Raspberry Pi system. The user sends a request from the web
application to the Raspberry Pi and a leaf image is taken and sent back to the web
application. The user can classify the imag e using the MATLAB functions. The Microsoft
SQL Server Database has as main function the storage of application data (namely images,
image data and classification results).

Figure 3.1. Block diagram of the application.
A more detailed description about the design and implementation of each
component is given in the following three subchapters. Information about design,
considered technologies and methodologies, implementation and testing are presented in
depth.

Design and Implementation
14 3.1 Raspberry Pi Image Acquisition System
3.1.1 Technology
The application, in its current form, was intended to be used in laboratory. Hence,
an image acquisition system was required. In order to make the system more easily to be
accessed by different clients and to make it easier to move, a device with wireless
connectivity was needed to perform this function. A Raspberry Pi was chosen due to its
small size, availability of a Wi -Fi module and the availability of free software (in the case
a Linux type operati ng system was installed).
For this application, a Raspberry Pi 3 Model B+ was used. A list of hardware
specifications is presented below:
• Processor : Broadcom BCM2837B0, Cortex -A53 64 -bit SoC @ 1.4GHz;
• Memory: 1GB LPDDR2 SDRAM;
• Connectivity:
o 2.4GHz and 5GHz IEEE 802.11.b/g/n/ac wireless LAN, Bluetooth
4.2, BLE;
o Gigabit Ethernet over USB 2.0 (maximum throughput 300Mbps);
o 4 × USB 2.0 ports.
• Access: Extended 40 -pin GPIO header;
• Video and sound:
o 1 × full size HDMI;
o MIPI DSI display port;
o MIPI CSI camera port;
o 4 pole stereo output and composite video port.
• Multimedia: H.264, MPEG -4 decode (1080p30); H.264 encode (1080p30);
OpenGL ES 1.1, 2.0 graphics;
• SD card: Micro SD format for loading operating system and data storage;
• Input power:
o 5V/2.5A DC via micro USB conn ector;
o 5V DC via GPIO header;
o Power over Ethernet (PoE) –enabled (requires separate PoE HAT).
***** Can I give a r eference to datasheet???
Based on these specifications, the following remarks were made. As a power
supply, both a dedicated power supply and a simple micro USB cable connected to a USB
port could be used. For image acquisition, a special Raspberry Pi camera could be used.
However, this option is more expensive than a standard web camera that has an USB
connection. Therefore, a standard USB camer a was used.
The operating system chosen for this application was Raspbian. It is a version of
Debian (Unix based operating system), that was optimized for the Raspberry Pi hardware.
It also contained more than 35000 software packages that are easy to be i nstalled. The
main advantage of this operating system is that it does not require a paid license to use it.

Design and Impleme ntation
15 It also provides powerful software and communication protocols suited for the current
application.

3.1.2 Hardware layout and setup
The following hardwar e components were used for this system (Figure 3.2) :
• Raspberry Pi 3 Model B+;
• Web camera with USB connection;
• Tripod;
• White background for the leaf;
• Raspberry Pi power supply (Micro USB, 5.1 V, 2.5 A).

Figure 3.2. Layout of the Image Acquisition System.
The Raspberry Pi require s a series of configurations to be done, before the proper
implementation of the image acquisition script. In order to perform these configurations,
two methods could be used: one base d on Remote Desktop Connection from a Windows
terminal and one based on direct configuration using a display and a keyboard.
The main advantage of the first method is that it does not require the extra hardware
components of the second method. However, it presents some drawbacks too, since it
requires more time to do it. It is possible to connect to the Raspberry Pi command line
using a laptop, an ethernet cable and PuTTY . PuTTY is an open source software that
provides an SSH and telnet client. An importan t prerequisite of using this approach is that
SSH must be enabled on the Raspberry Pi. One way to do this is to create an empty file,
without extension, having the name SSH on the micro SD card that contains the Raspbian
operating system before the first b oot of the system . This will automatically enable SSH

Design and Implementation
16 on the Raspberry Pi. After that, the xrdp package has to be installed from the command
line, using the sudo apt -get install xrdp command . xrdp is an open source remote
desktop protocol server. By obtain ing the IP address, a Remote Desktop Connection is
possible from the Windows terminal, giving the possibility to access the Raspbian UI and
make further configurations.
The second method represents an easier approach since it makes possible to directly
access the Raspbian desktop , without requiring the steps presented in the previous
paragraph . By making use of the HDMI port and one or two of the USB ports, a monitor, a
keyboard and a mouse can be connected to the Raspberry Pi, hence its usage becoming
simi lar to a desktop computer.
SSH can be enabled from the command line, by using the command sudo raspi –
config . A configuration screen will be opened and the SSH can be enabled from there.
In order to easily find the Raspberry Pi in the network, static IPs w ere configured
both for eth0 and wlan0 interfaces. These are interfaces for ethernet and Wi -Fi
connections. This configuration can be done by modifying the dhcpd.conf file, by using
the sudo nano /etc/dhcpcd.conf command. The contents of Figure 3. 3 should be
inserted in this file. The IP address can be set by modifying the value of the static
ip_address field. A valid address must be provided. The values depend on the router type.

Figure 3.3. Configuration o f static IP addresses for eth0 and wlan0 interfaces.
Another file that needs to be modified is the interfaces file. The main purpose is
to configure the supplicant for interface wlan0 . The file can be edited by using the
following command: sudo nano /etc/network/interfaces. In Figure 3. 4, the
supplicant for wlan0 is set to be the wpa_supplicant.conf file.

Figure 3.4. Contents of the interfaces file.

Design and Implementation
17 Finally, the contents of the wpa_supplicant.conf file can be adjusted. This step is
not usually necessary, since information related to Wi -Fi networks is automatically added
when the device successfully connects to a network. However, for the Raspberry Pi 3
Model B+ it is necessary to make sure that the file con tains the third line from Figure 3. 5
(country=RO ), since country settings are required by this model in particular. To
manually add a Wi -Fi network, the code should have a similar structure as presented
below, for each network (ssid, psk, key_mgmt). The fi le can be accessed using the
command sudo nano /etc/wpa_supplicant /wpa_supplicant.conf.

Figure 3.5. Structure of the wpa_supplicant.conf file.
Finally, the package for the web camera should be installed. F or a standard web
camera, the fswebcam package can be used. This software is used to generate images
from a web camera. It uses the GD Graphics Library to obtain images in PNG or JPEG
format. Installation can be done by using the sudo apt -get install fsweb cam
command. While the quality and the configurability of the standard web camera are lower
than a Raspberry Pi camera module, the hardware costs are much lower.
Another requirement, this time on the Windows side, represents the installation
of PuTTY and its PSCP application. PSCP is a Secure Copy Protocol client that allows the
transfer of files between though SSH, by using the command line.

3.1.3 Image acquisition and transmission script s
Two scripts were developed: one for image acquisition on the Raspberry Pi and one
for image transmission to the Windows terminal. Both scripts are given a detailed
description in the following paragraphs.
For the Raspberry Pi, a Bash script has been written. A code snippet is available in
Figure 3.6. The first line defines th e type of script (bash script in this case). The second
line calls the fswebcam application in order to perform the image acquisition. The image
resolution (720 x 576 px), save location, name and format are given as parameters to this
command. It is important that the provided save path does exist. It is important that the
acquired image has the same resolution as the ones used for creating the image database
for classifier development. This aspect is further detailed in the following subchapter,
together with the acquisition conditions.

Design and Implementation
18
Figure 3.6. Bash script for image acquisition.
A sample image of this script’s output is shown in Figure 3.7. The white background
has as purpose the easier identification of the leaf by the image processing algorithm.

Figure 3.7. Image acquired by the Bash script.
The second step represents the image transmission from Raspberry Pi to
Windows. This is one through the execution of a Batch file , on the Windows terminal
(Figure 3.8). The first two lines ensures that the command prompt is located in the
installation directory of PuTTY, since that is the location of the PSCP application. The third
line conne cts to the Raspberry Pi (using the static IP of the wlan0 interface) and calls the
Bash script previously described, creating the leaf image on the Raspberry Pi storage
location. The last line invokes the PSCP application. A call is made on the Raspberry P i
using the same static IP as before in order to transfer the image file from the source
location to the destination location. The source location represents the path where the
image is stored on the Raspberry Pi. The destination location is the physical path where
the image should be stored in the Windows terminal.

Figure 3.8. Batch file for image transmission.

Design and Implementation
19 3.2 Image processing and classification

Design and Implementation
20 3.3 Web application
Aceasta parte a lucrării este flexibilă și depinde foarte mult de natura lucrării , poate
fi organizată în mai multe capitole și conține contribuț iile personale ale autorului .
Include ți:
– Detalii referitoare la analiză și proiectare:
▪ descrierea metodelor pe care le -ați aplicat pentru rezolvarea problemei,
▪ descrierea materialelor, procedurilor
▪ calcule, tehnici, descrierea echipamentelor
▪ metodologia d e proiectare
▪ informa țiile necesare pentru ca cineva s ă poata reface lucrarea
– Implementare :
▪ Descrieti detaliile tehnice ale implementarii aplicatiei: mediul de
implementare, modul de prezentare, modul de utilizare al aplicatiei, etc.
– Testare si validare :
▪ Descrie ți metodologia de testare a aplica ției ș i rezultatele
▪ Include ți experimentele pe care le -ați realizat, analiza rezultatelor pe care
le-ați obținut.

Conclusion
21
4 Conclu sion
4.1 Obtained results
Evidentiați toate rezultatele pe care le -ați obtinut și trageți conc luzii din ele. Puteți
prezenta o analiză critică a ceea ce ați realizat comparativ cu alte lucrări/studii anterioare.
Includeți o listă a contribuțiilor pe care le -ați avut în domeniul temei abordate.
4.2 Further development
Descrieți direcțiile posibile de de zvoltare.

Reguli de formatare
22
5 Reguli de formatare
5.1 Formatarea paginii
– Dimensiunea paginii: A4
– Margini: 2.5 cm (sus, jos, stânga, dreapta)
– Antet și subsol: 1.27 cm de la marginea paginii
– În antetul paginii (header): titlul capitolului, centrat , stil: Header_style
– În subsolu l paginii: numărul paginii, centrat
5.2 Titluri și stiluri
Titlurile capitolelor și subcapitolelor se marchează cu stilurile Heading 1 – 4,
conform documentului model anexat în format Word. Descrierea stilurilor utilizate în
document este prezentată în Tabelul 5.1.
Tabelul 5.1. Stiluri utilizate în acest document
Nr. Stil Utilizat pentru Format
1 Normal Text normal Font: (Default) Cambria, 12 pt,
Justified, Line spacing: Multiple 1.1 li,
Space After: 6 pt
2 Titlu Titlul proiectului,
prima pagină Font: 24 pt, Small caps, Centered
Line spacing: single, Space Before: 126pt,
After: 0 pt,
3 Titlu2 Titlul proiectului,
pagina de
prezentare Font:14pt, Bold, Centered
4 Heading 1 Titlurile capitolelor
(nivel 1) Font: 24 pt, Indent: Left: 0 cm
Hanging: 0.76 cm, Space Before: 24pt,
After: 12pt
5 Heading 2 Titlurile
subcapitolelor
(nivel 2) Font: 14 pt, Bold, Indent: Left: 0 cm
Hanging: 1.02 cm, Space Before: 18pt,
After: 12pt
6 Heading 3 Titlurile secțiunilor
(nivel 3) Font: Bold, Indent: Left: 0 cm
Hanging: 1.27 cm, Space Before: 6 pt,
After: 6pt
7 Heading 4 Titlurile secțiunilor
(nivel 4) Font: Italic, Indent: Left: 0 cm
Hanging: 1.52 cm, Space Before: 2 pt,
After: 0 pt
8 Caption Legenda figurilor și
tabe lelor Font: Italic, Font color: Text 1, Line
spacing: single, Space After: 10 pt,

Reguli de formatare
23 Nr. Stil Utilizat pentru Format
9 Header_style Antetul paginii Font: 10 pt, Italic, Centered, Border:
Bottom: (Single solid line, Background 1,
0.5 pt Line width)

5.3 Figuri, tabele și ecuații
5.3.1 Figuri
Figurile se inserează în text centrate, cu etichetă de numerotare și legendă (Caption )
în partea de jos a figurii. Numărul figurii include și numărul capitolului, după exemplul
prezentat în Figura 5.1.

Figura 5.1. Figură exemplu , stil: Caption
5.4 Tabele
Tabelele se inserează în text centrate, cu etichetă și legendă (Caption) în partea de
sus a tabelului, aliniată la stânga. Numărul tabelului include și numărul capitolului, după
cum este prezentat, de exemplu, în Tabelul 5.1.
5.5 Ecuații
Ecuațiile se inserează în text centrate, cu numerotare în partea dreaptă. Numărul
ecuației include și numărul capitolului , conform exemplului din relația (5.1) .
(𝑥+𝑎)𝑛=∑(𝑛
𝑘)𝑥𝑘𝑎𝑛−𝑘𝑛
𝑘=0 (5.1)

Reguli de formatare
24 5.6 Referințe bibliografice
Se recomandă ca citarea referințelor bibliografice să fie făcută în formatul IEEE.
În secțiunea Bibliografie sunt prezentate exemple pentru: o citare a unui capitol
dintr-o carte [23] , un articol publicat într -o revistă [24] și un articol publicat la o
conferință [25] .
Detalii cu privire la formatul citării diverselor tipuri de referințe pot fi găsite în [26]
sau [27] .
Referințele bibliografice se pot insera în text utilizând facilitățile Word de a adăuga
surse ș i bibliografie unui document (References -> Citations & Bibliography) . Dacă
formatul IEEE pentru b ibliografie nu este instalat implicit în Word, se poate descărca
gratuit de la:
https://bibword.codeplex.com/wikipage?title=Styles&referringTitle=Home
Instrucțiunile de instalare pentru diferite versiuni de Word se pot obține de la aceeași
adresă.

References
25
6 References

[1] F. Clinovschi, Dendrologie, Suceava: Editura Universitatii Suceava, 2005.
[2] J. Chaki and R. Parekh, "Plant Leaf Recognition using S hape based
Features and Neural Network classifiers," International Journal of Advanced
Computer Science and Applications, vol. 2, 2011.
[3] K. Singh, I. Gupta and S. Gupta, "SVM -BDT PNN and Fourier Moment
Technique for Classification of Leaf Shape," International Journal of Signal
Processing, Image Processing and Pattern Recognition, Vols. 3, no. 4, December
2010.
[4] T. Munisami, M. Ramsurn, S. Kishnah and S. Pudaruth, "Plant leaf
recognition using shape features and colour histogram with k -nearest
neighbor classifiers," Procedia Computer Science, no. 50, pp. 740 -747, 2015.
[5] S. Hati and G. Sajeevan, "Plant Recognition from Leaf Image through
Artificial Neural Network," International Journal of Computer Applications
(0975 – 8887), vol. 62, Januar y 2013.
[6] P. Novotny and T. Suk, "Leaf recognition of woody species in Central
Europe," Biosystems Engineering, vol. 115, no. 4, 2013.
[7] S. Wu, F. Bao, E. Xu, Y. -X. Wang, Y. -F. Chang and Q. -L. Xiang, "A Leaf
Recognition Algorithm for Plant Classi fication Using Probabilistic Neural
Network," 2007 IEEE International Symposium on Signal Processing and
Information Technology: 11 -16, 2007.
[8] B. Adsule and J. Bhattad, "Leaves Classification Using SVM and Neural
Network for Disease Detection," Inter national Journal of Innovative Research
in Computer and Communication Engineering, vol. 3, no. 6, June 2015.
[9] A. Ehsanirad and S. Kumar Y. H., "Leaf recognition for plant classification
using GLCM and PCA methods," Oriental Journal of Computer Scienc e &
Technology, Vols. Vol. 3(1), 31 -36, 2010.
[10] T. Vijayashree and A. Gopal, "Authentication of leaf image using image
processing technique," ARPN Journal of Engineering and Applied Sciences, Vols.
10, no. 9, May 2015.
[11] T. Jin, X. Hou and F. Z hou, "A Novel Method of Automatic Plant Species
Identification Using Sparse Representation of Leaf Tooth Features," PLoS ONE
10(10): e0139482. doi:10.1371/journal., 2015.

References
26 [12] C. Gwo and C. Wei, "Plant identification through images: using feature
extrac tion of key points on leaf contours," Applications in Plant Sciences, 2011.
[13] V. Satti, A. Satya and S. Shama, "An automatic leaf recognition system for
plant identification using machine vision technology," International Journal of
Engineering Scien ce and Technology, vol. 5.4, pp. 874 -879, 2013.
[14] J. Chaki, R. Parekh and S. Bhattacharya, "Plant leaf recognition using
texture and shape features with neural classifiers," Elsevier: Pattern
Recognition Letters, vol. 58, pp. 61 -68, 2015.
[15] M. Larese, R. Namias, R. Craviotto, M. Arango, C. Gallo and P. Granitto,
"Automatic classification of legumes using leaf vein image features," Elsevier:
Pattern Recognition, vol. 47.6, pp. 158 -168, 2014.
[16] X. Wang, B. Feng, X. Bai, W. Liu and L. Latecki , "Bag of contour fragments
for robust shape classification," Elsevier: Pattern Recognition, vol. 47.6, pp.
2116 -2125, 2014.
[17] C. Arun, W. Sam Emmanuel and D. Durairaj, "Texture feature exraction
for identification of medicinal plans and comparison o f different classifiers,"
International Journal of Computer Applications, vol. 62.12, pp. 1 -9, 2012.
[18] A. Amin and A. Khan, "One shot classification of 2 -D leaf shapes using
distributed hierarchical graph neuron scheme with k -NN classifier," Elsevier :
17th Asia Pacific Symposium on Intelligent and Evolutionary Systems, pp. 84 -96,
2013.
[19] C. Uluturk and A. Ugur, "Recognition of leaves based on morphological
features derived from two half -regions," IEEE International Symposium on
Innovations in In telligent Systems and Applications, pp. 1 -4, 2012.
[20] S. Zhang and Y. Lei, "Modified locally linear discriminant embedding for
plant recognition," Elsevier: Neurocomputing, vol. 74.14, pp. 2284 -2290, 2011.
[21] A. Kadir, L. Nugroho, A. Susanto and P. Santosa, "Experiments of Zernike
moments for leaf identification," Journal of Theoretical and Applied
Information Technology (JATIT), vol. 41.1, pp. 82 -93, 2012.
[22] J. Hossain and M. Amin, "Leaf shape identification based plant
biometrics," IEEE Pr oceedings of 13th International Conference on Computer
and Information Technology, pp. 458 -463, 2010.
[23] P. Nume, "Titlul capitolului," in Titlul cartii , Oras, Editura, 2016, pp. 1 -24.
[24] P. Nume, "Titlul articolului," Titlul revistei, vol. 1, no. 2, pp. 22 -30, 2016.
[25] P. Nume, "Titlul articolului," in Numele conferintei , Oras, 2015.

References
27 [26] "IEEE Citation Reference," 2009. [Online]. Available:
https://www.ieee.org/documents/ieeecitationref.pdf.
[27] "IEEE Editorial Style Manual," 2016. [On line]. Available:
https://www.ieee.org/documents/style_manual.pdf.
[28] R. P. J. Chaki, "Plant Leaf Recognition using Shape based Features and
Neural Network classifiers," International Journal of Advanced Computer
Science and Applications, vol. 2, 2011.