A Fast Leaf Recognition Algorithm based on SVM Classifier and High
Dimensional Feature Vector
Cecilia Di Ruberto and Lorenzo Putzu
Department of Mathematics and Computer Science, University of Cagliari, via Ospedale 72, 09124 Cagliari, Italy
Keywords:
Image Analysis, Feature Extraction, Leaf Recognition, Plant Classification, Support Vector Machine.
Abstract:
Plants are fundamental for human beings, so it’s very important to catalog and preserve all the plants species.
Identifying an unknown plant species is not a simple task. Automatic image processing techniques based on
leaves recognition can help to find the best features useful for plant representation and classification. Many
methods present in literature use only a small and complex set of features, often extracted from the binary
images or the boundary of the leaf. In this work we propose a leaf recognition method which uses a new
features set that incorporates shape, color and texture features. A total of 138 features are extracted and used
for training a SVM model. The method has been tested on Flavia dataset (Wu et al., 2007), showing excellent
performance both in terms of accuracy that often reaches 100%, and in terms of speed, less than a second to
process and extract features from an image.
1 INTRODUCTION
Hundreds of thousands of plants exist on earth, which
share a very close relationship to human beings and
they play a very important role for human life in many
areas, such as food, medical science, industry and en-
vironment. For this reason it is very important to
be able to catalog and preserve all species of exist-
ing plants, many of which are endangered just be-
cause of human activity and pollution. Traditionally
the cataloging activities of an unknown plant species
requires the collection of a sample of the plant, which
must be done dry before being analyzed and pro-
cessed by specialists for comparison with other sam-
ples present in the herbaria. A faster way to catalog
the plants is based on leaves analysis which, unlike
flowers and fruits, are present for most of the year
and provide comprehensive morphological character-
istics for classification. Morphological analysis of the
leaf, although more rapid, can not be a simple task
especially for a non-expert user, as it needs for ac-
curate observations of the leaf itself and the compar-
ison with other samples of the same species, requir-
ing prior knowledge or the ability to navigate through
complex databases. For this reasons it’s important to
develop an automatic method in order to extract the
best leaves’ features and able to catalog a leaf from
an unknown plant species. Plant species classifica-
tion based on plant leaves images has become a very
important area of active research and many methods
for leaves recognition are present in literature. Typ-
ically these methods use features extracted from the
binary images or the boundary of the leaf. Some re-
searchers used part of those features only, while oth-
ers tried different and more complex approaches (Gao
et al., 2010a). In addition only few works make use
of color and texture information, such as first and sec-
ond order statistics (Kadir et al., 2011). However, in
almost all works in literature has been used a fairly
narrow set of feature, or in some cases techniques like
the PCA were used to reduce this set of features (Wu
et al., 2006). In this work we propose a plant classifi-
cation system based on leaf recognition which makes
use of a new and larger features set, that incorporates
shape, color and texture features, extractable from the
leaves images easily and quickly, in order to ensure
not only a remarkable accuracy but also speed in cat-
aloging an unknown leaf species.
The rest of this paper is organized as follows. Sec-
tion 2 describes how the image is pre-processed from
the input image. Section 3 illustrates the features
we propose. Section 4 describes the classification
step. The experimental results are presented in Sec-
tion 5 and conclusions in Section 6. Each phase of the
method is analyzed in detail and compared with other
approaches present in literature. The whole process
can be schematized as showed in Figure 1.
601
Di Ruberto C. and Putzu L..
A Fast Leaf Recognition Algorithm based on SVM Classifier and High Dimensional Feature Vector.
DOI: 10.5220/0004740606010609
In Proceedings of the 9th International Conference on Computer Vision Theory and Applications (VISAPP-2014), pages 601-609
ISBN: 978-989-758-003-1
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
Figure 1: Proposed method diagram.
2 IMAGE PRE-PROCESSING
The identification of leaves from Flavia dataset im-
ages is not a difficult task, in fact all the images
present a white and almost uniform background. First
of all the image is scaled in order to make the sub-
sequent stage faster and then is converted from RGB
to gray scale. To make faster also the segmentation
step we use a predetermined threshold, like that used
in (Wu et al., 2007), but with a value of 0.92, which
from the experiments carried out we observed that is
optimal for all the images present in this dataset. Then
a complement image is calculated in order to obtain
a white leaf on a dark background. However, this
threshold operation does not produce a binary image
cleaned in all its parts, due to the presence of dark par-
ticles in the background or light particles inside the
leaf, such as the vein. Therefore an image cleaning
is performed through an area opening operation, that
removes all the objects in the image except one, that
presents the maximum size, which is the leaf. After
that an hole filling operation is performed in order to
fill all the holes inside the leaf created by the threshold
operation. At this point it is possible to use the binary
image just obtained to crop the background from the
original image in gray scale, useful for the color and
texture feature extraction. The last operation required
before the feature extraction step is the leaf boundary
extraction, which is obtained by a simple removal op-
eration, that removes from the binary images all those
pixel belonging to the interior of the object, which
have as neighbors only white pixels. At this point we
have all the images (shown in Fig. 2) necessary for the
extraction of our set of features. Unlike other works
in literature, our method does not present a vein ex-
traction step, as this approach seems very linked to
the quality of the segmentation and therefore suscep-
tible to many errors, but it also seems very expensive
in terms of time, given that this operation consists of
repeated operations of opening followed by as many
threshold operations, according to the approach pro-
posed in (Wu et al., 2006), or by repeated operations
of edge detection exploiting different threshold val-
ues, according to the approach proposed in (Du et al.,
2006).
3 FEATURE EXTRACTION
Speaking about feature extraction in this context
means to transform the images into data, then extract
information reflecting the visual patterns to which
botanists refer to, but at the same time it is neces-
sary to extract the descriptors that are most relevant
to the subsequent classification process. As said ear-
lier, methods of feature extraction from leaf images
present in literature, typically, use features extracted
from the binary images or the boundary of the leaf,
considering leaf external shape features the most im-
portant information for classification. For example,
Wang (Wang et al., 2003) represented the shape of
leaf with a centroid-contour distances curve. Im (Im
et al., 1998) and Du (Du et al., 2006) used polygo-
nal approximation to recognize plant species of Acer
family. Lee (Lee and Chen, 2006) and Arribas (Ar-
ribas et al., 2011) instead proposed a region-based
classification method for leaves. Other methods com-
bine the previous shape features with other features
that involve the analysis of many parts of the leaves,
such as edge, vein and skeleton considering also leaf
internal shape features. For example, Wu (Wu et al.,
2006) combined geometric features and vein features.
Cheng (Cheng et al., 2007) used a method that com-
bines leaf shape, leaf apex, leaf base, leaf margin, leaf
phyllotaxy, leaf vein and leaf size. Pauwels (Pauwels
VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications
602
Figure 2: Leaf image pre-processing: RGB scaled image, gray level image, binary image, complement binary image after
cleaning, gray scale cropped image, boundary image.
et al., 2009) used several shape features, including in-
dentation depth which is the distance from leaf con-
tour to the convex hull. Gao (Gao et al., 2010a), (Gao
et al., 2010b) used shape features in combination with
contour angles. Some researchers used part of those
features only, while others tried different and more
complex approaches, for example the invariant mo-
ments proposed by Hu (Hu, 1962) have been used by
Zulkifli (Zulkifli et al., 2011). Zhang (Zhang and Lei,
2011) proposed the use of Polar Fourier Transform.
Chaki (Chaki and Parekh, 2011) analyzed plant leaf
images using shape modeling techniques based on the
invariant moments and the centroid-radii model. Wu
(Wu et al., 2007) identified plants species using shape
features such as aspect ratio, leaf dent, leaf vein and
invariant moment. In addition only few works make
use of color and texture information, such as first and
second order statistics. Man (Man et al., 2008) in-
volved color moments based on mean, standard de-
viation and skewness. Ehsanirad (Ehsanirad, 2010)
used textural features extracted from the gray-level
co-occurrence matrices. Zhang (Zhang and Zhang,
2008) retrieved the leaf database using features of
color, texture and shape. Machado (Machado et al.,
2013) investigated the use of partial differential equa-
tions and fractal dimension theory to describe the tex-
ture present in leaves. In a similar way to that pro-
posed by Kadir (Kadir et al., 2011) who realized a leaf
classification using shape, color and texture feature,
from the images generated previously we have ex-
tracted different types of descriptors: external shape
features, regional shape features, invariants moments,
color features and texture features. Differently from
Kadir (Kadir et al., 2011) our method involves the use
of a large number of features but easy and quick to
extract, in order to guarantee a high level of perfor-
mance. So, starting from the boundary image we have
extracted 36 descriptors of the centroid radii. This
was done by calculating the centroid and after the dis-
tance of each boundary point from the centroid itself.
Maximum and minimum distance are then used to ro-
tate the contour, in such a way that the first centroid
radii descriptor represents for all the leaves the maxi-
mum distance. At this point the distances are sampled
to 36 values, in order to preserve a good approxima-
tion of the originals radii, preserving a radius at every
10 degrees interval. Finally, the sampled radii are nor-
malized using again the maximum distance value, in
order to obtain values between 0 and 1.
Then, starting from the binary image containing
only the leaf, we have extracted 7 Invariant Moments
proposed by Hu (Hu, 1962). Again from the binary
image we have extracted some classical shape de-
scriptors such as area, perimeter, convex area, convex
perimeter, minimum bounding box width and height
(showed in Fig. 3). These classical measures with the
maximum and minimum distance (often called excir-
cle radius and incircle radius, showed in Fig. 3), cal-
culated previously, are combined in order to calculate
8 regional shape descriptors that are elongation, rect-
angularity, solidity, eccentricity, compactness, round-
ness, convexity and sphericity. We have chosen these
shape descriptors for their ability to discriminate vari-
ous shapes and therefore different leaves species, pre-
serving, due to their calculation (for formulas refer to
Gonzalez (Gonzalez et al., 2004)), values always be-
AFastLeafRecognitionAlgorithmbasedonSVMClassifierandHighDimensionalFeatureVector
603
Figure 3: Leaf minimum bounding box, convex hull, excircle and incircle.
tween 0 and 1. Also, these descriptors together with
invariant moments are invariant to the operations of
rotation, translation and scaling.
The main disadvantage of these and others shape
features is that they are very susceptible to errors in
segmentation. For this reason we have used these de-
scriptors together with regional descriptors less sus-
ceptible to errors. Among these there are the color
descriptors which delineate the gray level color dis-
tribution of the images. So, starting from the gray
scale containing the cropped leaf, we have extracted
7 color descriptors that are mean, standard deviation,
smoothness, skewness, kurtosis, uniformity and en-
tropy. Often, however, the descriptors based only
on histogram have some drawbacks as they do not
give information on the mutual position of the pix-
els. Some objects, in particular leaves, have in fact a
repeating pattern as the primary visual characteristic
and so it is necessary to consider not only the inten-
sity distribution but also the positions of the pixels
having a similar gray level, as showed by Haralick
in (Haralick et al., 1973). Then we have also con-
sidered the descriptors applied to the gray level co-
occurrence matrix (GLCM) calculated starting from
the images containing the cropped leaf in gray scale.
These descriptors are autocorrelation, contrast, corre-
lation, cluster prominence, cluster shade, dissimilar-
ity, energy, entropy, homogeneity, maximum proba-
bility, variance, sum average, sum variance, sum en-
tropy, difference variance, difference entropy, infor-
mation measure of correlation1, information measure
of correlation2, inverse difference normalized and in-
verse difference moment normalized. These charac-
teristics have been calculated for angles of 0, 45, 90
and 135 degrees and distance 1, so we have extracted
80 texture features. The total number of extracted fea-
tures is then 138: 36 boundary descriptors, 7 Hu in-
variant moments, 8 shape descriptors, 7 color descrip-
tors and 80 texture descriptors.
4 CLASSIFICATION
Various types of classifiers have been used for leaves
recognition. In particular Neural Networks have been
attracted researchers in area pattern recognition be-
cause of their power to learn from training dataset.
For example, Chaki (Chaki and Parekh, 2011) pre-
sented a scheme for automated detection using sev-
eral Neural Network classifiers, while Zulkifli (Zulk-
ifli et al., 2011) used a General Regression Neural
Network. Probabilistic NN is another neural network
that has been used by several authors, such as Wu
(Wu et al., 2006) and Huang (Lin and Peng, 2008).
Du (Du et al., 2007) instead defined a classification
method called the move median centers (MMC) hy-
persphere classifier. In some cases, before the classi-
fication step, various authors also used classical di-
mensionality reduction approaches in order to find
a low-dimensional subspace in a high-dimensional
input space by linear or no linear transformation.
Among them, Principal Component Analysis PCA
used by Wu (Wu et al., 2006) and Kernel Princi-
ple Component Analysis KPCA used by Valliammal
(Valliammal and Geethalakshmi, 2011). On the con-
trary, in our work, the idea is to use the entire fea-
ture space for the training of a SVM model, both be-
cause a high number of feature guarantees the cre-
ation of a model with Support Vector able to achieve
an optimal separation among the various classes, but
especially because the process of reducing the num-
ber of features using linear transformation techniques,
such as PCA, or by feature selection techniques are
time consuming. The SVM classifier is trained fol-
lowing the one vs rest approach, which is preferred
to the one vs one approach, since the latter is very
expensive in terms of time and resources, necessary
both for the creation of an higher number of mod-
els and for the classification of new instances that re-
quires the comparison with all the models previously
created. Since the performance of the SVM, while
VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications
604
1 2 3 4 5 6 7 8
9 10 11 12 13 14 15 16
17 18 19 20 21 22 23 24
25 26 27 28 29 30 31 32
Figure 4: All leaves species in the Flavia dataset.
adopted with different kernel functions and different
parameters, are quite different, the classifier has been
modeled with different kernels and have been chosen
only those with better performances. For each kernel
functions the parameters have been tuned through op-
timization techniques in order to find the maximum
accuracy value.
5 EXPERIMENTAL RESULTS
The proposed method has been finally tested on Flavia
dataset, shared by Wu et al. (Wu et al., 2007) for aca-
demic research purpose, that contains a total of 1907
leaf images belonging to 32 plant species, showed in
Fig. 4. In Table 1 are then listed the names of the
plants and the number of samples for each species,
which on average is approximately 60. Given the con-
siderable size of the dataset we have decided to per-
form the validation of our method using a stratified
sampling, that allows us to maintain in the training
and test sets the same proportion between the classes
present in the original dataset. The performances of
the classification models were then evaluated by cal-
culating the accuracy, which gives us a good indica-
tion of the performance since it considers each class
of equal importance. The accuracy value has also
been calculated individually for each plant species, as
showed in Table 1, in which we have also listed the
most common recognition mistakes to indicate with
which other leaves classes are frequently confused.
In order to assess the robustness of our algorithm,
we have tested the classifier with different percent-
ages of training and test set. The first test was per-
formed using the largest possible number of samples
for the training, therefore using only one sample for
each species for the test, so the training set size is
1875 and the test set size is 32. The following tests
have been made using a decreasing number of sam-
ples for the training and an increasing number for the
test, up to a percentage of 50%. So, in the last test
we have used almost the same number of samples for
the training and test sets, which in this case have a
size of 953 and 954 respectively. The experimental
results in Table 2 show the excellent performance of
the proposed system in all tests performed. Despite
each kernel function used in the creation of the mod-
els leads often to a precision of 100%, the highest per-
formances are related undoubtedly to the Linear ker-
nel and the RBF kernel that uses a Gaussian radial
basis function. In particular, it can be noted that, also
varying significantly the percentage of samples used
for training and for the test, in both cases the system
is able to effectively classify the plants species, while
maintaining a consistent high accuracy value and a
very low standard deviation value, which goes to sta-
bilize even more with the use of a greater number of
samples for testing. The efficacy of the feature set
used can be seen from Table 3, which shows the per-
formance of the system using different feature subsets
among those listed previously. For this test we have
used a sample percentage for training set and test set,
respectively, equal to 80% and 20%.
From the results shown it is possible to observe
how each feature subset added leads to a greater level
of accuracy and contributes greatly to the stability of
AFastLeafRecognitionAlgorithmbasedonSVMClassifierandHighDimensionalFeatureVector
605
Table 1: Details about leaf name and number, mean accuracy for each species and common recognition mistakes.
Common name Label Species Samples Accuracy Common Mistake
Pubescent Bamboo 1 59 98.2% 20
Chinese Horse Chestnut 2 63 98.4% 6
Chinese Redbud 3 72 98.2% 31
True Indigo 4 73 100% no.
Japanese Maple 5 56 100% no.
Nanmu 6 62 98% 2
Castor Aralia 7 52 100% no.
Goldenrain Tree 8 59 100% no.
Chinese Cinnamon 9 55 100% no.
Anhui Barberry 10 65 98.2% 4
Big-fruited Holly 11 50 100% no.
Japanese Cheesewood 12 63 98% 24
Wintersweet 13 52 98% 11
Camphortree 14 65 100% no.
Japanese Viburnum 15 60 100% no.
Sweet Osmanthus 16 56 100% no.
Deodar 17 77 100% no.
Ginkgo Maidenhair Tree 18 62 100% no.
Crape Myrtle 19 61 100% no.
Oleander 20 66 98.3% 21
Yew Plum Pine 21 60 98% 20
Japanese Flowering Cherry 22 55 100% no.
Glossy Privet 23 55 100% no.
Chinese Toon 24 65 98.2% 2
Peach 25 54 98% 10
Ford Woodlotus 26 52 94% 1, 20
Trident Maple 27 53 100% no.
Beale’s Barberry 28 55 100% no.
Southern Magnolia 29 57 100% no.
Canadian Poplar 30 64 98% 3
Chinese Tulip Tree 31 53 100% no.
Tangerine 32 56 100% no.
Table 2: Experimental results with different sampling percent: for each test accuracy (acc) and standard deviation (SD) are
reported.
Linear SVM Quadratic SVM RBF SVM
Sampling perc. acc S D acc S D acc S D
1 s 0,994 0,013 0,977 0,028 0,994 0,013
10% 0,99 0,006 0,976 0,013 0,989 0,007
20% 0,99 0,004 0,97 0,007 0,988 0,006
30% 0,984 0,004 0,963 0,008 0,981 0,006
40% 0,983 0,004 0,957 0,006 0,982 0,004
50% 0,971 0,004 0,949 0,009 0,974 0,004
the system by reducing the standard deviation value.
Furthermore, one can observe the remarkable contri-
bution of color and texture features, which in most
of the works reported in the literature have not been
taken into account, but which obviously are very use-
ful in the discrimination of different leaves species,
in particular the feature of texture with regard to the
veins of the leaves. The Table 4 compares the results
obtained through our method with the results obtained
by various authors who have proposed a method of
leaf recognition tested on the same dataset Flavia. Wu
(Wu et al., 2007) used 12 shape and vein features, re-
duced to 5 principal components through PCA in or-
der to train a PNN classifier. Also Singh (Singh et al.,
2010) used shape and vein features but for the train-
ing of an SVM model arranged in a binary tree struc-
VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications
606
Table 3: Experimental results for different feature sets: for each test accuracy (acc) and standard deviation (SD) are reported.
Linear SVM RBF SVM
Feature Type acc S D acc S D
Shape 0,723 0,014 0,869 0,017
Shape+Inv.mom. 0,738 0,018 0,874 0,009
Shape+Inv.mom.+Cent.radii 0,784 0,017 0,907 0,012
Shape+Inv.mom.+Cent.radii+Color 0,918 0,015 0,97 0,009
Shape+Inv.mom.+Cent.radii+Color+Texture 0,99 0,004 0,988 0,006
Table 4: Comparison between our method and other meth-
ods present in literature.
Method Accuracy
Wu (Wu et al., 2007) 93%
Singh (Singh et al., 2010) 96%
Kadir (Kadir et al., 2011) 94,7%
Kulkarni (Kulkarni et al., 2013) 93,8%
Lee (Lee and Hong, 2013) 97,2%
Our method 99%
ture. Kadir (Kadir et al., 2011) extracted shape, vein,
color and texture feature and Polar Fourier Transform
to train a PNN classifier. Kulkarni (Kulkarni et al.,
2013) tested Zernike moments combined with shape,
color and texture feature. Finally, Lee (Lee and Hong,
2013) used shape, vein and contour feature and FFT
to recognize leaf species with an iterative score prior-
ity system. From the results it can be seen how the
proposed method leads to greater accuracy than that
achieved by other methods tested on Flavia dataset.
The tests carried out have also shown the excellent
performance in terms of speed of the proposed algo-
rithm. Despite the speed of a system of leaf recog-
nition depends on many factors, such as the size and
resolution of the image, the size of the leaf and the
complexity of texture and last but not least the config-
uration of the computer ( our computers was config-
ured respectively with a processor Intel(R) Core(TM)
i7 CPU, 960@3.20 GHz, RAM 4.00 GB size and with
a processor AMD Phenom(TM) II x6 1090T, RAM
4.00 GB size), in any case, we are able to extract the
feature vector from a leaf sample in less than a sec-
ond. Even, the SVM classifier with the one vs. rest
approach has proven itself sufficiently fast, being able
to create the model in the worst case, that is using the
RBF kernel, in about 12 seconds, while the predictor
step is performed instantaneously. So, we are able to
classify an unknown leaf species in less then a second.
6 CONCLUSIONS
In this work we have proposed an innovative method
for the completely automatic identification and classi-
fication of plant species based on leaf recognition, in
order to provide an automated procedure as support
for plant cataloging and preserving. Starting from
an original leaf image we can process it entirely, in
a fully automatic way without need for manual in-
tervention by the user. The proposed method makes
use of a new and larger features set, that incorporates
shape, color and texture features, extractable from the
leaves images easily and quickly. This set of features
is used entirely to train a SVM classifier with one vs
rest approach. The experimental results demonstrate
that the proposed leaf recognition system has excel-
lent performance, both in terms of accuracy and in
terms of speed, less than a second to catalog an un-
known leaf species. The accuracy often reaches 100%
and on average is around 99%, so it is better than the
methods reported in literature. Further developments
could affect the feature extraction phase, in order to
extract vein features easy to detect and more discrim-
inating than the features shown in the works reported
in literature. Thereafter, it will be important to test
the proposed method on other available datasets that
present greater variations in the type of images ac-
quired, both with regard to the dimensions, the res-
olution and the quality of the images themselves. In
this way we would be able to assess the quality of the
proposed method and make the changes necessary to
maintain a high level of performance and speed. Fi-
nally we will be able to focus in the development of
a real application for PC, which allows to catalog im-
ages of leaves acquired with a digital camera, and a
mobile application that allows to catalog the plants
directly in their natural habitat. This will require the
addition of new phases in the algorithm, in particular
for the segmentation of the images with the presence
of both uniform and non-uniform backgrounds.
ACKNOWLEDGEMENTS
This work has been funded by Regione Autonoma
della Sardegna (R.A.S.) Project CRP-17615 DE-
NIS: Dataspace Enhancing Next Internet in Sardinia.
Lorenzo Putzu gratefully acknowledges Sardinia Re-
gional Government for the financial support of his
AFastLeafRecognitionAlgorithmbasedonSVMClassifierandHighDimensionalFeatureVector
607
PhD scholarship (P.O.R. Sardegna F.S.E. Operational
Programme of the Autonomous Region of Sardinia,
European Social Fund 2007-2013 - Axis IV Human
Resources, Objective l.3, Line of Activity l.3.1.). We
wish to thank Wu et al. for having made available the
Dataset on which we could test our method.
REFERENCES
Arribas, J. I., Snchez-Ferrero, G. V., Ruiz-Ruiz, G., and
Gmez-Gil, J. (2011). Leaf classification in sunflower
crops by computer vision and neural networks. Com-
puters and Electronics in Agriculture, 78(1):9 – 18.
Chaki, J. and Parekh, R. (2011). Plant leaf recognition using
shape based features and neural network classifiers.
International Journal of Advanced Computer Science
and Applications, 2(10):41 – 47.
Cheng, S.-C., Jhou, J.-J., and Liou, B.-H. (2007). Pda plant
search system based on the characteristics of leaves
using fuzzy function. In New Trends in Applied Ar-
tificial Intelligence, volume 4570 of Lecture Notes in
Computer Science, pages 834–844. Springer.
Du, J.-X., Huang, D.-S., Wang, X.-F., and Gu, X. (2006).
Computer-aided plant species identification based on
leaf shape matching technique. Transactions of the
Institute of Measurement and Control, 28(3):275–285.
Du, J.-X., Wang, X.-F., and Zhang, G.-J. (2007). Leaf shape
based plant species recognition. Applied Mathematics
and Computation, 185(2):883 – 893.
Ehsanirad, A. (2010). Plant classification based on leaf
recognition. International Journal of Computer Sci-
ence and Information Security, 8(4):78–81.
Gao, L., Lin, X., Zhao, W., Chen, S., and Huang, H.
(2010a). An algorithm of excising leafstalk while
keeping its main body intact for leaf recognition. In
Image and Signal Processing (CISP), 2010 3rd Inter-
national Congress on, volume 6, pages 2732–2736.
Gao, L., Lin, X., Zhong, M., and Zeng, J. (2010b). A neural
network classifier based on prior evolution and itera-
tive approximation used for leaf recognition. In Nat-
ural Computation (ICNC), 2010 Sixth International
Conference on, volume 2, pages 1038–1043.
Gonzalez, R. C., Woods, R. E., and Eddins, S. L. (2004).
Digital Image Processing Using MATLAB. Pearson
Prentice Hall Pearson Education, New Jersey, USA,
1st edition.
Haralick, R., Shanmugam, K., and Dinstein, I. (1973).
Textural features for image classification. Systems,
Man and Cybernetics, IEEE Transactions on, SMC-
3(6):610–621.
Hu, M.-K. (1962). Visual pattern recognition by moment
invariants. Information Theory, IRE Transactions on,
8(2):179–187.
Im, C., Nishida, H., and Kunii, T. (1998). Recognizing plant
species by leaf shapes-a case study of the acer family.
In 14th International Conference on Pattern Recogni-
tion, volume 2, pages 1171–1173.
Kadir, A., Nugroho, L. E., Susanto, A., and Santosa, P. I.
(2011). A comparative experiment of several shape
methods in recognizing plants. International Jour-
nal of Computer Science & Information Technology,
3(3):256–263.
Kulkarni, A. H., Rai, D. H. M., Jahagirdar, D. K. A., and
Upparamani, P. S. (2013). A leaf recognition tech-
nique for plant classification using rbpnn and zernike
moments. International Journal of Advanced Re-
search in Computer and Communication Engineering,
2(1):82–93.
Lee, C.-L. and Chen, S.-Y. (2006). Classification of leaf
images. International Journal of Imaging Systems and
Technology, 16(1):15–23.
Lee, K. B. and Hong, K. S. (2013). An implementa-
tion of leaf recognition system using leaf vein and
shape. International Journal of Bio-Science and Bio-
Technology, 5(2):57–65.
Lin, H. and Peng, H. (2008). Machine recognition for
broad-leaved trees based on synthetic features of
leaves using probabilistic neural network. In Com-
puter Science and Software Engineering, 2008 Inter-
national Conference on, volume 4, pages 871–877.
Machado, B. B., Casanova, D., Gonalves, W. N., and Bruno,
O. M. (2013). Partial differential equations and frac-
tal analysis to plant leaf identification. Journal of
Physics: Conference Series, 410(1).
Man, Q.-K., Zheng, C.-H., Wang, X.-F., and Lin, F.-Y.
(2008). Recognition of plant leaves using support
vector machine. In Advanced Intelligent Computing
Theories and Applications. With Aspects of Contem-
porary Intelligent Computing Techniques, volume 15
of Communications in Computer and Information Sci-
ence, pages 192–199. Springer Berlin Heidelberg.
Pauwels, E. J., de Zeeuw, P. M., and Ranguelova, E. (2009).
Computer-assisted tree taxonomy by automated image
recognition. Engineering Applications of Artificial In-
telligence, 22(1):26–31.
Singh, K., Gupta, I., and Gupta, S. (2010). Svm-bdt pnn
and fourier moment technique for classification of leaf
shape. International Journal of Signal Processing, Im-
age Processing and Pattern Recognition, 3(4):67–78.
Valliammal, N. and Geethalakshmi, S. N. (2011). Hybrid
image segmentation algorithm for leaf recognition and
characterization. In International Conference on Pro-
cess Automation, Control and Computing, pages 1–6.
Wang, Z., Chi, Z., and Feng, D. (2003). Shape based leaf
image retrieval. Vision, Image and Signal Processing,
IEEE Proceedings, 150(1):34–43.
Wu, Q., Zhou, C., and Wang, C. (2006). Feature extraction
and automatic recognition of plant leaf using artificial
neural network. Avances en Ciencias de la Computa-
cion, pages 5–16.
Wu, S., Bao, F., Xu, E., Wang, Y.-X., Chang, Y.-F., and
Xiang, Q.-L. (2007). A leaf recognition algorithm for
plant classification using probabilistic neural network.
In IEEE International Symposium on Signal Process-
ing and Information Technology, pages 11–16.
Zhang, S. and Lei, Y.-K. (2011). Modified locally linear dis-
criminant embedding for plant leaf recognition. Neu-
rocomputing, 74(1415):2284 – 2290.
Zhang, X. and Zhang, F. (2008). Images features extraction
VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications
608
of tobacco leaves. In Congress on Image and Signal
Processing (CISP), volume 2, pages 773–776.
Zulkifli, Z., Saad, P., and Mohtar, I. (2011). Plant leaf iden-
tification using moment invariants amp; general re-
gression neural network. In 11th International Con-
ference on Hybrid Intelligent Systems (HIS), pages
430–435.
AFastLeafRecognitionAlgorithmbasedonSVMClassifierandHighDimensionalFeatureVector
609