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

 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