International Journal of Bio-Science and Bio-Technology

Vol. 5, No. 2, April, 2013

An Implementation of Leaf Recognition System using Leaf Vein and

Shape

Kue-Bum Lee and Kwang-Seok Hong

College of Information and Communication Engineering, Sungkyunkwan University,
300, Chunchun-dong, Jangan-gu, Suwon, Kyungki-do, 440-746, South Korea
leo0608@[Link], kshong@[Link]

Abstract
In this paper, we propose and implement a leaf recognition system using the leaf vein and
shape that can be used for plant classification. The proposed approach uses major main vein
and frequency domain data by using Fast Fourier Transform (hereinafter, FFT) methods with
distance between contour and centroid on the detected leaf image. Total 21 leaf features were
extracted for the leaf recognition, which they include ① the distance feature between centroid
and all points on the leaf contour, ② frequency domain data by FFT that was performed
using the distances. In summary, 10 features of all the 21 leaf features were extracted using
distance, FFT magnitude, and phase, the other 10 features were extracted using t the digital
morphological features using four basic geometric features and five vein features, and the
last 1 feature was extracted using the convex hull. To verify the validity of the approach,
images of 1907 leaves apply to classify 32 kinds of plants. In the experimental results, the
proposed leaf recognition system showed an average recognition rate of 97.19%, and we can
confirm that the recognition rate of the proposed leaf recognition system was better than that
of the existed leaf recognition method.

Keywords: Intelligent Scissor (IS) algorithm, Leaf Contour Extraction, Edge Point, color
gradient magnitude

1. Introduction
Approximately 350,000 species of plants exist on earth, and they share a very close
relationship to human beings. Plants play a major role in various areas, such as food, medical
science, industry, and the environment. However, many species of plants are endangered
because of environmental pollution due to the rapid development of human society. Therefore,
it is very important to study automatic plant classification and recognition for plant protection.
Leaf recognition technology plays an important role in plant classification and its key issue
lies in whether selected features are stable and have good ability to discriminate different
kinds of leaves.
Many recent studies exist on plant classification and recognition based on plant
components such as flowers, leaves, and barks. To handle such volumes of information,
realization of a quick and efficient classification method has become an area of active study
[1-9]. In particular, it is well known that the correct way to extract plant features involves
plant recognition based on leaf images. Two features, which are widely used for plant
recognition based on leaf image, are color and shape [1], [6-9]. In the color-based
conventional study, a simple color similarity between two images can be measured by
comparing their color histogram. Also in the shape based-conventional study, they used
region and contour-based simple features and features could be considered time domain data.

57
International Journal of Bio-Science and Bio-Technology
Vol. 5, No. 2, April, 2013

However, the recognition performance was limited due to leaf color was affected by the
seasons and there is a problem that user to directly specify both ends of the leaves.
Consequently, in this paper, we propose a leaf recognition system for plant classification
based on the leaf vein and shape. One of the notable points is that we can decide leaf direction
using leaf vein, and then use frequency domain data by using FFT on distance between
contour and centroid in the given leaf image. Figure 1 shows the flowchart for the proposed
leaf recognition system.

Figure 1. Flowchart for the proposed leaf recognition system

The remainder of this paper is organized as follows: Section 2 describes leaf contour
extraction from the input image, and we describe main vein extraction and leaf
direction decision method in Section 3. Section 4 describes the extraction of the 21 leaf
features, and Section 5 describes the proposed leaf recognition system. The
experimental results are presented in Section 6, and concluded in Section 7.

2. Leaf contour extraction

We describe on the leaf contour extraction in the given leaf image. Most leaves have
generally green color, while the color of leaves is changed by season or environmental
factors. The color change of leaf image can cause decline of recognition performance or
non-recognized problem. The color converting process on input image is the first step
for leaf contour extraction, and it can set foundation to improve recognition
performance irrelevant to the leaf color change.
Therefore, we convert the input color leaf image to gray scale image as follows:
Gray  0.299 * R  0.587 * G  0.114 * B (1)

58
 International Journal of Bio-Science and Bio-Technology
Vol. 5, No. 2, April, 2013

The converted gray scale leaf image is converted to a binary image once again. The
threshold conversion is performed as follows:

0 if f ( x, y) ≤ T
B( x, y)   (2)
255 if f ( x, y )  T

Where, B( x, y) and f ( x, y) are the intensity values of the gray scale image and the
binary image, respectively, at position ( x, y) , and T is the threshold value [10]. Figure
2 shows an example of leaf contour extraction.

Figure 2. Example of leaf contour extraction: (a) input image, (b) gray scale
image, (c) binary image, (d) extracted leaf contour

3. Leaf direction decision using main vein

3.1. Leaf vein extraction
We convert the input leaf image to a gray scale image and then perform opening operations
[1]. In other words, erosion operations are performed after dilation operations are performed.
We are obtained difference image of gray scale image and image of performed opening
operation. Then, we are obtained leaf vein image by convert the difference image to binary
image. Figure 3 shows an example of leaf vein extraction.

Figure 3. Example of leaf vein extraction: (a) input image, (b) gray scale image,
(c) performed opening operations, (d) difference image of b and c, (e) leaf vein
image

3.2. Main vein extraction and leaf direction decision

We describe on the method of main vein extraction and leaf direction decision from leaf
vein image, using projection histogram in the horizontal and vertical directions, in order to
measure the distribution of the leaf vein.

59
International Journal of Bio-Science and Bio-Technology
Vol. 5, No. 2, April, 2013

In the first step, we extract main vein of leaf using projections in the horizontal. We
decide the main vein to the point of maximum of the histogram while leaf vein image rotated
180 degrees. After extracting the main vein, we decided the direction of the leaf through the
projections in the vertical direction.
Figure 4 shows an example of main vein extraction and leaf direction determined. In
Figure 4(c), when the image is divided in half, we can confirm that the histogram
distribution of the leafstalk region is larger than leaf apex region.

Figure 4. Example of main vein extraction and leaf direction decision: (a) leaf
vein image, (b) projections in the horizontal direction, (c) projection in the
vertical direction

4. Leaf feature extraction

Total 21 leaf features were extracted for the leaf recognition, which they include ① the
distance feature between centroid and all points on the leaf contour, ② frequency domain
data by FFT that was performed using the distances. In summary, 10 features of all the 21 leaf
features were extracted using distance, FFT magnitude, and phase [2], the other 10 features
were extracted using t the digital morphological features using four basic geometric features
and five vein features [1], and the last 1 feature was extracted using the convex hull.

4.1. Leaf feature (1): Distance and FFT-based frequency domain analysis
The centroid of the detected leaf region was found as follows:

N N

 y )
1 1
C ( x, y )  C ( xn , n (3)
N n 1
N n 1

Where, C ( x, y) is the centroid coordinate of leaf region image and N is the number
of pixels on the detected leaf region.

60
 International Journal of Bio-Science and Bio-Technology
Vol. 5, No. 2, April, 2013

The distance is calculated by measuring the centroid of the leaf region to all points
on the leaf contour as follows:

2 2
D(i )  C x  E (i )x  C y  E (i )y (4)

Where, D(i) is the distance between the centroid of the leaf region and the i th leaf
contour pixel. C x , C y are the coordinates of the centroid of the leaf region, and, E (i) x ,
th
E (i) y are the coordinates of i leaf contour pixel.

FFT is then performed using calculated distance values. The distance is acquired by
measuring the longest distance point from the centroid in a clockwise direction. 10
features were then extracted based on the distance, FFT magnitude, and phase. The 10
features are as follows: average of the distance, standard deviation of the distance, Zero
Crossing Rate (ZCR) of the distance from the average of the distance, average of the
FFT magnitude, standard deviation of the FFT magnitude, number of peaks higher than
the average of the FFT magnitude, the priority of the top ten peaks of the FFT
magnitude, average of the FFT phase, standard deviation of the FFT phase, and ZCR of
the FFT phase from the average of the FFT phase.

4.2. Leaf feature (2): Geometric and digital morphological features

We describe geometric and digital morphological features in order to leaf feature
extraction.
We extract four basic geometric features as leaf length, leaf width, leaf area, leaf
perimeter. The leaf length is defined as the longest distance between the centroid and the two
ends on the margin of the leaf on opposite sides of the centroid. It is denoted by LL . The leaf
width is defined as the distance between the intersection point with LL at the centroid and its
opposite side on the margin of the leaf. It is denoted by LW . Figure 5 shows the procedure
for obtaining LL and LW . The leaf area is the number of pixels in the leaf region. It is
denoted by LA . The leaf perimeter is the number of pixels in the leaf contour. It is denoted by
LP .

Figure 5. Procedure for Obtaining Leaf Length and Leaf Width: (a) leaf length,
(b) leaf width

We extract ten features based on digital morphological features using four basic geometric
features and the study conducted by S. G. Wu, et al., [1] as aspect ratio, form factor,
rectangularity, perimeter ratio of the leaf length, perimeter ratio of the leaf length and leaf
width, and five vein features. The aspect ratio is calculated using the leaf length LL and leaf
width LW . It is defined as LL/ LW . The form factor is used to describe the difference

61
International Journal of Bio-Science and Bio-Technology
Vol. 5, No. 2, April, 2013

between a leaf and a circle. It is defined as 4LA / LP2 , where LA is the leaf area and LP is
the perimeter of the leaf margin. The rectangularity describes the similarity between a leaf
and a rectangle. It is defined as LL  LW / LA , where LL is the leaf length, LW is the leaf
width and LA is the leaf area. The ratio of perimeter to leaf length, representing the ratio of
the leaf perimeter LP and leaf length LL , is calculated by LL/ LP . The perimeter ratio of the
leaf length and leaf width is defined as the ratio of the leaf perimeter LP and the sum of the
leaf length LL and leaf width LW , thus LP /(LL  LW ) . The Vein features have been
extracted using the methods proposed by S. G. Wu, et al., [1] and the morphological openings
on the gray scale images [11]. The five features are as follows: Lv1/ A , Lv 2 / A , Lv 3 / A ,
Lv 4 / A , Lv 4 / Lv1 .

4.3. Leaf feature (3): Convex hull

In mathematics, the convex hull of a set X of points in the Euclidean plane or Euclidean
space is the smallest convex set that contains X . For instance, when X is a bounded subset
of the plane, the convex hull may be visualized as the shape formed by a rubber band
stretched around X [12]. In this paper, we use convex hull in order to reflect the complexity
of the leaf contour. We compute the rate of extent of original image and convex hull image in
order to extract leaf feature. Figure 6 shows an example of original image and convex hull
image for extract leaf feature.

Figure 6. Example of convex hull image: (a) input image, (b) convex hull image.

5. Leaf recognition system

In this paper, we go through 20 steps in the plant classification process, and 3 steps in the
leaf recognition process, using the extracted 21 leaf features. When only the plant of 1 species
is classified after the 20 steps of the plant classification process, the recognition results are
shown.
Three steps in the leaf recognition process using 10 peaks of the FFT magnitude
procedure is as follows: ① The priority of the top 10 peaks of the FFT magnitude, ② If
priorities of the top 10 peaks have the same score, we count the number of peaks with the
same number, ③ If the same number also exists, we count the number of peaks with
matching numbers and position. The score was calculated using the priority of the top 10
peaks. An example is shown in Figure 7, and Table 1 shows an example of the leaf
classification and recognition process.

62
 International Journal of Bio-Science and Bio-Technology
Vol. 5, No. 2, April, 2013

Figure 7. Example of how the score is calculated

Table 1. Example of the leaf classification and recognition process

Input : 2 Corresponding image
Step 1 2 3 6 13 15 19 22 25 29 32
Step 2 2 6 13 15 25 29 32
Step 3 2 6 13 15 25 32
Step 4 2 6 13 15 25 32
Step 5 2 6 15 25 32
Step 6 2 6 15 25 32
Step 7 2 6 15 25 32
Step 8 2 6 15 25 32
Step 9 2 15 25 32
Step 10 2 15 25 32
Step 11 2 15 25
Step 12 2 15 25
Step 13 2 15 25
Step 14 2 15 25
Step 15 2 15 25
Step 16 2 15 25
Step 17 2 15 25
Step 18 2 15 25
Step 19 2 15 25
Step 20 2 15 25
Step 21 2 : 86, 15 : 79, 25 : 86
Step 22 2 : 9, 25 : 9
Step 23 2 : 7, 25 : 5
Recognition Result : 2

As shown in Figure 7, we give 10 points when match number and position, and we give 9
points when 1 space difference of number and position of priority of the top 10 peaks of the
FFT magnitude of input image and recognition models. The total score is calculated by
summing the points of each position.
In Table 1, step 1~20 are plant classification process, and step 21~23 are leaf recognition
process, and corresponding image shows classification and recognition result of each steps.

63
International Journal of Bio-Science and Bio-Technology
Vol. 5, No. 2, April, 2013

6. Experiments and results

In this paper, we used 1907 leaf images of 32 species collected by Wu et al. [1]. Each plant
species has a minimum of 50 to a maximum of 77 sample leaves.
The proposed system was implemented using Microsoft Visual C++ 6.0 and the Intel
OpenCV library. Because the leaf image size and position of the dataset is not constant,
we normalized to the leaf image.
To evaluate the performance of the proposed leaf recognition system, a recognition
model was created using a range of values with twenty-one features for each plant
species. To each kind of plant, 10 pieces of leaves from testing sets are used to test.
Table 2 shows the experimental results for the proposed leaf recognition system, and
Figure 8 shows the example of leaf recognition system. The average recognition
accuracy of proposed system is 97.19%.

Table 2. Experimental result for proposed leaf recognition system

Number of Incorrect
Recognition system Number of Leaf Image Recognition rate
recognition
Existing system [1] 1800 31 90.31%
Existing system [2] 1907 15 95.31%
Proposed system 1907 9 97.19%

Figure 8. Example of leaf recognition system

7. Conclusions
In this paper, we proposed and implemented of leaf recognition system based on the leaf
vein and shape for plant classification. We extract main vein from the input image, and leaf
direction is determined using projection histograms of extracted main vein image. We were
extracted twenty-one leaf features as distance, FFT, and convex hull for the leaf recognition.

64
 International Journal of Bio-Science and Bio-Technology
Vol. 5, No. 2, April, 2013

In the experimental results, the proposed leaf recognition system showed a performance of
97.19%. From the experimental results, we can confirm that the recognition rate of the
proposed advanced leaf recognition system was better than that of the existing leaf
recognition system.
In future work, we improve the proposed system and further improve its recognition
performance. In addition, we are continuing to research to find a correct leaf contour
extraction method in the complex background.

Acknowledgements
This research was supported by MKE, Korea under ITRC NIPA-2012-(H0301-12-3001)
and PRCP through NRF of Korea, funded by MEST (2012-0005861).

References
[1] S. G. F. Wu, S. Bao, E. Y. Y. Xu, X. Y. Wang, F. Chang and Q. L. Xiang, “A Leaf Recognition Algorithm
for Plant Classification using Probabilistic Neural Network”, The 7th IEEE International Symposium on
Signal Processing and Information Technology, (2007) Cairo, Egypt.
[2] K. B. Lee and K. S. Hong, “Advanced Leaf Recognition based on Leaf Contour and Centroid for Plant
Classification”, The 2012 International Conference on Information Science and Technology, (2012) Shanghai,
China.
[3] J. H. Kim, R. G. Huang, S. H. Jin and K. S. Hong, “Mobile-based flower recognition system”, 2009 Third
International Symposium on Intelligent Information Technology Application , (2009), NanChang, China.
[4] T. H. Hsu, C. H. Lee and L. H. Chen, “An interactive flower image recognition system”, Multimed. Tools
Appl., vol. 53, no. 1, (2011).
[5] S. J. Kim, B. W. Kim and D. P. Kim, “Tree Recognition for Landscape Using by Combination of Features of
its Leaf, Flower and Bark”, The SICE Annual Conference 2012, (2011) Tokyo, Japan.
[6] A. W. M. Smeulders, M. Worring, S. Santini, A. Gupta and R. Jain, “Content-based image retrieval at the end
of the early years”, IEEE Trans. Pattern Anal. Mach., vol. 22, (2000).
[7] X. Wang, D. Huang, J. Dua, H. Xu and L. Heutte, “Classification of plant leaf images with complicated
background”, Appl. Math. Comput., vol. 205, (2008).
[8] J. Du, D. Huang, X. Wang and X. Gu, “Computer-aided plant species identification (CAPSI) based on leaf
shape matching technique”, Trans. Inst. Meas. Control, vol. 23, (2006).
[9] S. Zhang and Y. K. Lei, “Modified locally linear discriminant embedding for plant leaf recognition”,
Neurocomputing, vol. 74, (2011).
[10] R. C. Gonzalez and R. E. Woods, “Digital Image Processing”, Addison Wesley, Boston, (1992).
[11] R. C. Gonzalez, R. E. Woods and S. L. Eddins, “Digital Image Processing Using MATLAB”, Prentice Hall,
New Jersey, (2004).
[12] M. Berg, O. Cheong, M. Kreveld and M. Overmars, “Computational Geometry: Algorithms and
Applications”, Springer-Verlag Press, New York, (2000).

Authors

Kue-Bum Lee received his B. S. in computer science from the Seoul

Health College, Sungnam, Korea, in 2006, and M.S. in electronic and
electrical engineering from Sungkyunkwan University, Seoul, Korea, in
2008, respectively. He is presently a Ph.D candidate at the department of
Information and Communication Engineering Sungkyunkwan University.

65
International Journal of Bio-Science and Bio-Technology
Vol. 5, No. 2, April, 2013

Kwang-Seok Hong received his B.S., M.S., and Ph.D. in Electronic

Engineering from Sungkyunkwan University, Seoul, Korea in 1985, 1988,
and 1992, respectively. Since March 1995, he has been a professor at
Sungkyunkwan University, Suwon, Korea. His current research focuses
on human-computer interaction, five-sense recognition, interaction, and
representation.

66