Identification of Grape Varieties via Digital Leaf Image ... - Oiv2010.ge

Identification of Grape Varieties via Digital Leaf Image Processing by Computer
1 J. ZHANG, 2 P. YANNE * and 3 H. LI
1,2 College of Information Engineering, Northwest A&F University
Yangling, Shaanxi, China
pyanne@nwsuaf.edu.cn (corresponding author) *
3 College of Oenology, Northwest A&F University
lihuawine@nwsuaf.edu.cn
ABSTRACT
Grape variety identification is of great significance for resource statistics, new specie
detection and protection of genetic resources. Based on the classical ampelographic grape
identification method combined with machine learning and pattern recognition techniques in
computer science, we proposed a new cheap and fast identification method via leaf image
processing. We demonstrated its feasibility via the implementation of a prototype which could
classify 354 leaf images belonging to 20 varieties with an accuracy rate of 87%. Our
techniques can be applied to computer aided diagnosis of grape leaf diseases and new variety
discovery, as well as to quantify the classical ampelographic identification method. We
propose further work to transform this new method from a prototype into a practical software
product.
Keywords: Grape variety identification; Ampelography; Pattern recognition; Image
processing; Hu's moment invariants
RESUME
L‟identification des cépages est très importante pour les statistiques de ressources, la
détection des nouvelles espèces et la protection des ressources génétiques. Basés sur la
méthode classique de l‟identification ampélographique et grâce aux récents progrès en
informatique, notamment la reconnaissance des formes et l‟apprentissage automatique, nous
proposons une nouvelle méthode efficace et automatique par ordinateur pour l‟identification
des cépages. Nous démontrons sa faisabilité via la réalisation d‟un prototype qui a pu
classifier 354 fichiers de feuilles, appartenant à20 cépages avec une précision de 87%. Notre
technique peut s‟appliquer au diagnostic des maladies de vigne, à la découverte de nouvelles
espèces et àla quantification de la méthode classique de l‟identification ampélographique.
Nous suggérons des directions de recherche future afin de transformer notre prototype en
produit logiciel.

I. Introduction
There are over 10,000 grape varieties throughout the world. About 3000 of them are widely
cultivated in production and many are wine varieties [Zhai, 2001]. Grape variety
identification is of great significance for resource statistics, new specie detection and
protection of genetic resources. OIV‟s Strategic Framework includes the task for recognising
new viticultural varieties [OIV, 2005].
The classical identification method is based on ampelography [Galet,1990; Tassie and
Blieschke, 2008]. Some new methods have been also developed recently, using different
approaches such as DNA molecular genetic marker [Bower et al., 1993; Zhang et al., 1996;
Testier et al., 1999], pollen morphology [Wang and Li, 2000], anthocyanin analysis
[Wendelin and Barna, 1994], etc. All these methods need expert intervention and are hence
quite expensive. Some of them need special devices and take a long time. Today, computer
technologies have a wide range of applications in many fields including grape production.
There are many successful examples where the computer has been used for image processing
[Li et al., 2007; Barbu, 2009] and identification of plant species [Ye et al., 2004] based on
pattern recognition. We look for a new method for identifying grape varieties combining the
computer techniques and the classical ampelography. Based on the processing of digital grape
leaf image, this new method would be rapid, efficient and nearly automatic with little or even
no human intervention. Our research objective is to develop a software product, available on
web, which will be able to tell a browser the variety of the grape leaf image that s/he uploads.
The ampelographic identification of grape varieties is based on the observation of features
on some organs of a grape, such as flower, berry, shoot and leaf. OIV has produced 2 editions
[OIV, 1983; OIV, 2009] of the document “OIV descriptor list for grape varieties and Vitis
species” which defines as a standard the ampelographic characteristics for the identification of
Vitis varieties and species. Using the 128 characteristics selected by [OIV, 1983] where each
characteristics is signed a code and may take values from 1 to 9 for all grapes, [OIV, 2000]
describes 250 wine grape varieties of its member states, by assigning a values to descriptor
codes for each variety . For a given grape sample, if each its code has the same value as the
variety V of the 250 in [OIV, 2000], this grape‟s variety is classified as V. All ampelographic
experts agreed that the features of mature leaf are the most determinate for the varieties
identification. For the 128 codes, 35 of them are for leaf and 29 for mature leaf. [OIV, 2009]
adds another 18 codes from 601 to 618 on mature leaf. On the “Primary descriptor priority list”
of 14 codes, there are 9 on leaf.
In our new approach based, the main idea is to let computer calculate all the code values
instead of measuring them by a human being. Then the computer can compare these values
against the known ones as in [OIV, 2000] to find the right variety. However, on one hand, it‟s
not easy to calculate some code values and on the other hand, it is not necessary to know all
these values for the identification purpose. Furthermore, some features not selected by [OIV,
2009] may also contribute to distinguish or identify varieties, for example, Hu's moment
invariants for an image [Hu, 1962].

A digital image is composed of a pixel f(x, y) matrix where (x, y) is the index or coordinator
of the matrix. Each pixel f(x, y) represents an image dot and is described by a series of
numbers. For a binary image like a photo in an old news paper, a pixel f(x, y) is either 0 for
white or 1 for black. For a colour image taken by a digital camera, a pixel may be a
combination of three basic colours with different densities. Hu defined 7 moment invariants
for any digital image. Each invariant can be easily calculated as the function of its pixels f(x,
y). The 7 invariants‟ values are nearly independent of the rotation, position or size of the
image of the matrix. They have been successfully used in computer pattern recognition
applications such as car registration number [Liu and Lu, 2008], static hand gesture [Liu et al.,
2008], tiger variety [Xu and Qi, 2009], human face [Gan and Zhang, 2002] and corn leaf
disease recognition [Shen et al., 2008]. Yanhua YE and Chun CHEN of Hong Kong
Polytechnic University have developed a Computer Plant Species Recognition System,
CPSRS [Ye et al., 2004] which could provide a convenient and efficient way to search and
identify plant species from a digital image file.
Departing from the works mentioned above, which consist of the cornerstone of our method,
we present our method in detail and experiment it by the implementing a software prototype
on an ordinary personal computer. We then analyze our experiment results and discuss on
some choices that have been made, the remaining problems and possible improvements as
well as applications. We conclude on the feasibility of our new method and point out the
future work.
II. Materials and Methods
Our identification method is constructed on 4 steps: 1) collect typical mature grape sample
leaves for the varieties we want to identify, 2) scan the leaves into digital image files, 3) select
a set of characteristics or features useful for identification and computable by computer from
the images, 4) build a software classifier based on the features calculated from the sample
files.
1) Collect mature sample leaves
Following the requirements of OIV [OIV, 2009], for each variety, we collected about 10
mature leaves from different shoots at the third middle level, between berry set and veraison
time. These leaves were collected from the grape variety culture field of College of Oenology,
Northwest A&F University in Yangling, Shaanxi, China. There are a total of 500 leaves
belonging to 3 wild local varieties and 47 cultured ones including Sauvignon, Riesling,
Traminer, Sémillon, Chenin Blanc, Ugni Blanc, Müller-Thurgau, Cabernet Sauvignon,
Carignan, Gamay, Syrah, Muscat, etc.
2) Obtain digital leaf files
For each leaf, we scanned both leaf sides with the default parameters of 3 A4 size ordinary
scanners. We got a total of 1073 colour leaf image files at the resolution of 300 DPI (Dot Per
Inch). In our software prototype, we used 354 leaf files of 20 varieties.

3) Select features for identification
Naturally, the 47 features on mature leaves, coded by OIV [OIV, 2009] have been
considered. More researches have to be done for calculating some features, e.g. “density of
prostrate hairs between the main veins on lower side of blade”, OIV code 84. We have found
a way to calculate some of them, including size and circumference of blade, length of petiole,
length of veins, etc. We select also some features, neither considered by OIV nor
ampelography, which are easy to calculate and useful for identification, e.g. Hu‟s 7 moment
invariants. In order to quickly build our prototype, with the criteria of both computable and
useful, we finally selected the size and circumference of blade and Hu‟s 7 moment invariants
to form the feature set or vector of 9 dimensions.
4) Build a software classifier
Let‟s explain the mathematical basis of our method. Each leaf image is represented by a
feature vector Lj= (fj1, … ,fj9) where fji is a real number. Such vector Lj is a point in the 9
dimension feature space in mathematics. We imagine the Euclidean distance D L 1 , L 2 =
f 11 − f 12 2 + ⋯ + f 19 − f 29 2 between 2 grape leaves L1 and L2 of the same variety should
be in average smaller than that of 2 different varieties. For the variety i, its mass centre
Ci=( ci1, …, ci9) Where cim=
n
f jm
j=1 , fjm is the m-th feature of the j-th leaf sample Lji of the
n
variety Vi, and its radius R i =maximum of D(Ci, Lji) for j=1 to n. For a given leaf j‟s vector
Lj, if we only find one variety S which can satisfy D(Lj, C S ) ≤ R S , we can conclude that the
leaf Lj belongs to the variety S.
Unfortunately, for the 354 vectors of 20 varieties, their mass centres are so close and their
radiuses are so big that the 9 dimension sphere of a variety S i at centre C i with radius R i
has intersection with the spheres of other varieties. To reduce the space occupied by each
variety, we improve the above method by detecting the 9 dimension cube which inscribed the
sphere. This can be done by finding the value range of the vector‟s each dimension for every
variety. Some of leaves still cannot be distinguished. For this case, based on the fact that the
values of each dimension for each variety should satisfy the normal distribution, we introduce
the probability of a leaf L belonging to a variety S by the following formula:
P L, S = 1 −
9 2∗dL j
j=1 ∗ 1 , dL j = L j − 1 r S,j 9 2
∗ r(S, j) , r(S, j)=max(fsj)–min(fsj)
where max/min(fsj) means the maximum/minimum value of j-th dimension for all samples
of the variety S.
Finally, we build our classifier with the following algorithm:

1) Find the vector Li of a leaf image i and compare Li (fij) with the value range max/min(fsj)
of all varieties S (S=1 to 20 and j= 1 to 9).
2) If max(fsj)>=fij>=min(fsj) holds for only one variety S with j=1 to 9, then the leaf Li
belongs to the variety S.
3) Else, leaf Li satisfies the relation in step 2) for m varieties S 1, ⋯ , S m, . We find out all the
probabilities P(L i, , S j ). Li belongs to the variety S j with the probability P(L i, S j ) which
is the maximum of P L i, S j for j=1 to m.
III. Results and Discussion
We developed a software prototype in Matlab implementing the above algorithm to verify
our method. The following figure 1 shows an execution of our prototype under Matlab
environment. The user selects a leaf image and then asks for the classification. The prototype
displays the image in 3 modes and prints out the variety name:
Figure 1. Execution of classification prototype
We have tested our algorithm to classify the 354 images files belonging to 20 varieties. The
correct classification rate is of 87%. This rate is obtained by calling the classifier on all the
354 files and count the present of corrected classified ones.

The accuracy decreases when the number of varieties increases. We may resolve this
problem by increasing the feature vector‟s dimensions, i.e. to find more features, and use
better classification algorithm such as SVM [Wu et al., 2008]. The latter is a well known
machine learning method for classification. It is in fact the capacity of computer to learn from
examples. After having trained it by giving many leaf samples belonging to each grape
variety, the software can decide to which variety a new leaf belongs to.
Our prototype can only classify scanned images. In order to classify digital camera images,
we have to consider factors such as photography distance and focus. On the other hand, a
camera may take photos from different angles. This may allow us to distinguish the prostrate
hairs of a leaf from the erect hairs. However, this problem can be better resolved by a 3
dimensions camera or scanner. Based on the 3D image technique, we can more easily
calculate other leaf features such as the profile of leaf (OIV code 74 [OIV, 2009]). Light wave
lengths other than visible ones, such as infrared, microwave and terahertz [Lu, 2002; Xing
and Baerdemaeker, 2005] etc. can also be used to obtain digital leaf images. These images
should supply complementary features, useful for the variety identification. By combining
these mentioned techniques, we are expected to be able to calculate all the 45 codes selected
by OIV and hence to identify all grape varieties based on digital image processing by
computer.
This new identification method may not only simplify the identification procedure, but its
techniques can also improve the classical ampelographic identification method. The current
OIV Descriptor List [OIV, 2009] uses the code values in a qualitative way. For example, the
code 65 for the size of blade takes values 1, 3, 5, 7 and 9 which means respectively, very
small, small, medium, large and very large. Our method can calculate the size of blade in
inch 2 or cm 2 effectively and automatically. We may do this for all quantifiable codes of all
known varieties. With these quantitive values, we may give a value range for each current
qualitative value on one hand, and check if the code values for the 250 varieties in [OIV, 2000]
are coherent. These techniques can be used to detect new variety and guess the parent
varieties of a new hybrid variety. In fact, the feature vector of a new variety will not belong to
any known varieties, but a hybrid one should be close to its parents‟ ones. Computers
software can easily find out all the similar varieties and sort them according to the similitude.
IV. Conclusions
Based on the classical ampelographic grape identification method combined with machine
learning and pattern recognition techniques in computer science, we proposed a new cheap
and fast identification method via leaf image processing. We demonstrated its feasibility by
implementing a software prototype which could classify 354 leaf images belonging to 20
varieties with an accuracy of 87%.
We are continuing our research to increase both the number of grape varieties and the
accuracy by calculating more features from digital images and improving the classification
algorithms.

Acknowledgments
We would first express our gratitude to Dr Jean-Claude Ruf, head of OIV‟s vitiviniculture
science and techniques department for his advices during his visit to our university at the
occasion of the 6th International Symposium on Viticulture and Enology, in Yangling,
Shaanxi, China. Mrs A. Tsioli, head of OIV‟s viticulture unity, supplies us with a lot of useful
information. The whole research team for the project of grape variety identification in our
university contributed to this work, especially, Dr JF NING for his suggestion of adopting
Hu's moment invariants, Dr C CAI as well as his students for the leaf sample collecting and
image scanning and Dr Y ZHANG for his encouragement. At last, we would thank the
students of our team, ZG FENG, WJ HAN, Z SONG, H ZHANG and Y ZHANG.
Bibliography
Barbu T., 2009. Content-based image retrieval using Gabor filtering. In: Proceedings - 20th
International Workshop on Database and Expert Systems Applications. New York: Institute
of Electrical and Electronics Engineers Inc. 2009: 236-240.
Bower J.E., Bandman E.B., Meredith C.P., 1993. DNA fingerprinting characterization of
some wine grape cultivars. American Society for Enology and Viticulture, 44: 266-271.
Galet P., 1990. French grapevine varieties and vineyards. Volume 2. The French
ampelography. 2nd edition. Montpellier.
Gan J.Y., Zhang Y.W., 2002. Face Recognition Based on Moment Invariants and Neural
Networks. Computer Engineering and Applications, 38(7): 53-57.
Hu M.K., 1962. Visual Pattern Recognition by Moment Invariants. IRE Transaction
Information Theory, 8(2): 179-187.
Li Y., Chi Z.R., Feng D.D., 2007. Leaf vein extraction using independent component analysis.
In: 2006 IEEE International Conference on Systems, Man and Cybernetics. New York:
Institute of Electrical and Electronics Engineers Inc. 5: 3890-3894.
Liu J.M., Lu K., 2008. The identification of car logo based on Hu‟s invariant moments.
scientific and technical information (Academic Edition), 36: 76-81.
Liu Y., Gan Z.J., Sun Y., 2008. Static Hand Gesture Recognition and its Application based on
Support Vector Machines. In: Software Engineering, Artificial Intelligence, Networking,
and Parallel/Distributed Computing. Ninth ACIS International Conference, Phuket:
2008(9): 517-521.
Lu R., 2002. Detection of bruises on apples using near-infrared hyperspectral imaging.
Information & Electrical Technologies Division of ASAE, vol.46(2): 1-8.

OIV, 1983. 1st Edtion of the OIV Descriptor list for grape varieties and Vitis species.
http://www.oiv.int.
OIV, 2000. Description of world vine varieties. http://www.oiv.int.
OIV, 2005. RESOLUTION AG 1/2005, OIV Strategic Framework 2005-2008. .
http://www.oiv.int.
OIV, 2009. 2nd Edtion of the OIV Descriptor list for grape varieties and Vitis species.
http://www.oiv.int.
Shen W.Z., Wu Y., Chen Z.L., Wei H.D., 2008. Grading method of leaf spot disease based on
image processing. In: Proceedings - International Conference on Computer Science and
Software Engineering. NJ: IEEE Computer Society. Vol. (6): 491-494.
Tassie L., Blieschke N., 2008. Ampelography: do you know what variety you are planting in
your vineyard or nursery. Australian & New Zealand Grape grower & Winemaker: national
journal of the grape and wine industry, No.537.
Testier C., Daivd J., This P., Boursiquot J. M., Charrier A., 1999. Optimization of the choice
of molecular markers for varietal identification in Vitis Vinifera L. Theoretical and Applied
Genetics, 98(1): 171-177.
Wang X.D., Li C.L., 2000. A study on pollen morphology of the genus Vitis L. Acta
Phytotaxonomica Sinica, 38(1): 43-52.
Wu D.K., Xie C.Y., Ma C.W., 2008. The SVM classification leafminer-infected leaves based
on fractal dimension. In: 2008 IEEE International Conference on Cybernetics and
Intelligent Systems, CIS 2008. NJ: Inst. of Elec. and Elec. Eng. Computer Society. 2008:
147-151.
Xing J., Baerdemaeker J.D., 2005. Bruise detection on „Jonagold‟ apples using hyperspectral
imaging. Postharvest Biology and Technology, 37(2005): 152-162.
Xu Q.J., Qi D.W., 2009. Parameters for Texture Feature of Panthera tigers altaica Based on
Gray Level Co-occurrence Matrix. Journal of Northeast Forestry University, 37(7):
125-130.
Ye Y.H., Chen C., Li C.T., Fu H., Chi Z.R., 2004. A Computerized Plant Species Recognition
System. In: Proceedings of 2004 International Symposium on Intelligent Multimedia,
Video and Speech Processing. Hong Kong: Institute of Electrical and Electronics
Engineers Inc. 2004: 723-726.
Zhai H., 2001. The main grape varieties used for processing. In: Wine Grape Growing and
Processing Techniques. Yang T.Q., 1st edition. Beijing: China Agriculture Press. 1: 48.
Zhang L.P., Lin B.N., Shen D.X., 1996. The extraction, purification and identification of
RFLP of the chromosomal DNA of Grape. Journal of Fruit Science, 13(2): 71-74.