Chapter Overview for Introduction
Plants are considered to be an essential part
of the ecosystem of the planet. Some of them provide edible eatables while some
of them have a medicinal value which can be used to curative purposes. In
ancient days our ancestors were good enough to identify the medicinal plants
which grow in back yards or alongside the roads and they used that knowledge
for curing various diseases. However in modern days people are unlikely to have
herbal medicines mainly due to few factors. Lack of knowledge about medicinal plants,
look for fast recovery and less interested to have traditional treatments due
to their smell, appearance.
According to incomplete statistics, there are
approximately over 3,210 kinds of
species in Sri Lanka and among such a vast range of plants it is never easy to
accurately identify certain plants with medicinal values. The report of
the Regional Meeting of the World Health Organization (2009) has stated that incorrect
identification of medicinal plants is one of the factors that make herbal
remedy unsafe. Owing to the ignorance of the exact identities of plants used in
ayurvedic practice, many exotics are being used mistakenly or as substitutes in
the absence of the plants originally recommended.
As asserted by Liwen and Xiaohua (2012), the
breeding, introduction, cultivation of medicinal plants, research and
development of medicine demand the professionals to spend a lot of time in
identifying the species. The existing recognition systems have a great
dependency on the knowledge accumulation and experimental operating skills of
humans and most of them are time consuming. Therefore, there is a clear need of
fast, reliable and convenient recognition method.
The author strongly believes that Sri
Lanka is gifted with a vast number of medicinal plants. Hence, having a good
knowledge in Herbal plants would be a great asset. In addition, it is
understood that several unfortunate incidents occurred due to lack of awareness
among medical practitioners and the public about Identifying herbal plant from
an adulterant. A most common instance is
that “Ranawara” tree is most often confused with another plant with yellow
flowers. The people who used this as a medication ended up in hospitals due to
the poisonous substance in the leaves. On the other hand, People spend a quite
a lot money on western medicine while they are unaware of the fact that they
could find medicinal solutions easily if they look around in the back yard or
alongside of the roads. Hence developing an intelligent application which could
address these two issues could be a worth and timely effort.
Chapter Overview for Literature Review
In previous chapter, Introduction and
motivation of the project were presented and in this chapter a comprehensive
literature survey on existing flower and leaf recognition systems followed by a
critical analysis is presented. In following analysis, author presents existing
computational models used in plant and flower analyzing systems and a critical
evaluation on each to identify strengths and weaknesses are included.
What are Medicinal Plants
World Health Organization has defined
the medicinal plant as a plant used for medicinal purposes. It is also
described as a plant which, in one or more of its organs, contains substance
that can be used for therapeutic purposes or which can be utilized in producing
useful drugs. Furthermore, the same
species of plants can have different medicinal values if they grow in different
places. The natural habitat, climate, available minerals, presence of other
plants and trees have influence on the medicinal properties of plants.
by Zou & Nagi (2004) unlike humans, the machines are unable to reliably
recognize a flower in several cases such as when images are taken are out of
focus, images contain multiple or overlapping flowers and where there are
complicated backgrounds. Thus, as shown in the following figure, shape features
have been derived using a modified rose curve model with the use of 6
parameters namely the center (x0, y0), the outer radius ro,
the inner radius ri, the number of petals n, and the phase ?.
Figure1: Rose curve Example
Furthermore, petal number and the ro /ri
ratio have been taken as the shape features. The aforementioned model is
used in CAVIAR, which is an interactive flower recognition system which
involves both human and machine interactions. Moreover, the rose curve integral
model acts effectively as the mediator for the interaction between human and
computer (Zou & Nagy, 2004). Therefore, it has a higher accuracy and
efficiency over machine alone and human alone flower recognition.
However, as illustrated in the above
figure, the rose curve is placed over the unknown picture (a). Subsequently
allows the user to change its parameters along the recognition process until
the user herself or himself concludes the recognition task in order to obtain
an accurate result (b). However, Zou & Nagi (2004) have stated that, on
average, 52% of the samples are immediately confirmed. Hence, it implies that
the remaining 48% was obtained as a result of changing the rose curve integral
model parameters by the user. The fact that this system has been user-dependent
could result in negatively impacting the overall accuracy of the system.
Intelligent Scissors (IS) is a
well-known method for object boundary extraction (Saitoh ,2004). A number of
manually selected points are required to draw the boundary. However according
to Saitoh (2004) the limitation is that it requires many manually selected
points particularly for an object with a complicated boundary. Hence, left
picture shown in the following Figure 9 requires about 42 points where the
right picture requires only about 8 points.
Figure2: Input flower images
(Saitoh , 2004)
Principle Component Analysis (PCA)
PCA is a
widely used tool in computer vision where its main objective is to reduce the
dimensionality of data while retaining as much information as possible.
Moreover, this is achieved with the use of projection that maximizes the
variance and conversely minimizes the mean squared reconstruction error. Even
though PCA outperform many other techniques in image processing, there exist
pitfalls such as computational cost and dimensionality. Following are the key
aspects of recognition process of PCA.
Kernel Methods are a class of
algorithms for pattern analysis. According to Thorstensen (2008), Kernel
methods are widely used in the image processing area where it has highly
contributed in pattern recognition. Nevertheless, Kernel PCA is an outstanding
Kernel approach which applies PCA on the mapped training samples.
Color is a vital aspect in some
recognition processes. As stated by (Zhao, 2009), a higher recognition rate
which is 4.4% higher than the traditional PCA approach has been obtained by
incorporating color aspect with PCA. Nevertheless, use of three color
components in extracting pattern has improved recognition accuracy over
traditional PCA. Furthermore, color is one of the main features in recognizing
flower. Therefore, this approach can be followed in the process of recognizing
medicinal flower in order to achieve a higher accuracy level.
The drawback with PCA is it contains
unnecessary information such as poses and lightings and This unnecessary
information affect the recognition accuracy and that makes a conclusion that
PCA is not the optimal rule for discrimination.
In order to process eigenfaces calculation, each two dimensional
(NxN) image in a set of training images, will be converted into a column vector
of dimension (NxN). Subsequently, all columns are joint in order to compute the
average image face vector. This will be followed by obtaining mean face and
co-variance matrix to generate new eigenvectors (Khan & Alizai, 2006).
According to Turk & Pentland (1991), “eigenvectors can be thought of as a
set of features which together characterize the variation between face images.
Each image location contributes more or less to each eigenvector. Thus the
eigenvectors can be displayed as ghostly faces. Nevertheless, aforementioned
vector contains information on each pixel after projecting the image into
When a new image is
given, it is also represented by its vector of weights. Hence, recognition of
the test image is done by locating the image in the database which has the
closest weight to the test image (Zhao et al., 2003). An image vector will
contain w*h pixels of information when width and height contain w and h number
of pixels respectively (Den, 2001).
However, Kirby and Sirovich (1990) have stated that the
eigenvectors of general images are called eigenpictures (cited Szeliski, 2009,
p.623). Therefore, set of eigenpictures which are usually 2D flower
arrangements of light and dark areas are made by combining all the pictures and
analyzing what is common to a group of individual flowers and where those
differ most. Principal component analysis is used to extract these
eigenpictures. Conversely, original images can be represented as the sum of
eigenpictures (Kirby & Sirovich, 1990).