Shape and Texture Based Plant Leaf Classification - Advanced Concepts for Intelligent Vision Systems - page 348

points in the contour distance signature. Each point in the signature is compared
to the 3 points either side of it. If the point is either less than all these neighbours,
or greater than them, then the point is an inflection point. Once every inflection
point has been detected, they are counted and if the total number is above some
threshold, the leaf is considered lobed.
Using this method, serrated leaves would be identified as having many lobes.
To prevent this, the signature is first smoothed by using a Gaussian filter. The
difference between a lobed and a serrated leaf, as well as their contour graphs
(normal and smoothed), can be observed in figure 1. The normal graph would
give a lot of inflection points for these two leaves and would classify both in the
lobed category although only the first one actually is.

3.2 Results

The results of the contour signature method can be seen in table 1. All the leaves
in the dataset were compared to all others, and classified as the same species
as the closest other leaf. The overall correct classification rate is 69.2%. Whilst
some of the species achieved a high recognition rate (with 3 at 100%), many did
much worse, with 6 under 50%. Part of reason for this is the high intra-species
variation present within some lobed species, and the low inter-species variation
between species with ovate leaves. Another cause of errors appears when leaves
have overlapping regions, which cause the contours to be incorrectly traced, as
shown in figure 2a. Figure 2b shows that petiole (stems) cut that different lengths
before imaging the leaves can also cause problems.

4 Texture Analysis Using Sobel

The results for the contour signatures suggest that leaves cannot be adequately
classified based on shape alone. The texture is also an important feature of the
leaf. Two types of texture can be defined: the micro-texture at the microscopic
scale and the macro-texture which is the pattern formed by the venation of the
leaf. The venation is specific to every leaf, similar to a fingerprint. In this chapter,
the concept of macro-texture is quantified using edge gradients.

4.1 Histogram of the Gradient Intensity

For each image, we calculate a histogram of the gradient orientations, whereby
for the angle θ :

h ( θ )=
x

M ( x, y )if Θ ( x, y )= θ, 0 otherwise (5)

y

Where M ( x, y ) is the gradient magnitude at pixel ( x, y )and Θ ( x, y ) is the gra-
dient direction, calculated using the Sobel operator. This histogram provides
a description of the relative directions of the main veins. Examples of these
histograms for four leaves from the species Quercus Ilex can be seen in figure 3.