Shape and Texture Based Plant Leaf Classification - Advanced Concepts for Intelligent Vision Systems - page 345

The texture-based method is presented in Section 4. The most common tech-
niques of texture description are, in general, based on the statistical analysis of
the pixels (co-occurence matrices, etc.) [8,21,5,17], and their spectral analysis
(Fourier Transform, Wavelet Transform, Gabor filters, etc.)[20,11,4,22,12,7,1].
Although there are numerous techniques for texture classification, few of them
have been applied to leaves [6,10,18,2]. The technique implemented by the au-
thors makes use of the Sobel operator to analyse the macro-texture of the leaf.
Finally, Section 5, will present an incremental algorithm used to combine the
results of the previous methods using probability density functions.

2 Data Pre-processing

The leaves used in this work were collected in the Royal Botanic Gardens, Kew,
UK. The dataset contains 3 to 10 leaves from each of 18 different species.
As the colour of the leaves cannot be used as reliable information, since it
varies depending on the period of the year as well as other factors, the data has
been transformed into greyscale images. The image background, the paper on
which the leaf is mounted, is removed using Otsu's thresholding method [16].

3 Analysis of the Contour Signature

Two contour signatures are calculated for analysing leaf shapes. For each leaf,
first the outline is extracted by selecting from the image the foreground pixels
which neighbour a background pixel on at least one of their four main sides
(N,S,E,W). Moving in a clockwise direction, for every n

th contour pixel, where
l is the length of the outline and n is the number of points to be sampled, two
values, f ( i )and g ( i ) are calculated:
f ( i )= ( cont x ( j ) cent x ) 2 +( cont y ( j ) cent y ) 2 − − (1)

tan( cont x ( j )

− cent x 2 iπ

n |

g ( i )= | ) − (2)

cont y ( j ) − cent y

Where, j = i×l
n

, cont x ( j ), cont y ( j )arethe x and y co-ordinates respectively for
the j th contour pixel, and cent x , cent y are the x and y co-ordinates of the leaf's
centroid.
The first of the resulting signatures f , gives the distances between the contour
point and the centre of the leaf. The second, g , is the absolute difference between
the angle at the leaf centre between the starting point and the current point,
and the corresponding angle on a circle. Together, these two signatures provide
a significant amount of information about the leaf's shape.
These signatures are treated like probability density functions (pdfs) by divid-
ing each value by the sum of all the values in the signature. Doing this provides
us with scale-invariance. The difference between the signatures for two leaves