| Abstract: |  The topic of this work is the recognition of natural objects from
images. Thanks to the recent development of mobile devices it is easy to
take a picture of a natural object (e.g. plant, animal, fungi) but its
identification is difficult and it might require expert knowledge even
with a proper identification key. In this work, we address the problem
of identification of plants from images of their leaves and bark. Unlike
moving animals, they are easy and safe to photograph, they can be found
everywhere and identification of a significant number of species does
not require any special equipment like a microscope or a DNA sequencer.
To perform recognition from images of leaves, we make use of the Inner
Distance Shape Context (Ling and Jacobs [10]) and recognition from
images of bark utilises Multi-Block Local Binary Patterns (Liao et al.
[44]). Both methods are efficient enough to be implemented as an
application for a mobile phone. Recognition of one leaf does not take
more than 2 seconds with a database with 954 items and one image of bark
can be classified in 3 seconds with a database with 543 items.
Classification accuracy was measured on two datasets: The Flavia dataset
(leaves [19]) and the Österreichische Bundesforste AG dataset (bark
[22]). On the Flavia dataset, top one, two and three candidates included
correct class in 83.3%, 91.0% and 94.3% of cases respectively. On the
Österreichische Bundesforste AG dataset, top one, two and three
candidates included correct class in 70.1%, 87.8% and 93.9% of cases
respectively.
|
|---|