You are on page 1of 1

AUTOMATED CLASSIFICATION OF ANT DIGITAL IMAGES

USING ARTIFICIAL NEURAL NETWORK


Nurshafira M. N. (ID: 012012110944)
Supervisor: Dr Radha Swathe Priya
Faculty of Health & Life Sciences (FHLS)
Management & Science University (MSU)

INTRODUCTION

METHODOLOGY

Ants are typically sampled ex situ to identify the


specific condition of a habitat. Although this method
of classification is generally reliable, it is inefficient
when it involves a large number of samples. Hence,
it is important to introduce an automated
classification system that is capable for the
recognition and classification of ant miniscule sizedbodies within an optimum time. In this study, an
automated classification system is proposed to
detect different sub-families of ants from their digital
images using machine learning method. This
process also involves pre-processing techniques
such as thresholding filter, contrast enhancement,
reduction of dimensionality of image, and
binarization. At the classification stage, backpropagation artificial neural network classification
method was employed, whereby the ranges of
feature values were input into the system to classify
the ant images based on its subfamilies. The overall
percentage of accuracy is 21.67%, in which 13
out of 60 testing images were predicted
correctly.
In conclusion, a new automated
classification system of ant digital images was
simplemented.

DISCUSSION

IMAGE CLASSIFICATION FRAMEWORK

Table 3

RESULT

Table 1

Figure 2: Raw images after pre-processing


stage.

Figure 3
Example of rejected image(s).
RESEARCH POSTER PRESENTATION DESIGN 2015

www.PosterPresentations.com

Dolichoderinae

Dorylinae

14

Ectatomminae

Formicinae

18

Myrmicinae

11

Ponerinae

13

13

Based on the system testing, only the


Ponerinae subfamily has been accurately
classified (100%).
Hence, the overall percentage of accuracy is
21.67%, in which 13 out of 60 testing images
were predicted correctly.

The automated classification of ant digital images


was successfully implemented. The training and
testing of 235 ant images were performed using
R language, with an overall accuracy of 21.67%.
The future works proposed for this study are:

Figure 2

Observed

CONCLUSION

THE RESULTS FOR THE ANT IMAGE


CLASSIFICATTION USING ARTIFICIAL NEURAL
NETWORK

MATERIAL

Figure 1
Figure 1: Example of the raw images of
Borneo ants from Antweb database.

Expected

Figure 4

OBJECTIVES
General objective:
To develop an automated image classification
system using artificial neural network to identify
ant subfamilies.
Specific objective(s):
1.To equalize the datasets for ant images.
2.To create training and testing field for the
datasets using back propagation algorithm.
3.To classify the datasets into the respective
subfamilies.
4.To measure the error and accuracy for the
classification system.

STUDY SAMPLE IMAGES:

Target

The parameter used in this classifier was set to 0.01


threshold value with 20 hidden layer numbers.
The ideal hidden layer numbers to be used for this
classifier was obtained using trial and error method.
The least error was obtained at 20 hidden layers.

Table 2
Target

Dolichoderinae

Formicinae

Ponerinae

Dolichoderinae

Dorylinae

Ectatomminae

Formicinae

18

Myrmicinae

Ponerinae

13

Net result of classifier:


Kappa statistics:
0.03169014085

Rand index (measure of the


similarity between two data
clustering): 0.397740113

Comparison of ANN and other classifier(s) such


as Random Forest, k-Means Neighboring, SVM.
Feature extraction of ant specific body parts to be
compared and labeled before classification.
Building of R packages for better image
manipulation process.

REFERENCE(S)
1.Izwan, N., & Amirrudin, A. (2014). Diversity of
Ants (Hymenoptera: Formicidae) at Kuala
Lompat, Krau Wildlife Reserve, Pahang,
Malaysia. Journal of Wildlife and Parks, 28, 31
REFERENCES
39.
2.Lai, M. Y. (2006). Classification of Lichen
Species using Artificial Neural Networks.
3.Lantz, B. (2013). Machine learning with R. Packt
Publishing Ltd.
4.Myers, G. (2012). Why bioimage informatics
matters. Nature Methods, 9(7), 65960.
doi:10.1038/nmeth.2024
5.Soda, P. (2014, July). BioImage Informatics: The
challenge of knowledge extraction from biological
images. In Digital Technologies (DT), 2014 10th
International Conference on (pp. 311-320). IEEE.

You might also like