978-1-4799-3351-8/14/$31.00 ©2014 IEEE
A Cortex-like Model For Animal Recognition Based
On Texture Using Feature-Selective Hashing
Sahar Seifzadeh
Department Of Electrical,Computer And Biomedical
Engineering, Qazvin Branch ,Islamic University
,Qazvin,Iran
Seifzadeh.sahar5@yahoo.com
Sahar.seifzadeh1986@gmail.com
Karim Faez
Electrical Engineering Department, AmirKabir University
of Technology Tehran, Iran
kfaez@aut.ac.ir
Abstract— Building a model that can mimic the brain's cortex
has always been a major goal, because the human brain
recognizes objects in terms of speed, reliability and flexibility
that are always unique pattern for machine vision systems. In
this paper, we are inspired by neuroscience and computer
science that have designed a framework that can be fast and
accurate emulation of the inferior temporal cortex with feature
selective hashing to recognize animals. We worked on KTH
database containing 1239 images in 13 classes that took photos
from animals in wild.
Keywords: Cortex-like model; animal recognition; textural
fearure extraction; feature-selective Hashing
I. INTRODUCTION
Human ability in vision always is a great pattern for machine
vision systems. Research shows that it takes 12.5
milliseconds to detect an object in the monkey’s brain. In
computer science when we have a big data for finding a
nearest neighborhood typically, the data are categorized by
constructing decision tree or hashed table. This is to avoid
wasting time. But in the neurosciences, neurons that give the
same answer to a specific object are categorized in same
category. This Framework are constructed by hierarchical
feed forward model and the appropriate hash function
(feature-selective hashing (FSH)) to recognize animals. The
human visual system is illustrated in Figure 1. As you can see
visual information through the optic nerve to the visual
cortex of the brain, and eventually it reaches the bottom of
the cortex or the Inferior Temporal (IT) (Fig.2) that Is
responsible for recognizing objects. And in this paper we
imitate this part of the cortex for animal recognition with
FSH based on texture features. We are inspired by the Yu-Ju
Lee et al [1] paper that they contrast the framework based on
cortex-model to object recognition and they test their
framework on different objects database, While we work on
the wild animal’s database by textural feature extractors. Use
of these extractors is that animals cannot be categorized
based on color or shape, Because of these two aspects they
are very similar, And we have a great error rate. The best
features which animals can be classified is textural properties
or combinations of texture and shape, according to these
reasons, we use Segmentation-based Fractal Texture Analysis
(SFTA), Haralick texture features and Haar Of Oriented
Gradient (HOOG) for extract textural features.
We first review the works that we inspired by them in this
paper, Then we discuss our proposed method step by step,
and finally we examine our experiment to test the efficiency
and accuracy of our method.
II. BACKGROUND
Here, we review previous works that we’re inspired by them
to write this paper:
Henry et al [2] used a neural network to recognize faces by
using intensities in a 20*20 sub-window. Haar wavelet
feature has been very popular since the Viola and Jones [3]
proposed their real-time face detection system. Their method
consists of a cascade classifier with AdaBoost algorithm on a
set of features that were trained by Haar wavelet. Dalal and
Triggs in 2005 [4], proposed higher human detection system
by training a SVM classifier using the histogram of gradient
features (HOG).
Alceu Ferraz Costa et al [5] presented a paper that proposed a
new and efficient texture feature extraction method:
SFTA. The extraction algorithm consists in analysis the input
image into a set of binary images from which the fractal
dimensions of the resulting regions are computed in order to
describe segmented texture patterns. The analysis of the input
image is achieved by the Two-Threshold Binary
Decomposition (TTBD) algorithm, which they also proposed
in their work. SFTA uses thresholding as part of its extraction
algorithm. However, the input image to the SFTA function
does not need to be segmented.