Proceedings of 2010 IEEE Student Conerence on Research and Development (SCOReD 2010),
13 - 14 Dec 2010, Putrajaya, Malaysia
GLCM Texture Feature Reduction for
EEG Spectrogram Image using PCA
Mahzah Mutafa1,2
I
Faculty of Electrical & Electronics Engineering
Ui versity Malaysia Pahang
Kuantan, Pahng, Malaysia
mahuzah@ump. edu. my
Abstract-In Electroencephalography (EEG) research, the
analysis using is time or frequency signals are very popular.
However, it has been shown elsewhere, that any feature rich
signals can be examined using time-frequency components. This
paper proposes a new technique of extracting Gray-level Co
occurrence Matrices (GLCM) texture via time-frequency analysis
of EEG signals. The output of this technique produces a big
feature matrix and it is reduced by applying Principal
Components Analysis (PCA). The resuls of this experiment
shows that EEG signals can be analysed or described using ive
major components of the GLCM.
Kywords-EEG, specrogram iage, GLC, txture feature,
eA
I. NTRODUCTION
he brain is an important organ in he human body. n its
centre there are cells called neurons. hese neurons produce
electric sinals nd can be measured using EEG. n practice,
EEG sinals are in time-based format. To nalyze in
requency-based, usually the sinals need to be transformed
into Fourier rnsform (F). n this paper, EEG sinals were
analyzed based on time-requency image processing technique
or called spectrogram. he most oten technique used to
analyze sinal in time-requency based is Short Time Fourier
Transform (SF). he SFT is to perform a FT on the sinal,
then mapping the sinal into a two-dimensional unction of
requency and time.
A lot of research has been conducted on time-requency
based but mainly in sinal processing area [1,2]. here are
studies, for example using spectrogram image in time
requency to classiY heart abnormalities rom
electrocrdioram (ECG) [3]. hey extracted the heiht and
width of pulses and Euler number rom the spectroram image.
Back-propagation NN was chosen as classiier and gives
100% accuracy in heart abnormalities classiication.
Once an image is obtained, rious image processing tools,
such as texture analysis, could be used. Most popular technique
for the textural classiication is GLCM. he GLCM is a
tabulation of rey levels requency occuring in an image. n
978-1-4244-8648-9/10/$26.00 ©2010 IEEE 426
Mohd Nasir Taib2,3,
Zunairah Hj. Muraf,3, Sahrim Liai,3
2
Faculty of Electrical Engineering
3
Biomedical Research Laboratory for Human Potential
University Teknologi MARA Malaysia
Shah Alam, Selangor, Malaysia
dr.nasir@ieee.org
[4], they extract spectrogram image using normalized rey
level co-occurrence matrices (NGLCM) rom ECG. his study
is about detecting sleep disorder breathing in hn het.
hey are using Fuzzy with result 79.29% accuracy in training
nd 75. 88% accuracy in testing.
Ater GLCM process done, texture features need to be
extracted. he most popular teture feature extraction is
proposed by Haralick [5] and Soh [6]. Haralick proposed 14
texture features to apply to the photomicroraphs of
sandstones, photoraphs of land-use and satellite photoraphs
of land-use while Soh popose 10 teture features for satellite
photoraphs of sea ice.
Principal components analysis (PCA) was used for data
reduction and classiication puposes. n [7], they choose the
irst three principal components which are contrast,
dissimilarity and homogeneity, out of eiht GLCM texture
features. As a result, the irst three components rom PCA
give better accuracy in classiication than all eiht GLCM
texture features. n this paper, a study of dimension reduction
of GLCM textre features is presented. he feature is
extracted rom EEG spectroram image.
II. METHODS
n this project, the EEG data were collected with scalp
electrodes by using standard gold disc electrodes, with 2-
channel Fpl and Fp2. his is set up using 10-20 ntenational
system with 256z sampling rate. n this experiment, 28
males and 23 females agree to be volunteers as sample in this
experiment. he data collections were perform at Biomedical
Research and Development Laboratory for Human Potential,
Faculty of Electrical Engineering, Universiti Teknologi
MARA Malaysia.
Fig. 1 shows the experiment low chart for this study. EEG
was collected then artefacts were removed. he artefact is
regarded as noise when the EEG sinals more than 1 00l V and
less than -100 l V. Next, the sinal was iltered using Band
pass ilter with Hamming window with 50% overlapping