mainly parietal and temporal are
found to take significant active participation during the experiment. The paper
6 reveals that temporal lobe and frontal lobe are highly associated with
human olfactory signal processing. Therefore, we select F3, F4, F7, F8 and FZ
(for frontal lobe) P3 and P4 (for parietal lobe), T3
and T4(for temporal lobe) and FP1 and FP2 (for
prefrontal lobe)for extracting necessary information using signal processing
techniques.Fig.6shows the electrode positions over the scalp during experiment,
where the red colored electrodes are used to collect data. Electrodes are
placed over the scalp using 10-20 electrode placement system 26.
Figure.6. Electrode position
of our experiment
Best services for writing your paper according to Trustpilot
A. EEG
Feature Extraction and Feature Selection
Selection
of required features is important for EEG classification problem for accurate
decoding of mental tasks. There are a variety of time domain feature extraction
techniques (e.g., Hjorth parameters 18, Autoregressive parameters 19,
frequency domain feature extraction techniques (e.g., power spectral density)
and time-frequency correlated feature extraction techniques (e.g., discrete
wavelet transform 20). We first plot the raw EEG signal pattern recorded from
the specific brain regions during the experiment. Figure 7(a) represents the
raw EEG signals acquired from the temporal lobes during low, medium and high
concentration respectively for four smell stimuli and Figure 7(b) shows the
filtered EEG signals for different odor stimuli.
Power
spectral density (PSD), is a well-known frequency-domain EEG feature extraction
technique to extract signal power distribution. PSD is applied
to the filtered EEG signal acquired from prefrontal, parietal, temporal and
frontal regions. Besides PSD, It is important to mention here that filtering
of EEG signal is done by using a standard Chebyshev 21 band pass infinite
impulse response (IIR) filter of order 10, which has the pass band frequency of
0.5-13 Hz. The selection is made so because of the superior performance of
Chebyshev filter as compared to its standard counterparts including Butterworth
and Elliptic filters. Now, for each subject and each odor concentration, PSD
extract 10×12×257 feature sets (since,
here, experiment is repeated 10 times and number of selected electrodes 12).
Fig.8(a) and Fig.8(b)present the PSD features extracted from the above brain
regions.
It
has been observed from Fig.7(a) and (b) that although PSD extracts 257 features
for a particular odor sample, only fewer features (e.g. 2nd, 6th,
and 14th) can discriminate odors jointly. In this manner, we finally obtain 12
such features that can be fed to the classifier to decode the smell stimulus.
The
feature dimension 10×12×257 is large enough. To reduce the feature
dimension, we use PCA to select the most significant features and we get 24
optimal features which are used in the classification.
II.
Classifier Performance
We examine the classification accuracy of the proposed classifier
techniques by observing i) individual class performance during the classifier
training and ii) overall classifier performance during testing phase.
A. Individual
Class Performance during Training
For individual
class performance of different genres, the proposed classification algorithms
are trained with 10 trials, one for each odor stimuli repeated ten times on
each of 10 subjects. A standard ten-fold cross validation technique is employed
to check the consistency of the data, where nine out of ten folds are applied
for training purposes and the remaining one fold is used for the validation
purposes. Table I provides the individual class performance of 4 odor stimulus.
The concentration level
should be maintained as Low (25% aroma with 75% water), medium (50% aroma and
50% water) and High (75% aroma and 25% water).
Table I. Classification Performance analysis with proposed
Classifier
Odor Type
Concentration level(%) with water
Classification Accuracy (%) using Model 1 for
Worst
Average
Best
Odor1
(Perfume)
High
Medium
Low
71
69
68
85
83
82
94
89
87
Odor2
(Dettol)
High
Medium
Low
75
72
71
86
84
82
93
90
89
Odor3
(Ascetic acid)
High
Medium
Low
69
65
64
79
75
73
92
88
86
Odor4
(Alchohol)
High
Medium
Low
77
75
74
86
83
87
95
90
89
B. Overall
Classifier Performance during Testing Phase
To study the relative performance, we consider the following two
standard classifiers: 1) support vector machine (SVM) 22 and back propagation
neural network (BPNN) (V.N.P) along with our two proposed method. Table II
provides the average percentage classification accuracies, where from it can be
concluded that the proposed model II outperforms the existing BPNN, SVM and
Model by a significant margin.
Table II.
Proposed classifier comparison with
existing standard classifier
Odor type
Concentration level(%) with water
SVM
BPNN
Proposed Model
Odor1
(Perfume)
High
Medium
Low
86
84
83
89
87
84
90
88
85
Odor2
(Dettol)
High
Medium
Low
89
86
82
87
83
81
93
90
89
Odor3
(Ascetic acid)
High
Medium
Low
90
87
85
90
88
86
89
86
84
Odor4
(Alchohol)
High
Medium
Low
79
76
75
92
90
87
94
91
89
Table III
present the statistical test results by applying well-known McNemar’s test24using our
model. The z value in McNemar’s test
is given by the equation 4:
(4)
Table III.Statistical
performance test for Model I algorithm with
McNemar’s Test
Classifier
name
McNemar’s
constant
(m)
McNemar’s
constant
(n)
Z
P
SVM
5
9
1.65
<0.0011 BPNN 6 11 1.82 <0.0014 From Table III, z value describes that our proposed model outperforms the above two standard classifiers with a wider margin. Table IV shows Friedman test25 performance using the proposed algorithms. Friedman test is done with four classifiers (SVM, BPNN, and proposed model) for each of twelve concentration levels of four different odor datasets. The statistical measure is given by (5). (5) is the rank of each odor database. is number of databases, here it is 12. is number of classifiers, here it is 3. This Friedman test is done on ten subject databases. Table IV describe the rank of each classifier, which is evaluated according to the average classifier accuracy with all stimuli. From Table IV, it can be concluded that our proposed model classifier also provide superior performance than SVM and BPNN classifier in Friedman test.