Transcript Facial Image Retrieval Through Compound

Md. Zia Uddin
Bio-Imaging Lab, Department of Biomedical Engineering
Kyung Hee University
Contents
Abstract
Introduction
Spectral Filtering
Special Filtering
Classification
Abstract

This presentation is about signal processing and machine
learning techniques and their applications to BCI.

Overview of general signal processing and classification
methods as used in single-trial EEG analysis is given.

For further study, original publications are encouraged.
Why ML for BCI

Subject wise experiments
◦

Session wise experiments
◦

Subject to subject result variances for same kind of experiments.
Session to session huge result variability for the same person.
Real time experiments
◦
◦
The system needs to identify the subjects mental state from single trial.
Much more complexity arises.
Solution

A session and user brain signature adaptable system is necessary
to overcome the subject to subject and session to session huge
variability.
Why Preprocessing?

Relevant information extraction is difficult because of large dimensional
data (i.e., Curse of dimensionality).

Dimensionality has to be reduced keeping the discriminative information
and eliminating undiscriminative information.
Most of the classification methods calculate covariance matrix of the data
for further feature analysis. Huge covariance matrix is required in the case
of large dimensional data.


Thus, Prepropcessiong steps regarding dimensionality reduction is required

In some cases
◦
◦
A priory knowledge is used (e.g., spatial Laplace filter at predefined scalp locations)
Automatic methods (e.g., spatial filters determined by common spatial pattern
analysis)
Spectral Filter: FIR & IIR

Common approach is to use digital frequency filter

To consider desired frequency range
◦
Two sequences of poles (a) and zeroes (b) with length na and nb are necessary
that can be calculated by Butterworth or elliptic.
The source signal x is filtered to y as
◦
a(1)y(t)=b(1)x(t) + b(2)x(t-1) +...+ b( nb )x(t- nb -1) – a(2)y(t-1) -...- a( na )y(t- na -1)

Where a and na are constrained to be 1, is called FIR filter (i.e., considering
all zeros).

Advantage of FIR
◦

Produce steeper slopes in between pass and stop band.
In most of the BCI applications, band pass filter is required to consider
specific frequency range.
Spectral Filter: Fourier-Based Filter
A good alternative than FIR and IIR is to use temporal Fourier-based
filtering in BCI.

◦


A signal switches from temporal to the spectral domain.
The filtered signal is obtained by choosing suitable weighting to the
relevant frequency components and applying Inverse Fourier Transformation
(IFT).
The short time window determines the frequency resolution
Spatial Filter: Bipolar Filtering

EEG channels are measured as voltage potential relative to a standard
reference (referential recording).

Also, it is possible to record all the channels as voltage difference between
the electrode pairs.

From referential EEG, bipolar channels can be obtained by subtracting the
respective channels
FC4-CP4=(FC4-ref) - (CP4-ref)=FC4ref -CP4ref

Reduces the effect of local smearing by computing local gradient.

Focuses on the local activity while contributions of more distant sources are
attenuated
Spatial Filter: Common Average Reference

The mean of all EEG channels are subtracted from each channel to get the
common average reference signals.

Reduces the influence of far field sources but may introduce some
undesired spatial smearing

Artifacts of one channel may spread to all other channels.
Spatial Filter: Laplace Filtering





More localize filter can be obtained through this.
Laplace signals are obtained by subtracting the average of surrounding
electrodes from each individual channel.
C4Lap =C4ref- ¼(C2ref + C6ref + FC4ref + CP4ref)
The choice of surrounding channels determine the characteristics of the
filter.
Usually, small Laplacians are used (as example given above).
Large Laplacians use neighbors at 20% distance as defined in international
10-20 system.
Spatial Filter: Principle Component Analysis(1)




Represent multidimensional data with fewer number of variables
retaining main features of the data.
It is inevitable that by reducing dimensionality some features of the
data will be lost. It is hoped that these lost features are comparable
with the “noise” and they do not tell much about underlying
population.
The method PCA tries to project multidimensional data to a lower
dimensional space retaining as much as possible variability of the
data.
Its simplicity makes it very popular. But care should be taken in
applications. First it should be analyzed if this technique can be
applied.
Spatial Filter: Principle Component Analysis(2)
Original Variable
B


PC 2
PC 1


Original Variable A

Orthogonal directions of greatest variance in data
Projections along PC1 discriminate the data most
along anyone axis
First principal component is the direction of
greatest variability (covariance) in the data.
Second is the next orthogonal (uncorrelated)
direction of greatest variability
◦ So first remove all the variability along the first
component, and then find the next direction of
greatest variability
And so on …
Spatial Filter: Principle Component Analysis(3)
25
Subtract the mean
Calculate the covariance
matrix
Calculate the
eigenvectors and
eigenvalues of the
covariance matrix
20
Variance (%)
Choosing top
components and forming
a feature vector
Get data
15
10
5
0
PC1 PC2 PC3 PC4 PC5 PC6 PC7 PC8 PC9 PC10
Basic Steps


Eigenplot
We can ignore the components of lesser significance.
We do lose some information, but if the eigenvalues are small, we don’t
lose much
 n dimensions in original data
 calculate n eigenvectors and eigenvalues
 choose only the first p eigenvectors, based on their eigenvalues
 final data set has only p dimensions

Problem Definition:
◦ Remove the noise to get VEP in the single trial 29 channels EEG data without
ensemble averaging

Technique adopted to solve the Problem:
◦ Selection of principal components as basis for the reconstruction of signal

Methodology
◦ Given signal is divided into an ensemble of signals, for each channel
◦ An ensemble average for each channel is obtained as a reference
◦ Apply PCA to find out the orthonormal eigenvectors which are used as basis for signal
approximation
◦ Selection of Principal components as basis by looking at the frequency components
present in the “prototype signal” i.e. the averaged signal
Channel # 9 signal
60
40
20
0
Original Signal
-20
-40
-60
0
2000
4000
6000
8000
10000
30
20
Filtered signal
Original signal
Template signal
20
15
50
10
10
5
100
0
0
-5
-10
150
-10
-15
-20
200
-20
-25
-30
-30
0
100
200
300
400
500
Single epoch after PCA filtering
100
200
300
400
Reconstructed epoch stacks
Spatial Filter: Independent Component Analysis(1)
Basically ICA is applied for Blind Source Separation (BSS)
Assume an observation (signal) is a linear mix of unknown
independent source signals
The mixing (not the signals) is stationary
We have as many observations as unknown sources
To find sources in observations






Need to define a suitable measure of independence
… For example - the cocktail party problem (sources are speakers):
Find Z


Formal Statement
◦ N independent sources … Zmn ( M xN )
◦ linear square mixing … Ann ( N xN )
◦ produces a set of observations … Xmn ( M xN )

….. XT = AZT
Spatial Filter: Independent Component Analysis(2)
‘demix’ observations … XT ( N xM ) into YT = WXT
W ( N xN )  A-1
How do we recover the independent sources?
(We are trying to estimate W  A-1 )
…. We require a measure of independence!
YT ( N xM )  ZT
Spatial Filter: Independent Component Analysis(3)




The source signals are mixed by random non orthogonal matrix
JADE algorithm was applied to demix the signals
After reordering and scaling, the demixed signals are very similar to sources.
PCA would fail here as the mixed signals are not orthogonal to each other, which is
the key assumption of PCA.

Other ICA algorithms
 Infomax
 FastICA
Spatial Filter: Independent Component Analysis(4)
Applying ICA to single-trial EEG epochs

Data Collection
EEG data were recorded from 31 scalp electrodes
29 placed at locations based on a modified International 10-20 system
one placed below the right eye (VEOG),
one placed at the left outer canthus (HEOG). All
31 channels were referred to the right mastoid and were digitally sampled for analysis at 256
Hz with a 0.01- to 100-Hz analog bandpass plus a 50-Hz lowpass filter.
Subjects participated in a 2-hour visual spatial selective attention task in which they were
instructed to attend to filled circles flashed in random order in five locations.
Component IC1, generated by blinks
IC4 generated by temporal muscle activity.
Spatial Filter: Independent Component Analysis(5)
Applying ICA to single-trial EEG epochs (2)
The scalp maps and power spectra of the 31 independent components derived from
target response epochs from a 32-year-old autistic subject.
Blink and eye movement artifact components (IC1 and IC9) had a typical strong low
frequency peak.
Temporal muscle artifact components (i.e., ICs 14, 22, 27, and 29) had characteristic
focal optima at temporal sites and power plateaus at 20 Hz and higher.
Conclusion

Next class
more classification techniques and some practical examples.
Thank you