Transcript Terrorists - u

Terrorists
Team members:
Ágnes Bartha
György Kovács
Imre Hajagos
Wojciech Zyla
The Project
•
•
•
•
What is our goal?
Who are they?
How to start?
How ro recognize face?
– Face detection
– Face feature detection
– Eyes, mouth, nose
related search
What is our goal?
• Find out if someone is a terrorist.
• Try to identify then even if they
are disguised
• We have a problem..
Who are they?
•
They are who
–
–
–
•
Blow up cars, buildings
Kill people
Undertake control
Enough reason to do something
How to start?
•
Database
–
–
Images of terrorist
Training images for identification ( by computer)
•
Take a picture of suspicious person
•
Start to do a program that decides if someone is a terrorist
How to recognize face?
• Problems
– Disguised person
– Other : rotated head, glasses.
• Use some algorithms
– PCA
– LDA
• OpenCV
– Haar object detection
– AdaBoot
PCA
•
Principal Component Analysis
•
reduce the dimensionality of the data while retaining as much
as possible of the variation present in the original dataset
•
implies information loss
•
The best low-dimensional space can be determined by
the "best„ eigenvectors of the covariance matrix
•
(i.e., the eigenvectors corresponding to the "largest"
eigenvalues, also called "principal components").
•
PCA projects the data along the directions where the
data varies the most.
Problems of Eigenface technique
• Sensitive to rotation, scale and translation.
• Sensitive to lighting variations
• Background interference
• Face images should be preprocessed to lessen the
effects of possible variations.
• Variations such as lighting and rotation can also be taken
into account during training. The training dataset may
include samples with such variations.
LDA
•
•
•
•
•
•
•
•
Linear Discriminant Analysis
The objective of LDA is to perform
dimensionality reduction while preserving as
much of the class discriminatory information
as possible
It seeks to find directions along which the
classes are best separated.
It does so by taking into consideration the
scatter within-classes but also the scatter
between-classes.
It is also more capable of distinguishing
image variation due to person identity from
variation due to other sources such as
illumination and expression.
μr mean feature vector for class r.
Kr number of training samples from class r.
LDA computes a transformation that
maximizes the between-class scatter while
minimizing the within-class scatter:
LDA 2.
•
Limitations:
–
–
at most R−1 nonzero eigenvalues.
matrix Sw-1 does not always exist.
•
•
•
need at least N +R training samples – not practical
Use PCA to reduce dimention
When the number of training samples is large and representative for each class,
LDA outperforms PCA.
OpenCV
•
Open Source Computer Vision Library
–
Extensive vision suport
•
•
•
•
Convolution, thresholding, floodfills, histogramming
Pyramidal-subsampling
Learning-based vision
Feature detection
–
–
–
Edge detection
Blob finders ,. ....
Haar cascade classifier
IplImage
OpenCV -- Haar
•
OpenCV has a Haar features based face
detection module.
•
Uses local features such as edges and line
patterns. It scans a given image at different
scales as in template matching.
•
Scale, translation and light invariant.
•
However it is sensitive to rotation.
–
Rotate image and run again
Advantages of using OpenCV
Haar object detection
•
Face detector already implemented
•
Its only argument is a xml file
•
Detection at any scale
•
Face detection (for videos) at 15 frames per second for 384*288 pixel images
•
90% objects detected – achievable doing 2 weeks training
Haar-Like Features
•
Each Haar-like feature consists of two or three jointed “black” and “white”
rectangles:
Figure 1: A set of basic Haar-like features.
Figure 2: A set of extended Haar-like features.
•
•
The value of a Haar-like feature is the difference between the sum of the
pixel gray level values within the black and white rectangular regions:
f(x)=Sumblack rectangle (pixel gray level) – Sumwhite rectangle (pixel gray level)
Compared with raw pixel values, Haar-like features can reduce/increase
the in-class/out-of-class variability, and thus making classification easier.
Haar-Like Features
•
•
The rectangle Haar-like features can be computed
rapidly using “integral image”.
Integral image at location of x, y contains the sum of the
pixel values above and left of x, y, inclusive:
P( x, y)   i( x' , y' )
x ' x , y ' y
P (x, y)
•
•
Haar features computed in constant time
The sum of pixel values within “D”:
P1  A, P2  A  B, P3  A  C, P4  A  B  C  D
P1  P4  P2  P3  A  A  B  C  D  A  B  A  C  D
A
P1
B
P2
D
C
P3
P4
Adaboost classifier
• Selects a small number of
critical visual features
• Combines a collection of weak
classification functions to form
a strong classifier
The first and second features
selected by AdaBoost for face
detection
2. Haar-Like Features (cont’d)
•
For example: to detect hand, the image is scanned by a sub-window
containing a Haar-like feature.
•
Based on each Haar-like feature fj , a weak classifier hj(x) is defined as:
where x is a sub-window, and θ is a threshold. pj indicating the direction
of the inequality sign.
Adaboost
•
The computation cost using Haar-like features:
Example: original image size: 320X240,
sub-window size: 24X24,
The total number of sub-windows with one Haar-like feature per second:
(320-24)X(240-24)=63,936
Considering the scaling factor and the total number of Haar-like features,
the computation cost is huge.
•
AdaBoost (Adaptive Boost) is an iterative learning algorithm to construct
a “strong” classifier using only a training set and a “weak” learning
algorithm. A “weak” classifier with the minimum classification error is
selected by the learning algorithm at each iteration.
•
AdaBoost is adaptive in the sense that later classifiers are tuned up in
favor of those sub-windows misclassified by previous classifiers.
Adaboost
•
The algorithm:
Adaboost
 Adaboost
starts
with
a
uniform
distribution of “weights” over training
examples. The weights tell the learning
algorithm the importance of the example.
 Obtain a weak classifier from the weak
learning algorithm, hj(x).
 Increase the weights on the training
examples that were misclassified.
 (Repeat)
 At the end, carefully make a linear
combination of the weak classifiers
obtained at all iterations.
ffinal (x)   final,1h1 (x)    final,n hn (x)
Adaboost
• Simple to implement
• But..
– Suboptimal solution
– Over fit in presence of noise
The Cascade of Classifiers
•
•
•
•
•
A series of classifiers are applied to every sub-window.
Increases speed
The first classifier eliminates a large number of negative sub-windows and
pass almost all positive sub-windows (high false positive rate) with very little
processing.
Subsequent layers eliminate additional negatives sub-windows (passed by
the first classifier) but require more computation.
After several stages of processing the number of negative sub-windows have
been reduced radically.
The Cascade of Classifiers
•
Negative samples: non-object images.
Negative samples are taken from arbitrary
images. These images must not contain
object representations.
•
Positive samples: images contain object
(hand in our case). The hand in the
positive samples must be marked out for
classifier training.
detecting
cropping
face
face
144x150
90x130
image
normalizing
0
1
0
creating
detecting
...
feature
vector
features
256x256
results
Database of terrorists
0
0
1
comparing vectors
...
is not in the
database
Eye detection with Haar
• eye_haarcascade_classifier
• create a growable
sequence of eyes
• detect the objects
• store them in
the sequence
Thank you for your attention