KLKKK

Intro
This repository implements EigenFaces paper. This is an algorithm that
tries to recognize and detect faces using PCA.
Usage
The src.py file contains the class EigenFaces, whose usage is given in
demo.py over the ATnT face dataset.
Run demo.py for a sample run.
Requirements
python 3.7.0
opencv-python 3.4.2.16
opencv-contrib-python 3.4.2.16
numpy 1.15.2
scipy 1.1.0
sklearn 0.19.2
Idea
This paper tries to use the idea of Principal Component Analysis to
reduce the number of dimensions of the image and hopefully only
measure those “Features” that vary to give different identities to people

faces.
Methodology
Training
Suppose we have images that are of the dimensions .
We first convert all the images into vectors of dimention
. Each of these vectors is denoted by
We next find the ‘mean-face’, i.e. the mean of all the vectors
as .
We find the offsets of the images from the mean face by
substracting every image as,
We next want to find the eigenvectors of the Covariance
matrix , where,
Since is of the dimention we don’t want to spend
so much compute power to find out eigenvectors. We work
around this by finding the eigenvectors of .
After we get all of the , we drop the all but the top
(according to thier eigenvalues) of them (PCA at work).
We convert these to using the equation,
We now normalize the we got and then these are exactly
the ‘eigenfaces’ that the paper talks about.
Next we find the weight vector by the following
equation,
We store the resulting weight vectors for each of the training
images.
Testing
During testing we find the of the testing image and try
to minimize the from all the of the training images. The
that is closest to set the label of the testing image, i.e.
the label that the argmin belonged to, is predicted the label of
the testing image.
There may be cases that we have an unknown face, in this case
we set a threshold . The value of the minimum norm we got
from subtracting and should be less than this, else the
testing image is considered to correspond to an unknownface,
whose we don’t have.
Other case is when we provide a non face image to the classifier,
in this case we try to reconstruct the testing image from
by the following equation
We want this reconstructed image to be withing , threshold
from the originial image , i.e.
To be considered a face image, else a non face image.
Testing
We tested the class that we implemented by only letting the
algorithm learn 38 subject images in the dataset above. We then
tested the algorithm on a collection of all the 40 subjects in the
dataset.
We set the values of the threshold by looking at the sample
values we were getting of the L2 norms of the different
differences in weights.
To test images that are non-faces, we used a random numpy
array generated by np.random.rand to test for non-face
detection by our algorithm.
We chose the values of the hyperparameters that resulted in a
good accuracy over all the accuracy measures, one can
obviously change the values of the hyperparameters to get a
different distribution of the accuracies according to a particular
use case.
Also we took the train_test_split ratio of 80-20 .
Results
We used the ATnT face data set to train our algorithm and then we
tested the implementation using some known faces, some unknown
faces, and some non-faces, in each of the case we received accuracies of
Accuracy_knownFaces: 0.7157894736842105 ,
Accuracy_unknownFaces: 0.8 and Accuracy_nonFaces: 1.0 . The

hyperparameters that we worked with were as below. One can change
the hyperparameter at the start of demo.py to get different results.
We print the accuracies to the console and the main class that has been
implemented resides in src.py . Have a look at that file to know the

implementation details.
Hyperparameters used:
K = 30 # dimension of face_space
b = 2 # number of classes to keep unseen
te = 2100 # Threshold for the L2 distance for training wei
ght vectors
tf = 13000 # Threshold for the L2 distance from the face s
pace
unknownface = -1 # label to denote an unknownface
nonface = -2 # label to denote an nonface
Top 3 eigenfaces:
Conclusions
Setting the hyperparameter of threshold te to a big value
ensures that we get good recognition when the faces are known
but keeping it low will increase the accuracy of detecting an
unknown face at the cost of accuracy of recognizing a known
face.
Keeping K bigger ensures better recognition, at the cost of
detecting an unknown face.
One needs to tune the values of te , tf , K according to a
particular use case. For example, taking K=120 and te = 3500
in our dataset testing results in Accuracy_knownFaces:

0.9263157894736842 , Accuracy_unknownFaces: 0.0 and
Accuracy_nonFaces: 1.0 .
Reference
F. Samaria and A. Harter
“Parameterisation of a stochastic model for human face
identification”
2nd IEEE Workshop on Applications of Computer Vision
December 1994, Sarasota (Florida).
http://www.vision.jhu.edu/teaching/vision08/Handouts/case_study_pca1
http://www.face-rec.org/algorithms/pca/jcn.pdf

KLKKK

Uploaded by

Copyright:

Available Formats

KLKKK

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

KLKKK

Uploaded by

Copyright:

Available Formats

Intro

This repository implements EigenFaces paper. This is an algorithm that

tries to recognize and detect faces using PCA.

Run demo.py for a sample run.

reduce the number of dimensions of the image and hopefully only

measure those “Features” that vary to give diﬀerent identities to people

Suppose we have images that are of the dimensions .

We ﬁrst convert all the images into vectors of dimention

. Each of these vectors is denoted by

We ﬁnd the oﬀsets of the images from the mean face by

substracting every image as,

We next want to ﬁnd the eigenvectors of the Covariance

Since is of the dimention we don’t want to spend

so much compute power to ﬁnd out eigenvectors. We work

around this by ﬁnding the eigenvectors of .

(according to thier eigenvalues) of them (PCA at work).

We convert these to using the equation,

We now normalize the we got and then these are exactly

the ‘eigenfaces’ that the paper talks about.

Next we ﬁnd the weight vector by the following

During testing we ﬁnd the of the testing image and try

to minimize the from all the of the training images. The

that is closest to set the label of the testing image, i.e.

the testing image.

There may be cases that we have an unknown face, in this case

we set a threshold . The value of the minimum norm we got

from subtracting and should be less than this, else the

testing image is considered to correspond to an unknownface,

whose we don’t have.

Other case is when we provide a non face image to the classiﬁer,

in this case we try to reconstruct the testing image from

by the following equation

We want this reconstructed image to be withing , threshold

from the originial image , i.e.

To be considered a face image, else a non face image.

We set the values of the threshold by looking at the sample

values we were getting of the L2 norms of the diﬀerent

To test images that are non-faces, we used a random numpy

array generated by np.random.rand to test for non-face

detection by our algorithm.

We chose the values of the hyperparameters that resulted in a

good accuracy over all the accuracy measures, one can

obviously change the values of the hyperparameters to get a

diﬀerent distribution of the accuracies according to a particular

Also we took the train_test_split ratio of 80-20 .

tested the implementation using some known faces, some unknown

faces, and some non-faces, in each of the case we received accuracies of

Accuracy_unknownFaces: 0.8 and Accuracy_nonFaces: 1.0 . The

the hyperparameter at the start of demo.py to get diﬀerent results.

implemented resides in src.py . Have a look at that ﬁle to know the

b = 2 # number of classes to keep unseen

te = 2100 # Threshold for the L2 distance for training wei

tf = 13000 # Threshold for the L2 distance from the face s

unknownface = -1 # label to denote an unknownface

nonface = -2 # label to denote an nonface

but keeping it low will increase the accuracy of detecting an

unknown face at the cost of accuracy of recognizing a known

Keeping K bigger ensures better recognition, at the cost of

detecting an unknown face.