23-08-2012, 02:43 PM
CONTENT BASED FACE RECOGNITION
CONTENT BASED FACE RECOGNITION.ppt (Size: 644 KB / Downloads: 34)
Introduction
Problem Statement :
Given an image, to identify it as a face and/or extract face images from it.
To retrieve the similar images (based on a heuristic) from the given database of face images.
Why face recognition ?
Various potential applications, such as
person identification.
human-computer interaction.
security systems.
Approach
Similar to Content Based Image Retrieval (CBIR).
Neural Networks and Self Organizing Maps (SOMs).
Principal Component Analysis (PCA).
Relevance feed back.
PCA
Main assumption of PCA approach:
Face space forms a cluster in image space.
PCA gives suitable representation.
Eigenfaces (1)
Calculation of Eigenfaces
(1) Calculate average face : v.
(2) Collect difference between training images and average face in matrix A (M by N), where M is the number of pixels and N is the number of images.
(3) The eigenvectors of covariance matrix C (M by M) give the eigenfaces.
M is usually big, so this process would be time consuming.
What to do?
Eigenfaces (2)
Calculation of Eigenvectors of C
If the number of data points is smaller than the dimension (N<M), then there will be only N-1 meaningful eigenvectors.
Instead of directly calculating the eigenvectors of C, we can calculate the eigenvalues and the corresponding eigenvectors of a much smaller matrix L (N by N).
if λi are the eigenvectors of L then A λi are the eigenvectors for C.
The eigenvectors are in the descent order of the corresponding eigenvalues.
Eigenfaces (3)
Representation of Face Images using Eigenfaces
The training face images and new face images can be represented as linear combination of the eigenfaces.
When we have a face image u :
Since the eigenvectors are orthogonal :
Eigenfaces (4)
Experiment and Results
Data used here are from the ORL database of faces. Facial images of 16 persons each with 10 views are used. - Training set contains 16×7 images.
- Test set contains 16×3 images.
First three eigenfaces :
Classification Using Nearest Neighbor
Save average coefficients for each person. Classify new face as the person with the closest average.
Recognition accuracy increases with number of eigenfaces till 15.
Later eigenfaces do not help much with recognition.
Best recognition rates
Training set 99%
Test set 89%
What are Neural Networks ?
Individual units to simulate Neurons
Parallel Processing
Many inputs and single output
Organization/structure of the TLU’s is important
What is SOM ?
TS-SOM :- Tree structure self-organizing maps
Competitive learning ANN
Each unit of map receives identical inputs
Units compete for selection
Modification of selected node and its neighbors
Training of SOM
Randomly initialized
Selection based on some query parameter
On selection a node and its neighbors are modified
Degree of modification reduces with each iteration
Algorithm
Calculate weight vector for first level.
Initialize weight vectors of other levels.
Calculate centroid associated to each node as mean of closest training samples.
Iterate to the next level.
Relevance Feedback
System content based retrieval.
Point of human intervention
User analysis of system output.
User selects most relevant
Query iterated if output not satisfactory
Interaction Between User & System
A random set of faces is presented to the user.
User interactive selection of faces.
System content-based face retrieval.
User analysis of retrieved faces.
Requested face was found -> Exit
Similar faces were found. -> Go to 2
No similar faces were found.
User tired -> Exit
User not tired (re initialization -> Go to 1
Comparison of the Two Approaches
Training time
Nearest neighbor is much faster.
Storage
About the same.
Classification time
Nearest neighbor is slightly slower.
Accuracy
Neural network is able to achieve the same accuracy using 5 eigenfaces with nearest neighbor using 15, and a higher accuracy when using 15.
Neural network models the problem better, but takes more training time.
Future Work
Face Detection in motion pictures.
Detailed study of the proposed system assuming PCA assumptions not to be true.
Investigate whether eigenfaces is a good solution for this problem by comparing with other feature extraction techniques such as DCT
References
Navarrete P. and Ruiz-del-Solar J. (2002), “Interactive Face Retrieval using Self-Organizing Maps”, 2002 Int. Joint Conf. on Neural Networks – IJCNN 2002, May 12-17, Honolulu, USA.
“A tutorial on Principal Components Analysis”, By Lindsay I Smith.
“Eigenfaces for Recognition”, Turk, M. and Pentland A., (1991)Journal of Cognitive Neuroscience, Vol. 3, No. 1, pp. 71-86.
Ruiz-del-Solar, J., and Navarrete, P. (2002). “Towards a Generalized Eigenspace-based Face Recognition Framework”, 4th Int. Workshop on Statistical Techniques in Pattern Recognition, August 6-9, Windsor, Canada.
Simulating Neural Networks by James A. Freeman.
Artificial Intelligence by Neil J. Nielsson.