CS298 Proposal
Recognition and Age Prediction with Digital Images of Missing Children
Wallun Chan (wallunchan@yahoo.com)
Advisor: Dr. Chris Pollett
Committee Members: Dr. Mark Stamp (stamp@cs.sjsu.edu) and Dr. Robert Chun (rchun@cs.sjsu.edu).
Abstract:
In this project, we use the eigenface technique [Turk91] to produce age-enhanced
face images of missing children. The eigenface approach is based on using PCA
(Principal Components Analysis) on image data. PCA allows orthogonal vectors
(eigenvectors) to be computed as a new coordinate system for multi-variate data.
The projection of the data onto the eigenvector with the greatest eigenvalue
(principal component axis) results in maximum variance about that axis.
For the remaining eigenvectors, decreasing eigenvalues correspond to
decreasing variances of the projected data. The lesser eigenvectors
that do not result in much scatter may be truncated without sacrificing
accuracy of data reconstruction. PCA has the property such that the
projected data results in a set of weights that represent linear proportions
of the principal components. Thus high-dimensionality data is reduced
significantly in dimensions, allowing efficient algorithms that use only
the required eigenvectors for the largest few corresponding eigenvectors.
The weights may then be used to reconstruct each original data point.
If a set of weights is used to reconstruct a data point, such that the
data point is one that the PCA classifier is not trained on, then the
classifier attempts to reconstruct the data point closest to one of the
training data points.
The idea is to implement software that takes color images of faces as input
and allows users to select and remove specified features of the images
(e.g. eyes, nose, and mouth) to be PCA trained. Prior to this, the
images are preprocessed in terms of size (all images of same size),
rotation (removal of in-plane rotation), and cropping (eliminate background
and maximize face space). For each selected feature, each color (RGB) channel
is extracted to produce three separate grayscale images that are concatenated
to form an overall grayscale image. This is done for original and age-enhanced
sets of images, results of which are also concatenated to form final training
grayscale images for each feature. These training images provide a mapping
between original and corresponding age-enhanced features. The face images
with the selected features removed are also PCA trained. An input face
image is preprocessed in a similar manner as mentioned above except that
the selected features are concatenated with themselves. Each selected
feature from the input image is then projected onto a corresponding
PCA trained feature space to produce a set of weights, which are
then used to reconstruct a predicted aged feature; this is also
performed on the feature-removed face images. The aged features
are then blended back into the age-predicted feature-removed faces
at the same locations.
CS297 Results
- Verified the eigenface approach by implementing a program that trains a PCA classifier to successfully recognize and reconstruct preprocessed digital images of missing children
- Implemented three Python scripts to support the downloading and preprocessing (image resizing, rotation, color format, and cropping) of digital images of missing children from a specific website (www.missingkids.com)
- Implemented a program to verify the feasibility of using shape contexts [Belongie01] and weighted bipartite matching [Jonker87] to locate features on a face image
- Further validated the eigenface approach by successfully reconstructing age-enhanced images using concatenated training images of original and aged images
Proposed Schedule
Week 1 - 6:
8/29 to 10/7 | Implement age-enhancement / reconstruction program. |
Week 7 - 8:
10/10 to 10/21 | Revise and improve on program based on tests and results. |
Week 9 - 11:
10/24 to 11/11 | Write CS298 report. |
Week 13 - 14:
11/14 to 11/25 | Complete draft report for committee review; prepare project presentation slides. |
Week 15:
11/28 to 12/2 | Official report completion. |
Week 16:
12/5 to 12/9 | Project defense / presentation. |
Key Deliverables:
- Software
- Allows user to train program to reconstruct age-enhanced images via a features-based approach
- Allows user-selected features on a template image to be automatically extracted from training and input images
- Relies on and ties in much of the code from CS297
- Downloading of digital images of missing children
- Preprocessing of digital images to be used by program
- Face recognition and reconstruction using eigenfaces
- Report
- Goals, results, and assumptions of project
- Demonstration of successful age-enhancement / reconstruction of known and unknown input images
- Theoretical and practical limitations
- Theory
- Eigenfaces
- Principal components analysis
- Image Data
- Image acquisition methods
- Image preprocessing techniques
- Image post-processing (image clean-up, filtering, etc.)
- Software architecture and operations
Innovations and Challenges
- Well designed user interface that supports image loading, feature selection, and display of age-enhanced / reconstructed images
- Initially rely on the preprocessed alignment of the training images for locating user-selected features; must research other alternatives to achieve feature location on training and input images if results are not satisfactory
- May need to due some image post-processing to clean up age-enhanced / reconstructed image, due to the fact that the aged features may have different pixel intensities when blended back into the corresponding faces
References:
[Belongie01] S. Belongie, J. Malik, and J. Puzicha. Shape matching and object recognition using shape contexts. Technical
Report UCB//CSD-00-1128, UC Berkeley, January 2001.
[Bishop 95] Neural Networks for Pattern Recognition. Christopher M. Bishop. Oxford University Press, 1995.
[Duda01] Pattern Classification. Richard O. Duda, Peter E. Hart, and David G. Stork. John Wiley & Sons, 2001.
[Gonzales02] Digital Image Processing. Rafael Gonzalez and Richard E. Woods. Prentice Hall, 2002.
[Jonker87] R. Jonker and A. Volgenant. "A Shortest Augmenting Path Algorithm for Dense and Sparse Linear Assignment Problems", Computing 38, 325-340, 1987.
[Press88] Numerical Recipes in C. William H. Press, Brian P. Flannery, Saul A. Teukolsky, William T. Vetterling. Cambridge University Press, 1988.
[Russell03] Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach. Pearson Education, Inc., 2003.
[Seul00] Practical Algorithms for Image Analysis: Description, Examples, and Code. Michael Seul, Lawrence O'Gorman, and Michael J. Sammon. Cambridge University Press, 2000.
[Smith02] A Tutorial on Principal Components Analysis. Lindsay I. Smith.
http://www.cs.otago.ac.nz/cosc453/student_tutorials/principal_components.pdf, 2002.
[Turk91] Eigenfaces for Recognition. Matthew Turk and Alex Pentland. Journal of Cognitive Neuroscience, Vol. 3, No. 1, 1991.
|