Description
1 Overview
In this project, you will experiment with the application of dimensionality reduction and
classification methods in face recognition. You will be given a well-known dataset of face
images, which consists of 10 different images for each of the 40 different subjects. On
this dataset, you will experience the use of PCA to extract the so-called eigenfaces and
use them in combination with linear classification methods for image reconstruction, face
recognition, and face identification. As the last project of the course, this project will
require more creativity and independent learning ability from you with only a minimal
introductory information given for the need of the projects.
2 Problems
The application of PCA in face recognition, first developed in 1980s, has been proved to
be a great success. The term eigenface represents a set of principal components generated
from PCA on a large set of images depicting different human faces. They can be considered a set of standardized face ingredients, a linear combination of which can then renders
the face of an individual human subject. In another word, the facial image of a person
is no longer represented on the image domain, but on the domain of the eigenfaces. In
this project, you will experience the application of the simple technique of PCA in a set of
classic problems.
3 Task
A dataset of face images are provided on MyCourses. This is also known as the AT&T face
database.
1
Project 4 – Dimensionality Reduction & Classification 2
This database consists of 400 facial images from 40 subject × 10 image per subject. After
figuring out how to load these images into your matlab, you will go through the
following tasks:
• Dimensionality reduction
1. Use PCA to extract eigenfaces on all 400 images, examine
(a) How do the (leading) eigenfaces look like as an image (show some examples in your report)?
(b) How does the importance of the eigenfaces decrease?
2. Face reconstruction with PCA: On a subset of images from different subjects,
reconstruct the face image using different numbers of principal components.
(a) Observe the difference between reconstructed and original images, as the
number of eigenfaces used in reconstruction increases.
(b) How many eigenfaces are required to recover an original face with reasonable errors?
• Classification: use one of the classification methods discussed in class for the following tasks. You will divide the dataset into a training set and a testset. Make
sure you training set has images from 35 subjects and 8 image per subject. Use the
remaining 2 image for each of the 35 subjects as well as all images for the other 5
subjects as test images.
1. Face recognition: Given a new image, recognize whether it is a face image or
not. Test it on the left-out 2 × 35 subjects, the left-out 10 × 5 subjects, and some
“non-face” images.
2. Face identification: Given a new face image, identify and label it to the correct
person in the training set
• Bonus question (10 on top of 50 for the content part): Based on the limitations observed above, introduce a new dimensionality reduction method and / or new classification method, and compare the results to the above.
4 Deliverables
1. Code. Please submit all the code, along with a brief documentation. I should be able
to run your code based on these.
2. Report. Three page maximum including figures, single column. Please be clear and
concise. You are expected to include (but not restricted to) to include the answers to
the above questions asked.
Project 4 – Dimensionality Reduction & Classification 3
3. Presentation. You will give a 8 minute presentation as a group.
5 Grading
• Content: 50%
• Report: 20%
• Presentation: 30%
• 10 points bonus for an extra method attempted