1. Age Prediction using
PCA, LDA and Hybrid
Methods
Mahdi Roozbahani and Shankar Vishwanath
CSE 6730 - Data and Visual analytics
2. Agenda
• 1. Objective
• 2. Applications and Use
• 3. Current Technologies
• 4. Methodology
• 5. Challenges
• 6. Results
• 7. Conclusion
3. Objective
• 1. Given an image, classify it to a certain age group.
• Using different methods (non-explicit)
• 1. PCA
• 2. LDA
• 3. PCA + LDA Combination
• Cross Validate Results
• Study effects of leaving out ‘first n’ Eigenvectors, to negate
shadow/light, camera effects,
4. Where Used?
• 1. Online Marketing and Advertisement
• 2. Security and Information Purposes
• 3. Age based Content Censorship.
5. Current Technologies/Papers
• 1. Belhumeur et al. seminal paper on ‘Recognition Using Class
Specific Linear Projection’.
• 2. They experimented with Fishers Linear Discriminant model
and Eigenface to project image to low dimensional subspace.
• 3. Gao and Ai studied classification based on Fuzzy LDA
method and Gabor features.
• 4. Still a lot to be explored.
6. Current Technologies
• Face.com has implemented age detection to its photo
scanning API.
• Crime department in Bristol experimented with Age
Classification – result not too encouraging.
7. Methodology: Idea
• Main idea: Use PCA and other approaches to find vectors that
best account for variation of face images in image space.
• Images of faces being similar, will not be randomly distributed
in space.
• Can be described by low dimensional subspace.
• We are looking at using PCA as a pre-processor and then use
LDA.
8. Methodology : Preprocessing
• 1. Data Acquisition and Preprocessing
• Obtain training data and sort them according to age groups
• Trim and resize the images to 61 X 49 pixels.
• Convert to grayscale.
9. Methodology: FaceSpace
• FaceSpace – PCA Approach
Normalize
Data
• Center data by
subtracting mean
of all images
Build
Covariance
Matrix
• Cov matrix is just
multiplication of
centered data
with its transpose
Find
Eigenvectors
• mutiply back by
original matrix
to get the eigen
vectors.
11. Methodology: Test Data
• Multiply eigenvectors with the original training data to get
class training data.
• Project test data to training data and apply Classifier to
determine age group.
• Use l2 Norm distance metric for classification
• KNN classifier works too.
13. Challenges
• Accuracy highly dependent on data.
• Very difficult to obtain standard data with uniform lighting
conditions, camera angles, facial expressions.
• Research papers used subjects with previous age tracked
photos under controlled conditions.
• Addition of accessories, spectacles, hairstyles also affected
results/accuracy.
• Perceived age not the same as actual age.
14. Conclusions
• Why PCA is Better than LDA at times?
• When would LDA+PCA work?
• How does leaving first few eigenvectors help?