1. Final Year Project Presentation --- Magic Paint Face
Supervisor Name: Korris CHUNG
Co-examiner: Simon S.F. Hau
Student Name: LI Deqing, Yuki
Student ID: d

2. Content Introduction & Background System Overview System Modules
Image Preprocessing Module
Image Transformation Module
Aging Simulation Module
Aging Visualization Module
Evaluations
Parameter Analysis
Evaluation of Aging Simulation Systems
Conclusion

3. Introduction & Background
Definition:
Magic paint face is a facial aging simulation system that predicts a facial look during growth and aging process
Application:
Prediction of facial appearance
Update of face images in identification documents
Prediction on characters’ aging faces
Face recognition robust to aging variation
Automatic age verification

4. Introduction & Background (Con)
Existing Researches:
Statistical Face Model
Wrinkle Extraction
Problem to be solved:
Can not study from new samples
Only predict from one age scope to another
Prediction of wrinkle rather than simulation
Neural Network:
Facial aging is an intelligent process which can be better performed by human brains rather than symbolic processors

5. System Overview Image Sample Database User Preprocessed Image
Eigen-faces of Adult Image
Aging Simulation Module
Image Preprocessing Module
Database
Aging Visualization Module
Generate Aging Image
User
Converged Network
User image
User selection
Aging Image
Preprocess Image
Image Sample
Train Neural Network System
Image Transformation Module
Preprocessed Image
Convert Image into eigen-faces
Eigen-faces of Youngster Image

6. Image Preprocessing Module
Standard:
frontal, full-face, plain expressive image with concolorous background
Face Database:
55 image pairs
128*128 pixels
male and female
white and black
Wrinkle Database
70 image pairs
64*32 pixels

7. Image Transformation Module
Problem: Raw pixels are too huge to be fed into the neural network system
Eigen-faces based on PCA
Eigen-vectors derived from principal component analysis but used to represent face images
Function:
To transform the M pixels of a face image into N (number of images in database) eigen-face images

8. Image Transformation Module (Con)
Each human face can be divided into N eigen-faces with different proportions
Each eigen-face represents a certain feature in human face
Eigen-face with higher weight means it contributes more in the original face
Compression can be done by ignoring the less weighted eigen-faces

9. Image Transformation Module (Con)

10. Aging Simulation Module
Neural network
An artificial neuron system which simulates biological neurons
Training:
Map the input image onto a pre-assigned output image
Testing:
Feed the converged network with unseen input image
Input image
Input node
Hidden node
Output node
Output image
Eigen-face Coefficients
Input vector
Output vector

11. Aging Visualization Module
Input: user image, selection of required section
Output: aging facial image compared with input image

12. Parameter Analysis Eigen-faces Evaluation
More eigen-faces, less error
Error reduced slower as the number of eigen-faces grows larger
30% to 10% error rate for image recognition, 5% to 10% error rate for image estimation

13. Parameter Analysis Eigen-faces Evaluation (Con)
No. of eigen-faces 5 10 15 20
Reconstructed image
Not recognizable 25 30 35 40
recognizable 45 50 55 60
Almost identical 65 70 75

14. Parameter Analysis Hidden Node Evaluation
Speed of error reduction is slower and slower
Error meets minimum at 40 hidden nodes
A proper number of hidden nodes = generalize
A too large number of hidden nodes = memorize

15. Parameter Analysis Noise Evaluation
Eigen-face no.
Input Image 1
Output Image 1
Input Image 2
Output Image 2
original 25 30 35 40 45
Able to tolerate noise when 25 eigen-face coefficients were distorted
Up from 30 eigen-face, the deviation is distinct
Deviation of the output face depends on the distortion of the facial features of the input image

16. Parameter Analysis General Evaluation
Performance of estimating female models was much poorer
may be because
females usually wear makeup when taking pictures
female’s facial features change much when they grow up
Sample ID
Original Input Image
Estimated by Training 1
Estimated by Training 2
Actual Output Image 1 2 3 4

17. Parameter Analysis Sample Evaluation
Faces estimated by 5 training samples look similar to each other
20 training samples can differentiate each of the input faces
Estimation of testing samples with small distance to each other may follow the same transformation pattern
Sample ID
Original Input Image
Estimated by Training 1
Estimated by Training 2
Estimated by Training 3
Actual Output Image 1 2 3 4 5

18. Evaluation of Aging Simulation Systems Child to Adult Evaluation
Similarity of facial features between estimated face and its desired face can be found subjectively
Able to generalize different races
Standard input images result in better performance
Sample ID
Original Input Image
Estimated by Training 1
Actual Output Image 1 2 3 4 5

19. Evaluation of Aging Simulation Systems Adult to Child Evaluation
Similarity of the facial features between estimated face and its desired face can be found subjectively
Able to estimate the target age
Sample ID
Original Input Image
Estimated by Training 1
Actual Output Image 1 2 3 4 5

20. Evaluation of Aging Simulation Systems Wrinkle Estimation
Able to predict a natural wrinkle image
Wrinkle estimation depends on the facial texture of different people
Therefore, Training 2 and Training 3 are proved to be successful
Original Input Image
Estimated by Training 1
Estimated by Training 2
Estimated by Training 3

21. Evaluation of Aging Simulation Systems Linear Kernel (Support Vector Regression)
Able to estimate normal aging images from unseen faces
Sensitive to the quality of both the training and the testing samples
Sample ID
Original Input Image
Estimated by Training 1
Actual Output Image 1 2 3 4 5

22. Evaluation of Aging Simulation Systems RBF Kernel (Support Vector Regression)
Performance of estimating aging images is not satisfactory
First eigen-face coefficient of each test sample is almost the same
Parameters need to be fine-tune
Sample ID
Original Input Image
Estimated by Training 1
Actual Output Image 1 2 3 4 5

23. Conclusion
Neural network is able to predict natural aging face based on the facial features of an unseen individual.
Neural network is able to generalize and predict the correct race and age.
Neural network finds application in both facial aging and wrinkle aging.
The performance depends on a lot of parameters.
Neural network system favors standard testing samples and produces better results on them.

24. End
Thank you for listening!
Q and A