Face classification with eigenfacesMAS622J/1.126J: Pattern Recognition and Analysis, Fall 2002 Class Project
Face Classification with Eigenfaces
Seha Kim Dec. 5, 2002
<<|>>
Face classification with eigenfaces
Nearest neighbor classification:
- gender, age, race, and facial expression classification
Parzen window:
- facial expression classification
|
|
Classes |
|
Gender |
male / female |
|
Age |
child / teen / adult / senior |
|
Race |
white / Hispanic / Asian / black / other |
|
Facial expression |
smiling / serious / funny |
|
Properties |
beard / non-beard, moustache / non-moustache |
Intensity data of the face image
Dimensionality reduction with PCA
Eigenfaces:
- Principal components of the distribution of faces
- Eigenvectors of the covariance matrix of the set of face images
- Characterize the variation between face images
2000 training data, 2000 test data
128*128 dimensional -> 99 dimensional
Selecting optimal k:
- Evidence curve v(k):
As a result, k = 1 is most appropriate to all cases
gender case age case >>
k=1 at maximum k=1 at maximum
race case facial expression case
k=1 at maximum k=1 at maximum
k=1
à
Nearest neighborhood method
Nearest neighborhood classification result
|
|
Accuracy rate =(#correctly classified)/(#data) |
|
Gender |
0.61 |
|
Age |
0.68 |
|
Race |
0.80 |
|
Facial expression |
0.52 |
|
Beard |
0.06 |
|
Moustache |
0.002 |
- Gender, age and race cases: Fairly well
- Facial expression, beard, and moustache cases:
Inadequate à Parzen window
Facial expression classification with Parzen window
Assumption of Gaussian window
Selecting optimal h with evidence curve v(h)
smiling face serious face funny face
h = 0.686 h = 0.486 h = 0.440
Facial expression classification with Parzen window
- Among the three classes, the highest density is selected
- Result
|
|
Accuracy rate |
|
Smiling face |
0.06 |
|
Serious face |
0.25 |
|
Funny face |
0.75 |
- Funny face: Considerably good result
- Smiling and Serious face: Too low accuracy
- Need another method: Simple dimensionality reduction
Dimensionality reduction in simple way
- Selecting the size of dimension of feature
just clipping the data
Facial expression classification with Parzen window
Various dimensionalities
smiling face serious face funny face
9 dimensions 4 dimensions 99 dimensions
- Need another features rather than the intensity of the image:
For example, the features can be the size of the mouse,
the height of the eyebrow, the size of the eyes and so on.
Facial expression classification with Parzen window
Comparison
|
|
Accuracy rate w/ 99-dimension |
Accuracy rate w/ optimal dimensionality |
|
Smiling face |
0.06 |
0.54(d=9) |
|
Serious face |
0.25 |
0.31(d=4) |
|
Funny face |
0.75 |
0.75(d=99) |
Facial expression classification with Parzen window
Comparison between Parzen window and the KNN
total accuracy = (# correctly classified/2000)=0.42
accuracy of nearest neighborhood = 0.52
In many papers to classify the facial expression, they used to compare the
facial templates such as nose width and length, mouth position and chin type.
Using templates can improve performance over the face-only templates.
In this project, we pursued the face classification. For classification, we applied nearest neighbor method to determine gender, age, race or facial expression and applied Parzen window technique to classify facial expression classes.
- In classification of gender, age and race classes, the nearest neighborhood method produced fair results.
- The large dimensional data does not always produce better result in the facial expression case. Also to classify the facial expression, we need not only the face-templates but also the templates of eyes, nose, mouth and so on.
- The human brain is a very fine face classifier. :-)
<<|>>