11/08 Notes¶
Face Recognition¶
Final Project is on Face Recognition
- What DPI is best
- Algorithm takes longer the higher the DPI
- Higher DPI can detect more differences
- A balence is required, depends on what you are doing
- Level 1
- General Shape
- Skin Tone
- Level 2
- Facial Components
- eyes/nose/mouth
- Relationship between the Components is revealed
- have 30 to 75 IPD
- IPD is Inter Pupilary Distance - Distance between centers of the 2 eyes
- the more features you expose, the more vulnerable the image is to intra class variations
- Sad vs happey is an example
- Low resolution images are affected by this less than high resolution
Most use Level 2
- Level 3
- Micro Features, which are vulnerable to intra class variarions as well
- Scars, freckles, skin discoloration, etc...
Face Detection¶
- Uses Adaboost
- set of weak classifiers that are combined
Input Patch (x) –> Classifier (h) –> Face(y=1) or Not Face(y=0)
Combine Weak Classifiers¶
Classifier = small red window
- Called Cascade Classifiers
- h_1(x) = 0 then non face
- h_1(x) = 1 then go to h_2(x)
- etc..
h_i( input patch, haar filter, polarity, theta)
polarity can be 1 or -1
this should output 1 or 0
1 if filter(x) > p x theta
0 if filter(x) <= p x theta
160,000 filters/classifiers to use
Training¶
Algorithm is in the book
- goal of Training
- Select T Classifiers
- given set of examples - face or not face
- Assign a weight to each of the examplars
- Classifiy the examplars using each of the 160,000 classifiers
- select the Classifier that gives the lowest weighted error rate
- revise the weight of the examplars as follows
- correctly classified examplars - no weight change
- incorrectly classified examples - increase weight
- Loop back to Classifiy the examplars using now 159,999 classifiers
- continue till it covers everything? or hits T selected classifiers
- SHOULD GO OVER INTEGRAL IMAGES
- sum of pixels above and to the left of a chosen pixel