Face Detection

• Find Faces → Discriminate between faces/non faces
• uses
  • preprocessing step to Face Recognition
  • Autofocus, Exposure, color balance in photography
  • surveillance, biometrics, monitoring
• invariant to
  • scale
  • rotation
  • illumination

Haar Features

a.k.a. Viola Jones Algorithm

slide windows of different scales across the image, extract features and classify the features as face/non-face

Haar Filters:

• Images are convolved with a set of Haar filters (24x24 base size) across different filter scales
• 
• Haar features are sensitive to direction of the filters
• The filters are themselves simple black and white (+1 and -1) filters
  • the convolution can be simplified to simple summation
  • 
  • filters are meant to capture certain facial features

Integral Image

• table that holds the inclusive sum of all pixel intensity values to the left and top of a given pixel
• 2D cumulative sum
• 
• integral images Allows for fast summation of image intensity of arbitrary rectangles
• 
• Can be used to quickly compute Haar filters
• 
• The computation cost remains constant regardless of the size of the Haar filter

Classification

• Haar features around a pixel need to be classified as face/non-face → Classification problem
  • Nearest Neighbour Classifier
    • large dataset needed for robustness
    • slow and computationally expensive
      • large dataset comparison
      • have to do it for every scale
  • Support Vector Classifier
• Large number of features
  • use Adaboost and Cascading to optimize classification

Adaboost

see Adaboost

• eliminate redundant/irrelevant features
• weak classifier: a relevant feature that does better than random guessing
• Adaboost constructs a strong classifier as a linear combination of weak classifiers

Cascading

• Strong features are formed into a binary classifier
  • +ve matches are sent along to the next feature
  • -ve matches are rejected
• Reduces the amount of computation time spent on false windows