Brunelli, R. & T. Poggio (1993), Face Recognition: Features versus Templates, IEEE Transactions on PAMI, 15(10):1042-1052.

@inproceedings{Brunelli:91IJ,
  AUTHOR = {R. Brunelli and T. Poggio},
  TITLE = {HyperBF Networks for Real Object Reccognition},
  YEAR = 1991,
  BOOKTITLE = {Proc. of the 12th IJCAI},
  ADDRESS = "Sidney, Australia",
  PAGES = {1278-1284},
  KEYWORDS = {}}
}

Author of the summary: Jim Davies and Alexander Stoytchev, 2000, {jimmyd, saho}@cc.gatech.edu

Cite this paper for:

• the template approach beats the feature based approach for face recognition.
• The eyes are the most discriminating part of the face.

Summary:

The results show that the second approach (template matching) performs better for the database of faces that was used (these results should be applied to a larger database.) It is also simpler.

As a general conclusion the authors suggest that successful face recognition systems should use a combination of the two.

Detailed Outline:

Feature-based approach: Find the distances and sizes of features in the face (eyes, distance to nose, etc.) and match on those feature properties.

Template approach: Match data to templates, stored as images themselves (an array of pixels with values)

Feature approach:

• normalize: each pixel is divided by the average intensity over a neighborhood around that pixel (neighborhood size not specified.)
• Eye templates at 5 different scales used to locate the eye in an image using hierarchical correlation.
• Image scale is preadjusted.
• Feature Extraction using the following fundamental face constraints:
  • bilateral symmetry
  • 2 eyes
  • 1 nose
  • 1 mouth
• 35 features extracted (35d numerical vector)
  • mouth: look for peak in the horizontal projection of the vertical gradient
  • mouth width: threshold the vertical projection of the vertical gradient at the average value
  • nose: look for peak inthe vertical
  • eyebrows: thickness and vertical position at the eye center
  • chin shape/face outline (11 features!)
• recognition
  • a covariance matrix is constructed
  • The distances from the test image to the database of images are found. Nearest neighbor is the match.
  • The robustness of the classification is evaluated with the Min/Max metric (minimal distance to a wrong correspondence over the maximal distance to a correct correspondence.) A minmax of over 1 is a perfect match. Each class is the point in the space that is the average of all in the class.
  • as size of db increaces, performance and minmax decrease monotonically. (fig 9)
  • a rejection threshold may improve performance.

• normalize as above
• each person is represented by a set of 4 masks: eyes, nose, mouth and face (region from eyebrows to below chin). Their location is relative to eye position.
• Correlations are found for each feature. (all 4) The new face is matched with the highest cumulative score.
• Correlation is sensitive to illumination gradients. 4 different normalization techniques were tried and compared: Gradient information was the best.
• Recognition as a function of the image resolution was tested. A 4 level Gaussian pyramid was constructed and was found that for the the first 3 levels recognition was still possible with templates as small as 36x36 pixels.
• Feauture performance by mask: (decreasing order)
  1. eyes
  2. nose
  3. mouth
  4. whole face
  The similarity scores for each feature were added to produce the global score.

Summary author's notes:

• none