Viirya

"Eigenfaces for recognition"

By Matthew Turk, Alex Pentland

--

Computational models for face recognition contribute not only to theoretical insights but also practical application. However, it is a difficult problem.

The eigenface approach proposed by Matthew Turk and Alex Pentland in this paper contributes a simple yet mathematically well-formed model based on solving eigenvectors of covariance matrix obtained from training face images.

This work turns its attention to the information content of face images, emphasizing the significant local and global features that may not be directly related to our intuitive notion of face features such as the eyes, nose, lip, and hair.

In mathematical terms, the approach called eigenface is to find principal components of the distribution of faces. For implementation, that are the eigenvectors of the covariance matrix of the set of face images. These eigenvectors then can be thought of as a set of features that characterize the variation between face images. Since each image in training set contributes more or less to each eigenvector, we can display the eigenvector as a sort of ghostly face we call an eigenface.

The approach of this work comprises two processes: training and recognition.

Training:

1. Acquire an initial set of face images for training.

2. Calculate the eigenfaces from the training set, keeping only the M images corresponding to the highest eigenvalues. These M images define the face space.

3. Calculate the corresponding distribution in M-dim weight space for each image in training set, by projecting the face images onto the face space.

Recognition:

1. Project the input image onto the face space to obtain a set of weights based on the input image and the M eigenfaces.

2. Classify the weights as either a known person or as unknown.