Computer Vision Demonstration Website

Electronics and Computer Science
University of Southampton

Face Recognition

There are many different approaches to face recognition, one of the more popular methods being the eigenface approach, which is demonstrated below. The approach is popular because it has small memory requirements and is robust to change in pose, and partial occlusion such as facial hair and glasses.

How it works

Before face detection can take place a 'training set' of face images is required, from which the eigenface algorithm extracts statistical features, which are later used for classification.

The first stage of training is to take all of the images in the training set and construct an average image. Difference images are then constructed by subtracting the average image from each original image.

The final stage of the training phase is to create an 'eigenface' for each image, which is a normalised eigenvector that is calculated from the difference images.

After the training it is possible to classify new images. For each new image we can say with a degree of certainty if it is the same face as one in our training set, or if the image is a face or not. The Euclidean distance function is used to compare one eigenface with another:

Links

Mark Nixon & Alberto Aguado, 2002, Feature Extraction & Image Processing, Newnes

ECS | Feature Extraction & Image Processing | © 2005 University of Southampton