Face Recognition: Robustness of the ‘Eigenface’ Approach

While face recognition is a fairly trivial task for humans, much of computer vision research has been dedicated to finding an algorithm to teach a computer how to recognize faces. This paper discusses the robustness of the Turk and Pentland ‘Eigenface’ algorithm [1]. The algorithm consists of two stages, the learning stage, which is done offline, and the recognition stage, which is done online. The learning stage consists of making a database of the principal components of all the images in the training set to which new images can be compared to. This database is called the “face space”. The recognition stage projects each new image of a face onto the “face space”, using principal component analysis, and compares it to known faces from the training set to find the best match. The algorithm has a high recognition rate when the images of the faces were upright and had similar lighting and feature conditions to the training images. By feature conditions, we mean smiling, winking, glasses, etc... However, once the faces were tilted, the lighting was changed or the face was obscured somehow, the algorithm suffered from a loss of recognition.