As an appearance-based approach, eigenface recognition method has several advantages: (1) Raw intensity data are used directly for learning and recognition without any significant low-level or mid-level processing; (2) No knowledge of geometry and reflectance of faces is required; (3) Data compression is achieved by the low-dimensional subspace representation; (4) Recognition is simple and efficient compared to other matching approaches.
These advantages reflect the power of appearance-based approach in ease of implementation. However, the experimental results also demonstrate some serious limitations of eigenface representation method for face recognition under different conditions.
First, the method is very sensitive to scale, therefore, a low-level preprocessing is still necessary for scale normalization. Secondly, since the eigenface representation is, in a least-squared sense, faithful to the original images, its recognition rate decreases for recognition under varying pose and illumination. The fisherface projection approach [1] is aimed to solve the illumination problem by maximizing the ratio of between-class scatter to within-class scatter, however, finding an optimum way of projection that is able to simultaneously separate multiple face classes is almost impossible. Third, though the eigenface approach is shown to be robust when dealing with expression and glasses, these experiments were made only with frontal views. The problem can be far more difficult when there exists extreme change in pose as well as in expression and disguise. Fourth, all the face images tested in the experiments are taken with a uniform background. However, this condition may not be satisfied in most natural scenes, which will seriously deteriorate the recognition performance. In such cases, a segmentation process has to be considered.
Additionally, the eigenface recognition method bears some common disadvantages due to its ``appearance-based'' nature. First, learning is very time-consuming, which makes it difficult to update the face database. Second, recognition is efficient only when the number of face classes is larger than the dimensions of the face space; otherwise, the projection of an unknown image requires pixel-by-pixel multiplication of the input image by all eigenfaces, which is equivalent to or worse than template-matching with respect to computation time since an extra distance calculation is needed in the subspace. However, the occurrence of class overlapping increases when more face classes are represented by the same face space, thus lowering the recognition rate.
Wei Sun