Comparing Robustness of Two-Dimensional PCA and Eigenfaces for Face Recognition Springer
Résumé
In this paper, we aim at evaluating the robustness of 2D-PCA for face recognition, and comparing it with the classical eigenfaces method. For most applications, a sensory gap exists between the images collected and those used for training. Consequently, methods based upon statistical projection need several preprocessing steps: face detection and segmentation, rotation, rescaling, noise removal, illumination correction, etc... This paper determines, for each preprocessing step, the minimum accuracy required in order to allow successful face recognition with 2D-PCA and compares it with the eigenfaces method. A series of experiments was conducted on a subset of the FERET database and digitally-altered versions of this subset. The tolerances of both methods to eight different artifacts were evaluated and compared. The experimental results show that 2D-PCA is significantly more robust to a wide range of preprocessing artifacts than the eigenfaces method.