Combining ICA Representations for Recognizing Faces

Abstract

Independent Component Analysis (ICA) is a generalization of Principal Component Analysis (PCA), and it looks for components that are both statistically independent and non-Gaussian. ICA is sensitive to high-order statistic and it expected to outperform PCA in finding better basis images. Moreover, with face recognition, high-order relationships among pixels may have more important information than those of pairwise relationships on which base images found by PCA depend. Two different representations can be applied by ICA; ICA architecture I and ICA architecture II. A new classifier that combines the two ICA architectures is proposed for face recognition. By the new classifier, the similarity measure vector was employed in which the similarity measure vectors for both ICA representations were resorted in descending order and then integrated by merging the corresponding values of each vector. The new classifier was performed on face images in the AR Face Database. Cumulative Match Characteristic was taken as a measure for evaluating the performance of the new classifier with illumination variation, expression, and Occlusion. The proposed classifier outperforms both ICA architectures in all cases especially in later ranks. Keywords— ICA; PCA; face recognition; ICA representations