Abstract:
Two-dimensional principal component analysis (2DPCA) and its variants have been successfully used for the task of face recognition (FR). However, one of the major limitations of 2DPCA-based FR methods is that they only consider the holistic information of a given training dataset, ignoring both class-specific discriminant information and class-separation components, which could further improve recognition performance. To address this limitation, this paper suggests a class-wise 2DPCA (CW2DPCA) framework that seeks to model class-specific subspaces, where each subspace retains the discriminatory information of a particular class, as well as class separability information. In this way, CW2DPCA not only feeds discriminative representations of facial images to the classification model, but also enables a high degree of separation between the different classes present in the training dataset. Experimental evaluation on two face datasets proved the effectiveness of the CW2DPCA in FR.