Weighted Modular Image Principal Component Analysis for face recognition

• We propose two feature extraction methods for face recognition: MIMPCA and wMIMPCA.
• The proposed methods use modular PCA to minimize local variation.
• The proposed methods deal with changes in illumination and head pose.
• The proposed methods obtained better results compared with other methods.

This paper proposes two feature extraction techniques that minimizes the effects of distortions generated by variations in illumination, rotation and, head pose in automatic face recognition systems. The proposed techniques are Modular IMage Principal Component Analysis (MIMPCA) and weighted Modular Image Principal Component Analysis (wMIMPCA). Both techniques are based on PCA and they use the modular image decomposition to minimize local variation. Also, the covariance matrix is calculated directly from the original image matrix. This strategy generates a smaller matrix compared with traditional PCA and reduces the computational effort. wMIMPCA assumes that parts of the face are more discriminatory than others, so a Genetic Algorithm is used to obtain weights for each region in the face image. The proposed techniques are compared with Modular PCA and two-dimensional PCA using three well-known databases, showing better results.