Relative frequency maps for face recognition

Single face-image comparisons are extremely challenging, particularly in the context of pose, expression variations and scene illumination changes. Most of the existing schemes are sub-space learning based, where dominant eigen-directions are determined from the covariance matrix computed over the entire face space. In this paper we propose a simple hashing method based on the relative magnitudes of selective frequencies obtained from the intensity histogram and use this indicator function as an elastic representation of the face, termed as the Order Preserving Selective Relative frequency map (OPSRFM). Despite being a histogram derivative, the OPSRFM has been found to be robust to contrast stretching operations and pose variations in faces, while remaining discriminative across face classes. Recognition rates obtained for the ORL and YALE databases were 87.63% and 76.36% respectively which are comparable to computationally intensive sub-space learning and hashing methods.