Landmarks-SIFT face representation for gender classification

Abstract

Existing methods for gender classification from facial images mostly rely on either shape or texture cues. This paper presents a novel face representation that combines both shape and texture information for gender classification. We propose extracting the Scale Invariant Feature Transform (SIFT) descriptors at specific facial landmarks positions, hence encoding both the face shape and local-texture information. Moreover, we propose a decision-level fusion framework combining this Landmarks-SIFT with Local Binary Patterns (LBP) descriptor extracted for the whole face image. LBP is known of being tolerant against uncontrolled image capturing conditions. Competitive correct classification rates for both controlled (97% for FERET) and uncontrolled (95% and 94% for LFW and KinFace) benchmark datasets were achieved using our proposed decision-level fusion. © 2013 Springer-Verlag.