Sparse reconstruction of facial expressions with localized gabor moments

Facial expression recognition depends on the detection of a few subtle facial feature traces. EMFACS (Emotion Facial Action Coding System) is a taxonomy of face muscle movements and positions called Action Units (AU) [1]. AUs can be combined to describe complex facial expressions. We propose to (1) deconstruct facial expressions into face regions, grouping AUs by their proximity and contour direction; (2) recognize facial expressions by combining sparse reconstruction methods with face regions. We aim at finding a minimal set of AU to represent a given expression and apply l₁ reconstruction to compute the deviation from the average face as an additive model of facial micro-expressions (the AUs). We compared our proposal to existing methods on the CK+ [2] and JAFFE datasets [3]. Our experiments indicate that sparse reconstruction with l₁ penalty outperforms SVM and k-NN baselines. On the CK+ dataset, the best accuracy (89.8%) was obtained using sparse reconstruction.