Human Emotion Classification: An Expression Specific Geometric Approach

Abstract

Human face emotions are generally classified in six different expressions such as Anger, Disgust, Fear, Happiness, Sadness, and Surprise. The authors propose a novel method for selecting an expression specific set of salient landmark points out of 68 landmark points produced by applying an Active Appearance Model (AAM) on an input face image. The salient Landmark points are selected by training a MultiLayer Perceptron network using a Histogram oriented Gradient (HoG) feature of neighboring pixels of a Landmark point. Next, a shape signature vector is constructed by forming triangulation using those salient landmarks for each expression. This is trained with six Multilayered Perceptron (MLP) network for classification of each of the six basic expressions. The suggested algorithm is tested on CK+, JAFFE, MMI, and MUG database. The outcomes are found extremely promising.