Abstract

One class of computational approaches to facial expression recognition focuses on mapping face images into six basic emotions: anger, disgust, fear, joy, sadness, and surprise. When tested in parallel with human behavioral data, these models can provide useful information about which attributes of visual input are used by humans to make judgments of emotional expression. Of particular interest are models in which similarity is not built into the model but is learned in the mapping from visual input to expression category. The model we are testing uses support vector machines to classify spatial-frequency filtered (Gabor) face images into the six basic emotions above plus neutral, without a predefined similarity schema or predefined facial features. The model achieves 93% accuracy on multiple datasets, and performs in real-time. Because the model makes a 7-way forced choice, it provides judgments that at times are mixtures between facial expressions, allowing the model to make implicit similarity judgments. Comparing model and human performance on generalizing to expressions such as contempt provides information about whether expressions not present in the training set can be represented as combinations of the visual features that make up the pre-defined categories. Expressing contempt as a mixture of basic emotions is of particular interest in the controversy over whether contempt is itself a basic emotion or a subset of emotions such as anger and disgust.