Quantitative study of facial expression asymmetry using objective measure based on neural networks

Previous studies have been reported that facial expressions on the left side of face appear stronger than these on the right side. We described an algorithm of an effective feature selection method based on supervised learning of multi-layer neural networks for facial expression recognition. We extracted the emotion masks focusing on perceptually significant pixels in a face image by using exhaustive searches based on the backward feature selection method. It provided an objective measure for evaluating the facial asymmetry. We demonstrated effectiveness of our approach in qualitative experiments for rating the asymmetric facial expressions. In the experiment, the left-right asymmetry of facial expressions has been proved objectively by using perceptually significant pixels within the emotion masks. The facial expression recognition rate using the emotions masks was improved from 78.8% to 83.1%.