Facial Emotion Recognition Using Active Shape Models and Statistical Pattern Recognizers

This paper investigates various emotion recognition techniques from the facial expression of human subjects. To describe human facial expressions, a number of characteristic points are extracted from input face images using active shape models (ASMs), and translated 49 scalar features so that they are invariant to scale and position changes. The scalar feature values then construct a 49-dimensional feature vector for each still image. Statistical pattern recognizers, such as support vector machine (SVM) and multi-layer perceptron (MLP), are used to identify various emotions from the feature vectors. To analyze the performances of the various pattern recognizers on the limited amount of image data, 5-fold cross-validation is carried out, with varying numbers of emotions from 3 to 6. Evaluation results show that SVM is the most stable and best in terms of emotion classification rates.