Video Key Techniques Of Facial Expression Recognition

In recent years, researchers have paid more attention on facial expressionrecognition, and have achieved some success. For the reason that the face is a non-rigidorgan and facial expression is sensitive to age, gender, race, hair, accessories, lights,there is not an accurate and efficient recognition algorithm.This thesis researched the feature extraction algorithms and the classificationalgorithm of facial expression. Because the facial expression recognition based onsingle image has the limitation of information, we focus on the image sequenceprocessing in this thesis.At first, this thesis proposed a new feature extraction method by using bothdynamic and static extraction algorithm, on the study of traditional algorithms. In ouralgorithm, we track the AAM feature points in expression image sequence and recordtheir trajectories, which are dynamical feature. The LPQ histogram of Local PhaseQuantization is also extracted as static feature.Secondly, we do not use Traditional Action Unit (AU) based local feature methodbecause it may generate unreasonable output, this thesis proposed an organ-basedmethod that can handle this problem. Yes or no classifier just gives a binary result in aclassification problem that lead to the loss of information, we use SVM andBradley-Terry Model to predict the probability of each facial organ’s action which cankeep more classification information.This thesis studies the learning and inference algorithms of Bayesian Network. Webuild a single Bayesian Network with just7units to identify the6basic expressions(happy, surprise, angry, disgust, sad, fear).Experiments on Cohn-Kanade database shows that our algorithm can recognizeexpression in real-time and can achieve90%recognition ratio. The algorithm output isthe probability of each expression, which can provide richer information to emotioncomputing than before.