Facial Expression Analysis Based On Visual Perception Mechanism

Affect analysis would facilitate research in areas such as behavior science,anthropology, neurology and psychiatry. And various application areas andtechnologies could benefit from efforts to affect analysis such as intelligentsurveillance, intelligent human-machine interaction, emotional cues analysis,multimedia retrieval, intelligent toys and games. Therefore, many theories andalgorithms on facial feature extraction and facial expression classification have beenproposed in recent years.This thesis studies the biologically inspired facial expression recognition based onvisual perception mechanism. Particularly,(1) it proposes sparse representation basedfacial expression analysis algorithm considering compression sensing theory, andapplies the algorithm for the analysis of mixture facial expression and facialexpression space;(2) it proposes discriminant dictionary learning (D-KSVD) basedfacial expression recognition algorithm;(3) it proposes Hessian regularized facialexpression recognition algorithm considering manifold learning theory;(4) itproposes a HMAX model for facial expression recognition. The details are presentedas follows:1. Sparse representation based facial expression analysis consists of facialexpression recognition based sparse representation, mixture facial expressionanalysis and facial expression space analysis. The facial expression can beexpressed as the sparse linear combination of an over-completed expressiondictionary which consists of all facial expression classes. The maximumcoefficients determine the facial expression class. A mixture facial expressioncan be regarded as the combination of several facial expression classes. Andfacial expression space can also be analyzed via the sparse coefficients.2. Facial expression recognition based on dictionary learning. It has been foundthat the sparse coefficient is difficult to obtain when the number of facial expression class increases and unsupervised dictionary learning lacksdiscrimination. This thesis incorporates label information into K-SVD andproposes facial expression recognition based on D-KSVD to significantlyimprove the ability of representation and discrimination.3. Facial expression recognition based on manifold learning. Manifold learningis one of the best solutions to cover the great number of unlabelled examplesin training. Considering the disadvantage of Laplacian regularization whichlacks exploitation, this thesis proposes Hessian regularized support vectormachine for facial expression recognition. Experimental results demonstratethat Hessian regularization can significantly improve the performance.4. Facial expression recognition based on HMAX model. HMAX is one of theexcellent biologically inspired computational models. However, HMAXcannot be directly implied to facial expression recognition. Considering thecharacteristics of facial expression, this thesis proposes a new HMAX modelto extract facial expression features and conducts facial expression recognitionexperiments on different facial expression datasets. Experimental results showthat the proposed HMAX can extract general facial features and achieve wellperformance on different facial expression datasets.