Feature selection for facial expression recognition based on optimization algorithm

This paper presents a wrapper approach to feature selection from image sequences and applies it to the facial expression classification problem. The pre-processing phase automatically scans image sequences and detects frames with maximum intensity of facial expression. The features are generated using the log-Gabor filters. A global optimization algorithm genetic algorithm (GA) is adopted to select a sub-set of features based on minimization of the classification error. The wrapper approach is compared with two previously known filter-based feature selection methods: MID-mRMR and MIQ-mRMR. The features are classified using the naive Bayesian (NB) classifier. The average classification rates are: 79% (MIQ-mRMR), 78% (wrapper) and 64% (MID-mRMR). The results from the filter methods did not appear to be significantly effected by the size of the feature subset.