Relevant mRMR features for visual speech recognition

To improve the accuracy of visual speech recognition systems, forming a subset of relevant visual features, from a large set of extracted visual cues, is of fundamental importance. In this paper, two feature selection techniques, Principal Component Analysis (PCA) and a relatively recent method, Minimum Redundancy Maximum Relevance (mRMR), are separately applied on the extracted visual features. Prominent attributes are selected by each to form a feature vector for classification. Experimental results show that recognition accuracy for an isolated word database is not affected when a few selected mRMR features from the complete visual feature set are used for classification. This considerably reduces computation and storage overheads. It is also seen that features determined by mRMR perform better than PCA features. Both techniques yield inner mouth area segments as principal features as compared to other geometrical parameters.