Human mouth-state recognition based on learned discriminative dictionary and sparse representation combined with homotopy

Abstract

In order to detect the number of audio sources and improve the speech recognition capability of an intelligent robot auditory system, recognizing human mouth-states, open or closed, is studied in this paper. A discriminative dictionary and sparse representation combined with homotopy based human mouth-state recognition algorithm is proposed. In the algorithm, a label consistent K-SVD (LC-KSVD) algorithm is used to learn a discriminative single over-complete dictionary and an optimal linear classifier simultaneously. Meanwhile, homotopy algorithm is used at the sparse decomposition stage. Experiments are carried out with the database established with the ROI images localized and extracted from the face images downloaded from Google online. Compared with several state-of-the-art methods, the proposed method obtains higher classification rates (CRs), costs less time for recognizing a test sample and has good noise immunity performance. Particularly, superior performance is attained when the training samples are extremely limited, even one sample per class.