Signer-independent sign language recognition based on manifold and discriminative training

Signer-independent sign language recognition is an urgent problem for the practicability of sign language recognition system. Currently, there is still a huge gap between signer-independent sign language recognition and signer-dependent sign language recognition system owing to the variation of sample data and the insufficient of training samples. Discriminative training can well compensate the recognition shortages caused by insufficient training samples and the similarity of sign language models. This paper applied the HMM (hidden Markov model) training parameter model modified by DT (discriminative training) method to recognize signer-independent sign language. The modified HMM model reduced the effects of small training samples to signer-independent sign language recognition. Furthermore, this paper proposed a tangent vectors method of manifold (TV/HMM) concept to improve the statistical model of sign language recognition considering the learning and reasoning capability of manifold concept in sign language recognition field. The proposed model reduced the variation impact of sign language to signerindependent sign language recognition. Finally, a novel statistical training model (DT+TV/HMM) combining discriminative training and manifold methods was established to solve the data variation and small samples problems of sign language recognition system. Experiments show that the discrimination rate of the integrated DT+TV/HMM recognition system is 82.43% increasing by 15.07% compared with traditional MLE recognition system.