Automatic Synthesis of Training Data for Sign Language Recognition Using HMM
The paper describes a method of synthesizing sign language samples for training HMM. First face and hands regions are detected, and then features of sign language are extracted. For generating HMM, training data are automatically synthesized from a limited number of actual samples. We focus on the common hand shape in different word. The database hand shapes is generated and the training data of each word is synthesized by replacing the same shape in the database. Experiments using real image sequences are shown.
KeywordsTraining Data Hide Markov Model Sign Language Sign Language Recognition Hand Shape
Unable to display preview. Download preview PDF.
- 1.Kinscher, J., Trebbe, H.: The Munster Taging Project - Mathematical Background, Arbeitabereich Linguistik, University of Munster, D-58149 Munster (1995)Google Scholar
- 2.Toda, T., Kawai, H., et al.: Unit Selection Algorithm For Japanese Speech Synthesis Based on Both Phoneme Unit and Diphone Unit. In: Proc. Int. Conf. Acoustics, Speech and Signal Processing (ICASSP 2002), pp. 3088–3091 (2002)Google Scholar
- 4.Fukuda, Y., Kimura, H., et al.: Expression through hand and finger actions used in ‘words in continuous signing’: A study in reference to the dialogue in Japanese Sign Language by the deaf persons. Technical Report of IEICE (in Japanese) 97(586), 81–85 (1998)Google Scholar