Visual speech recognition from lip image sequences using a neocognitron-type neural net

摘要

This paper describes a neocognitron-type neural net model for discriminating image Sequences. This model is constructed by extending the architecture of our proposed Neocognitron-type Temporal Data Recognition Model. The structure, in which cells are arranged on three-dimensional arraies, allows the model to discriminate image sequences like lip movies. We apply the model to visual speech recognition from lip image sequences for one subjects saying names of train stations: 'Shinjuku', 'Ikebukuro', 'Tokyo', 'Shinagawa'. The model gives good discrimination result for each category.