Temporal series recognition using a new neural network structure T-CombNET

We present a new neural network structure dedicated to the temporal data series recognition, called T-CombNET (Temporal CombNET). The T-CombNET is a modified and improved version of the CombNET-II structure, which was designed to deal with very large vocabulary character recognition. We incorporated a time normalization stage and used recurrent neural networks to enable CombNET-II to do the temporal analysis. Then we applied this new model T-CombNET to the hand gesture recognition task. The preliminary results show that the proposed structure is very efficient, obtaining a recognition rate of 99.41% in visual based kana hand alphabet recognition experiments.