Increasing manual sign recognition vocabulary through relabelling

In this paper we present the results of relabelling a self organizing map (SOM) to increase the dynamic manual signs it can recognize. Relabelling exploits the global ordering of self organizing map and abrogates the need for retraining, thereby reducing the computational costs and increasing the recognition ability of the network. This relabelling technique was applied to a dynamic sign recognition system to increase the recognition vocabulary from 10 to 14 signs. The data was collected from a person wearing a DataGlove with a Polhemus sensor and signing the 14 signs. The sampled hand data over the duration of sign was fed to phonemic recognition modules and the collective outputs of these modules were fed to the sign recognition module consisting of a relabelled self organizing network. The results showed that the overall recognition rate of the relabelled network was 84% as compared to 86% for the retrained network. Further, it was found that the dynamic sampling of the signs made the movement phoneme module unnecessary