Automatic recognition of finger spelling for LIBRAS based on a two-layer architecture

Abstract

Different feature extraction techniques have been applied to the problem of automatic finger spelling (and gesture) recognition problem. However, different hand postures and gestures with different complexities have been given the same space representation. Our approach tries to get rid off the assumption that one size fits all. A two level architecture was investigated where signs with similar hand postures were grouped together for a preliminary artificial neural network (ANN) classification. A second ANN was applied to disambiguate the confusions among symbols, using another space representation. Our results indicate that it is possible to improve recognition rates with this approach.