Handwritten Script Identification: Fusion based Approaches

Script identification is one of the preprocessing steps in any document image processing task. Script identification in printed documents has achieved a greater attention whereas script identification in handwritten documents has achieved less attention from document research community. Almost all the existing works have made attempts on identifying suitable features or classifiers for handwritten script identification. On the other hand, currently an attention is given towards fusion of features and/or classifiers. To the best of our knowledge only a few works are reported on study of fusion strategies for script identification of handwritten documents. In this work, we present two different models, based on fusion strategies for handwritten script identification of trilingual documents (Kannada, Hindi and English). The first model is feature level fusion where Gabor, LBPV and wavelet features are fused and second model is on decision level fusion where decisions from nearest neighbor and the support vector machine classifiers are used for fusion. Experimentation has been carried out on a data set of three classes each with 100 documents and experimental results ensure that fusion strategies perform better than conventional methods.