Multi-lingual character segmentation and recognition based on adaptive projection profiles and composite feature vectors

In recent years, multi-lingual character segmentation and recognition has attracted the wide range of researchers. However, the variations in the structures of characters in different languages, writing styles of different people and font sizes challenges that need further enhancement. Towards such objective, in this paper, we have proposed a novel multi-lingual handwritten character recognition framework for three different languages such as Bangla, Kannada and Telugu. At segmentation phase, this framework performs both word and character segmentation with the help of an Adaptive Projection Profiling (APP) and Edge Density Filter (EDF) respectively. Further at the recognition phase, we propose to use gradient based feature descriptors to extract a Composite Feature Vector (CFV) from handwritten character images, which are then fed to Support Vector Machine (SVM) algorithm for recognition. At experimental evaluation, we have simulated the proposed model over three different language scripts. The experimental results show that the proposed model outperforms the conventional method with an average improvement in the recognition accuracy of 3% for both cross validation and test simulations.