Recurrent neural networks based Indic word-wise script identification using character-wise training.

摘要: This paper presents a novel methodology of Indic handwritten script recognition using Recurrent Neural Networks and addresses the problem in poor data scenarios, such as when only character level online is available. It based on hypothesis that curves comprise sufficient information for prediction at word level. Online used to train RNNs BLSTM architecture which are then make predictions data. These results test set par with models trained data, while training model much less intensive takes time. Performance binary-script 5 reported, along comparison HMM models.The system extended offline prediction. Raw lacks temporal available required To overcome this, stroke recovery implemented strokes utilized predicting models. The performance reported.