Strike off removal in Indic scripts with transfer learning

Strike-off text poses major challenges in handwritten text recognition as it changes the semantic and structural information of the image. Although significant results have been achieved in identifying and removing such strike-off data using deep learning methodologies, most have been done for Roman scripts only. Deep learning approaches require a large amount of data with a high cost of training for every script individually to derive effective performance. Due to its complex nature and non-availability of sufficient data, research in strike-off removal in Indic scripts is limited. To address this problem, we propose reducing the requirement of a huge amount of data and minimizing the training cost through transfer learning. With the objective of strike-off removal in multiple Indic scripts, we leverage the experiences of a pre-trained model (trained on the Roman script) for strike-off removal in different domains (Indic scripts). We consider handwritten text documents of 10 different Indic scripts and introduce 7 different strike-offs in these documents. We implement Few-Shot Learning (FSL) and Zero-Shot Learning (ZSL) to train various state-of-the-art deep generative models on a few samples of the mentioned Indic texts. An extensive analysis of the results for ZSL and FSL has been presented with the perspective of source hypothesis generalization capability and the strength of relatedness of source and target domains. The results show that the degree of adaptability of the source hypothesis is significant for the right amount of transfer to take place. The scripts with angular structure have performed better than the round structured scripts as there is a higher degree of relatedness of angular scripts with the Roman script (source script). FSL and ZSL approaches promise to reduce data requirements and training costs for strike-off removal.