Towards Accurate Identification and Removal of Shirorekha from Off-line Handwritten Devanagari word Documents

Shirorekha identification and removal is an important and a challenging pre-processing stage in almost all machine interpretations for handwritten Devanagari documents. Within this area of investigation, all studies are designed based on traditional image processing techniques. Which are mainly based on hand-engineering and learn local transformations only. However, it can also be viewed as a supervised classification task in which each pixel, in a document, is examined/ queried so that those classified as shirorekha are removed. For this purpose, we extended this area of investigation by designing an encoder-decoder based convolutional neural network (EDCNN). Which have demonstrated, from various studies, that they learn image intricacies very well. The contribution of this work is three-fold, first, we created our own handwritten word dataset comprising of words with and without shirorekha, such that, effective training takes place. Next, we trained the proposed network with binary as well as in gray scale formats. Finally, we demonstrated that the proposed approach is accurate and generalizable.