Ai+ Training

Data being the most valuable resource on earth in today’s date there is a pressing need to transform all data digitally. The world has been historically and still in multiple sectors operating based on handwritten text. In the last few decades, a considerable amount of work has been done in this field and various approaches are proposed based on conventional image processing and Deep Neural Network. Researches from different authors present an application wherein they take in word crop images generated from a document manually and extract the text from it. Current methods work best only when a word crop is fed with perfect offset, as these models are trained on IAM dataset where crops are with perfect offset. Also, these researches do not present an end-to-end solution of text extraction from a handwritten page.

In this talk, we focus on an end to end solution wherein an image with handwritten text from a local user will be uploaded and handwritten text from the entire image will be extracted as it is. The three main contributions of our thesis are: 1. Improved text localizer trained on a various handwritten dataset, 2. Classification model to segregate printed and handwritten words, 3. Effective image pre-processing techniques applied to handwritten word crops to make them eligible to be fed to the deep learning model for improved accuracy.