Optical Character Recognition For Handwritten Forms With Dynamic Layout

Optical Character Recognition (OCR) is an established problem statement in machine learning and artificial intelligence. While most believe it is an open and shut case the challenge lies when the data present is rather ambiguous and unregulated, which is precisely the case in handwritten text recognition. This paper stresses on the major setbacks faced while dealing with such forms of multifarious data and how a finite machine can accommodate for this inconsistency. The paper specifically proposes a localised zonal method of character detection which is seen to significantly improve accuracy levels for recognition. This implementation accounts for the contextual placement of characters by making use of two separate custom convolutional neural networks(alphanumeric and numeric) which were trained on the EMNIST balanced dataset and gave test accuracies of 97.2% and 76.8% respectively.