CNN BASED APPROACH FOR HANDWRITING RECOGNITION USING TENSORFLOW

Abstract

Handwriting Chinese Character Recognition (HCCR) has been area of research for over a decade. Because of its wide range of characters, similarities and complexities it is difficult to classify them as compared to other languages or numerical representations. It is very important to recognize handwriting as we are moving towards the automated world. It is a field of machine learning or more specifically the deep learning. Deep learning is showing great results in the fields of visual and speech recognitions. But still we are not able to match the accuracy of human vision and advancements will be continuing till we reach or cross that limit. It is also very significant today as we want to replace human beings with machines in every field one such example could be automatic cars. But for all this we need high accuracy because if there will be any error there could be a disaster. In this thesis we are using deep learning techniques to improve the accuracy of HCCR. We are modifying the existing model to a more deeper and slimmer one i.e. having less number of overall parameters. We are evaluating our proposed model on ICDAR (International Conference on Document Analysis and Research) test dataset (HWDB1.1). There are many applications of reading human writing in the real world like it can be used in banks or post offices to read the cheques and envelops. We can directly submit them to machine and machine can do the specific tasks. This is our attempt to close the gap between machine and human vision.