DEEP SCENE END TO END OCR TOOL

Author:

Lv, Weining

Graduate Program:

Computer Science and Engineering

Degree:

Master of Science

Document Type:

Master Thesis

Date of Defense:

November 14, 2019

Committee Members:

Lee Giles, Thesis Advisor/Co-Advisor
Jesse Louis Barlow, Committee Member
Chitaranjan Das, Program Head/Chair

Keywords:

deep learning ocr system

Abstract:

As we know , how to accurately recognize the meaning of words from the natural world is a big issue in computer vision . Due to the development of high-performance computers , neural networks model could be used in research and open up new horizons for people . This paper attempts to pull from neural networks model in order to solve deep scene text detection and text recognition. I start with synthetic data in order to give infinite amounts of training data . After that , I write an automatic crop program to help evaluate data set in some specific situation or work for augmenting data set . I did two parts in text detection . On the one hand , I use the traditional maximally stable extremal regions(MSER) algorithm with deep convolutional neural network to get the bounding box of text . On the other hand , I use a recent famous model, Efficient and Accurate Scene Text Detector (EAST) model, to get more accurate detection results . Based on the EAST results , I use RCNN structure of text recognition . Experiments show that this method can be applied to several data sets, ICDAR2013 , SVT and ICDAR2015 and small screenshot data set .