Off-line Handwriting Recognition Using Recurrent Neural Networks

I wrote this thesis for my Ph.D. while a research student at Cambridge University Engineering Department, in the Speech, Vision and Robotics Group.


Computer handwriting recognition offers a new way of improving the human-computer interface and of enabling computers to read and process the many handwritten documents that must currently be processed manually. This thesis describes the design of a system that can transcribe handwritten documents. First, a review of the aims and applications of computer handwriting recognition is presented, followed by a description of relevant psychological research. Previous researchers' approaches to the problems of off-line handwriting recognition are then described. A complete system for automatic, off-line recognition of handwriting is then detailed, which takes word images scanned from a handwritten page and produces word-level output. Methods for the normalization and representation of handwritten words are described, including a novel technique for detecting stroke-like features. Three probability estimation techniques are described, and their application to handwriting recognition investigated. The method of combining the probability estimates to choose the most likely word is described, and performance improvements are made by modelling the lengths of letters and the frequency of words in the corpus. The system is tested on a database of transcripts from a corpus of modern English and recognition results are shown. Recognition is described both with the search constrained to a fixed vocabulary and with an unlimited vocabulary. The final chapter summarizes the system and highlights the advances made before assessing where future work is most likely to bring about improvements.

Key words

Off-line cursive script, handwriting recognition, OCR, recurrent neural networks, forward-backward algorithm, hidden Markov models, duration modelling.

Complete text

You can download my thesis: pdf or gzipped postscript.


I have made the data which I collected as part of my PhD work publicly available. Read a brief description first, then get a sample. After that you can download the whole database of numbers data (11MB) or LOB corpus data (35MB).

The data is also available from Back to my list of papers

Andrew Senior