Comparing natural and synthetic training data for off-line cursive handwriting recognition

In this paper, a perturbation model for the generation of synthetic textlines from existing cursively handwritten lines of text, produced by human writers, is presented. The goal of synthetic textline generation is to improve the performance of an off-line cursive handwriting recognition system by providing it with additional, synthetic training data. In earlier papers, it has been shown that it is possible to improve the recognition performance by using such synthetically expanded training sets. In this paper, we investigate the suitability of synthetically generated handwriting when enlarging the training set of a handwriting recognition system in a more rigorous way. In particular, the improvements achieved with synthetic training data are compared to those achieved by expanding the training set using natural, i.e. human written, textlines.

1. Introduction

Handwriting recognition systems need to be trained. It is well known that the performance of a handwriting recognition system is strongly affected by the size and quality of the training data [1]. As a rule of thumb says, the classifier that is trained on the most data wins. This was empirically confirmed in a number of experiments [2], [8], [10], [15].