Image-to-image Translation for Handwriting Generation using GANs

Bakker, Thomas (2019) Image-to-image Translation for Handwriting Generation using GANs. Bachelor's Thesis, Artificial Intelligence.

Preview	Text AI_BA_2019_TPJBAKKER.pdf Download (776kB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (139kB)

Abstract

: In this paper it is investigated whether machines are able to synthesize human-like handwriting through deep learning that can be used to augment datasets. For this purpose the relatively novel generative adversarial networks (GANs) are deployed and experimented with. GANs are generative neural networks, consisting of two competing networks, that adversarially learn through a loss function to generate new, non-existing samples. There exist several potentially useful loss functions such as L1 and L2. By feeding the network images with machine-print words and simultaneously feeding it images with handwritten words the system learns to mimic the handwriting. These synthesized handwritten words can be used to supplement datasets used to train Handwritten Text Recognition (HTR) systems. Several experiments demonstrate what the optimal settings are in terms of network architecture to achieve machine-print handwriting resembling human writing.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Schomaker, L.R.B.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	13 Aug 2019
Last Modified:	15 Aug 2019 09:28
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/20654

Actions (login required)

View Item