Improving Offline Handwritten Text Recognition Using Conditional Writer-Specific Knowledge

Werff, Tobias van der (2022) Improving Offline Handwritten Text Recognition Using Conditional Writer-Specific Knowledge. Master's Thesis / Essay, Artificial Intelligence.

Preview	Text mAI_2022_VanderWerffT.pdf Download (2MB) | Preview
	Text toestemming.pdf Restricted to Registered users only Download (122kB)

Abstract

Modern neural networks suffer from a major flaw: an inability to deal with changing data distributions. In the field of handwritten text recognition (HTR), this shows itself in poor recognition accuracy for writers that are not similar to those seen during training. In this research, the question is asked whether explicit information regarding writer identity can be used to condition a neural network to a writer-specific distribution. For example, writer information can be presented in the form of a small batch of labeled examples originating from a particular writer. Two state-of-the-art HTR architectures are used as baseline models (writer-unaware), using a ResNet backbone along with either an LSTM or Transformer sequence decoder. Using these base models, various methods are proposed to make them writer-adaptive, based primarily on 1) an idea originating from automatic speech recognition known as speaker codes, and 2) model-agnostic meta-learning (MAML), an algorithm commonly used for tasks such as few-shot classification that has gained much attention in recent years. Results show that an HTR-specific version of MAML known as MetaHTR improves performance compared to the baseline with a 1.4 to 2.0 improvement in word error rate. Furthermore, it is shown that a deeper model lends itself better to adaptation using MetaHTR than a shallower model. Speaker codes do not show a concrete benefit for writer-aware adaptation.

Item Type:	Thesis (Master's Thesis / Essay)
Supervisor name:	Schomaker, L.R.B. and Dhali, M.A.
Degree programme:	Artificial Intelligence
Thesis type:	Master's Thesis / Essay
Language:	English
Date Deposited:	05 Jul 2022 10:14
Last Modified:	05 Jul 2022 10:14
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/27604

Actions (login required)

View Item