Generalised, multilingual, optical Braille recognition models

Optical Braille recognition (OBR) is the computer vision task of extracting and classifying character cells from a Braille document. This study investigates the impact of class imbalance and adverse data quality conditions on OBR performance, by training standard multiclass (well adopted methodology) and novel multilabel (proposed in this work) models on different scenarios with resampled training data. These models are evaluated on unseen test data, both in-distribution and out-of-distribution, as well as on simulated adverse conditions. The results show that the proposed multilabel modelling approach outperforms the current standard multiclass approach across different evaluations. Multilabel models further generalise better than multiclass models to severe adverse data quality conditions, and achieve more stable performance across both majority and minority classes. Additionally, the use of data resampling to reduce imbalance in training data leads to improved performance generalisation, for both modelling approaches. In particular, a multi-objective resampling strategy minimising both the imbalance in character classes and the correlations between Braille dots, yields exceptionally stable performance across different conditions. In contrast, multiclass models trained on imbalanced, non-resampled data consistently exhibit poor generalisation to adverse conditions and minority Braille characters. In summary, these findings strongly support the conclusion that the application of the multilabel classification approach, in combination with strategic data resampling, yields robust OBR models exhibiting generalised performance across diverse Braille datasets.

Grantor dc:publisher

Stellenbosch : Stellenbosch University

Year dc:date.issued

2026

Author dc:creator

• Van der Linden, Wicus Jan

Advisors dc:contributor.advisor

• Grobler, T. L.
• Van Zijl, L.

Language dc:language.iso

en

Repository record dc:identifier.uri

https://scholar.sun.ac.za/handle/10019.1/135695

OAI identifier oai:identifier

oai:scholar.sun.ac.za:10019.1/135695