Semi-automatic Segmentation & Alignment of Handwritten

Loghi transcription software

Loghi (van der Veen 2023) is a newly released transcription software based on machine learning developed by Rutger van Koert of the Digital Infrastructure Department of the KNAW Humanities Cluster in cooperation with the Nationaal Archief in The Hague. The goal was to make scanned historical documents digitally readable and searchable. The authors report an error rate of under 4 % when the software is trained on a specific data collection. This software segments and recognizes text, in contrast to the algorithm developed in this thesis, which segments and looks at a GT of the document to align.

5. Machine learning for image segmentation

Machine learning is the subject of making a computer program learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T , as measured by P, improves with experience E (Mitchell 1997).

In the context of the segmentation of an image, machine learning could be interpreted as making an algorithm that learns a specific task of segmenting desired objects from the image. In this thesis, that would be simplified to make the segmentation of text lines and words self-learning. Learning tasks tied to image segmentation often include optimising parameters for better performance. Self-learning can be achieved in different ways, such as using a recurrent-convolutional neural network as performed by Wilkinson & Brun (2015) , or using a structured support vector machine as accomplished by Ryu et al. (2015).

Another approach to parameter optimising is through the use of Bayesian optimisation. Using Bayesian optimisation, the goal is to maximize the output of an unknown function by iterating through different probability-based selections of input parameters. In the case of this thesis, this would mean tuning parameters that affect techniques used in the segmentation, thus giving a different output.

3.