Semi-automatic Segmentation & Alignment of Handwritten

HTR systems

An HTR System is a system that, through the use of different statistical and machine learning techniques, is able to transcribe handwritten text. This can impose a challenge due to varieties in handwriting. For example, there will always be some variability in a particular sentence written by even the same author, let alone another author. A tradi- tional HTR system often utilized N-gram language models and Hidden Markov Model optical modelling with Gaussian mixture emission distributions. However, recent tech- nical advancements have allowed for the usage of neural networks due to their adapt- ability in more complex tasks (Neto et al. 2020). A typical workflow of a state-of-the-art HTR system is seen in Figure 2; a short description of the workflow is as follows:

• 
  Preprocessing: Image preprocessing and normalisation to prepare the image for extraction of features.
  

• 
  Feature Extraction: Extract features in the image using a Convolutional and Recurrent Neural Network.
  

• 
  Feature Propagation: Propagate the selected features over the sequence by re- constructing missing node features with a Recurrent Neural Network.
  

• 
  Temporal Classification + Dictionary + Language Model: Connectionist Tem- poral Classification is used for loss function calculation, together with the tradi- tional language model, dictionary included, and Hidden Markov Model, to get a final output.
  

4. Related Work

Applications found in HTR nowadays are vast, and there are many approaches for tran- scribing documents. With the recent advances in deep learning and HTR research, mak- ing your own model from scratch requires a certain amount of knowledge in the field. The importance of having easy-to-use virtual research environments accessible to the public has grown as these new techniques have emerged (Souibgui et al. 2022). In the following subchapter, some research related to the subject of this thesis is described.