Semi-automatic Segmentation & Alignment of Handwritten

Closing Remarks

1. Conclusion

It was found that there are many factors to consider for the segmentation and alignment to be perfect, namely, the amount of noise present, the quality of the GT, and the limitations of the used techniques. The performance results from Labour’s Memory are hard to draw a conclusion from as the GT is insufficient for comparison. This could be further investigated for a more equitable evaluation. Nevertheless, the evaluation of the IAM data set shows promising results which could be further improved to hopefully serve as a useful tool in the future. According to performance metrics, the developed method 2 generally proved to be performing better than method 1 for both data sets. Self-learning was integrated with the use of Bayesian optimisation to set parameters. This technique is shown to be reasonable for more time-efficient processing but will under-segment more frequently than setting parameters manually.

2. Future work

The work accomplished in this project could be seen as the groundwork for the future de- velopment of a useful tool for the alignment of text documents. Due to time constraints, there are improvements to be made to both the algorithms as well as the evaluation. To give a fair assessment of the performance of the algorithms on Labour’s Memory, a more precise GT would have to be produced for the experiments. Furthermore, the performance has not been compared to others’ work, which would add credibility to the functionality of the algorithms. The evaluation could be improved by being compared to the transcription software Loghi (van der Veen 2023). Further development of a suitable application could be beneficial for the usability and simplicity of the algorithms; Omeka (Roy-Rosenzweig-Center 2008) could be used as a reasonable alternative platform.