Publicação
Optimizing digital archiving: An artificial intelligence approach for OCR error correction
| Resumo: | This thesis research scopes the knowledge gap for effective ways to address OCR errors and the importance to have training datasets adequated size and quality, to promote digital documents OCR recognition efficiency. The main goal is to examine the effects regarding the following dimensions of sourcing data: input size vs performance vs time efficiency, and to propose a new design that includes a machine translation model, to automate the errors correction caused by OCR scan. The study implemented various LSTM, with different thresholds, to recover errors generated by OCR systems. However, the results did not overcomed the performance of existing OCR systems, due to dataset size limitations, a step further was achieved. A relationship between performance and input size was established, providing meaningful insights for future digital archiving systems optimisation. This dissertation creates a new approach, to deal with OCR problems and implementation considerations, that can be further followed, to optimise digital archive systems efficiency and results. |
|---|---|
| Autores principais: | Fernandes, Bruno Daniel Alho |
| Assunto: | Optical Character Recognition Machine Translation Neural Networks |
| Ano: | 2023 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | This thesis research scopes the knowledge gap for effective ways to address OCR errors and the importance to have training datasets adequated size and quality, to promote digital documents OCR recognition efficiency. The main goal is to examine the effects regarding the following dimensions of sourcing data: input size vs performance vs time efficiency, and to propose a new design that includes a machine translation model, to automate the errors correction caused by OCR scan. The study implemented various LSTM, with different thresholds, to recover errors generated by OCR systems. However, the results did not overcomed the performance of existing OCR systems, due to dataset size limitations, a step further was achieved. A relationship between performance and input size was established, providing meaningful insights for future digital archiving systems optimisation. This dissertation creates a new approach, to deal with OCR problems and implementation considerations, that can be further followed, to optimise digital archive systems efficiency and results. |
|---|