Extrakce strukturovaných dat z fotokopií historických dokumentů

This bachelor’s thesis deals with the digitization of structured historical documents. It first describes the general challenges encountered by text conversion tools and then focuses on a specific case study. This case study consists of population census data from the Czechoslovak Republic from 1921 and 1930. Before the practical task, the thesis defines the theoretical framework for the digitization of historical structured documents. It focuses primarily on the principles of optical character recognition, document layout analysis, and current technological possibilities in this field. It distinguishes between two key phases of data extraction: the recognition of document structure and the recognition of textual characters themselves. This framework then serves as the basis for selecting the tested tools, creating the reference dataset, and interpreting the results obtained. The first step in the attempt to digitize the dataset was to create a reference dataset that could be used to evaluate individual optical character recognition tools. It was then necessary to examine the currently available options and select specific tools for pilot testing of data conversion into a machine-readable form. When selecting platforms and services, the author focused on ensuring that both large commercial and smaller specialized services in this field were covered. Even the results of the most accurate model from the pilot testing, calculated using CER and WER metrics, were not sufficiently satisfactory. Therefore, an additional attempt was made to improve the results using an automated process designed to correct the OCR output through logical aggregations based on the data structure. However, the automated process encountered the lack of robustness inherent in low-code solutions. Based on these results, the discussion addresses the reasons why the extraction was not successful and how further research in the field of digitizing structured historical documents could proceed. It also presents and discusses an example of using OCR tools on simpler contemporary documents.