Optical Character Recognition (OCR) serves as the foundational step for numerous document processing applications, commonly built upon available general-purpose OCR models. However, given a vast heterogeneity of document sources and applications, specialising the general-purpose OCR to the specific use-cases might deliver substantial improvements in the quality of the entire application. The objective of this thesis is explore the potential of adapting general-purpose OCR models to specific domains and applications.

The thesis will deliver the following outcomes:
(1) Survey of the methods for fine-tuning general-purpose OCR models to specific domains;
(2) Application of selected existing, or proposed adaptation methods to the domain of Czech financial documents and evaluation of the quality of resulting OCR models;
(3) Extrinsic evaluation of the existing and proposed methods in a document processing pipeline on the end task of automated invoices processing.

The methods will be evaluated on the datasets collected within the Intelligent Back Office grant project.