Review of Modern Handwritten Text Recognition Methods Applicable for Old Church Slavonic Manuscripts Processing

Abstract

The study presents a systematization and critical analysis of modern approaches to handwritten text recognition, focused on their applicability to Old Slavonic manuscripts. The methodology involves the analysis of scientific publications, dissertation research, and technical documentation of leading HTR platforms based on the principles of scientific relevance for the period 2010-2025. Particular attention is paid to projects for the recognition of historical texts in Cyrillic - both international ones (Transkribus, eScriptorium, Tesseract) and Russian developments (RECO,DigitalPetr, “Handwritten Heritage of Ancient Rus”). A comparative analysis of classical and neural network recognition methods has been carried out. Classical algorithms (template matching, SVM, k-NN) have shown low efficiency due to the need for precise segmentation, dependence on feature quality, and poor adaptability to handwriting variability. In turn, neural network approaches have demonstrated higher performance: CNNs are effective for extracting spatial features but they do not model sequences; CRNNs with CTC allow recognition of lines without prior character segmentation; transformers (TrOCR, SATRN) achieve the best results due to the self-attention mechanism, which helps capture long-term dependencies and contextual links - this is especially important for processing ligatures, diacritics, and abbreviations. Key challenges in recognizing Old Slavonic manuscripts have been identified: the lack of large annotated corpora and insufficient modeladaptation to the linguistic and graphic features of Slavonic writing. It has been established that transformer architectures open up new possibilities for HTR of historical texts, allowing many limitations of previous approaches to be overcome. The research findings can be used in corpus creation, selection of model architectures, and development of new solutions in the field.