Old land registers created during the Japanese colonial period constitute important administrative records that document the formation and transformation of modern land ownership systems in Korea. These records contain essential
information such as la...
Old land registers created during the Japanese colonial period constitute important administrative records that document the formation and transformation of modern land ownership systems in Korea. These records contain essential
information such as land location, parcel number, land category, ownership, and detailed histories of ownership changes. However, the practical utilization of old land registers has been limited due to their complex document structures, the mixture of classical Chinese characters and Japanese-style scripts, and the heavy reliance on manual interpretation during Korean-language conversion processes.
This study proposes a deep learning-based processing system for the Korean-language conversion of old land registers, focusing on document- 58 classification and history-centered automated processing. Unlike conventional approaches that treat documents as uniform textual entities, the proposed system adopts a multi-stage document understanding framework that reflects the structural characteristics of old land registers.
The system first classifies input documents to determine whether they conform to the old land register format. For documents identified as old land registers, the system automatically detects and crops the history area, which contains the core information related to land ownership changes. A history type classification step is then introduced to distinguish different structural types of history records, enabling adaptive processing strategies. To ensure the reliability of subsequent processing, a crop quality assessment module evaluates whether the extracted history area is suitable for automated processing or requires reprocessing or manual inspection.
Furthermore, this study extends history processing to a finer semantic level by introducing the classification of individual history seals, referred to as “Young-History” units. Assuming a dataset consisting of 100 seal types with 1,000 samples per class, a multi-class deep learning-based classification experiment was conducted to verify the feasibility of semantic-level history interpretation. Experimental results demonstrate that the proposed system achieves stable performance across document classification, history area detection, quality assessment, and history seal classification tasks, while significantly reducing the need for manual intervention.
The findings of this study indicate that the proposed system effectively supports scalable and reliable Korean-language conversion of old land registers. By shifting from document-level processing to history-centered and semantic-unit-based automation, this research contributes to advancing document understanding methodologies for administrative records and provides a practical foundation for the digital utilization of large-scale historical land records