Computational Visual Cataloguing

Digitisation of handwritten heritage outpaces cataloguing. Manual workflows remain slow, automatic text recognition struggles with heterogeneous scripts and layouts, and text-only transcriptions omit visual cues such as orientation, colour and writing implements. This limits discovery and reuse in libraries, archives and museums, especially for notebooks with complex, irregular layouts. We propose Computational Visual Cataloguing (CVC), an extensible computer vision framework that performs content-bearing page detection, word localisation irrespective of script or language, and word-level visual-attribute classification (orientation, colour, writing implement). Our framework is designed to learn robustly from densely written manuscript pages and limited human annotations, reflecting the natural challenges of historical handwritten collections. These outputs populate structured, searchable catalogue fields. The contributions of this work are: (1) the formulation of CVC and a word-level visual-attribute schema for handwritten heritage; (2) task-specific methods and publicly released models for page detection, cross-script word localisation and visual-attribute classification; (3) a manually annotated dataset and full computational catalogue for Rainer Maria Rilke’s notebooks; and (4) ScriptSight, a tool for exploring and exploiting these catalogues. We evaluate CVC on Rilke’s notebooks, a culturally significant corpus with multilingual content, varied orientations and colours, diverse implements and irregular layouts. we report quantitative performance for individual components (including F1-based evaluation and ablation studies) and qualitative assessment of the integrated system, which together demonstrate its applicability to complex pages and its practical utility for exploratory cataloguing.