Enhancing Handwritten Text Recognition in Historical Documents Using Spectral Imaging

Overview

Many historical documents are difficult to transcribe due to different types of degradation, where conventional RGB imaging is often insufficient [1]. This master’s thesis focuses on the generation and use of synthetic hyperspectral (HSI) and multispectral (MSI) data for handwritten text recognition (HTR). By transforming existing annotated RGB document datasets into synthetic spectral data, the thesis evaluates whether spectral information can improve HTR performance compared to RGB approaches.

Figure 1: False RGB images of historical documents extracted from the HYPERDOC dataset [2].

Objectives

Review state-of-the-art
Generate synthetic HSI and MSI data from annotated RGB datasets
Adapt an existing HTR model to spectral data
Compare HTR performance across RGB, MSI, and HSI modalities

Methodology

Literature Review - Spectral imaging basics and HTR models
Synthetic Data Generation - Unmixing-based [3] generation of HSI from annotated RGB datasets
Model Adaptation - Train an HTR model with RGB, MSI, and HSI inputs
Evaluation - Quantitative comparison using standard HTR metrics
Reporting - Final thesis, presentation, and optional publication

Skills Required

Python programming
Deep learning
Interest in document analysis, historical handwritting and spectral imaging

References

[1] L. Pardo et al., Advanced imaging to recover illegible text in historic documents. The challenge of past chemical treatments for ink enhancement, Journal of Cultural Heritage, 2024.

[2] A. B. López-Baldomero et al., Hyperspectral dataset of historical documents (HYPERDOC), Scientific Data, 2025.

[3] E. M. Valero et al., Unmixing and pigment identification using visible and short-wavelength infrared: Reflectance vs logarithm reflectance hyperspaces, Journal of Cultural Heritage, 2023.

Created: 01.02.2026