Benchmarking Classical Preprocessing Pipelines for Multilingual Archival Handwriting Analysis

Abstract

The digital processing of archival handwritten documents is challenging due to degradation, heterogeneous backgrounds, noise, and variability of multilingual graphemes. This paper presents a systematic comparison of seven classical preprocessing methods for handwritten image analysis on a multilingual dataset comprising 87 character classes. A unified experimental framework was designed in which preprocessing methods were evaluated while keeping the feature extraction (HOG) and classification (SVM) pipeline fixed. The methods were assessed using Accuracy, Macro-Precision, Macro-Recall, and Macro-F1, with repeated experiments and statistical validation. The results show that certain approaches, particularly Non-Local Means combined with Sauvola binarization, can provide local improvements in recognition quality. However, their performance is not stable across different data splits. The baseline approach without preprocessing and the Gaussian-Otsu method demonstrated comparable or more consistent results in repeated evaluations. Additionally, preprocessing effectiveness varies across linguistic-graphic subsets, highlighting the limitations of universal binarization techniques for multilingual archival data. The findings indicate that classical preprocessing methods have a limited and inconsistent impact on recognition performance. This motivates the development of specialized, structure-oriented preprocessing approaches aimed at preserving stroke details and improving discrimination between visually similar graphemes in historical manuscripts.