Contribution to historical manuscript dating: A hybrid approach employing hand-crafted features with vision transformers

Abstract

Precisely dating historical manuscripts represents a paramount endeavor in the comprehension and the interpretation of their historical significance as well as in the preservation of our cultural heritage; however, despite the strides made in computer-based dating methodologies, the quest for heightened robustness persists. Recent advancements in vision transformers, renowned for their success across diverse image processing domains, have stimulated our inquiry into their potential applicability in the domain of historical manuscript dating. In our pursuit of achieving efficient refinement of the manuscript and ensuring pristine datasets for subsequent analysis, we initiate a meticulous dataset preprocessing step. This involves employing accurate methodologies for denoising through the Non-Local-Means algorithm, and binarization using the Canny-edge detector. Following these preprocessing steps, we delve into the intricate realm of feature detection using the Harris-corner detector. This detector is employed to extract keypoints from the manuscript, and we subsequently apply clustering to these keypoints using the k-means algorithm. Our dual objective here is to extract significant patches of specific dimensions and engage in a systematic augmentation of our dataset. Through this process, we aim to amplify the depth and diversity of our dataset, thereby empowering our models with an enriched corpus of historical knowledge. The final phase of our proposed system unfolds as we leverage the latent power of the sophisticated deep learning architecture known as vision transformers. The model is finetuned to our specific task and re-learned with the automatically extracted handcrafted features, emerges as a formidable classification framework. As an added layer of refinement, we deploy a majority vote mechanism on image patches, meticulously engineered to heighten system accuracy. Our rigorous testing regimen, carried out on the well-known MPS historical document dataset, has yielded results of remarkable caliber. Our system's prowess is vividly reflected in its performance metrics, boasting an impressive Mean Absolute Error (MAE) of 3.97 for document-level evaluations and a resounding 7.42 MAE for patch-level assessments. The significance of this work lies in its potential to revolutionize historical manuscript dating, not only by enhancing precision but also by providing a versatile framework adaptable to diverse historical contexts and document types. Beyond its immediate applications, this research paves the way for a deeper understanding of historical narratives, cultures, and the invaluable insights that lie within the annals of our past.