Combination of Local Features and Deep Learning to Historical Manuscripts Dating

While paleographers face various challenges in the dating process of historical manuscripts, computer scientists encounter multiple obstacles in automating them. To address this problem, machine learning and deep learning techniques which have proven effective in other domains have been used. This study introduces a system that integrates two primary methods of feature extraction - deep learning and hand-crafted features. The Harris-detector was utilized to extract key-points from the manuscripts, and the K-means algorithm was applied to cluster them. From these clusters, patches of size nxn were extracted. Then, a densnet model was trained on these patches using transfer learning, finally we used majority voting on document patches to determine the date based on document level. The effectiveness of this approach was evaluated on two datasets, MPS and CLaMM, and achieved a CS of 96% and 65% and an MAE of 4.01 years and 20.1 years respectively for both datasets.