Clustering thermograms of Schubert autographs based on chain and laid line features

Abstract We propose a novel approach towards the automated clustering of handmade paper. Specifically, we acquired a large dataset of thermographic images and developed a feature extraction and automated clustering method to identify leaves of paper made with the same mold. Our method is based on the structure of laid lines and chain lines that are imprinted on the paper during production. While only relying on these features already gives a good clustering performance, a semi-supervised approach with manual consideration of edge cases gives an excellent performance, surpassing even previously existing expert classification. This enables new findings in paper research, such as identifying new mold types, non-mirrored twin molds, classification of empty pages, and the use of different watermarks on the same mold. Furthermore, the systematic analysis of a large corpus allows us to find leaves of paper belonging to the same original sheet across different manuscripts.