Classification of Metal Type for Identical Letters in Old Metal Type Printed Books

The 『Jikji』, created during the Goryeo Dynasty, is known as the oldest extant metal type-printed book in the world. This study proposes a novel algorithm to effectively classify the typesetting of this historically significant metal type-printed book. The proposed method, Iterative Integer Linear Programming Refinement (IILPR), combines feature extraction using a Graph Neural Network (GNN) with Integer Linear Programming (ILP) to identify and classify subtle differences in typefaces. IILPR significantly enhances clustering accuracy by iteratively performing the three steps of re-adjustment, merging, and splitting. Unlike traditional clustering methods, IILPR operates flexibly without prior knowledge of the number of clusters kkk, demonstrating superior performance. The GNN converts character images into graph structures to extract features, while IILPR performs optimized clustering based on these features. Experimental results show that the IILPR method achieves a performance improvement of 11.4% over the previous best performance in terms of ACC. Notably, it recorded a high accuracy of 91.38% for the Chinese character '然'.