Text region detection in historical astronomical diagrams

A research team has released a dataset of 948 historical astronomical diagrams annotated with 10,940 oriented polygonal text regions, spanning ten centuries and seven linguistic traditions, alongside a detection model that sets new performance records on two established benchmarks. The open-access corpus, described in a paper submitted on 14 June 2026, draws from documents produced between the 8th and 18th centuries [1][2]. It includes diagrams in Arabic and Persian (115), Chinese (332), Byzantine (233), Latin (185), Hebrew (48), and Sanskrit (35) [2]. Each text instance is labeled with ordered polygons that capture both the precise boundary of the text and its reading direction [1][2]. For the 2,293 regions found in Latin diagrams, annotators also assigned one of 20 class labels [1][2]. The work targets a gap in document analysis. While text-detection benchmarks exist for manuscripts and maps, mathematical and astronomical diagrams have received little systematic attention [1][2]. Historical astronomical diagrams often combine symbols, labels, and multi-line paragraphs arranged in non-linear layouts, making them a distinct challenge for computer-vision systems [2]. To establish a baseline, the authors evaluated several detection architectures, including TESTR and DeepSolo++, and introduced Poly-DETR, an extension of DINO-DETR designed to predict ordered polygon vertices directly [1][2]. Poly-DETR achieved state-of-the-art results on the MTHv2 and cBAD2019 benchmarks, two standard evaluations for text detection in historical documents [1][2]. The authors describe the model as a solid, simple baseline for their new astronomical-diagram dataset [2]. The dataset and accompanying code have been made available online [1][2]. The submission file is 4,029 KB [1]. The paper was posted by Zeynep Sonat Baltaci [1]. Astronomical diagrams have long served as essential tools for recording and communicating celestial knowledge. The field of astrophysics, as a discipline, applies the methods of physics and chemistry to study the nature of heavenly bodies rather than merely their positions [4]. Diagrams such as the Hertzsprung–Russell plot, which classifies stars by color and brightness into a main sequence where they spend most of their lives fusing hydrogen, are foundational to modern stellar astronomy [3]. The new dataset captures a much broader historical range of such visual traditions, from medieval Byzantine manuscripts to early-modern Chinese treatises, offering a resource for both computer scientists and historians of science [1][2].