
Historical scripts are structured visual systems whose internal organization has traditionally been explored through qualitative paleographic analysis. However, many of the principles shaping their evolution remain difficult to test empirically across corpora. One such principle, formulated by Jean Irigoin, is the hypothesis of graphic compensation, according to which graphic systems tend to maximize visual distinctiveness between letters that share similar phonetic functions while allowing greater visual proximity in other cases. In this work, we develop a computational approach to explore the visual structure of historical scripts and apply it to the cursive Greek writing of Hellenistic papyri. Building on a previously introduced representation-learning framework for ancient Greek letterforms, we analyze the behavior of a ConvNeXt-V2 character recognizer trained with lacuna-based fragmentation and dynamically learned supervised contrastive loss in order to extract the similarity relations that structure the script. Rather than treating classification errors as poor technical results, we interpret them as paleographically meaningful patterns that reveal the internal organization of the graphic system. The resulting analysis shows that several of the strongest confusions occur between visually similar but phonetically distinct letters, including Alpha–Lambda, Omicron–Theta, Iota–Rho, Epsilon–Sigma, Tau–Upsilon, and Eta–Pi, providing quantitative support for Irigoin's hypothesis and suggesting that the principle of graphic compensation continues to operate even in the less controlled domain of cursive handwriting. This tendency is supported by a global pairwise comparison of same-category and vowel–consonant confusions, by stability checks across alternative training variants, by evaluation on damaged characters, and by a papyrus-level split designed to reduce leakage between train and test material. More broadly, this study proposes a methodological framework for computational paleography, in which model behavior can be used to extract interpretable descriptions of script structure from large collections of cultural heritage material. Finally, we suggest that the learned relationship between phonetic difference and visual confusion may itself serve as paleographic background knowledge and be incorporated into future HTR pipelines, helping recognition systems resolve visual ambiguity when visual evidence and linguistic expectations compete during decoding.
Reconstructing fragmented archaeological artifacts is a central challenge in the study and preservation of ancient cultural heritage. A crucial step in this process often involves identifying the stylistic origin of isolated fragments, in part to dissociate fragments that could not originate from the same object. However, this task is inherently difficult: fragments often preserve only limited visual information, such as partial motifs or degraded surface details, and are frequently detached from their original archaeological context. Ancient Pompeian wall paintings, traditionally categorized into four styles, provide a compelling and historically significant case study. Over time, natural disasters and human activity caused many of the decorated structures to collapse, dispersing and mixing fragments from diverse sources. As a result, stylistic attribution has become a complex and labor-intensive task. In this work, we demonstrate how recent advances in deep image classification can be leveraged to develop highly accurate systems for fragment style recognition, addressing key challenges faced by archaeologists on site. To support this task, we introduce a new dataset for fragment-style classification based on the shapes of real fragments recovered from Pompeii excavations. Our research spans large-scale foundation models down to lightweight, edge-compatible networks, significantly improving on previous state-of-the-art methods while maintaining a balance between accuracy and computational efficiency.
The study of archaeological signs and non-verbal markings requires navigating complex relationships between artefacts, symbolic forms, and contextual metadata. While digital heritage platforms increasingly provide cataloguing and metadata services, they still offer limited support for visual exploration and comparison, particularly when signs are small, spatially unconstrained, and embedded within deteriorated artefacts. This paper presents the International Etruscan Sigla Project (IESP), an integrated visual heritage system that combines structured cataloguing, intelligent document-based knowledge extraction, and AI-assisted visual similarity retrieval within a unified archaeological exploration environment. Rather than introducing these capabilities as independent tools, IESP supports the complete archaeological workflow, from knowledge acquisition and catalogue navigation to image-based exploration and comparative analysis. The system stores and models artefacts and sign elements, automatically extracts structured information from publicly available archaeological documents, and enables similarity-driven exploration through a dual-encoder architecture specifically designed for fine-grained comparison of archaeological signs. The proposed retrieval strategy aligns global sign representations with spatial feature maps extracted from artefact images, enabling similarity-based retrieval while preserving the local visual evidence required for archaeological interpretation. Users can interact with the system through multiple retrieval modalities, including sketch-based queries, image-based sign retrieval, and support-based exploration. The paper also presents the system's software architecture and reports the results of an expert-centred user study conducted through realistic archaeological tasks. The evaluation combines task performance, usability, cognitive workload, technology acceptance, and qualitative observations, demonstrating that the proposed interaction paradigm effectively supports archaeological exploration while preserving human interpretation and scholarly transparency. Overall, IESP demonstrates how integrating structured knowledge management, document understanding, and interpretable AI can transform digital heritage repositories into interactive environments that actively support archaeological research and visual analysis.
Multi-Light Image Collections (MLIC) are fundamental for virtual re-lighting and detailed surface analysis. In cultural heritage, the most common acquisition approach is free-form moving-light capture, where images taken under varying illumination are later processed to derive shape, material properties, or relightable image models. Despite its flexibility, this workflow remains limited by the need for trained operators, calibration targets, and often significant manual processing. To address these challenges, we present M \({{}^{2}}\) LIC (Mobile Multi-Light Image Collection), an automated, target-free pipeline utilizing video recordings from a fixed camera and a moving camera attached to an illuminator. Realizing this setup with just two commodity smartphones lowers hardware costs and technical barriers to entry by leveraging devices that users already own and operate daily. Our framework combines video synchronization, estimation of light directions directly from the scene structure, relighting-aware frame processing, and automatic generation of relightable models ready for web viewing. The protocol is fast and simple, requiring only the simultaneous recording of short videos, allowing operators to produce usable relightable models with minimal training and effort. Validated on a diverse set of artifacts, this open-source system helps lower the barrier to MLIC acquisition and aims to support more scalable MLIC digitization workflows.
Replicating the style of historical artistic traditions remains a notoriously challenging task for vision-based generative models, as artistic style encodes complex visual, cultural, and contextual dimensions that extend beyond surface-level appearance. Despite rapid advances in generative modeling, it remains unclear to what extent recent systems have achieved substantive improvements in stylistic emulation across diverse historical traditions. This article presents a systematic evaluation of contemporary generative models in artistic style imitation, comparing their outputs to those of earlier model generations. The study builds upon the AI-Pastiche dataset by incorporating six additional recent state-of-the-art models, thereby enabling a longitudinal assessment of performance progression over time. To ensure methodological consistency and comparability, the evaluation adopts the same experimental protocol used in the original release of the dataset. The analysis combines controlled prompt-based image generation with extensive user studies designed to assess multiple dimensions of aesthetic and stylistic performance. These dimensions include: (i) visual plausibility and perceptual realism; (ii) stylistic coherence and fidelity to historical conventions; (iii) versatility across a broad range of artistic traditions; (iv) adherence to prompt-specified content and stylistic constraints; and (v) the absence of structural artifacts or semantic inconsistencies. The results indicate substantial and measurable improvements in stylistic emulation in recent models. Nevertheless, persistent limitations remain, particularly in the consistent reproduction of deeper compositional structures and culturally embedded stylistic features. These findings suggest that progress in surface-level resemblance does not necessarily equate to structural or contextual understanding.