SESHAT: A MODULAR PIPELINE FOR SEGMENTATION, ENCODING, AND SEQUENTIAL HIEROGLYPHIC ANALYSIS AND TRANSLATION

Ancient Egyptian hieroglyphic inscriptions remain largely inaccessible to non-specialists due to the scarcity of automated tools capable of jointly performing visual recognition and linguistic translation. This thesis presents SESHAT (Segmentation, Encoding, and Sequential Hieroglyphic Analysis and Translation), a modular neural pipeline for the automatic recognition and translation of Egyptian hieroglyphs, developed and evaluated on real photographic material from heterogeneous sources: the HLA (Egyptian Hieroglyphic Layout Analysis) dataset, comprising photographs of museum artifacts and archaeological sites, and the Glyphnet dataset, containing high-resolution photographs of inscribed reliefs from the Pyramid of Unas with fine-grained, sign-level annotations. SESHAT is organized into three sequential components. A Multi-Class Layout Segmenter, trained on the HLA dataset, localizes and classifies hieroglyphic text regions, namely lines of text (Lines) and cartouches, directly from raw photographs of inscribed surfaces, using instance segmentation. A two-stage recognition module then receives each classified region and encodes it into a sequence of Gardiner sign-list codes: a dedicated glyph detector first localizes the individual signs and arranges them in reading order according to the region’s structural class (top-to-bottom for Lines, cadrat order for Cartouches), and a Document Vision-Language Model then reads each sign in isolation and assigns it its Gardiner code. Both models are trained on annotated crops from the Glyphnet dataset, establishing a visual vocabulary of isolated hieroglyphic signs that is then composed, sign by sign, into the region’s complete code sequence. Finally, a Context-Conditioned Neural Machine Translation module converts the encoded sign sequences into natural language; conditioned on the structural class of each region, it distinguishes between standard semantic translation for running text and phonetic transliteration for cartouches, which encode proper names rather than ordinary lexical content. The resulting components are evaluated on held-out, stratified real photographic data from both sources, reporting detection, sequence-level, and character-level metrics, and discussing the broader trade-offs involved in designing modular, low-resource visual sequence recognition systems for historical and archaeological document analysis.