In our project, we deal with translating texts recorded by the ancient Egyptian civilization into contemporary English. In this paper, we focus particularly on identification of text structures consisting of hieroglyphs. The identification is based on classification of segmented image blobs and their spatial relations using graph neural networks. We reached 99.8% accuracy on a dataset of facsimiles created for the shaft tomb of Menekhibnekau. The high accuracy is due to a combination of precise results achieved by the CRAFT method, additional features like the size of the hieroglyphs, including very robust topological properties of blob adjacency weighted by the learned nonlinear graph neural network scheme not relying on simple horizontal or vertical projections as used in standard OCR approaches. The graph of blob spatial relations is built using distance transform. We also propose an algorithm for a separation of hieroglyphs from mostly linear structures delineating the strips of hieroglyphs if they touch. Strips of hieroglyphs identified this way can be used to extract blobs of glyphs into reading sequences before their classification to Gardiner’s codes, transliteration and translation.