Graph Neural Network for Symbol Detection on Document Images

Guillaume Renton (1, 2) , Pierre Héroux (2, 1) , Sébastien Adam (2, 1) , Benoit Gaüzère (2, 1)

1 LITIS - Laboratoire d'Informatique, de Traitement de l'Information et des Systèmes
2 DocApp - LITIS - Equipe Apprentissage

In this paper, we propose a new method to simultaneously detect and classify symbols in floorplan images. This method relies on the very recent developments of Graph Neural Networks (GNN). In the proposed approach, floorplan images are first converted into Region Adjacency Graphs (RAGs). Within those graphs, each node corresponds to a white region in the original image, and each edge indicates an adjacency relationship between two regions encoded by incident nodes. Nodes are attributed using Zernike moments, and edges are characterised using the distance between centers of gravity of connected components. Then, graphs are fed to a dedicated neural network which has been learned to classify the nodes of unknown graphs using both node attributes and topology. The method is evaluated on the ILPIso dataset and obtains very promising results. These results show the interest of using graphs for such a task, especially when input data are noisy.

Mots clés: Graph Neural Network; Graphs; Floorplans

Domaines: Traitement du texte et du document; Vision par ordinateur et reconnaissance de formes [cs.CV]; Réseau de neurones [cs.NE]; Apprentissage [cs.LG]

Guillaume Renton, Pierre Héroux, Sébastien Adam, Benoit Gaüzère. Graph Neural Network for Symbol Detection on Document Images. 13th IAPR International Workshop on Graphics Recognition, Sep 2019, Sydney, Australia. ⟨hal-02490877⟩