Article Dans Une Revue Signal Processing: Image Communication Année : 2026

Semantic image segmentation using multi-view graph neural network

Résumé

Semantic image segmentation with standard CNNs often misses long-range structure and topological cues, leading to boundary leaks, holes, or label inconsistencies. We address this with a Multi-View Graph Neural Network (MVGNN) that makes the structural assumptions explicit: each predicted region becomes a node, and multiple complementary graphs (views) encode distinct spatial relations such as geometric proximity for long-range smoothing, boundary connectivity for contour fidelity, and hierarchical inclusion for part-whole consistency. A learned fusion module aggregates these views into a unified multigraph on which a lightweight GNN refines region labels. We evaluate MVGNN on synthetic data and on FASSEG, comparing against U-Net and U-Net+CRF. Across datasets, MVGNN improves Dice Score and reduces Hausdorff Distance, with the largest gains on cluttered scenes and near thin or intricate boundaries. Ablations confirm that (i) each view contributes additively to quality, (ii) choosing appropriate graph operators (NNConv, ChebConv) matters for capturing complementary relations, and (iii) fusing views is superior to any single-graph variant. These results indicate that explicitly modelling multi-relation structure offers a robust and efficient alternative for segmentation, yielding sharper boundaries and more globally consistent masks.

Fichier non déposé

Dates et versions

hal-05581192 , version 1 (05-04-2026)

Identifiants

Citer

Elie Karam, Nisrine Jrad, Patty Coupeau, Jean-Baptiste Fasquel, Fahed Abdallah, et al.. Semantic image segmentation using multi-view graph neural network. Signal Processing: Image Communication, 2026, 145, pp.117539. ⟨10.1016/j.image.2026.117539⟩. ⟨hal-05581192⟩
60 Consultations
0 Téléchargements

Altmetric

Partager

  • More