Résumé : Advanced seismic interpretation most commonly rests on transforming original data representations by considering more or less numerous seismic attributes, which bear no explicit relation with geology. For this reason, they hardly allow fully solving problems such as reassembling sparse geological surface elements or specifying chronological or topological relationships between surfaces such as unconformity, on lap, interruption by fault. The present work intends to make further progress in geology-based interpretation of seismic data by using artificial intelligence tools based on cognitive vision. We propose a cognitive vision workflow for seismic interpretation based on a visual ontology and on three associated module dealing for data management, visual characterisation and geological correlation. An example of results is given showing the possibilities of the method for easily merging disconnected reflectors within one stratigraphical horizon taking into account simple geological criteria (amplitude, thickness, dip, vertical distance between reflectors).