Nouvelle méthode d'extraction automatique de routes dans des images satellitaires
In the present thesis, a new automatic method to extract roads from satellite imagery is proposed. This new method called Tridimensional Multilayer (3DM) is part of the global methods of linear feature extraction and is based on the Radon transform concept. The 3DM method eliminates simultaneously the three restrictions of the linear Radon transform for line extraction. This method allows the extraction of lines with different lengths and curvatures even in a noisy context. The 3DM method allows also to establish a geometrical database relative to extracted lines like the length and the endpoints of extracted lines. This database can be integrated into a Geographic Information System (GIS) and it can be used in diverse applications. The methodological approach of this study is divided in two phases: mathematical and algorithm developments.In the first phase, we generalized the Radon transform for a continuous second-degree polynomial function (Tridimensional Radon Transform 3DRT) for extracting lines with different curvatures. The second phase consists first in elaborating a new concept of acquisition and analysis of information adapted to the methods of linear feature extraction (Multilayer method ). Then, we developed the 3DM method by combining 3DRT and MM. The 3DM method was applied to a binary noisy image for extracting the lines that represent roads with different lengths and the river borders with different curvatures. The performance of the 3DM method is evaluated by comparing the result obtained from the reference image (input image without noise). The evaluation of the 2DM method shows that 88% of the lines are correctly extracted. Meanwhile the percentage of omitted lines is 12% and committed lines reach 4%. The extraction success rate of this method is consequently quantified at 82%. These measurements show the improvement brought by the 3DM method in the extraction of the different curve lines. Implementation of the 3DM method onto images obtained by the binarisation of different real images shows moreover the potential of this method in diverse applications for extracting the different types of linear features from various types of imagery.