Abstract :

The subject of this thesis is the development of a new method for shape extraction and recognition. Most of the techniques used in our algorithms are not new, but the use we make of them is. Shapes in the images are defined as local, contrast invariant features, i.e. pieces of level lines. These level lines can be extracted from the original digital image or after bilinear interpolation of its gray levels. An inclusion tree of curves can be created in both cases, based on the monotonicity property of level sets. The next step consists of a regularization of the extracted curves. An affine invariant smoothing, equivalent to a curvature driven PDE (AMSS), is applied to all the curves in the image. From the regularized curves, local information is extracted (inflexion points, bitangents, etc.) that allows for the encoding of pieces of curves with different degrees of invariance (rotation-translation, euclidean or affine). A dictionary of codes is then created and, from now on, the search of a piece of curve in the image reduces to the search of a word code in the dictionary.