Représentation et reconnaissance d'objets par champs réceptifs
Vincent Colin de Verdière · HAL (Le Centre pour la Communication Scientifique Directe) · 1999
This thesis belong to the field of modelisation and recognition of objects using their appearance. Each object is modeled by a collection of images and recognition is obtained by the matching between a new image and a model image. Images are modeled by measures on local features. Several local descriptor bases are theoretically and experimentally evaluated and a Gaussian derivative basis is selected for its properties which include~: high discriminability for a concise description, scalability and steerability. Invariance to orientation is obtained by setting derivative directions according to the local gradient. Invariance to scale is obtained by a new technique which locally selects a characteristic scale for describing a neighborhood. This scale correspond to a maximum over scale of a Laplacian operator. These invariances are experimentally validated. In our system, an image is decomposed in a grid of overlapping windows which is represented by a corresponding grid of local features computed on these windows. This highly redundant representation enables us to design two robust recognition techniques~: the first one is based on a simple vote and the second one based on the prediction--verification method consists in splitting the recognition process in an hypothesis generation step for a single window followed by a verification step which checks the generated hypothesis on the neighboring windows with inclusion of geometric coherence constraints.