We propose a new approach for the representation and recognition of patterns. The primitives extraction process is based on the properties of an original contour profile. Corners, curves and line sebments can easily be detected. The main characteristic of the method is the possible overlap among primitives, which allows multiple descriptions of ambiguous parts. The relation "is followed by" is the most important of the graph. As the graph representation is very simple, model graphs can be intuitively defined by hand. The recognition stage consists in finding the greatest common structure between a model graph and an input graph. Some results illustrate the method.