Résumé : The design of a graphics recognition method is a complex task due the multiplity of symbols and its configurations in real applications. In this paper, we propose a new approach based on a new concept called the Galois lattice which is often used for supervised recognition of graphical symbols. The originality of this paper is to build a concept lattice not in terms of all the features of symbols, but in terms of the most frequent or the most relevant ones. Our approach consists firstly, to build a visual vocabulary containing the frequent visual words from descriptors calculted on the symbols learning set. These visual vectors are calculated at the stable neighborhood of random points chosen from a given symbol. Secondly, we propose a new approach for symbol recognition based on a frequent Galois lattice. The concept lattice obtained based on frequent patterns is trained and used as a supervised classifier for symbols. The experimental results show recognition capabilities of our method and its robustness to noise.