General-purpose exhaustive graph mining algorithms have seldom been used in real life contexts due to the high complexity of the process that is mostly based on costly isomorphism tests and countless expansion possibilities. In this paper, we explain how to exploit grid-based representations of problems to efficiently extract frequent grid subgraphs and create Bag-of-Grids which can be used as new features for classification purposes. We provide an efficient grid mining algorithm called GriMA which is designed to scale to large amount of data. We apply our algorithm on image classification problems where typical Bag-of-Visual-Words-based techniques are used. However, those techniques make use of limited spatial information in the image which could be beneficial to obtain more discriminative features. Experiments on different datasets show that our algorithm is efficient and that adding the structure may greatly help the image classification process.