Owing to the rapid mounting of massive image data, image classification has attracted lots of research efforts. Several diverse research disciplines have been confluent on this important theme, looking for more powerful solutions. In this paper, we propose a novel image representation method B2S (Bag to Set) that keeps all frequency information and is more discriminative than traditional histogram based bag representation. Based on B2S, we construct two different image classification approaches. First, we apply B2S to a state-of-the-art image classification algorithm SPM in computer vision. Second, we design a framework DisIClass (Discriminative Frequent Pattern-Based Image Classification) to utilize data mining algorithms to classify images, which was hardly done before due to the intrinsic differences between the data of computer vision and data mining fields. DisIClass adapts the locality property of image data, and apply sequential covering method to induce the most discriminative feature sets from a closed frequent item set mining method. Our experiments with real image data show the high accuracy and good scalability of both approaches.