Image classification has been an interesting research issue in multimedia content analysis due to the wide applications. In this paper, we observe that images can be classified (or annotated) in two ways: i) Classify by some main object, ii) Classify by multiple objects with their relations. These two types of images usually exist concurrently in real-life image databases. Although a number of image classification methods have been propose, they can only handle one certain type of images well and fail to deal with both types of images correctly at the same time. In this paper, we propose a hybrid image classification method, namely "CBROA" (Classify By Representative Or Associations), that can effectively classify both types of images at the same time. CBROA integrates the decision tree and association rules mining method in an adaptive manner with construction of a virtual semantic ontology. Experimental results show that CBROA outperforms other classification methods in terms of classification accuracy in classifying mixed types of images.