We propose a similarity-based learning style algorithm by regarding each image as a multi-instance (MI) sample for
image classification. An image featured as vectorial representation interesting regions is transferred to a MI sample.
Then a similarity like matrix is constructed using MI kernel between given images and some carefully selected base
images, as the new representation of given images. Three selection strategies are proposed to build the base images set to
find an optimal solution. A Weka implementation decision tree is used as the main learner in this paper. Experiments on
image data repository ALOI and Corel Image 2000 show the effectiveness of the proposed algorithm compared to some
previous based line methods.