In multi-instance learning, the core instances play an important role on the prediction of bags' label. And if two instances have different numbers of instances with the same category around them, they have different levels of representative. In order to improve the classification accuracy, multi-instance learning with instance-level covering algorithm (MILICA) is proposed by which we could select the most representative instances to form the core instance set. Firstly, with the max Hausdorffdistance and the covering algorithm, the initial core instance set is constructed. Then, the final core instance set and the number of instances in a cover are obtained. Finally, a similarity measure function is used to convert a bag into a single sample for classification. Experimental results on two-category datasets and multi-category image datasets demonstrate that the proposed MILICA method has perfect classification capability.