Object-Level Representation Learning for Few-Shot Image Classification