Abstract: The bag-of-visual words model needs to learn visual vocabulary and classifier from the beginning when it learns a novel image category,and it cannot make use of learned visual vocabulary.This paper proposed a transfer lear-ning algorithm based on visual phrases.The visual phrases contain not only local invariable features,but also local spatial information,which can describe the common characteristics among different image categories.Our algorithm can make the bag-of-visual words model obtain good performance in the novel image category by transferring visual phrases from source visual vocabulary.The experimental results validate that the algorithm effectively utilizes learned knowledge and gains better performance in the novel image category even when there are a few training images.

Key words: Image categorization,Bag-of-visual words model,Transfer learning