Object Categorization Using Bag-of-Synonyms Model

Object categorization is one of the hottest research topics and key basic issues in computer vision, which aims to decide whether some instances of object categories appear in the visual media such as images or videos. In my thesis, I will mainly focus on image-based object categorization techniques. Generally speaking, there are three basic steps involved in the whole categorization process: (1) representing the images, (2) learning the visual models for object categories, and (3) performing the object categorization based on these visual models. Considering this, first I review the Bag-of-Words model—a popular, simple and efficient image representation. Then based on it, a novel representation, Bag-of-Synonyms model, is proposed and applied to object categorization tasks. Finally, my work is summarized.So far some effective image representations have been proposed, such as Bag-of-Words models, part-based models, hierarchical models, etc. To associated with these representations, some powerful probabilistic (e.g. graphical models) and discriminative (e.g. kernel methods) categorization approaches have been developed. The Bag-of-Synonyms model tries to improve the Bag-of-Words model by involving more information (e.g. locations and scales) of the image patches and organizing it more efficiently (e.g. hierarchically or not). It has been applied to object categorization tasks and the experimental results are quite successful and impressive, which can demonstrate that the Bag-of-Synonyms model is quite suitable and useful for object categorization.