earticle

영어

Object categorization based on hierarchical context modeling has shown to be useful in large database of object categories, especially, when a large number of object classes needs to be recognized from a range of different scene categories. However, average precision of categorization is still low compared to other existing methods. This may reflect that the contribution of underlying relations between objects has not been fully considered. In this paper, we improve average precision of contextual object recognition by taking advantage of objects co-occurrence information. Our method consists of two main phases. In the first phase, object representation is derived by considering the frequency of objects appeared in each image. The second phase is focused on classification of objects by applying a decision tree algorithm. We use SUN09 database to evaluate our proposed method. This database consists of images spanning from different scene categories and object instances. Our experimental results demonstrate that our proposed method achieves a higher average precision in comparison to a recent similar method by encoding contextual information in an efficient way.