Spatial context for visual vocabulary construction

The bag-of-visual-words model has been widely used in many applications, such as object recognition, image categorization, and visual information retrieval. However, most existing approaches construct a visual vocabulary by simply clustering image regions represented with low-level visual features, where spatial context of image regions has not been well utilized. In this paper, we present two techniques to take such a context into account. One is based on the Self-Organizing Map for Adaptive Processing of Structured Data (SOM-SD), and the other is based on our proposed Hierarchical Fuzzy C-means with Spatial Constraints (FCM-HS). We have employed these two methods together with language modeling for image categorization. Experimental results obtained on Caltech dataset have demonstrated that these two methods can achieve better classification performance than those without considering spatial context. The comparison of these two methods is also discussed in this paper.