Using Quadtree-Based Semantic Link Network for Image Classification

Bag-of-words model is widely utilized for representing images in a semantic intermediate manner. However traditional visual words are orderless without information with regard to co-occurrences of them as well as their spatial distributions. Spatial pyramid matching provides an effective way to preserve partly spatial information within images, but ignores geometric relations between visual words. By means of quadrant division, this paper proposed a novel methodology called quadtree-based Semantic Link Network(qt-SLN) to represent images in the form of semantic network to keep the structural and topological information of an image. The task of image classification then becomes one of classifying graphs which can be implemented with graph kernels defined on different structured data. In addition, co-occurrences of visual words are also modeled using Pointwise Mutual Information(PMI), which is exploited as a substitution matrix in approximate graph matching after normalization. The experimental results show that incorporating structural relations and co-occurrences of visual words allows for a more semantical framework of classification task.