Exploiting the Spatial Distribution of Interest Points Using Hierarchical Clustering for Improved Scene Retrieval
Today's methods for doing content-based image retrieval fail to take spatial relations of local points into consideration. These relations convey important information to determine what is represented within an image. This thesis demonstrates how the spatial information can be effectively conveyed using an agglomerative hierarchical clustering (AHC) based on the spatial proximity of interest points. Spatial relationships are hence ensured by clustering locally adjacent interest points together, albeit not forcing any particular arrangement upon them. This approach provides robustness to affine and scale transformations. Moreover it can be used independently of any feature description algorithm. The AHC allows the improvement of performance in a CBIR system while keeping processing and memory usage almost unaffected. During experiments, a performance increase of up to 12.47% with respect to a reference system using bag of features was achieved. The AHC also proved to be an adequate structure by outperforming simpler variants such as a pyramid of grid partitions over the surface of an image and a hierarchical random clustering of interest points.