Fusion of region and image-based techniques for automatic image annotation

We propose a concept-centered approach that combines region- and image-level analysis for automatic image annotation (AIA). At the region level, we group regions into separate concept groups and perform concept-centered region clustering separately. The key idea is that we make use of the inter- and intra-concept region distribution to eliminate unreliable region clusters and identify the main region clusters for each concept. We then derive the correspondence between the image region clusters and concepts. To further enhance the accuracy of AIA task, we employ a multi-stage kNN classification using the global features at the image level. Finally, we perform fusion of region- and image-level analysis to obtain the final annotations. Our results have been found to improve the performance significantly, with gains of 18.5% in recall and 8.3% in "number of concepts detected", as compared to the best reported AIA results for the Corel image data set.