MULTIMEDIA '06 Proceedings of the 14th annual ACM international conference on Multimedia
Attention-driven image interpretation with application to image retrieval
Pattern Recognition
Image retrieval model based on weighted visual features determined by relevance feedback
Information Sciences: an International Journal
ConVeS: a context verification framework for object recognition system
Proceedings of the 2009 conference on Information Science, Technology and Applications
Hi-index | 0.00 |
Several recent works attempt to automaticallyannotate image collection by exploiting the links betweenvisual information provided by segmented image featuresand semantic concepts provided by associated text. Themain limitation of such approaches, however, is thatsemantically meaningful segmentation is in generalunavailable. This paper proposes a novel statisticallearning-based approach to overcome this problem. Weemploy two different segmentation methods to segmentthe image into two sets of regions and learn theassociation between each set of regions with textconcepts. Given a new image, the idea is to first employ agreedy strategy to annotate the image with conceptsderived from different sets of overlapping and possiblyconflicting regions. We then incorporate a decisionmodel to disambiguate the concepts learned using thevisual features of the overlapping regions. Experimentson a mid-sized image collection demonstrate that the useof our disambiguation approach could improve theperformance of the system by about 12-16% on averagein terms of F1 measures as compared to system that usesonly one segmentation method.