Kernel on Bags for multi-object database retrieval.

Philippe-Henri Gosselin (1) , Matthieu Cord (2) , Sylvie Philipp-Foliguet (3)

1 MIDI - Multimedia Indexation and Data Integration
2 MALIRE - Machine Learning and Information Retrieval
3 ETIS - UMR 8051 - Equipes Traitement de l'Information et Systèmes

In this paper, a kernel-based method for multi-object re- trieval in large image database is presented. First, our approach exploits a fuzzy region segmentation ap- proach in order to get robust local feature extraction and characterization. All the region features are summarized in bags representing the image index. The main part of this work concerns the kernel functions to deal with sets of fea- tures. Based on the linear combination of minor kernels, a family of kernels on bags is introduced. Several weighting schemes and combinations are proposed. Their introduction are motivated in the specific context of dealing with multi- object recognition with heterogeneous background. Com- bined with SVMs classification and interactive online learn- ing framework, the resulting algorithm satisfies the robust- ness requirements for representation and classification of ob- jects. Experiments and comparisons demonstrate the good performances of our multi-object retrieval technique.

Domains: Machine Learning [stat.ML]

Philippe-Henri Gosselin, Matthieu Cord, Sylvie Philipp-Foliguet. Kernel on Bags for multi-object database retrieval.. ACM International Conference on Image and Video Retrieval, Jul 2007, Amsterdam, Netherlands. pp.226-231, ⟨10.1145/1282280.1282317⟩. ⟨hal-00520303⟩