Towards Social Media as a Data Source for Opportunistic Sensor Networking

Chathurani, N.W.U.D., Geva, S., Chandran, V. and Chappell, T.

    Retrieving relevant images from a large, diversified collection using visual queries (image content) as search argument is a challenging and important open problem. It requires an efficient and effective content-based image retrieval (CBIR) system. Image representation has a profound effect on the performance of CBIR. This paper presents a CBIR system based on a novel image representation using a new approach to the generation of image signatures (CBIR-ISIG). Image signatures are generated by applying random indexing (RI) to a Bag-ofvisual Words (BoW) representation of the images. RI is an efficient and scalable approach to dimensionality reduction, based on random projection which avoids the computational cost of matrix factorization. Most importantly, it can be performed incrementally as new data arrives, as is crucial for online systems. The retrieval quality of the proposed approach is evaluated using a benchmark dataset for image classification (a subset of the Corel dataset). The proposed approach shows promising results with comparable retrieval quality to state of the art approaches while retaining the benefits of the highly efficient representational scheme..
Cite as: Chathurani, N.W.U.D., Geva, S., Chandran, V. and Chappell, T. (2014). Towards Social Media as a Data Source for Opportunistic Sensor Networking. In Proc. Twelfth Australasian Data Mining Conference (AusDM14) Brisbane, Australia. CRPIT, 158. Li, X., Liu, L., Ong, K.L. and Zhao, Y. Eds., ACS. 175-182
pdf (from crpit.com) pdf (local if available) BibTeX EndNote GS