BPT Using Multi-dimensional Information for Semi-Automatic Content Retrieval
Shirin Ghanbari, John Woods, Simon Mark Lucas · 2009
This paper presents a semi-automatic tool for content management systems for use in visual media production which is based on a Binary Partition Tree (BPT). In contrast to traditional content based image retrieval systems that work with entire images, the proposed tool analyses individual objects. Defined objects are produced by performing the segmentation inside a multi-dimensional BPT. The tree structure is colour based with the advantage of incorporating spatial frequency to form semantically meaningful tree nodes. The BPTs can be intuitively browsed and stored within XML documents. To perform retrieval when post processing, a node of the query image is matched against the nodes of the BPT in the database representation. These are matched according to a combination of MPEG-7 descriptors, including colour, texture and edge histograms. The paper illustrates how the use of multi-dimensional information stored within BPTs, can significantly enhance content retrieval results for natural images. Performance evaluation results show how influential the quality segmentations are within retrieval results.