Content area
Full Text
A similarity-matching technique for a content-based image retrieval system derived from the MPEG-7 texture descriptor is described. This proposed technique is especially invariant with regard to the rotated images. For the rotational invariance, a hierarchical similarity measurement method was employed to simplify the computational complexity. To verify this method, experiments with MPEG-7 data sets were performed.
(ProQuest: ... denotes formulae omitted.)
Introduction: More efficient systems for storing, indexing, and retriev- ing visual information are required in areas involving vast amounts of digital images [1]. A content-based image indexing and retrieval system, which has recently received a lot of attention, not only eliminates man- ual processing in image indexing but also provides automatic indexing according to the contents of the images [1, 2]. For contents-based fea- tures, texture is a fundamental objective, since it provides important information for interpreting the scenes and classifying the image [3].
The current MPEG-7 texture descriptor, which is now at the working draft stage for international standardisation, uses texture featuring and description techniques based on the human visual system (HVS) [4, 5]. For texture featuring, the best-adopted sub-band representation of the texture is a division of the spatial frequency domain with five octave- band divisions in the radial direction and six equal-width divisions in an angular direction [5]. The image can therefore be decomposed into thirty sub-bands or feature channels to describe the texture. In this Let- ter, we investigate the rotational invariance for the current MPEG-7 descriptor. To reduce the computational complexity associated with the rotational invariance, a hierarchical rotational similarity measurement which is suitable for the current MPEG-7 texture descriptor is used.
MPEG-7 texture descriptor: The features of the current MPEG-7 descriptor are calculated in the frequency domain of the feature channels with energy vector [e1, e2, e3, ..., e30] and energy deviation vector [d1, d2, d3, ..., d30]. By adding the average (fDC) and the standard deviation (fSD) of the entire image, the final homogeneous texture descriptor can be written as [6]
... (1)
Rotational invariant similarity measurement: The similarity distance between two texture images is measured by summing the weighted and absolute differences between two sets of texture descriptors. We denote the homogeneous texture descriptor of the query image TDquery and the texture descriptor of...