Low-Complexity Rotation-Invariant Image Retrieval Based on Steerable Sub-Gaussian Modeling

This paper addresses issues that arise in the design of a rotation-invariant content-based image retrieval system. In our proposed procedure, we first construct a steerable multivariate sub-Gaussian model, which associates the fractional lower-order moments (FLOMs) of an image, transformed via a steerable pyramid, with those of its rotated versions. The feature extraction step consists of estimating the covariations between the orientations of the pyramid at the same or at adjacent decomposition levels. The similarity measurement between the two images is performed by inverting the steerable model, and then by applying a rotation-invariant version of a normbased deterministic function. We illustrate how this retrieval scheme achieves a low probability of retrieval error when the distribution of the pyramid coefficients exhibits a distinct heavytailed behavior, as compared to previous methods, and at the same time, it maintains a low computational complexity.