Multivariate statistical modeling for stereo image retrieval

Ongoing developments in stereoscopic display technologies have led to the proliferation of huge stereo image databases. Therefore, the design of an appropriate Content Based Image Retrieval (CBIR) system for stereo images is an important emerging issue. In this paper, we propose a novel retrieval method which exploits simultaneously the spatial and cross-view dependencies of the stereo images. Within each subband, the joint distribution of the resulting wavelet coefficients of the two views located at the same spatial position as well as those of the neighboring pixels, is modeled by a multivariate statistical model based on Spherically Invariant Random Vectors (SIRV). The parameters of the SIRV model are selected as relevant signatures of the stereo pair. Experimental results show the benefits which can be drawn from the proposed retrieval approach.