Omnidirectional Integral Photography images compression using the 3D-DCT N. P. Sgouros, D. P. Chaikalis, P. G. Papageorgas and M. S. Sangriotis Department of Informatics and Telecommunications, University of Athens Panepistimiopolis, Athens, Greece, 15784 nsg@di.uoa.gr , rtsimage@di.uoa.gr , ppapag@teipir.gr , sagri@di.uoa.gr Abstract: Integral Photography images exhibit high intra-pixel as well as inter-elemental-image correlation. In this work, we present an efficient, omnidirectional Integral Photography compression scheme based on a Hilbert curve scan and a three dimensional transform technique. ©2007 Optical Society of America OCIS codes: (999.9999) Image compression; (100.6890) Three-dimensional image processing; (110.3000) Image quality assessment. 1. Introduction There are many compression schemes proposed for Integral Photography (IP) [1] image coding based on disparity estimation techniques [2,3] or higher order transforms [4]. However, disparity estimation encoders, based on the MPEG standard [5], are inefficient in cases of IP images that contain small sized elemental images, while the proposed higher order transform techniques are used in unidirectional IP compression. In this work, we describe a three-dimensional discrete cosine transform (3D-DCT) [6] encoder for use in omnidirectional IP image compression. The encoder utilizes the 2D Hilbert curve [7], shown in Fig.1, to rearrange the two-dimensional (2D) elemental image lattice. The Hilbert space-filling curve has excellent locality preservation properties and is used in an effort to maximize the correlation of the elemental images contained within each transformation cube. The performance of the encoder is also evaluated for a number of different scanning topologies [3]. Fig. 1. Elemental images traversal scheme based on the Hilbert space filling curve. 2. Compression using the 3D-DCT and 2D scanning topologies The 2D lattice of elemental images is initially transformed to a 1D series of elemental images, using the Hilbert curve or one of the scanning topologies described in [3]. Consecutive elemental images are grouped together to form volumes on which the 3D-DCT will be applied. In our approach, the 3D-DCT is applied on groups of eight elemental images as a compromise between good quality and computational efficiency. Improved quality can be pursued using adaptive schemes [4,6] on the expense of higher computational complexity. In detail each group of elemental images is further segmented in 8 x 8 x 8 pixel volumes and the 3D-DCT is applied on each of these volumes, producing 512 coefficients denoted as F(u,v,w), with u,v,w=1,…,8. The 3D-DCT quantized coefficients are determined using Eq. 1.