Full Parallax 3D Video Content Compression - Novel 3D Media Technologies - page 57

4.2.4.6 Depth Blending

This technique is a combination of the two previous methods, hence includes
varying patch sizes (estimated from the disparity between MIs) and blending. It
aims to produce artifacts-free and full-focused viewpoint images. This is the
method that provides the most accurate extracted views because it eliminates
most of the artifacts reported with less advanced methods.

4.3 Full Parallax Content Compression

4.3.1 State of the Art

4.3.1.1 Integral Imaging Compression

Several methods related to integral imaging compression have been presented in the
scientific literature. In the following we group them in five categories: DCT-based
methods, Wavelet-based methods, methods based on the processing of micro-
images or viewpoint images, methods based on a multi-view approach, and finally
the self-similarity (SS) approach.

DCT-Based Methods The most natural approach consists in applying the Discrete
Cosine Transform (DCT) to the micro-images, followed by quantization and
lossless coding; possibly, a differential coding between MI can be used [ 20 ]. The
differential coding can also be used for video sequences in order to remove the
temporal correlations [ 21 , 22 ]. Inter-MI correlation can be removed using the
3D-DCT on stacked MIs. Several scanning orders are tested in order to create the
MIs 3D structure. An optimization of the quantization step (for 3D-DCT-based
compression algorithms) is proposed in [ 23 ]. This optimization is done by gener-
ating a matrix of quantization coefficients which depends on the content on the
image.

Wavelet-Based Methods These methods use a scheme based on a Discrete Wave-
let Transform (DWT) applied to the viewpoint images. In [ 24 ], a hybrid four-
dimensions transform based on DWT and DCT is described (4D hybrid
DWT-DCT coding scheme). The 2D DWT is applied to the MIs, followed by a
2D DCT applied to the resulting blocks of coefficients. In [ 25 ], the integral image is
decomposed in viewpoint images. A 2D transform is performed by using 1D
transforms on the lines and rows of the viewpoint images, resulting in four
frequency sub-bands. The lower band is a coarse approximation of the original
viewpoint image. The 2D transform is applied recursively to increase the level of
decomposition at a coarser scale. The sub-bands are then grouped in 8

8 elements volumes and processed by a 3D-DCT. As in the previous methods, the
coefficients are then quantized and arithmetically coded. In [ 26 ], the transform is

8