Image Processing Reference
In-Depth Information
sub-images by extracting one pixel with the same position from each micro-lens
and each of these sub-images is decomposed using a 2D DWT, resulting in an array
of coefficients corresponding to several frequency bands. The lower frequency
bands of the sub-images are arranged and transformed using a 3D DCT while the
remaining higher bands are simply quantized and the resulting data is entropy
encoded.
Encoding schemes for 3DHV are quite scarce in the literature in [
61
-
65
], and the
authors decompose the 3DHV sequence into several sub-image video sequences
and then jointly exploit motion (temporal prediction) and disparity between adja-
cent sub-images to perform compression. In these schemes, the spatial redundancy
is obviously exploited by the disparity estimation part of the scheme, similarly to
what is done in MVC; as such, a precise knowledge of the image structure is
needed, notably the sub-image dimensions.
3.3.4.2 Self-Similarity Predictive Encoding
A more efficient alternative approach to 3D holoscopic content encoding is to
explore the intrinsic redundancy (or self-similarity) of the repetitive structure of
this type of images, without needing explicit knowledge of the micro-image
structure.
Similarly to motion-compensated prediction encoding, in self-similarity predic-
tive encoding the encoder finds the “best” prediction for each coding unit through a
self-similarity estimation process restricting the allowed search area to the already
coded and reconstructed areas of the current picture (see Fig.
3.15
). The quantized
residue and the self-similarity vector representing the relative position between
current coding unit and its predictor are then conveyed to the decoder.
The concept of self-similarity predictive encoding was firstly proposed in
[
66
,
67
], to improve, respectively, the performance of H.264/AVC for 3DHV
coding and to take advantage of the flexible partition patterns used in the emerging
HEVC video coding standard.
Fig. 3.15 Allowed search areas for estimation of self-similarity vector in two different prediction
stages [
67
]
