Image Processing Reference
In-Depth Information
combined with a Principal Component Analysis (PCA, also called Karhunen-
Loeve Transform or Hotelling Transform). DWT is applied to viewpoint images,
and then PCA is applied to the resulting coefficients. Several kinds of DWT filters
are proposed (e.g., Dauchechies wavelets). In [
27
], the SPIHT (Set Partitioning in
Hierarchical Trees) method allows to display/transmit progressively the integral
image as a quality scalable bitstream. Two algorithms (2D and 3D) are proposed.
The first one is a 2D-DWT applied to the integral image and followed by the
2D-SPIHT. The second is based on the creation of a volume of viewpoint images on
which a 3D-DWT is applied and followed by 3D-SPIHT.
Elemental Images-Based Methods Another approach consists in encoding the
viewpoint images or the MIs of a still integral image as if they were a video
sequence (called pseudo video sequence) and then exploiting the temporal com-
pression tools of traditional video coders [
28
,
29
]. This approach shares some ideas
with the next two: multi-view and self-similarity. The method proposed in [
30
]
exploits the inter-MIs redundancies (using the optical characteristic that MIs have
overlapping zones). In [
31
], a KLT is applied to the viewpoint images.
Multi-view We consider here the integral image as a group of viewpoint images
that is encoded as a multi-view sequence (using inter-view prediction). In [
32
] and
[
33
], the viewpoint images are encoded using MVC encoder [
34
]. The exploitation
of temporal correlation and inter-view correlations induces an increase in complex-
ity. The evolutionary strategy (ES) proposed in [
35
] is based on the evolution theory
and allows an optimization of the coding scheme. In [
36
], ES is also applied and
combined to a half-pixel precision for
the motion/disparity prediction and
compensation.
Self-Similarity The method described in this part exploits the non-local spatial
correlation between MIs. The algorithm is mainly the same as for the inter predic-
tion modes (of H.264/AVC [
37
] and HEVC [
38
]) but within one frame. A block
matching algorithm is used to find a block similar to the current block in the causal
zone in the current frame (which corresponds to the blocks that have already been
coded and reconstructed). This similar block is then used as a predictor in the same
manner as for a temporal prediction and compensation. In [
39
] and [
40
], the
implementation in H.264/AVC of the INTRA_SS (for INTRA Self Similarity)
modes is described. These publications show the BD-rate gain brought by the SS
mode and also the interest of a precise partitioning in macro-blocks. In [
41
], the SS
mode is implemented in HEVC and the interest of the CU partitioning is shown.
Conti et al. [
42
] shows the BD-rate gain brought by this method for video sequences
(i.e., with a competition with inter-prediction modes). In [
43
] a scalable coding
scheme is described as follows: the layer 0 corresponds to an extracted view, the
layer 1 corresponds to a set of extracted views, and the layer 2 is the integral image.
Layers 0 and 1 are encoded, respectively, with reference HEVC and MV-HEVC
encoders. For layer 2 the self-similarity method is used. An inter-layer prediction is
